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 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 505 {
 506         LASSERT(!list_empty(&qunit->lq_hash));
 507         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 508
 509         list_del_init(&qunit->lq_hash);
 510         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 511         qunit_put(qunit);
 512 }
 513
 514 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 515                                  (limit = count) : (limit += count)
 516
 517
 518 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 519 {
 520         return qctxt->lqc_handler ? 1 : 0;
 521 }
 522
 523 static int
 524 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 525                struct qunit_data *qdata, int opc, int wait,
 526                struct obd_trans_info *oti);
 527
 528 static int
 529 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 530                  struct qunit_data *qdata, int rc, int opc)
 531 {
 532         struct lustre_qunit *qunit = NULL;
 533         struct super_block *sb = qctxt->lqc_sb;
 534         int err = 0;
 535         struct quota_adjust_qunit *oqaq = NULL;
 536         int rc1 = 0;
 537         ENTRY;
 538
 539         LASSERT(qdata);
 540         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 541                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 542
 543         /* do it only when a releasing quota req more than 5MB b=18491 */
 544         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 545                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 546
 547         /* update local operational quota file */
 548         if (rc == 0) {
 549                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 550                 struct obd_quotactl *qctl;
 551                 __u64 *hardlimit;
 552
 553                 OBD_ALLOC_PTR(qctl);
 554                 if (qctl == NULL)
 555                         GOTO(out, err = -ENOMEM);
 556
 557                 /* acq/rel qunit for specified uid/gid is serialized,
 558                  * so there is no race between get fs quota limit and
 559                  * set fs quota limit */
 560                 qctl->qc_cmd = Q_GETQUOTA;
 561                 qctl->qc_id = qdata->qd_id;
 562                 qctl->qc_type = QDATA_IS_GRP(qdata);
 563                 err = fsfilt_quotactl(obd, sb, qctl);
 564                 if (err) {
 565                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 566                         GOTO(out_mem, err);
 567                 }
 568
 569                 if (QDATA_IS_BLK(qdata)) {
 570                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 571                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 572                 } else {
 573                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 574                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 575                 }
 576
 577                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 578
 579                 if (*hardlimit == 0)
 580                         goto out_mem;
 581
 582                 switch (opc) {
 583                 case QUOTA_DQACQ:
 584                         INC_QLIMIT(*hardlimit, count);
 585                         break;
 586                 case QUOTA_DQREL:
 587                         LASSERTF(count < *hardlimit,
 588                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 589                                  "count("LPU64") qd_qunit("LPU64") "
 590                                  "hardlimit("LPU64").\n",
 591                                  qdata->qd_id, qdata->qd_flags,
 592                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 593                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 594                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 595                         *hardlimit -= count;
 596                         break;
 597                 default:
 598                         LBUG();
 599                 }
 600
 601                 /* clear quota limit */
 602                 if (count == 0)
 603                         *hardlimit = 0;
 604
 605                 qctl->qc_cmd = Q_SETQUOTA;
 606                 err = fsfilt_quotactl(obd, sb, qctl);
 607                 if (err)
 608                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 609
 610                 QDATA_DEBUG(qdata, "%s completion\n",
 611                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 612 out_mem:
 613                 OBD_FREE_PTR(qctl);
 614         } else if (rc == -EDQUOT) {
 615                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 616         } else if (rc == -EBUSY) {
 617                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 618         } else {
 619                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 620         }
 621 out:
 622         /* remove the qunit from hash */
 623         spin_lock(&qunit_hash_lock);
 624
 625         qunit = dqacq_in_flight(qctxt, qdata);
 626         /* this qunit has been removed by qctxt_cleanup() */
 627         if (!qunit) {
 628                 spin_unlock(&qunit_hash_lock);
 629                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 630                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 631                 RETURN(err);
 632         }
 633
 634         LASSERT(opc == qunit->lq_opc);
 635         /* remove this qunit from lq_hash so that new processes cannot be added
 636          * to qunit->lq_waiters */
 637         remove_qunit_nolock(qunit);
 638         spin_unlock(&qunit_hash_lock);
 639
 640         compute_lqs_after_removing_qunit(qunit);
 641
 642         if (rc == 0)
 643                 rc = QUOTA_REQ_RETURNED;
 644         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 645         /* wake up all waiters */
 646         wake_up_all(&qunit->lq_waitq);
 647
 648         /* this is for dqacq_in_flight() */
 649         qunit_put(qunit);
 650         /* this is for alloc_qunit() */
 651         qunit_put(qunit);
 652         if (rc < 0 && rc != -EDQUOT)
 653                  RETURN(err);
 654
 655         /* don't reschedule in such cases:
 656          *   - acq/rel failure and qunit isn't changed,
 657          *     but not for quota recovery.
