fs/xfs/xfs_dquot.c

   1 /*
   2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_format.h"
  21 #include "xfs_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_alloc.h"
  27 #include "xfs_quota.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_inode.h"
  31 #include "xfs_bmap.h"
  32 #include "xfs_bmap_util.h"
  33 #include "xfs_rtalloc.h"
  34 #include "xfs_error.h"
  35 #include "xfs_itable.h"
  36 #include "xfs_attr.h"
  37 #include "xfs_buf_item.h"
  38 #include "xfs_trans_space.h"
  39 #include "xfs_trans_priv.h"
  40 #include "xfs_qm.h"
  41 #include "xfs_cksum.h"
  42 #include "xfs_trace.h"
  43
  44 /*
  45  * Lock order:
  46  *
  47  * ip->i_lock
  48  *   qi->qi_tree_lock
  49  *     dquot->q_qlock (xfs_dqlock() and friends)
  50  *       dquot->q_flush (xfs_dqflock() and friends)
  51  *       qi->qi_lru_lock
  52  *
  53  * If two dquots need to be locked the order is user before group/project,
  54  * otherwise by the lowest id first, see xfs_dqlock2.
  55  */
  56
  57 #ifdef DEBUG
  58 xfs_buftarg_t *xfs_dqerror_target;
  59 int xfs_do_dqerror;
  60 int xfs_dqreq_num;
  61 int xfs_dqerror_mod = 33;
  62 #endif
  63
  64 struct kmem_zone                *xfs_qm_dqtrxzone;
  65 static struct kmem_zone         *xfs_qm_dqzone;
  66
  67 static struct lock_class_key xfs_dquot_group_class;
  68 static struct lock_class_key xfs_dquot_project_class;
  69
  70 /*
  71  * This is called to free all the memory associated with a dquot
  72  */
  73 void
  74 xfs_qm_dqdestroy(
  75         xfs_dquot_t     *dqp)
  76 {
  77         ASSERT(list_empty(&dqp->q_lru));
  78
  79         mutex_destroy(&dqp->q_qlock);
  80         kmem_zone_free(xfs_qm_dqzone, dqp);
  81
  82         XFS_STATS_DEC(xs_qm_dquot);
  83 }
  84
  85 /*
  86  * If default limits are in force, push them into the dquot now.
  87  * We overwrite the dquot limits only if they are zero and this
  88  * is not the root dquot.
  89  */
  90 void
  91 xfs_qm_adjust_dqlimits(
  92         struct xfs_mount        *mp,
  93         struct xfs_dquot        *dq)
  94 {
  95         struct xfs_quotainfo    *q = mp->m_quotainfo;
  96         struct xfs_disk_dquot   *d = &dq->q_core;
  97         int                     prealloc = 0;
  98
  99         ASSERT(d->d_id);
 100
 101         if (q->qi_bsoftlimit && !d->d_blk_softlimit) {
 102                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 103                 prealloc = 1;
 104         }
 105         if (q->qi_bhardlimit && !d->d_blk_hardlimit) {
 106                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 107                 prealloc = 1;
 108         }
 109         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 110                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 111         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 112                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 113         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 114                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 115         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 116                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 117
 118         if (prealloc)
 119                 xfs_dquot_set_prealloc_limits(dq);
 120 }
 121
 122 /*
 123  * Check the limits and timers of a dquot and start or reset timers
 124  * if necessary.
 125  * This gets called even when quota enforcement is OFF, which makes our
 126  * life a little less complicated. (We just don't reject any quota
 127  * reservations in that case, when enforcement is off).
 128  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 129  * enforcement's off.
 130  * In contrast, warnings are a little different in that they don't
 131  * 'automatically' get started when limits get exceeded.  They do
 132  * get reset to zero, however, when we find the count to be under
 133  * the soft limit (they are only ever set non-zero via userspace).
