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_log_format.h"
  22 #include "xfs_shared.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_bit.h"
  25 #include "xfs_mount.h"
  26 #include "xfs_inode.h"
  27 #include "xfs_bmap.h"
  28 #include "xfs_bmap_util.h"
  29 #include "xfs_alloc.h"
  30 #include "xfs_quota.h"
  31 #include "xfs_error.h"
  32 #include "xfs_trans.h"
  33 #include "xfs_buf_item.h"
  34 #include "xfs_trans_space.h"
  35 #include "xfs_trans_priv.h"
  36 #include "xfs_qm.h"
  37 #include "xfs_cksum.h"
  38 #include "xfs_trace.h"
  39 #include "xfs_log.h"
  40 #include "xfs_bmap_btree.h"
  41
  42 /*
  43  * Lock order:
  44  *
  45  * ip->i_lock
  46  *   qi->qi_tree_lock
  47  *     dquot->q_qlock (xfs_dqlock() and friends)
  48  *       dquot->q_flush (xfs_dqflock() and friends)
  49  *       qi->qi_lru_lock
  50  *
  51  * If two dquots need to be locked the order is user before group/project,
  52  * otherwise by the lowest id first, see xfs_dqlock2.
  53  */
  54
  55 #ifdef DEBUG
  56 xfs_buftarg_t *xfs_dqerror_target;
  57 int xfs_do_dqerror;
  58 int xfs_dqreq_num;
  59 int xfs_dqerror_mod = 33;
  60 #endif
  61
  62 struct kmem_zone                *xfs_qm_dqtrxzone;
  63 static struct kmem_zone         *xfs_qm_dqzone;
  64
  65 static struct lock_class_key xfs_dquot_group_class;
  66 static struct lock_class_key xfs_dquot_project_class;
  67
  68 /*
  69  * This is called to free all the memory associated with a dquot
  70  */
  71 void
  72 xfs_qm_dqdestroy(
  73         xfs_dquot_t     *dqp)
  74 {
  75         ASSERT(list_empty(&dqp->q_lru));
  76
  77         kmem_free(dqp->q_logitem.qli_item.li_lv_shadow);
  78         mutex_destroy(&dqp->q_qlock);
  79
  80         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
  81         kmem_zone_free(xfs_qm_dqzone, dqp);
  82 }
  83
  84 /*
  85  * If default limits are in force, push them into the dquot now.
  86  * We overwrite the dquot limits only if they are zero and this
  87  * is not the root dquot.
  88  */
  89 void
  90 xfs_qm_adjust_dqlimits(
  91         struct xfs_mount        *mp,
  92         struct xfs_dquot        *dq)
  93 {
  94         struct xfs_quotainfo    *q = mp->m_quotainfo;
  95         struct xfs_disk_dquot   *d = &dq->q_core;
  96         struct xfs_def_quota    *defq;
  97         int                     prealloc = 0;
  98
  99         ASSERT(d->d_id);
 100         defq = xfs_get_defquota(dq, q);
 101
 102         if (defq->bsoftlimit && !d->d_blk_softlimit) {
 103                 d->d_blk_softlimit = cpu_to_be64(defq->bsoftlimit);
 104                 prealloc = 1;
 105         }
 106         if (defq->bhardlimit && !d->d_blk_hardlimit) {
 107                 d->d_blk_hardlimit = cpu_to_be64(defq->bhardlimit);
 108                 prealloc = 1;
 109         }
 110         if (defq->isoftlimit && !d->d_ino_softlimit)
 111                 d->d_ino_softlimit = cpu_to_be64(defq->isoftlimit);
 112         if (defq->ihardlimit && !d->d_ino_hardlimit)
 113                 d->d_ino_hardlimit = cpu_to_be64(defq->ihardlimit);
 114         if (defq->rtbsoftlimit && !d->d_rtb_softlimit)
 115                 d->d_rtb_softlimit = cpu_to_be64(defq->rtbsoftlimit);
 116         if (defq->rtbhardlimit && !d->d_rtb_hardlimit)
 117                 d->d_rtb_hardlimit = cpu_to_be64(defq->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         xfs_dqid_t      curid;
 239         int             i;
 240
 241         ASSERT(tp);
 242         ASSERT(xfs_buf_islocked(bp));
 243
 244         d = bp->b_addr;
 245
 246         /*
 247          * ID of the first dquot in the block - id's are zero based.