 658          *   - local dqacq/dqrel.
 659          *   - local disk io failure.
 660          */
 661          OBD_ALLOC_PTR(oqaq);
 662          if (!oqaq)
 663                  RETURN(-ENOMEM);
 664          qdata_to_oqaq(qdata, oqaq);
 665          /* adjust the qunit size in slaves */
 666          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 667          OBD_FREE_PTR(oqaq);
 668          if (rc1 < 0) {
 669                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 670                  RETURN(rc1);
 671          }
 672          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 673                  RETURN(err);
 674
 675          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 676              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 677                  RETURN(err);
 678
 679         /* reschedule another dqacq/dqrel if needed */
 680         qdata->qd_count = 0;
 681         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 682         rc1 = check_cur_qunit(obd, qctxt, qdata);
 683         if (rc1 > 0) {
 684                 int opc;
 685                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 686                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 687                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 688         }
 689         RETURN(err);
 690 }
 691
 692 struct dqacq_async_args {
 693         struct lustre_quota_ctxt *aa_ctxt;
 694         struct lustre_qunit *aa_qunit;
 695 };
 696
 697 static int dqacq_interpret(const struct lu_env *env,
 698                            struct ptlrpc_request *req, void *data, int rc)
 699 {
 700         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 701         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 702         struct obd_device_target *obt = qctxt->lqc_obt;
 703         struct lustre_qunit *qunit = aa->aa_qunit;
 704         struct obd_device *obd = req->rq_import->imp_obd;
 705         struct qunit_data *qdata = NULL;
 706         int rc1 = 0;
 707         ENTRY;
 708
 709         LASSERT(req);
 710         LASSERT(req->rq_import);
 711
 712         /* there are several forms of qunit(historic causes), so we need to
 713          * adjust qunit from slaves to the same form here */
 714         OBD_ALLOC(qdata, sizeof(struct qunit_data));
 715         if (!qdata)
 716                 RETURN(-ENOMEM);
 717
 718         down_read(&obt->obt_rwsem);
 719         /* if a quota req timeouts or is dropped, we should update quota
 720          * statistics which will be handled in dqacq_completion. And in
 721          * this situation we should get qdata from request instead of
 722          * reply */
 723         rc1 = quota_get_qdata(req, qdata,
 724                               (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 725                               QUOTA_IMPORT);
 726         if (rc1 < 0) {
 727                 DEBUG_REQ(D_ERROR, req,
 728                           "error unpacking qunit_data(rc: %d)\n", rc1);
 729                 GOTO(exit, rc = rc1);
 730         }
 731
 732         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 733         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 734
 735         if (qdata->qd_id != qunit->lq_data.qd_id ||
 736             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 737                 CDEBUG(D_ERROR, "the returned qd_id isn't expected!"
 738                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 739                        qunit->lq_data.qd_id);
 740                 qdata->qd_id = qunit->lq_data.qd_id;
 741                 rc = -EPROTO;
 742         }
 743         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 744                 CDEBUG(D_ERROR, "the returned grp/usr isn't expected!"
 745                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 746                        qunit->lq_data.qd_flags);
 747                 if (QDATA_IS_GRP(&qunit->lq_data))
 748                         QDATA_SET_GRP(qdata);
 749                 else
 750                         QDATA_CLR_GRP(qdata);
 751                 rc = -EPROTO;