 134  */
 135 void
 136 xfs_qm_adjust_dqtimers(
 137         xfs_mount_t             *mp,
 138         xfs_disk_dquot_t        *d)
 139 {
 140         ASSERT(d->d_id);
 141
 142 #ifdef DEBUG
 143         if (d->d_blk_hardlimit)
 144                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 145                        be64_to_cpu(d->d_blk_hardlimit));
 146         if (d->d_ino_hardlimit)
 147                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 148                        be64_to_cpu(d->d_ino_hardlimit));
 149         if (d->d_rtb_hardlimit)
 150                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 151                        be64_to_cpu(d->d_rtb_hardlimit));
 152 #endif
 153
 154         if (!d->d_btimer) {
 155                 if ((d->d_blk_softlimit &&
 156                      (be64_to_cpu(d->d_bcount) >
 157                       be64_to_cpu(d->d_blk_softlimit))) ||
 158                     (d->d_blk_hardlimit &&
 159                      (be64_to_cpu(d->d_bcount) >
 160                       be64_to_cpu(d->d_blk_hardlimit)))) {
 161                         d->d_btimer = cpu_to_be32(get_seconds() +
 162                                         mp->m_quotainfo->qi_btimelimit);
 163                 } else {
 164                         d->d_bwarns = 0;
 165                 }
 166         } else {
 167                 if ((!d->d_blk_softlimit ||
 168                      (be64_to_cpu(d->d_bcount) <=
 169                       be64_to_cpu(d->d_blk_softlimit))) &&
 170                     (!d->d_blk_hardlimit ||
 171                     (be64_to_cpu(d->d_bcount) <=
 172                      be64_to_cpu(d->d_blk_hardlimit)))) {
 173                         d->d_btimer = 0;
 174                 }
 175         }
 176
 177         if (!d->d_itimer) {
 178                 if ((d->d_ino_softlimit &&
 179                      (be64_to_cpu(d->d_icount) >
 180                       be64_to_cpu(d->d_ino_softlimit))) ||
 181                     (d->d_ino_hardlimit &&
 182                      (be64_to_cpu(d->d_icount) >
 183                       be64_to_cpu(d->d_ino_hardlimit)))) {
 184                         d->d_itimer = cpu_to_be32(get_seconds() +
 185                                         mp->m_quotainfo->qi_itimelimit);
 186                 } else {
 187                         d->d_iwarns = 0;
 188                 }
 189         } else {
 190                 if ((!d->d_ino_softlimit ||
 191                      (be64_to_cpu(d->d_icount) <=
 192                       be64_to_cpu(d->d_ino_softlimit)))  &&
 193                     (!d->d_ino_hardlimit ||
 194                      (be64_to_cpu(d->d_icount) <=
 195                       be64_to_cpu(d->d_ino_hardlimit)))) {
 196                         d->d_itimer = 0;
 197                 }
 198         }
 199
 200         if (!d->d_rtbtimer) {
 201                 if ((d->d_rtb_softlimit &&
 202                      (be64_to_cpu(d->d_rtbcount) >
 203                       be64_to_cpu(d->d_rtb_softlimit))) ||
 204                     (d->d_rtb_hardlimit &&
 205                      (be64_to_cpu(d->d_rtbcount) >
 206                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 207                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 208                                         mp->m_quotainfo->qi_rtbtimelimit);
 209                 } else {
 210                         d->d_rtbwarns = 0;
 211                 }
 212         } else {
 213                 if ((!d->d_rtb_softlimit ||
 214                      (be64_to_cpu(d->d_rtbcount) <=
 215                       be64_to_cpu(d->d_rtb_softlimit))) &&
 216                     (!d->d_rtb_hardlimit ||
 217                      (be64_to_cpu(d->d_rtbcount) <=
 218                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 219                         d->d_rtbtimer = 0;
 220                 }
 221         }
 222 }
 223
 224 /*
 225  * initialize a buffer full of dquots and log the whole thing
 226  */
 227 STATIC void
 228 xfs_qm_init_dquot_blk(
 229         xfs_trans_t     *tp,
 230         xfs_mount_t     *mp,
 231         xfs_dqid_t      id,
 232         uint            type,
 233         xfs_buf_t       *bp)
 234 {
 235         struct xfs_quotainfo    *q = mp->m_quotainfo;
 236         xfs_dqblk_t     *d;
 237         int             curid, i;
 238
 239         ASSERT(tp);
 240         ASSERT(xfs_buf_islocked(bp));
 241
 242         d = bp->b_addr;
 243
 244         /*
 245          * ID of the first dquot in the block - id's are zero based.
 246          */
 247         curid = id - (id % q->qi_dqperchunk);
 248         ASSERT(curid >= 0);
 249         memset(d, 0, BBTOB(q->qi_dqchunklen));
 250         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 251                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 252                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 253                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 254                 d->dd_diskdq.d_flags = type;
 255                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 256                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_uuid);
 257                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
 258                                          XFS_DQUOT_CRC_OFF);
 259                 }
 260         }
 261
 262         xfs_trans_dquot_buf(tp, bp,
 263                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 264                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 265                              XFS_BLF_GDQUOT_BUF)));
 266         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 267 }
 268
 269 /*
 270  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
 271  * watermarks correspond to the soft and hard limits by default. If a soft limit
 272  * is not specified, we use 95% of the hard limit.
 273  */
 274 void
 275 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
 276 {
 277         __uint64_t space;
 278
 279         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
 280         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
 281         if (!dqp->q_prealloc_lo_wmark) {
 282                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
 283                 do_div(dqp->q_prealloc_lo_wmark, 100);
 284                 dqp->q_prealloc_lo_wmark *= 95;
 285         }
 286
 287         space = dqp->q_prealloc_hi_wmark;
 288
 289         do_div(space, 100);
 290         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
 291         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
 292         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
 293 }
 294
 295 STATIC bool
 296 xfs_dquot_buf_verify_crc(
 297         struct xfs_mount        *mp,
 298         struct xfs_buf          *bp)
 299 {
 300         struct xfs_dqblk        *d = (struct xfs_dqblk *)bp->b_addr;
 301         int                     ndquots;
 302         int                     i;
 303
 304         if (!xfs_sb_version_hascrc(&mp->m_sb))
 305                 return true;
 306
 307         /*
 308          * if we are in log recovery, the quota subsystem has not been
 309          * initialised so we have no quotainfo structure. In that case, we need
 310          * to manually calculate the number of dquots in the buffer.
