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