 752         }
 753         if (qdata->qd_count > qunit->lq_data.qd_count) {
 754                 CDEBUG(D_ERROR, "the returned qd_count isn't expected!"
 755                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 756                        qunit->lq_data.qd_count);
 757                 rc = -EPROTO;
 758         }
 759
 760         rc = dqacq_completion(obd, qctxt, qdata, rc,
 761                               lustre_msg_get_opc(req->rq_reqmsg));
 762
 763 exit:
 764         up_read(&obt->obt_rwsem);
 765         OBD_FREE(qdata, sizeof(struct qunit_data));
 766
 767         RETURN(rc);
 768 }
 769
 770 /**
 771  * check if quota master is online
 772  */
 773 int check_qm(struct lustre_quota_ctxt *qctxt)
 774 {
 775         int rc;
 776         ENTRY;
 777
 778         spin_lock(&qctxt->lqc_lock);
 779         /* quit waiting when mds is back or qctxt is cleaned up */
 780         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 781         spin_unlock(&qctxt->lqc_lock);
 782
 783         RETURN(rc);
 784 }
 785
 786 /* wake up all waiting threads when lqc_import is NULL */
 787 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 788 {
 789         struct lustre_qunit *qunit, *tmp;
 790         int i;
 791         ENTRY;
 792
 793         spin_lock(&qunit_hash_lock);
 794         for (i = 0; i < NR_DQHASH; i++) {
 795                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 796                         if (qunit->lq_ctxt != qctxt)
 797                                 continue;
 798
 799                         /* Wake up all waiters. Do not change lq_state.
 800                          * The waiters will check lq_rc which is kept as 0
 801                          * if no others change it, then the waiters will return
 802                          * -EAGAIN to caller who can perform related quota
 803                          * acq/rel if necessary. */
 804                         wake_up_all(&qunit->lq_waitq);
 805                 }
 806         }
 807         spin_unlock(&qunit_hash_lock);
 808         EXIT;
 809 }
 810
 811 static int got_qunit(struct lustre_qunit *qunit)
 812 {
 813         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 814         int rc = 0;
 815         ENTRY;
 816
 817         spin_lock(&qunit->lq_lock);
 818         switch (qunit->lq_state) {
 819         case QUNIT_IN_HASH:
 820         case QUNIT_RM_FROM_HASH:
 821                 break;
 822         case QUNIT_FINISHED:
 823                 rc = 1;
 824                 break;
 825         default:
 826                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 827         }
 828         spin_unlock(&qunit->lq_lock);
 829
 830         if (!rc) {
 831                 spin_lock(&qctxt->lqc_lock);
 832                 rc = !qctxt->lqc_import || !qctxt->lqc_valid;
 833                 spin_unlock(&qctxt->lqc_lock);
 834         }
 835
 836         RETURN(rc);
 837 }
 838
 839 static int
 840 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 841                struct qunit_data *qdata, int opc, int wait,
 842                struct obd_trans_info *oti)
 843 {
 844         struct lustre_qunit *qunit, *empty;
 845         struct l_wait_info lwi = { 0 };
 846         struct ptlrpc_request *req;
 847         struct dqacq_async_args *aa;
 848         struct obd_import *imp = NULL;
 849         struct lustre_qunit_size *lqs = NULL;
 850         struct timeval work_start;
 851         struct timeval work_end;
 852         long timediff;
 853         int rc = 0;
 854         ENTRY;
 855
 856         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 857         do_gettimeofday(&work_start);
 858         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
 859                 RETURN(-ENOMEM);
 860
 861         spin_lock(&qunit_hash_lock);
 862         qunit = dqacq_in_flight(qctxt, qdata);
 863         if (qunit) {
 864                 spin_unlock(&qunit_hash_lock);
 865                 qunit_put(empty);
 866
 867                 goto wait_completion;
 868         }
 869         qunit = empty;
 870         qunit_get(qunit);
 871         insert_qunit_nolock(qctxt, qunit);
 872         spin_unlock(&qunit_hash_lock);
 873
 874         quota_search_lqs(qdata, NULL, qctxt, &lqs);
 875         if (lqs) {
 876                 spin_lock(&lqs->lqs_lock);
 877                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
 878                 /* when this qdata returned from mds, it will call lqs_putref */
 879                 lqs_getref(lqs);
 880                 spin_unlock(&lqs->lqs_lock);
 881                 /* this is for quota_search_lqs */
 882                 lqs_putref(lqs);
 883         } else {
 884                 CDEBUG(D_ERROR, "Can't find the lustre qunit size!\n");
 885         }
 886
 887         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 888                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 889         /* master is going to dqacq/dqrel from itself */
 890         if (is_master(qctxt)) {
 891                 int rc2;
 892                 QDATA_DEBUG(qdata, "local %s.\n",
 893                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 894                 QDATA_SET_CHANGE_QS(qdata);