 311          */
 312         if (mp->m_quotainfo)
 313                 ndquots = mp->m_quotainfo->qi_dqperchunk;
 314         else
 315                 ndquots = xfs_qm_calc_dquots_per_chunk(mp,
 316                                         XFS_BB_TO_FSB(mp, bp->b_length));
 317
 318         for (i = 0; i < ndquots; i++, d++) {
 319                 if (!xfs_verify_cksum((char *)d, sizeof(struct xfs_dqblk),
 320                                  XFS_DQUOT_CRC_OFF))
 321                         return false;
 322                 if (!uuid_equal(&d->dd_uuid, &mp->m_sb.sb_uuid))
 323                         return false;
 324         }
 325         return true;
 326 }
 327
 328 STATIC bool
 329 xfs_dquot_buf_verify(
 330         struct xfs_mount        *mp,
 331         struct xfs_buf          *bp)
 332 {
 333         struct xfs_dqblk        *d = (struct xfs_dqblk *)bp->b_addr;
 334         xfs_dqid_t              id = 0;
 335         int                     ndquots;
 336         int                     i;
 337
 338         /*
 339          * if we are in log recovery, the quota subsystem has not been
 340          * initialised so we have no quotainfo structure. In that case, we need
 341          * to manually calculate the number of dquots in the buffer.
 342          */
 343         if (mp->m_quotainfo)
 344                 ndquots = mp->m_quotainfo->qi_dqperchunk;
 345         else
 346                 ndquots = xfs_qm_calc_dquots_per_chunk(mp, bp->b_length);
 347
 348         /*
 349          * On the first read of the buffer, verify that each dquot is valid.
 350          * We don't know what the id of the dquot is supposed to be, just that
 351          * they should be increasing monotonically within the buffer. If the
 352          * first id is corrupt, then it will fail on the second dquot in the
 353          * buffer so corruptions could point to the wrong dquot in this case.
 354          */
 355         for (i = 0; i < ndquots; i++) {
 356                 struct xfs_disk_dquot   *ddq;
 357                 int                     error;
 358
 359                 ddq = &d[i].dd_diskdq;
 360
 361                 if (i == 0)
 362                         id = be32_to_cpu(ddq->d_id);
 363
 364                 error = xfs_qm_dqcheck(mp, ddq, id + i, 0, XFS_QMOPT_DOWARN,
 365                                        "xfs_dquot_buf_verify");
 366                 if (error)
 367                         return false;
 368         }
 369         return true;
 370 }
 371
 372 static void
 373 xfs_dquot_buf_read_verify(
 374         struct xfs_buf  *bp)
 375 {
 376         struct xfs_mount        *mp = bp->b_target->bt_mount;
 377
 378         if (!xfs_dquot_buf_verify_crc(mp, bp) || !xfs_dquot_buf_verify(mp, bp)) {
 379                 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 380                 xfs_buf_ioerror(bp, EFSCORRUPTED);
 381         }
 382 }
 383
 384 /*
 385  * we don't calculate the CRC here as that is done when the dquot is flushed to
 386  * the buffer after the update is done. This ensures that the dquot in the
 387  * buffer always has an up-to-date CRC value.
 388  */
 389 void
 390 xfs_dquot_buf_write_verify(
 391         struct xfs_buf  *bp)
 392 {
 393         struct xfs_mount        *mp = bp->b_target->bt_mount;
 394
 395         if (!xfs_dquot_buf_verify(mp, bp)) {
 396                 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
 397                 xfs_buf_ioerror(bp, EFSCORRUPTED);
 398                 return;
 399         }
 400 }
 401
 402 const struct xfs_buf_ops xfs_dquot_buf_ops = {
 403         .verify_read = xfs_dquot_buf_read_verify,
 404         .verify_write = xfs_dquot_buf_write_verify,
 405 };
 406
 407 /*
 408  * Allocate a block and fill it with dquots.
 409  * This is called when the bmapi finds a hole.
 410  */
 411 STATIC int
 412 xfs_qm_dqalloc(
 413         xfs_trans_t     **tpp,
 414         xfs_mount_t     *mp,
 415         xfs_dquot_t     *dqp,
 416         xfs_inode_t     *quotip,
 417         xfs_fileoff_t   offset_fsb,
 418         xfs_buf_t       **O_bpp)
 419 {
 420         xfs_fsblock_t   firstblock;
 421         xfs_bmap_free_t flist;
 422         xfs_bmbt_irec_t map;
 423         int             nmaps, error, committed;
 424         xfs_buf_t       *bp;
 425         xfs_trans_t     *tp = *tpp;
 426
 427         ASSERT(tp != NULL);
 428
 429         trace_xfs_dqalloc(dqp);
 430
 431         /*
 432          * Initialize the bmap freelist prior to calling bmapi code.
