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