 895                 rc = qctxt->lqc_handler(obd, qdata, opc);
 896                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
 897                 /* this is for qunit_get() */
 898                 qunit_put(qunit);
 899
 900                 do_gettimeofday(&work_end);
 901                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 902                 if (opc == QUOTA_DQACQ)
 903                         lprocfs_counter_add(qctxt->lqc_stats,
 904                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
 905                                             timediff);
 906                 else
 907                         lprocfs_counter_add(qctxt->lqc_stats,
 908                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
 909                                             timediff);
 910                 RETURN(rc ? rc : rc2);
 911         }
 912
 913         spin_lock(&qctxt->lqc_lock);
 914         if (!qctxt->lqc_import) {
 915                 spin_unlock(&qctxt->lqc_lock);
 916                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
 917
 918                 spin_lock(&qunit_hash_lock);
 919                 remove_qunit_nolock(qunit);
 920                 spin_unlock(&qunit_hash_lock);
 921
 922                 compute_lqs_after_removing_qunit(qunit);
 923
 924                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
 925                 wake_up_all(&qunit->lq_waitq);
 926
 927                 /* this is for qunit_get() */
 928                 qunit_put(qunit);
 929                 /* this for alloc_qunit() */
 930                 qunit_put(qunit);
 931                 spin_lock(&qctxt->lqc_lock);
 932                 if (wait && !qctxt->lqc_import) {
 933                         spin_unlock(&qctxt->lqc_lock);
 934
 935                         LASSERT(oti && oti->oti_thread &&
 936                                 oti->oti_thread->t_watchdog);
 937
 938                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
 939                         CDEBUG(D_QUOTA, "sleep for quota master\n");
 940                         l_wait_event(qctxt->lqc_wait_for_qmaster,
 941                                      check_qm(qctxt), &lwi);
 942                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
 943                         lc_watchdog_touch(oti->oti_thread->t_watchdog);
 944                 } else {
 945                         spin_unlock(&qctxt->lqc_lock);
 946                 }
 947
 948                 RETURN(-EAGAIN);
 949         }
 950         imp = class_import_get(qctxt->lqc_import);
 951         spin_unlock(&qctxt->lqc_lock);
 952
 953         /* build dqacq/dqrel request */
 954         LASSERT(imp);
 955
 956         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
 957                                         LUSTRE_MDS_VERSION, opc);
 958         class_import_put(imp);
 959         if (req == NULL) {
 960                 CDEBUG(D_ERROR, "Can't alloc request\n");
 961                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
 962                 /* this is for qunit_get() */
 963                 qunit_put(qunit);
 964                 RETURN(-ENOMEM);
 965         }
 966
 967         ptlrpc_request_set_replen(req);
 968         req->rq_no_resend = req->rq_no_delay = 1;
 969         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
 970         if (rc < 0) {
 971                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
 972                 ptlrpc_req_finished(req);
 973                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
 974                 /* this is for qunit_get() */
 975                 qunit_put(qunit);
 976                 RETURN(rc);
 977         }
 978
 979         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
 980         aa = ptlrpc_req_async_args(req);
 981         aa->aa_ctxt = qctxt;
 982         aa->aa_qunit = qunit;
 983
 984         req->rq_interpret_reply = dqacq_interpret;
 985         ptlrpcd_add_req(req, PSCOPE_OTHER);
 986
 987         QDATA_DEBUG(qdata, "%s scheduled.\n",
 988                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 989 wait_completion:
 990         if (wait && qunit) {
 991                 struct qunit_data *p = &qunit->lq_data;
 992
 993                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
 994                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
 995                 /* rc = -EAGAIN, it means the quota master isn't ready yet
 996                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
 997                  * rc = -EDQUOT, it means out of quota
 998                  * rc = -EBUSY, it means recovery is happening
 999                  * other rc < 0, it means real errors, functions who call
1000                  * schedule_dqacq should take care of this */
1001                 spin_lock(&qunit->lq_lock);
1002                 rc = qunit->lq_rc;
1003                 spin_unlock(&qunit->lq_lock);
1004                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1005                        qunit, rc);
1006         }
1007
1008         qunit_put(qunit);
1009         do_gettimeofday(&work_end);
1010         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1011         if (opc == QUOTA_DQACQ)