 433          */
 434         xfs_bmap_init(&flist, &firstblock);
 435         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 436         /*
 437          * Return if this type of quotas is turned off while we didn't
 438          * have an inode lock
 439          */
 440         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 441                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 442                 return (ESRCH);
 443         }
 444
 445         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 446         nmaps = 1;
 447         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 448                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 449                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 450                                 &map, &nmaps, &flist);
 451         if (error)
 452                 goto error0;
 453         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 454         ASSERT(nmaps == 1);
 455         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 456                (map.br_startblock != HOLESTARTBLOCK));
 457
 458         /*
 459          * Keep track of the blkno to save a lookup later
 460          */
 461         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 462
 463         /* now we can just get the buffer (there's nothing to read yet) */
 464         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 465                                dqp->q_blkno,
 466                                mp->m_quotainfo->qi_dqchunklen,
 467                                0);
 468
 469         error = xfs_buf_geterror(bp);
 470         if (error)
 471                 goto error1;
 472         bp->b_ops = &xfs_dquot_buf_ops;
 473
 474         /*
 475          * Make a chunk of dquots out of this buffer and log
 476          * the entire thing.
 477          */
 478         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 479                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 480
 481         /*
 482          * xfs_bmap_finish() may commit the current transaction and
 483          * start a second transaction if the freelist is not empty.
 484          *
 485          * Since we still want to modify this buffer, we need to
 486          * ensure that the buffer is not released on commit of
 487          * the first transaction and ensure the buffer is added to the
 488          * second transaction.
 489          *
 490          * If there is only one transaction then don't stop the buffer
 491          * from being released when it commits later on.
 492          */
 493
 494         xfs_trans_bhold(tp, bp);
 495
 496         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 497                 goto error1;
 498         }
 499
 500         if (committed) {
 501                 tp = *tpp;
 502                 xfs_trans_bjoin(tp, bp);
 503         } else {
 504                 xfs_trans_bhold_release(tp, bp);
 505         }
 506
 507         *O_bpp = bp;
 508         return 0;
 509
 510       error1:
 511         xfs_bmap_cancel(&flist);
 512       error0:
 513         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 514
 515         return (error);
 516 }
 517 STATIC int
 518 xfs_qm_dqrepair(
 519         struct xfs_mount        *mp,
 520         struct xfs_trans        *tp,
 521         struct xfs_dquot        *dqp,
 522         xfs_dqid_t              firstid,
 523         struct xfs_buf          **bpp)
 524 {
 525         int                     error;
 526         struct xfs_disk_dquot   *ddq;
 527         struct xfs_dqblk        *d;
 528         int                     i;
 529
 530         /*
 531          * Read the buffer without verification so we get the corrupted
 532          * buffer returned to us. make sure we verify it on write, though.
 533          */
 534         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 535                                    mp->m_quotainfo->qi_dqchunklen,
 536                                    0, bpp, NULL);
 537
 538         if (error) {
 539                 ASSERT(*bpp == NULL);
 540                 return XFS_ERROR(error);
 541         }
 542         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 543
 544         ASSERT(xfs_buf_islocked(*bpp));
 545         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 546
 547         /* Do the actual repair of dquots in this buffer */
 548         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 549                 ddq = &d[i].dd_diskdq;
 550                 error = xfs_qm_dqcheck(mp, ddq, firstid + i,
 551                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 552                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 553                 if (error) {
 554                         /* repair failed, we're screwed */
 555                         xfs_trans_brelse(tp, *bpp);
 556                         return XFS_ERROR(EIO);
 557                 }
 558         }
 559
 560         return 0;
 561 }
 562
 563 /*
 564  * Maps a dquot to the buffer containing its on-disk version.
 565  * This returns a ptr to the buffer containing the on-disk dquot
 566  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 567  */
 568 STATIC int
 569 xfs_qm_dqtobp(
 570         xfs_trans_t             **tpp,
 571         xfs_dquot_t             *dqp,
 572         xfs_disk_dquot_t        **O_ddpp,
 573         xfs_buf_t               **O_bpp,
 574         uint                    flags)
 575 {
 576         struct xfs_bmbt_irec    map;
 577         int                     nmaps = 1, error;
 578         struct xfs_buf          *bp;
 579         struct xfs_inode        *quotip = xfs_dq_to_quota_inode(dqp);
 580         struct xfs_mount        *mp = dqp->q_mount;
 581         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 582         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 583
 584         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 585
 586         xfs_ilock(quotip, XFS_ILOCK_SHARED);
 587         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 588                 /*
 589                  * Return if this type of quotas is turned off while we
 590                  * didn't have the quota inode lock.
 591                  */
 592                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 593                 return ESRCH;
 594         }
 595
 596         /*
 597          * Find the block map; no allocations yet
 598          */
 599         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 600                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 601
 602         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 603         if (error)
 604                 return error;
 605
 606         ASSERT(nmaps == 1);
 607         ASSERT(map.br_blockcount == 1);
 608
 609         /*
 610          * Offset of dquot in the (fixed sized) dquot chunk.