 248          */
 249         curid = id - (id % q->qi_dqperchunk);
 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_meta_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;
 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         if (!bp) {
 358                 error = -ENOMEM;
 359                 goto error1;
 360         }
 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         error = xfs_bmap_finish(tpp, &flist, NULL);
 386         if (error)
 387                 goto error1;
 388
 389         /* Transaction was committed? */
 390         if (*tpp != tp) {
 391                 tp = *tpp;
 392                 xfs_trans_bjoin(tp, bp);
 393         } else {
 394                 xfs_trans_bhold_release(tp, bp);
 395         }
 396
 397         *O_bpp = bp;
 398         return 0;
 399
 400 error1:
 401         xfs_bmap_cancel(&flist);
 402 error0:
 403         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 404
 405         return error;
 406 }
 407
 408 STATIC int
 409 xfs_qm_dqrepair(
 410         struct xfs_mount        *mp,
 411         struct xfs_trans        *tp,
 412         struct xfs_dquot        *dqp,
 413         xfs_dqid_t              firstid,
 414         struct xfs_buf          **bpp)
 415 {
 416         int                     error;
 417         struct xfs_disk_dquot   *ddq;
 418         struct xfs_dqblk        *d;
 419         int                     i;
 420
 421         /*
 422          * Read the buffer without verification so we get the corrupted
 423          * buffer returned to us. make sure we verify it on write, though.
 424          */
 425         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 426                                    mp->m_quotainfo->qi_dqchunklen,
 427                                    0, bpp, NULL);
 428
 429         if (error) {
 430                 ASSERT(*bpp == NULL);
 431                 return error;
 432         }
 433         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 434
 435         ASSERT(xfs_buf_islocked(*bpp));
 436         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 437
 438         /* Do the actual repair of dquots in this buffer */
 439         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 440                 ddq = &d[i].dd_diskdq;
 441                 error = xfs_dqcheck(mp, ddq, firstid + i,
 442                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 443                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 444                 if (error) {
 445                         /* repair failed, we're screwed */
 446                         xfs_trans_brelse(tp, *bpp);
 447                         return -EIO;
 448                 }
 449         }
 450
 451         return 0;
 452 }
 453
 454 /*
 455  * Maps a dquot to the buffer containing its on-disk version.
 456  * This returns a ptr to the buffer containing the on-disk dquot
 457  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 458  */
 459 STATIC int
 460 xfs_qm_dqtobp(
 461         xfs_trans_t             **tpp,
 462         xfs_dquot_t             *dqp,
 463         xfs_disk_dquot_t        **O_ddpp,
 464         xfs_buf_t               **O_bpp,
 465         uint                    flags)
 466 {
 467         struct xfs_bmbt_irec    map;
 468         int                     nmaps = 1, error;
 469         struct xfs_buf          *bp;
 470         struct xfs_inode        *quotip;
 471         struct xfs_mount        *mp = dqp->q_mount;
 472         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 473         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 474         uint                    lock_mode;
 475
 476         quotip = xfs_quota_inode(dqp->q_mount, dqp->dq_flags);
 477         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 478
 479         lock_mode = xfs_ilock_data_map_shared(quotip);
 480         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 481                 /*
 482                  * Return if this type of quotas is turned off while we
 483                  * didn't have the quota inode lock.
 484                  */
 485                 xfs_iunlock(quotip, lock_mode);
 486                 return -ESRCH;
 487         }
 488
 489         /*
 490          * Find the block map; no allocations yet
 491          */
 492         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 493                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 494
 495         xfs_iunlock(quotip, lock_mode);
 496         if (error)
 497                 return error;
 498
 499         ASSERT(nmaps == 1);
 500         ASSERT(map.br_blockcount == 1);
 501
 502         /*
 503          * Offset of dquot in the (fixed sized) dquot chunk.