1012                 lprocfs_counter_add(qctxt->lqc_stats,
1013                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1014                                     timediff);
1015         else
1016                 lprocfs_counter_add(qctxt->lqc_stats,
1017                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1018                                     timediff);
1019
1020         RETURN(rc);
1021 }
1022
1023 int
1024 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1025                    uid_t uid, gid_t gid, __u32 isblk, int wait,
1026                    struct obd_trans_info *oti)
1027 {
1028         int rc = 0, i = USRQUOTA;
1029         __u32 id[MAXQUOTAS] = { uid, gid };
1030         struct qunit_data qdata[MAXQUOTAS];
1031         ENTRY;
1032
1033         CLASSERT(MAXQUOTAS < 4);
1034         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1035                 RETURN(0);
1036
1037         for (i = 0; i < MAXQUOTAS; i++) {
1038                 qdata[i].qd_id = id[i];
1039                 qdata[i].qd_flags = i;
1040                 if (isblk)
1041                         QDATA_SET_BLK(&qdata[i]);
1042                 qdata[i].qd_count = 0;
1043
1044                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1045                 if (rc > 0) {
1046                         int opc;
1047                         /* need acquire or release */
1048                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1049                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1050                                             wait,oti);
1051                         if (rc < 0)
1052                                 RETURN(rc);
1053                 } else if (wait == 1) {
1054                         /* when wait equates 1, that means mds_quota_acquire
1055                          * or filter_quota_acquire is calling it. */
1056                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1057                         if (rc < 0)
1058                                 RETURN(rc);
1059                 }
1060         }
1061
1062         RETURN(rc);
1063 }
1064
1065 int
1066 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1067                          unsigned short type, int isblk)
1068 {
1069         struct lustre_qunit *qunit = NULL;
1070         struct qunit_data qdata;
1071         struct timeval work_start;
1072         struct timeval work_end;
1073         long timediff;
1074         struct l_wait_info lwi = { 0 };
1075         int rc = 0;
1076         ENTRY;
1077
1078         do_gettimeofday(&work_start);
1079         qdata.qd_id = id;
1080         qdata.qd_flags = type;
1081         if (isblk)
1082                 QDATA_SET_BLK(&qdata);
1083         qdata.qd_count = 0;
1084
1085         spin_lock(&qunit_hash_lock);
1086         qunit = dqacq_in_flight(qctxt, &qdata);
1087         spin_unlock(&qunit_hash_lock);
1088
1089         if (qunit) {
1090                 struct qunit_data *p = &qunit->lq_data;
1091
1092                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1093                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
1094                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1095                        qunit, qunit->lq_rc);
1096                 /* keep same as schedule_dqacq() b=17030 */
1097                 spin_lock(&qunit->lq_lock);
1098                 rc = qunit->lq_rc;
1099                 spin_unlock(&qunit->lq_lock);
1100                 /* this is for dqacq_in_flight() */
1101                 qunit_put(qunit);
1102                 do_gettimeofday(&work_end);
1103                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1104                 lprocfs_counter_add(qctxt->lqc_stats,
1105                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1106                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1107                                     timediff);
1108         } else {
1109                 do_gettimeofday(&work_end);
1110                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1111                 lprocfs_counter_add(qctxt->lqc_stats,
1112                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1113                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1114                                     timediff);
1115         }
1116
1117         RETURN(rc);
1118 }
1119
1120 int
1121 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1122 {
1123         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1124         struct obd_device_target *obt = &obd->u.obt;
1125         struct super_block *sb = obt->obt_sb;
1126         int rc = 0;
1127         ENTRY;
1128
1129         LASSERT(qctxt);
1130
1131         rc = ptlrpcd_addref();
1132         if (rc)
1133                 RETURN(rc);
1134
1135         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1136         spin_lock_init(&qctxt->lqc_lock);
1137         spin_lock(&qctxt->lqc_lock);
1138         qctxt->lqc_handler = handler;
1139         qctxt->lqc_sb = sb;
1140         qctxt->lqc_obt = obt;
1141         qctxt->lqc_import = NULL;
1142         qctxt->lqc_recovery = 0;