 611          */
 612         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 613                 sizeof(xfs_dqblk_t);
 614
 615         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 616         if (map.br_startblock == HOLESTARTBLOCK) {
 617                 /*
 618                  * We don't allocate unless we're asked to
 619                  */
 620                 if (!(flags & XFS_QMOPT_DQALLOC))
 621                         return ENOENT;
 622
 623                 ASSERT(tp);
 624                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 625                                         dqp->q_fileoffset, &bp);
 626                 if (error)
 627                         return error;
 628                 tp = *tpp;
 629         } else {
 630                 trace_xfs_dqtobp_read(dqp);
 631
 632                 /*
 633                  * store the blkno etc so that we don't have to do the
 634                  * mapping all the time
 635                  */
 636                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 637
 638                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 639                                            dqp->q_blkno,
 640                                            mp->m_quotainfo->qi_dqchunklen,
 641                                            0, &bp, &xfs_dquot_buf_ops);
 642
 643                 if (error == EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 644                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 645                                                 mp->m_quotainfo->qi_dqperchunk;
 646                         ASSERT(bp == NULL);
 647                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 648                 }
 649
 650                 if (error) {
 651                         ASSERT(bp == NULL);
 652                         return XFS_ERROR(error);
 653                 }
 654         }
 655
 656         ASSERT(xfs_buf_islocked(bp));
 657         *O_bpp = bp;
 658         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 659
 660         return (0);
 661 }
 662
 663
 664 /*
 665  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 666  * and release the buffer immediately.
 667  *
 668  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 669  */
 670 int
 671 xfs_qm_dqread(
 672         struct xfs_mount        *mp,
 673         xfs_dqid_t              id,
 674         uint                    type,
 675         uint                    flags,
 676         struct xfs_dquot        **O_dqpp)
 677 {
 678         struct xfs_dquot        *dqp;
 679         struct xfs_disk_dquot   *ddqp;
 680         struct xfs_buf          *bp;
 681         struct xfs_trans        *tp = NULL;
 682         int                     error;
 683         int                     cancelflags = 0;
 684
 685
 686         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 687
 688         dqp->dq_flags = type;
 689         dqp->q_core.d_id = cpu_to_be32(id);
 690         dqp->q_mount = mp;
 691         INIT_LIST_HEAD(&dqp->q_lru);
 692         mutex_init(&dqp->q_qlock);
 693         init_waitqueue_head(&dqp->q_pinwait);
 694
 695         /*
 696          * Because we want to use a counting completion, complete
 697          * the flush completion once to allow a single access to
 698          * the flush completion without blocking.
 699          */
 700         init_completion(&dqp->q_flush);
 701         complete(&dqp->q_flush);
 702
 703         /*
 704          * Make sure group quotas have a different lock class than user
 705          * quotas.
 706          */
 707         switch (type) {
 708         case XFS_DQ_USER:
 709                 /* uses the default lock class */
 710                 break;
 711         case XFS_DQ_GROUP:
 712                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 713                 break;
 714         case XFS_DQ_PROJ:
 715                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 716                 break;
 717         default:
 718                 ASSERT(0);
 719                 break;
 720         }
 721
 722         XFS_STATS_INC(xs_qm_dquot);
 723
 724         trace_xfs_dqread(dqp);
 725
 726         if (flags & XFS_QMOPT_DQALLOC) {
 727                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 728                 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_attrsetm,
 729                                           XFS_QM_DQALLOC_SPACE_RES(mp), 0);
 730                 if (error)
 731                         goto error1;
 732                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 733         }
 734
 735         /*
 736          * get a pointer to the on-disk dquot and the buffer containing it
 737          * dqp already knows its own type (GROUP/USER).
 738          */
 739         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 740         if (error) {
 741                 /*
 742                  * This can happen if quotas got turned off (ESRCH),
 743                  * or if the dquot didn't exist on disk and we ask to
 744                  * allocate (ENOENT).
 745                  */
 746                 trace_xfs_dqread_fail(dqp);
 747                 cancelflags |= XFS_TRANS_ABORT;
 748                 goto error1;
 749         }
 750
 751         /* copy everything from disk dquot to the incore dquot */
 752         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 753         xfs_qm_dquot_logitem_init(dqp);
 754
 755         /*
 756          * Reservation counters are defined as reservation plus current usage
 757          * to avoid having to add every time.
 758          */
 759         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 760         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 761         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 762
 763         /* initialize the dquot speculative prealloc thresholds */
 764         xfs_dquot_set_prealloc_limits(dqp);
 765
 766         /* Mark the buf so that this will stay incore a little longer */
 767         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 768
 769         /*
 770          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 771          * So we need to release with xfs_trans_brelse().
 772          * The strategy here is identical to that of inodes; we lock
 773          * the dquot in xfs_qm_dqget() before making it accessible to
 774          * others. This is because dquots, like inodes, need a good level of
 775          * concurrency, and we don't want to take locks on the entire buffers
 776          * for dquot accesses.
 777          * Note also that the dquot buffer may even be dirty at this point, if
 778          * this particular dquot was repaired. We still aren't afraid to
 779          * brelse it because we have the changes incore.