 504          */
 505         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 506                 sizeof(xfs_dqblk_t);
 507
 508         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 509         if (map.br_startblock == HOLESTARTBLOCK) {
 510                 /*
 511                  * We don't allocate unless we're asked to
 512                  */
 513                 if (!(flags & XFS_QMOPT_DQALLOC))
 514                         return -ENOENT;
 515
 516                 ASSERT(tp);
 517                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 518                                         dqp->q_fileoffset, &bp);
 519                 if (error)
 520                         return error;
 521                 tp = *tpp;
 522         } else {
 523                 trace_xfs_dqtobp_read(dqp);
 524
 525                 /*
 526                  * store the blkno etc so that we don't have to do the
 527                  * mapping all the time
 528                  */
 529                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 530
 531                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 532                                            dqp->q_blkno,
 533                                            mp->m_quotainfo->qi_dqchunklen,
 534                                            0, &bp, &xfs_dquot_buf_ops);
 535
 536                 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 537                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 538                                                 mp->m_quotainfo->qi_dqperchunk;
 539                         ASSERT(bp == NULL);
 540                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 541                 }
 542
 543                 if (error) {
 544                         ASSERT(bp == NULL);
 545                         return error;
 546                 }
 547         }
 548
 549         ASSERT(xfs_buf_islocked(bp));
 550         *O_bpp = bp;
 551         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 552
 553         return 0;
 554 }
 555
 556
 557 /*
 558  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 559  * and release the buffer immediately.
 560  *
 561  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 562  */
 563 int
 564 xfs_qm_dqread(
 565         struct xfs_mount        *mp,
 566         xfs_dqid_t              id,
 567         uint                    type,
 568         uint                    flags,
 569         struct xfs_dquot        **O_dqpp)
 570 {
 571         struct xfs_dquot        *dqp;
 572         struct xfs_disk_dquot   *ddqp;
 573         struct xfs_buf          *bp;
 574         struct xfs_trans        *tp = NULL;
 575         int                     error;
 576
 577         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 578
 579         dqp->dq_flags = type;
 580         dqp->q_core.d_id = cpu_to_be32(id);
 581         dqp->q_mount = mp;
 582         INIT_LIST_HEAD(&dqp->q_lru);
 583         mutex_init(&dqp->q_qlock);
 584         init_waitqueue_head(&dqp->q_pinwait);
 585
 586         /*
 587          * Because we want to use a counting completion, complete
 588          * the flush completion once to allow a single access to
 589          * the flush completion without blocking.
 590          */
 591         init_completion(&dqp->q_flush);
 592         complete(&dqp->q_flush);
 593
 594         /*
 595          * Make sure group quotas have a different lock class than user
 596          * quotas.
 597          */
 598         switch (type) {
 599         case XFS_DQ_USER:
 600                 /* uses the default lock class */
 601                 break;
 602         case XFS_DQ_GROUP:
 603                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 604                 break;
 605         case XFS_DQ_PROJ:
 606                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 607                 break;
 608         default:
 609                 ASSERT(0);
 610                 break;
 611         }
 612
 613         XFS_STATS_INC(mp, xs_qm_dquot);
 614
 615         trace_xfs_dqread(dqp);
 616
 617         if (flags & XFS_QMOPT_DQALLOC) {
 618                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
 619                                 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
 620                 if (error)
 621                         goto error0;
 622         }
 623
 624         /*
 625          * get a pointer to the on-disk dquot and the buffer containing it
 626          * dqp already knows its own type (GROUP/USER).
 627          */
 628         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 629         if (error) {
 630                 /*
 631                  * This can happen if quotas got turned off (ESRCH),
 632                  * or if the dquot didn't exist on disk and we ask to
 633                  * allocate (ENOENT).
 634                  */
 635                 trace_xfs_dqread_fail(dqp);
 636                 goto error1;
 637         }
 638
 639         /* copy everything from disk dquot to the incore dquot */
 640         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 641         xfs_qm_dquot_logitem_init(dqp);
 642
 643         /*
 644          * Reservation counters are defined as reservation plus current usage
 645          * to avoid having to add every time.
 646          */
 647         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 648         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 649         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 650
 651         /* initialize the dquot speculative prealloc thresholds */
 652         xfs_dquot_set_prealloc_limits(dqp);
 653
 654         /* Mark the buf so that this will stay incore a little longer */
 655         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 656
 657         /*
 658          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 659          * So we need to release with xfs_trans_brelse().
 660          * The strategy here is identical to that of inodes; we lock
 661          * the dquot in xfs_qm_dqget() before making it accessible to
 662          * others. This is because dquots, like inodes, need a good level of
 663          * concurrency, and we don't want to take locks on the entire buffers
 664          * for dquot accesses.
 665          * Note also that the dquot buffer may even be dirty at this point, if
 666          * this particular dquot was repaired. We still aren't afraid to
 667          * brelse it because we have the changes incore.