1143         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1144         qctxt->lqc_valid = 1;
1145         qctxt->lqc_cqs_boundary_factor = 4;
1146         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1147         qctxt->lqc_cqs_least_iunit = 2;
1148         qctxt->lqc_cqs_qs_factor = 2;
1149         qctxt->lqc_flags = 0;
1150         QUOTA_MASTER_UNREADY(qctxt);
1151         qctxt->lqc_bunit_sz = default_bunit_sz;
1152         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1153         qctxt->lqc_iunit_sz = default_iunit_sz;
1154         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1155         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1156                                           * after the last shrinking */
1157         qctxt->lqc_sync_blk = 0;
1158         spin_unlock(&qctxt->lqc_lock);
1159
1160         qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH", 7, 7,
1161                                                &lqs_hash_ops, 0);
1162         if (!qctxt->lqc_lqs_hash) {
1163                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1164                 RETURN(-ENOMEM);
1165         }
1166
1167 #ifdef LPROCFS
1168         rc = lquota_proc_setup(obd, is_master(qctxt));
1169         if (rc)
1170                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1171 #endif
1172
1173         RETURN(rc);
1174 }
1175
1176 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1177 {
1178         struct lustre_qunit *qunit, *tmp;
1179         struct list_head tmp_list;
1180         struct obd_device_target *obt = qctxt->lqc_obt;
1181         int i;
1182         ENTRY;
1183
1184         CFS_INIT_LIST_HEAD(&tmp_list);
1185
1186         spin_lock(&qctxt->lqc_lock);
1187         qctxt->lqc_valid = 0;
1188         spin_unlock(&qctxt->lqc_lock);
1189
1190         spin_lock(&qunit_hash_lock);
1191         for (i = 0; i < NR_DQHASH; i++) {
1192                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1193                         if (qunit->lq_ctxt != qctxt)
1194                                 continue;
1195                         remove_qunit_nolock(qunit);
1196                         list_add(&qunit->lq_hash, &tmp_list);
1197                 }
1198         }
1199         spin_unlock(&qunit_hash_lock);
1200
1201         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1202                 list_del_init(&qunit->lq_hash);
1203                 compute_lqs_after_removing_qunit(qunit);
1204
1205                 /* wake up all waiters */
1206                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1207                 wake_up_all(&qunit->lq_waitq);
1208                 qunit_put(qunit);
1209         }
1210
1211         down_write(&obt->obt_rwsem);
1212         lustre_hash_exit(qctxt->lqc_lqs_hash);
1213         qctxt->lqc_lqs_hash = NULL;
1214         up_write(&obt->obt_rwsem);
1215
1216         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1217          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1218         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1219                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1220                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1221                                      cfs_time_seconds(1));
1222         }
1223
1224         ptlrpcd_decref();
1225
1226 #ifdef LPROCFS
1227         if (lquota_proc_cleanup(qctxt))
1228                 CERROR("cleanup proc error!\n");
1229 #endif
1230
1231         EXIT;
1232 }
1233
1234 struct qslave_recov_thread_data {
1235         struct obd_device *obd;
1236         struct lustre_quota_ctxt *qctxt;
1237         struct completion comp;
1238 };
1239
1240 /* FIXME only recovery block quota by now */
1241 static int qslave_recovery_main(void *arg)
1242 {
1243         struct qslave_recov_thread_data *data = arg;
1244         struct obd_device *obd = data->obd;
1245         struct lustre_quota_ctxt *qctxt = data->qctxt;
1246         unsigned int type;
1247         int rc = 0;
1248         ENTRY;
1249
1250         ptlrpc_daemonize("qslave_recovd");
1251
1252         complete(&data->comp);
1253
1254         if (qctxt->lqc_recovery)
1255                 RETURN(0);
1256         qctxt->lqc_recovery = 1;
1257
1258         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1259                 struct qunit_data qdata;
1260                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1261                 struct list_head id_list;
1262                 struct dquot_id *dqid, *tmp;
1263                 int ret;
1264
1265                 LOCK_DQONOFF_MUTEX(dqopt);
1266                 if (!sb_has_quota_enabled(qctxt->lqc_sb, type)) {
1267                         UNLOCK_DQONOFF_MUTEX(dqopt);
1268                         break;
1269                 }
1270
1271                 LASSERT(dqopt->files[type] != NULL);
1272                 CFS_INIT_LIST_HEAD(&id_list);
1273 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1274                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1275 #else
1276                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1277 #endif
1278                 UNLOCK_DQONOFF_MUTEX(dqopt);
1279                 if (rc)
1280                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1281