 780          */
 781         ASSERT(xfs_buf_islocked(bp));
 782         xfs_trans_brelse(tp, bp);
 783
 784         if (tp) {
 785                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 786                 if (error)
 787                         goto error0;
 788         }
 789
 790         *O_dqpp = dqp;
 791         return error;
 792
 793 error1:
 794         if (tp)
 795                 xfs_trans_cancel(tp, cancelflags);
 796 error0:
 797         xfs_qm_dqdestroy(dqp);
 798         *O_dqpp = NULL;
 799         return error;
 800 }
 801
 802 /*
 803  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 804  * a locked dquot, doing an allocation (if requested) as needed.
 805  * When both an inode and an id are given, the inode's id takes precedence.
 806  * That is, if the id changes while we don't hold the ilock inside this
 807  * function, the new dquot is returned, not necessarily the one requested
 808  * in the id argument.
 809  */
 810 int
 811 xfs_qm_dqget(
 812         xfs_mount_t     *mp,
 813         xfs_inode_t     *ip,      /* locked inode (optional) */
 814         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 815         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 816         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 817         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 818 {
 819         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 820         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 821         struct xfs_dquot        *dqp;
 822         int                     error;
 823
 824         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 825         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 826             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 827             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 828                 return (ESRCH);
 829         }
 830
 831 #ifdef DEBUG
 832         if (xfs_do_dqerror) {
 833                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 834                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 835                         xfs_debug(mp, "Returning error in dqget");
 836                         return (EIO);
 837                 }
 838         }
 839
 840         ASSERT(type == XFS_DQ_USER ||
 841                type == XFS_DQ_PROJ ||
 842                type == XFS_DQ_GROUP);
 843         if (ip) {
 844                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 845                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 846         }
 847 #endif
 848
 849 restart:
 850         mutex_lock(&qi->qi_tree_lock);
 851         dqp = radix_tree_lookup(tree, id);
 852         if (dqp) {
 853                 xfs_dqlock(dqp);
 854                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 855                         xfs_dqunlock(dqp);
 856                         mutex_unlock(&qi->qi_tree_lock);
 857                         trace_xfs_dqget_freeing(dqp);
 858                         delay(1);
 859                         goto restart;
 860                 }
 861
 862                 dqp->q_nrefs++;
 863                 mutex_unlock(&qi->qi_tree_lock);
 864
 865                 trace_xfs_dqget_hit(dqp);
 866                 XFS_STATS_INC(xs_qm_dqcachehits);
 867                 *O_dqpp = dqp;
 868                 return 0;
 869         }
 870         mutex_unlock(&qi->qi_tree_lock);
 871         XFS_STATS_INC(xs_qm_dqcachemisses);
 872
 873         /*
 874          * Dquot cache miss. We don't want to keep the inode lock across
 875          * a (potential) disk read. Also we don't want to deal with the lock
 876          * ordering between quotainode and this inode. OTOH, dropping the inode
 877          * lock here means dealing with a chown that can happen before
 878          * we re-acquire the lock.
 879          */
 880         if (ip)
 881                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 882
 883         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 884
 885         if (ip)
 886                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 887
 888         if (error)
 889                 return error;
 890
 891         if (ip) {
 892                 /*
 893                  * A dquot could be attached to this inode by now, since
 894                  * we had dropped the ilock.
 895                  */
 896                 if (xfs_this_quota_on(mp, type)) {
 897                         struct xfs_dquot        *dqp1;
 898
 899                         dqp1 = xfs_inode_dquot(ip, type);
 900                         if (dqp1) {
 901                                 xfs_qm_dqdestroy(dqp);
 902                                 dqp = dqp1;
 903                                 xfs_dqlock(dqp);
 904                                 goto dqret;
 905                         }
 906                 } else {
 907                         /* inode stays locked on return */
 908                         xfs_qm_dqdestroy(dqp);
 909                         return XFS_ERROR(ESRCH);
 910                 }
 911         }
 912
 913         mutex_lock(&qi->qi_tree_lock);
 914         error = -radix_tree_insert(tree, id, dqp);
 915         if (unlikely(error)) {
 916                 WARN_ON(error != EEXIST);
 917
 918                 /*
 919                  * Duplicate found. Just throw away the new dquot and start
 920                  * over.
 921                  */
 922                 mutex_unlock(&qi->qi_tree_lock);
 923                 trace_xfs_dqget_dup(dqp);
 924                 xfs_qm_dqdestroy(dqp);
 925                 XFS_STATS_INC(xs_qm_dquot_dups);
 926                 goto restart;
 927         }
 928
 929         /*
 930          * We return a locked dquot to the caller, with a reference taken
 931          */
 932         xfs_dqlock(dqp);
 933         dqp->q_nrefs = 1;
 934
 935         qi->qi_dquots++;
 936         mutex_unlock(&qi->qi_tree_lock);
 937
 938  dqret:
 939         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 940         trace_xfs_dqget_miss(dqp);
 941         *O_dqpp = dqp;
 942         return (0);
 943 }
 944
 945
 946 STATIC void
 947 xfs_qm_dqput_final(
 948         struct xfs_dquot        *dqp)
 949 {
 950         struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 951         struct xfs_dquot        *gdqp;
 952         struct xfs_dquot        *pdqp;
 953
 954         trace_xfs_dqput_free(dqp);
 955
 956         if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 957                 XFS_STATS_INC(xs_qm_dquot_unused);
 958
 959         /*
 960          * If we just added a udquot to the freelist, then we want to release
 961          * the gdquot/pdquot reference that it (probably) has. Otherwise it'll
 962          * keep the gdquot/pdquot from getting reclaimed.