 668          */
 669         ASSERT(xfs_buf_islocked(bp));
 670         xfs_trans_brelse(tp, bp);
 671
 672         if (tp) {
 673                 error = xfs_trans_commit(tp);
 674                 if (error)
 675                         goto error0;
 676         }
 677
 678         *O_dqpp = dqp;
 679         return error;
 680
 681 error1:
 682         if (tp)
 683                 xfs_trans_cancel(tp);
 684 error0:
 685         xfs_qm_dqdestroy(dqp);
 686         *O_dqpp = NULL;
 687         return error;
 688 }
 689
 690 /*
 691  * Advance to the next id in the current chunk, or if at the
 692  * end of the chunk, skip ahead to first id in next allocated chunk
 693  * using the SEEK_DATA interface.
 694  */
 695 static int
 696 xfs_dq_get_next_id(
 697         xfs_mount_t             *mp,
 698         uint                    type,
 699         xfs_dqid_t              *id,
 700         loff_t                  eof)
 701 {
 702         struct xfs_inode        *quotip;
 703         xfs_fsblock_t           start;
 704         loff_t                  offset;
 705         uint                    lock;
 706         xfs_dqid_t              next_id;
 707         int                     error = 0;
 708
 709         /* Simple advance */
 710         next_id = *id + 1;
 711
 712         /* If new ID is within the current chunk, advancing it sufficed */
 713         if (next_id % mp->m_quotainfo->qi_dqperchunk) {
 714                 *id = next_id;
 715                 return 0;
 716         }
 717
 718         /* Nope, next_id is now past the current chunk, so find the next one */
 719         start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
 720
 721         quotip = xfs_quota_inode(mp, type);
 722         lock = xfs_ilock_data_map_shared(quotip);
 723
 724         offset = __xfs_seek_hole_data(VFS_I(quotip), XFS_FSB_TO_B(mp, start),
 725                                       eof, SEEK_DATA);
 726         if (offset < 0)
 727                 error = offset;
 728
 729         xfs_iunlock(quotip, lock);
 730
 731         /* -ENXIO is essentially "no more data" */
 732         if (error)
 733                 return (error == -ENXIO ? -ENOENT: error);
 734
 735         /* Convert next data offset back to a quota id */
 736         *id = XFS_B_TO_FSB(mp, offset) * mp->m_quotainfo->qi_dqperchunk;
 737         return 0;
 738 }
 739
 740 /*
 741  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 742  * a locked dquot, doing an allocation (if requested) as needed.
 743  * When both an inode and an id are given, the inode's id takes precedence.
 744  * That is, if the id changes while we don't hold the ilock inside this
 745  * function, the new dquot is returned, not necessarily the one requested
 746  * in the id argument.
 747  */
 748 int
 749 xfs_qm_dqget(
 750         xfs_mount_t     *mp,
 751         xfs_inode_t     *ip,      /* locked inode (optional) */
 752         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 753         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 754         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 755         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 756 {
 757         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 758         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 759         struct xfs_dquot        *dqp;
 760         loff_t                  eof = 0;
 761         int                     error;
 762
 763         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 764         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 765             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 766             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 767                 return -ESRCH;
 768         }
 769
 770 #ifdef DEBUG
 771         if (xfs_do_dqerror) {
 772                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 773                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 774                         xfs_debug(mp, "Returning error in dqget");
 775                         return -EIO;
 776                 }
 777         }
 778
 779         ASSERT(type == XFS_DQ_USER ||
 780                type == XFS_DQ_PROJ ||
 781                type == XFS_DQ_GROUP);
 782         if (ip) {
 783                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 784                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 785         }
 786 #endif
 787
 788         /* Get the end of the quota file if we need it */
 789         if (flags & XFS_QMOPT_DQNEXT) {
 790                 struct xfs_inode        *quotip;
 791                 xfs_fileoff_t           last;
 792                 uint                    lock_mode;
 793
 794                 quotip = xfs_quota_inode(mp, type);
 795                 lock_mode = xfs_ilock_data_map_shared(quotip);
 796                 error = xfs_bmap_last_offset(quotip, &last, XFS_DATA_FORK);
 797                 xfs_iunlock(quotip, lock_mode);
 798                 if (error)
 799                         return error;
 800                 eof = XFS_FSB_TO_B(mp, last);