1282                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1283                         list_del_init(&dqid->di_link);
1284                         /* skip slave recovery on itself */
1285                         if (is_master(qctxt))
1286                                 goto free;
1287                         if (rc && rc != -EBUSY)
1288                                 goto free;
1289
1290                         qdata.qd_id = dqid->di_id;
1291                         qdata.qd_flags = type;
1292                         QDATA_SET_BLK(&qdata);
1293                         qdata.qd_count = 0;
1294
1295                         ret = check_cur_qunit(obd, qctxt, &qdata);
1296                         if (ret > 0) {
1297                                 int opc;
1298                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1299                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1300                                                     0, NULL);
1301                                 if (rc == -EDQUOT)
1302                                         rc = 0;
1303                         } else {
1304                                 rc = 0;
1305                         }
1306
1307                         if (rc)
1308                                 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1309                                        "qslave recovery failed! (id:%d type:%d "
1310                                        " rc:%d)\n", dqid->di_id, type, rc);
1311 free:
1312                         kfree(dqid);
1313                 }
1314         }
1315
1316         qctxt->lqc_recovery = 0;
1317         RETURN(rc);
1318 }
1319
1320 void
1321 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1322 {
1323         struct qslave_recov_thread_data data;
1324         int rc;
1325         ENTRY;
1326
1327         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1328                 goto exit;
1329
1330         data.obd = obd;
1331         data.qctxt = qctxt;
1332         init_completion(&data.comp);
1333
1334         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1335         if (rc < 0) {
1336                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1337                 goto exit;
1338         }
1339         wait_for_completion(&data.comp);
1340 exit:
1341         EXIT;
1342 }
1343
1344
1345 /**
1346  * lqs<->qctxt hash operations
1347  */
1348
1349 /**
1350  * string hashing using djb2 hash algorithm
1351  */
1352 static unsigned
1353 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1354 {
1355         struct quota_adjust_qunit *lqs_key;
1356         unsigned hash;
1357         ENTRY;
1358
1359         LASSERT(key);
1360         lqs_key = (struct quota_adjust_qunit *)key;
1361         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1362
1363         RETURN(hash & mask);
1364 }
1365
1366 static int
1367 lqs_compare(void *key, struct hlist_node *hnode)
1368 {
1369         struct quota_adjust_qunit *lqs_key;
1370         struct lustre_qunit_size *q;
1371         int rc;
1372         ENTRY;
1373
1374         LASSERT(key);
1375         lqs_key = (struct quota_adjust_qunit *)key;
1376         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1377
1378         spin_lock(&q->lqs_lock);
1379         rc = ((lqs_key->qaq_id == q->lqs_id) &&
1380               (QAQ_IS_GRP(lqs_key) == LQS_IS_GRP(q)));
1381         spin_unlock(&q->lqs_lock);
1382
1383         RETURN(rc);
1384 }
1385
1386 static void *
1387 lqs_get(struct hlist_node *hnode)
1388 {
1389         struct lustre_qunit_size *q =
1390             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1391         ENTRY;
1392
1393         atomic_inc(&q->lqs_refcount);
1394         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1395                q, atomic_read(&q->lqs_refcount));
1396
1397         RETURN(q);
1398 }
1399
1400 static void *
1401 lqs_put(struct hlist_node *hnode)
1402 {
1403         struct lustre_qunit_size *q =
1404             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1405         ENTRY;
1406
1407         LASSERT(atomic_read(&q->lqs_refcount) > 0);
1408         atomic_dec(&q->lqs_refcount);
1409         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1410                q, atomic_read(&q->lqs_refcount));
1411
1412         RETURN(q);
1413 }
1414
1415 static void
1416 lqs_exit(struct hlist_node *hnode)
1417 {
1418         struct lustre_qunit_size *q;
1419         ENTRY;
1420
1421         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1422         /*
1423          * Nothing should be left. User of lqs put it and
1424          * lqs also was deleted from table by this time
1425          * so we should have 0 refs.
1426          */
1427         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1428                  "Busy lqs %p with %d refs\n", q,
1429                  atomic_read(&q->lqs_refcount));
1430         OBD_FREE_PTR(q);
1431         EXIT;
1432 }
1433
1434 static lustre_hash_ops_t lqs_hash_ops = {
1435         .lh_hash    = lqs_hash,
1436         .lh_compare = lqs_compare,
1437         .lh_get     = lqs_get,
1438         .lh_put     = lqs_put,
1439         .lh_exit    = lqs_exit
1440 };
1441 #endif /* HAVE_QUOTA_SUPPORT */