 963          */
 964         gdqp = dqp->q_gdquot;
 965         if (gdqp) {
 966                 xfs_dqlock(gdqp);
 967                 dqp->q_gdquot = NULL;
 968         }
 969
 970         pdqp = dqp->q_pdquot;
 971         if (pdqp) {
 972                 xfs_dqlock(pdqp);
 973                 dqp->q_pdquot = NULL;
 974         }
 975         xfs_dqunlock(dqp);
 976
 977         /*
 978          * If we had a group/project quota hint, release it now.
 979          */
 980         if (gdqp)
 981                 xfs_qm_dqput(gdqp);
 982         if (pdqp)
 983                 xfs_qm_dqput(pdqp);
 984 }
 985
 986 /*
 987  * Release a reference to the dquot (decrement ref-count) and unlock it.
 988  *
 989  * If there is a group quota attached to this dquot, carefully release that
 990  * too without tripping over deadlocks'n'stuff.
 991  */
 992 void
 993 xfs_qm_dqput(
 994         struct xfs_dquot        *dqp)
 995 {
 996         ASSERT(dqp->q_nrefs > 0);
 997         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 998
 999         trace_xfs_dqput(dqp);
1000
1001         if (--dqp->q_nrefs > 0)
1002                 xfs_dqunlock(dqp);
1003         else
1004                 xfs_qm_dqput_final(dqp);
1005 }
1006
1007 /*
1008  * Release a dquot. Flush it if dirty, then dqput() it.
1009  * dquot must not be locked.
1010  */
1011 void
1012 xfs_qm_dqrele(
1013         xfs_dquot_t     *dqp)
1014 {
1015         if (!dqp)
1016                 return;
1017
1018         trace_xfs_dqrele(dqp);
1019
1020         xfs_dqlock(dqp);
1021         /*
1022          * We don't care to flush it if the dquot is dirty here.
1023          * That will create stutters that we want to avoid.
1024          * Instead we do a delayed write when we try to reclaim
1025          * a dirty dquot. Also xfs_sync will take part of the burden...
1026          */
1027         xfs_qm_dqput(dqp);
1028 }
1029
1030 /*
1031  * This is the dquot flushing I/O completion routine.  It is called
1032  * from interrupt level when the buffer containing the dquot is
1033  * flushed to disk.  It is responsible for removing the dquot logitem
1034  * from the AIL if it has not been re-logged, and unlocking the dquot's
1035  * flush lock. This behavior is very similar to that of inodes..
1036  */
1037 STATIC void
1038 xfs_qm_dqflush_done(
1039         struct xfs_buf          *bp,
1040         struct xfs_log_item     *lip)
1041 {
1042         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
1043         xfs_dquot_t             *dqp = qip->qli_dquot;
1044         struct xfs_ail          *ailp = lip->li_ailp;
1045
1046         /*
1047          * We only want to pull the item from the AIL if its
1048          * location in the log has not changed since we started the flush.
1049          * Thus, we only bother if the dquot's lsn has
1050          * not changed. First we check the lsn outside the lock
1051          * since it's cheaper, and then we recheck while
1052          * holding the lock before removing the dquot from the AIL.
1053          */
1054         if ((lip->li_flags & XFS_LI_IN_AIL) &&
1055             lip->li_lsn == qip->qli_flush_lsn) {
1056
1057                 /* xfs_trans_ail_delete() drops the AIL lock. */
1058                 spin_lock(&ailp->xa_lock);
1059                 if (lip->li_lsn == qip->qli_flush_lsn)
1060                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1061                 else
1062                         spin_unlock(&ailp->xa_lock);
1063         }
1064
1065         /*
1066          * Release the dq's flush lock since we're done with it.
1067          */
1068         xfs_dqfunlock(dqp);
1069 }
1070
1071 /*
1072  * Write a modified dquot to disk.
1073  * The dquot must be locked and the flush lock too taken by caller.
1074  * The flush lock will not be unlocked until the dquot reaches the disk,
1075  * but the dquot is free to be unlocked and modified by the caller
1076  * in the interim. Dquot is still locked on return. This behavior is
1077  * identical to that of inodes.
1078  */
1079 int
1080 xfs_qm_dqflush(
1081         struct xfs_dquot        *dqp,
1082         struct xfs_buf          **bpp)
1083 {
1084         struct xfs_mount        *mp = dqp->q_mount;
1085         struct xfs_buf          *bp;
1086         struct xfs_disk_dquot   *ddqp;
1087         int                     error;
1088
1089         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1090         ASSERT(!completion_done(&dqp->q_flush));
1091
1092         trace_xfs_dqflush(dqp);
1093
1094         *bpp = NULL;
1095
1096         xfs_qm_dqunpin_wait(dqp);
1097
1098         /*
1099          * This may have been unpinned because the filesystem is shutting
1100          * down forcibly. If that's the case we must not write this dquot
1101          * to disk, because the log record didn't make it to disk.