 801         }
 802
 803 restart:
 804         mutex_lock(&qi->qi_tree_lock);
 805         dqp = radix_tree_lookup(tree, id);
 806         if (dqp) {
 807                 xfs_dqlock(dqp);
 808                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 809                         xfs_dqunlock(dqp);
 810                         mutex_unlock(&qi->qi_tree_lock);
 811                         trace_xfs_dqget_freeing(dqp);
 812                         delay(1);
 813                         goto restart;
 814                 }
 815
 816                 /* uninit / unused quota found in radix tree, keep looking  */
 817                 if (flags & XFS_QMOPT_DQNEXT) {
 818                         if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 819                                 xfs_dqunlock(dqp);
 820                                 mutex_unlock(&qi->qi_tree_lock);
 821                                 error = xfs_dq_get_next_id(mp, type, &id, eof);
 822                                 if (error)
 823                                         return error;
 824                                 goto restart;
 825                         }
 826                 }
 827
 828                 dqp->q_nrefs++;
 829                 mutex_unlock(&qi->qi_tree_lock);
 830
 831                 trace_xfs_dqget_hit(dqp);
 832                 XFS_STATS_INC(mp, xs_qm_dqcachehits);
 833                 *O_dqpp = dqp;
 834                 return 0;
 835         }
 836         mutex_unlock(&qi->qi_tree_lock);
 837         XFS_STATS_INC(mp, xs_qm_dqcachemisses);
 838
 839         /*
 840          * Dquot cache miss. We don't want to keep the inode lock across
 841          * a (potential) disk read. Also we don't want to deal with the lock
 842          * ordering between quotainode and this inode. OTOH, dropping the inode
 843          * lock here means dealing with a chown that can happen before
 844          * we re-acquire the lock.
 845          */
 846         if (ip)
 847                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 848
 849         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 850
 851         if (ip)
 852                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 853
 854         /* If we are asked to find next active id, keep looking */
 855         if (error == -ENOENT && (flags & XFS_QMOPT_DQNEXT)) {
 856                 error = xfs_dq_get_next_id(mp, type, &id, eof);
 857                 if (!error)
 858                         goto restart;
 859         }
 860
 861         if (error)
 862                 return error;
 863
 864         if (ip) {
 865                 /*
 866                  * A dquot could be attached to this inode by now, since
 867                  * we had dropped the ilock.
 868                  */
 869                 if (xfs_this_quota_on(mp, type)) {
 870                         struct xfs_dquot        *dqp1;
 871
 872                         dqp1 = xfs_inode_dquot(ip, type);
 873                         if (dqp1) {
 874                                 xfs_qm_dqdestroy(dqp);
 875                                 dqp = dqp1;
 876                                 xfs_dqlock(dqp);
 877                                 goto dqret;
 878                         }
 879                 } else {
 880                         /* inode stays locked on return */
 881                         xfs_qm_dqdestroy(dqp);
 882                         return -ESRCH;
 883                 }
 884         }
 885
 886         mutex_lock(&qi->qi_tree_lock);
 887         error = radix_tree_insert(tree, id, dqp);
 888         if (unlikely(error)) {
 889                 WARN_ON(error != -EEXIST);
 890
 891                 /*
 892                  * Duplicate found. Just throw away the new dquot and start
 893                  * over.
 894                  */
 895                 mutex_unlock(&qi->qi_tree_lock);
 896                 trace_xfs_dqget_dup(dqp);
 897                 xfs_qm_dqdestroy(dqp);
 898                 XFS_STATS_INC(mp, xs_qm_dquot_dups);
 899                 goto restart;
 900         }
 901
 902         /*
 903          * We return a locked dquot to the caller, with a reference taken
 904          */
 905         xfs_dqlock(dqp);
 906         dqp->q_nrefs = 1;
 907
 908         qi->qi_dquots++;
 909         mutex_unlock(&qi->qi_tree_lock);
 910
 911         /* If we are asked to find next active id, keep looking */
 912         if (flags & XFS_QMOPT_DQNEXT) {
 913                 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 914                         xfs_qm_dqput(dqp);
 915                         error = xfs_dq_get_next_id(mp, type, &id, eof);
 916                         if (error)
 917                                 return error;
 918                         goto restart;
 919                 }
 920         }
 921
 922  dqret:
 923         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 924         trace_xfs_dqget_miss(dqp);
 925         *O_dqpp = dqp;
 926         return 0;
 927 }
 928
 929 /*
 930  * Release a reference to the dquot (decrement ref-count) and unlock it.
 931  *
 932  * If there is a group quota attached to this dquot, carefully release that
 933  * too without tripping over deadlocks'n'stuff.
 934  */
 935 void
 936 xfs_qm_dqput(
 937         struct xfs_dquot        *dqp)
 938 {
 939         ASSERT(dqp->q_nrefs > 0);
 940         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 941
 942         trace_xfs_dqput(dqp);
 943
 944         if (--dqp->q_nrefs == 0) {
 945                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 946                 trace_xfs_dqput_free(dqp);
 947
 948                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 949                         XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
 950         }
 951         xfs_dqunlock(dqp);
 952 }
 953
 954 /*
 955  * Release a dquot. Flush it if dirty, then dqput() it.
 956  * dquot must not be locked.
 957  */
 958 void
 959 xfs_qm_dqrele(
 960         xfs_dquot_t     *dqp)
 961 {
 962         if (!dqp)
 963                 return;
 964
 965         trace_xfs_dqrele(dqp);
 966
 967         xfs_dqlock(dqp);
 968         /*
 969          * We don't care to flush it if the dquot is dirty here.
 970          * That will create stutters that we want to avoid.
 971          * Instead we do a delayed write when we try to reclaim
 972          * a dirty dquot. Also xfs_sync will take part of the burden...
 973          */
 974         xfs_qm_dqput(dqp);
 975 }
 976
 977 /*
 978  * This is the dquot flushing I/O completion routine.  It is called
 979  * from interrupt level when the buffer containing the dquot is
 980  * flushed to disk.  It is responsible for removing the dquot logitem
 981  * from the AIL if it has not been re-logged, and unlocking the dquot's
 982  * flush lock. This behavior is very similar to that of inodes..
 983  */
 984 STATIC void
 985 xfs_qm_dqflush_done(
 986         struct xfs_buf          *bp,
 987         struct xfs_log_item     *lip)
 988 {
 989         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 990         xfs_dquot_t             *dqp = qip->qli_dquot;
 991         struct xfs_ail          *ailp = lip->li_ailp;
 992
 993         /*
 994          * We only want to pull the item from the AIL if its
 995          * location in the log has not changed since we started the flush.
 996          * Thus, we only bother if the dquot's lsn has
 997          * not changed. First we check the lsn outside the lock
 998          * since it's cheaper, and then we recheck while
 999          * holding the lock before removing the dquot from the AIL.
1000          */
1001         if ((lip->li_flags & XFS_LI_IN_AIL) &&
1002             lip->li_lsn == qip->qli_flush_lsn) {
1003
1004                 /* xfs_trans_ail_delete() drops the AIL lock. */
1005                 spin_lock(&ailp->xa_lock);
1006                 if (lip->li_lsn == qip->qli_flush_lsn)
1007                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1008                 else
1009                         spin_unlock(&ailp->xa_lock);
1010         }
1011
1012         /*
1013          * Release the dq's flush lock since we're done with it.
1014          */
1015         xfs_dqfunlock(dqp);
1016 }
1017
1018 /*
1019  * Write a modified dquot to disk.
1020  * The dquot must be locked and the flush lock too taken by caller.
1021  * The flush lock will not be unlocked until the dquot reaches the disk,
1022  * but the dquot is free to be unlocked and modified by the caller
1023  * in the interim. Dquot is still locked on return. This behavior is
1024  * identical to that of inodes.
1025  */
1026 int
1027 xfs_qm_dqflush(
1028         struct xfs_dquot        *dqp,
1029         struct xfs_buf          **bpp)
1030 {
1031         struct xfs_mount        *mp = dqp->q_mount;
1032         struct xfs_buf          *bp;
1033         struct xfs_disk_dquot   *ddqp;
1034         int                     error;
1035
1036         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1037         ASSERT(!completion_done(&dqp->q_flush));
1038
1039         trace_xfs_dqflush(dqp);
1040
1041         *bpp = NULL;
1042
1043         xfs_qm_dqunpin_wait(dqp);
1044
1045         /*
1046          * This may have been unpinned because the filesystem is shutting
1047          * down forcibly. If that's the case we must not write this dquot
1048          * to disk, because the log record didn't make it to disk.