1102          *
1103          * We also have to remove the log item from the AIL in this case,
1104          * as we wait for an emptry AIL as part of the unmount process.
1105          */
1106         if (XFS_FORCED_SHUTDOWN(mp)) {
1107                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
1108                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1109
1110                 spin_lock(&mp->m_ail->xa_lock);
1111                 if (lip->li_flags & XFS_LI_IN_AIL)
1112                         xfs_trans_ail_delete(mp->m_ail, lip,
1113                                              SHUTDOWN_CORRUPT_INCORE);
1114                 else
1115                         spin_unlock(&mp->m_ail->xa_lock);
1116                 error = XFS_ERROR(EIO);
1117                 goto out_unlock;
1118         }
1119
1120         /*
1121          * Get the buffer containing the on-disk dquot
1122          */
1123         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1124                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp, NULL);
1125         if (error)
1126                 goto out_unlock;
1127
1128         /*
1129          * Calculate the location of the dquot inside the buffer.
1130          */
1131         ddqp = bp->b_addr + dqp->q_bufoffset;
1132
1133         /*
1134          * A simple sanity check in case we got a corrupted dquot..
1135          */
1136         error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1137                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1138         if (error) {
1139                 xfs_buf_relse(bp);
1140                 xfs_dqfunlock(dqp);
1141                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1142                 return XFS_ERROR(EIO);
1143         }
1144
1145         /* This is the only portion of data that needs to persist */
1146         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1147
1148         /*
1149          * Clear the dirty field and remember the flush lsn for later use.
1150          */
1151         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1152
1153         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1154                                         &dqp->q_logitem.qli_item.li_lsn);
1155
1156         /*
1157          * copy the lsn into the on-disk dquot now while we have the in memory
1158          * dquot here. This can't be done later in the write verifier as we
1159          * can't get access to the log item at that point in time.
1160          *
1161          * We also calculate the CRC here so that the on-disk dquot in the
1162          * buffer always has a valid CRC. This ensures there is no possibility
1163          * of a dquot without an up-to-date CRC getting to disk.
1164          */
1165         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1166                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1167
1168                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1169                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1170                                  XFS_DQUOT_CRC_OFF);
1171         }
1172
1173         /*
1174          * Attach an iodone routine so that we can remove this dquot from the
1175          * AIL and release the flush lock once the dquot is synced to disk.
1176          */
1177         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1178                                   &dqp->q_logitem.qli_item);
1179
1180         /*
1181          * If the buffer is pinned then push on the log so we won't
1182          * get stuck waiting in the write for too long.
1183          */
1184         if (xfs_buf_ispinned(bp)) {
1185                 trace_xfs_dqflush_force(dqp);
1186                 xfs_log_force(mp, 0);
1187         }
1188
1189         trace_xfs_dqflush_done(dqp);
1190         *bpp = bp;
1191         return 0;
1192
1193 out_unlock:
1194         xfs_dqfunlock(dqp);
1195         return XFS_ERROR(EIO);
1196 }
1197
1198 /*
1199  * Lock two xfs_dquot structures.
1200  *
1201  * To avoid deadlocks we always lock the quota structure with
1202  * the lowerd id first.
1203  */
1204 void
1205 xfs_dqlock2(
1206         xfs_dquot_t     *d1,
1207         xfs_dquot_t     *d2)
1208 {
1209         if (d1 && d2) {
1210                 ASSERT(d1 != d2);
1211                 if (be32_to_cpu(d1->q_core.d_id) >
1212                     be32_to_cpu(d2->q_core.d_id)) {
1213                         mutex_lock(&d2->q_qlock);
1214                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1215                 } else {
1216                         mutex_lock(&d1->q_qlock);
1217                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1218                 }
1219         } else if (d1) {
1220                 mutex_lock(&d1->q_qlock);
1221         } else if (d2) {
1222                 mutex_lock(&d2->q_qlock);
1223         }
1224 }
1225
1226 int __init
1227 xfs_qm_init(void)
1228 {
1229         xfs_qm_dqzone =
1230                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1231         if (!xfs_qm_dqzone)
1232                 goto out;
1233
1234         xfs_qm_dqtrxzone =
1235                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1236         if (!xfs_qm_dqtrxzone)
1237                 goto out_free_dqzone;
1238
1239         return 0;
1240
1241 out_free_dqzone:
1242         kmem_zone_destroy(xfs_qm_dqzone);
1243 out:
1244         return -ENOMEM;
1245 }
1246
1247 void
1248 xfs_qm_exit(void)
1249 {
1250         kmem_zone_destroy(xfs_qm_dqtrxzone);
1251         kmem_zone_destroy(xfs_qm_dqzone);
1252 }