1049          *
1050          * We also have to remove the log item from the AIL in this case,
1051          * as we wait for an emptry AIL as part of the unmount process.
1052          */
1053         if (XFS_FORCED_SHUTDOWN(mp)) {
1054                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
1055                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1056
1057                 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
1058
1059                 error = -EIO;
1060                 goto out_unlock;
1061         }
1062
1063         /*
1064          * Get the buffer containing the on-disk dquot
1065          */
1066         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1067                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
1068                                    &xfs_dquot_buf_ops);
1069         if (error)
1070                 goto out_unlock;
1071
1072         /*
1073          * Calculate the location of the dquot inside the buffer.
1074          */
1075         ddqp = bp->b_addr + dqp->q_bufoffset;
1076
1077         /*
1078          * A simple sanity check in case we got a corrupted dquot..
1079          */
1080         error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1081                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1082         if (error) {
1083                 xfs_buf_relse(bp);
1084                 xfs_dqfunlock(dqp);
1085                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1086                 return -EIO;
1087         }
1088
1089         /* This is the only portion of data that needs to persist */
1090         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1091
1092         /*
1093          * Clear the dirty field and remember the flush lsn for later use.
1094          */
1095         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1096
1097         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1098                                         &dqp->q_logitem.qli_item.li_lsn);
1099
1100         /*
1101          * copy the lsn into the on-disk dquot now while we have the in memory
1102          * dquot here. This can't be done later in the write verifier as we
1103          * can't get access to the log item at that point in time.
1104          *
1105          * We also calculate the CRC here so that the on-disk dquot in the
1106          * buffer always has a valid CRC. This ensures there is no possibility
1107          * of a dquot without an up-to-date CRC getting to disk.
1108          */
1109         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1110                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1111
1112                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1113                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1114                                  XFS_DQUOT_CRC_OFF);
1115         }
1116
1117         /*
1118          * Attach an iodone routine so that we can remove this dquot from the
1119          * AIL and release the flush lock once the dquot is synced to disk.
1120          */
1121         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1122                                   &dqp->q_logitem.qli_item);
1123
1124         /*
1125          * If the buffer is pinned then push on the log so we won't
1126          * get stuck waiting in the write for too long.
1127          */
1128         if (xfs_buf_ispinned(bp)) {
1129                 trace_xfs_dqflush_force(dqp);
1130                 xfs_log_force(mp, 0);
1131         }
1132
1133         trace_xfs_dqflush_done(dqp);
1134         *bpp = bp;
1135         return 0;
1136
1137 out_unlock:
1138         xfs_dqfunlock(dqp);
1139         return -EIO;
1140 }
1141
1142 /*
1143  * Lock two xfs_dquot structures.
1144  *
1145  * To avoid deadlocks we always lock the quota structure with
1146  * the lowerd id first.
1147  */
1148 void
1149 xfs_dqlock2(
1150         xfs_dquot_t     *d1,
1151         xfs_dquot_t     *d2)
1152 {
1153         if (d1 && d2) {
1154                 ASSERT(d1 != d2);
1155                 if (be32_to_cpu(d1->q_core.d_id) >
1156                     be32_to_cpu(d2->q_core.d_id)) {
1157                         mutex_lock(&d2->q_qlock);
1158                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1159                 } else {
1160                         mutex_lock(&d1->q_qlock);
1161                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1162                 }
1163         } else if (d1) {
1164                 mutex_lock(&d1->q_qlock);
1165         } else if (d2) {
1166                 mutex_lock(&d2->q_qlock);
1167         }
1168 }
1169
1170 int __init
1171 xfs_qm_init(void)
1172 {
1173         xfs_qm_dqzone =
1174                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1175         if (!xfs_qm_dqzone)
1176                 goto out;
1177
1178         xfs_qm_dqtrxzone =
1179                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1180         if (!xfs_qm_dqtrxzone)
1181                 goto out_free_dqzone;
1182
1183         return 0;
1184
1185 out_free_dqzone:
1186         kmem_zone_destroy(xfs_qm_dqzone);
1187 out:
1188         return -ENOMEM;
1189 }
1190
1191 void
1192 xfs_qm_exit(void)
1193 {
1194         kmem_zone_destroy(xfs_qm_dqtrxzone);
1195         kmem_zone_destroy(xfs_qm_dqzone);
1196 }