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_defer.h"
  27 #include "xfs_inode.h"
  28 #include "xfs_bmap.h"
  29 #include "xfs_bmap_util.h"
  30 #include "xfs_alloc.h"
  31 #include "xfs_quota.h"
  32 #include "xfs_error.h"
  33 #include "xfs_trans.h"
  34 #include "xfs_buf_item.h"
  35 #include "xfs_trans_space.h"
  36 #include "xfs_trans_priv.h"
  37 #include "xfs_qm.h"
  38 #include "xfs_cksum.h"
  39 #include "xfs_trace.h"
  40 #include "xfs_log.h"
  41 #include "xfs_bmap_btree.h"
  42
  43 /*
  44  * Lock order:
  45  *
  46  * ip->i_lock
  47  *   qi->qi_tree_lock
  48  *     dquot->q_qlock (xfs_dqlock() and friends)
  49  *       dquot->q_flush (xfs_dqflock() and friends)
  50  *       qi->qi_lru_lock
  51  *
  52  * If two dquots need to be locked the order is user before group/project,
  53  * otherwise by the lowest id first, see xfs_dqlock2.
  54  */
  55
  56 struct kmem_zone                *xfs_qm_dqtrxzone;
  57 static struct kmem_zone         *xfs_qm_dqzone;
  58
  59 static struct lock_class_key xfs_dquot_group_class;
  60 static struct lock_class_key xfs_dquot_project_class;
  61
  62 /*
  63  * This is called to free all the memory associated with a dquot
  64  */
  65 void
  66 xfs_qm_dqdestroy(
  67         xfs_dquot_t     *dqp)
  68 {
  69         ASSERT(list_empty(&dqp->q_lru));
  70
  71         kmem_free(dqp->q_logitem.qli_item.li_lv_shadow);
  72         mutex_destroy(&dqp->q_qlock);
  73
  74         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
  75         kmem_zone_free(xfs_qm_dqzone, dqp);
  76 }
  77
  78 /*
  79  * If default limits are in force, push them into the dquot now.
  80  * We overwrite the dquot limits only if they are zero and this
  81  * is not the root dquot.
  82  */
  83 void
  84 xfs_qm_adjust_dqlimits(
  85         struct xfs_mount        *mp,
  86         struct xfs_dquot        *dq)
  87 {
  88         struct xfs_quotainfo    *q = mp->m_quotainfo;
  89         struct xfs_disk_dquot   *d = &dq->q_core;
  90         struct xfs_def_quota    *defq;
  91         int                     prealloc = 0;
  92
  93         ASSERT(d->d_id);
  94         defq = xfs_get_defquota(dq, q);
  95
  96         if (defq->bsoftlimit && !d->d_blk_softlimit) {
  97                 d->d_blk_softlimit = cpu_to_be64(defq->bsoftlimit);
  98                 prealloc = 1;
  99         }
 100         if (defq->bhardlimit && !d->d_blk_hardlimit) {
 101                 d->d_blk_hardlimit = cpu_to_be64(defq->bhardlimit);
 102                 prealloc = 1;
 103         }
 104         if (defq->isoftlimit && !d->d_ino_softlimit)
 105                 d->d_ino_softlimit = cpu_to_be64(defq->isoftlimit);
 106         if (defq->ihardlimit && !d->d_ino_hardlimit)
 107                 d->d_ino_hardlimit = cpu_to_be64(defq->ihardlimit);
 108         if (defq->rtbsoftlimit && !d->d_rtb_softlimit)
 109                 d->d_rtb_softlimit = cpu_to_be64(defq->rtbsoftlimit);
 110         if (defq->rtbhardlimit && !d->d_rtb_hardlimit)
 111                 d->d_rtb_hardlimit = cpu_to_be64(defq->rtbhardlimit);
 112
 113         if (prealloc)
 114                 xfs_dquot_set_prealloc_limits(dq);
 115 }
 116
 117 /*
 118  * Check the limits and timers of a dquot and start or reset timers
 119  * if necessary.
 120  * This gets called even when quota enforcement is OFF, which makes our
 121  * life a little less complicated. (We just don't reject any quota
 122  * reservations in that case, when enforcement is off).
 123  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 124  * enforcement's off.
 125  * In contrast, warnings are a little different in that they don't
 126  * 'automatically' get started when limits get exceeded.  They do
 127  * get reset to zero, however, when we find the count to be under
 128  * the soft limit (they are only ever set non-zero via userspace).
 129  */
 130 void
 131 xfs_qm_adjust_dqtimers(
 132         xfs_mount_t             *mp,
 133         xfs_disk_dquot_t        *d)
 134 {
 135         ASSERT(d->d_id);
 136
 137 #ifdef DEBUG
 138         if (d->d_blk_hardlimit)
 139                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 140                        be64_to_cpu(d->d_blk_hardlimit));
 141         if (d->d_ino_hardlimit)
 142                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 143                        be64_to_cpu(d->d_ino_hardlimit));
 144         if (d->d_rtb_hardlimit)
 145                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 146                        be64_to_cpu(d->d_rtb_hardlimit));
 147 #endif
 148
 149         if (!d->d_btimer) {
 150                 if ((d->d_blk_softlimit &&
 151                      (be64_to_cpu(d->d_bcount) >
 152                       be64_to_cpu(d->d_blk_softlimit))) ||
 153                     (d->d_blk_hardlimit &&
 154                      (be64_to_cpu(d->d_bcount) >
 155                       be64_to_cpu(d->d_blk_hardlimit)))) {
 156                         d->d_btimer = cpu_to_be32(get_seconds() +
 157                                         mp->m_quotainfo->qi_btimelimit);
 158                 } else {
 159                         d->d_bwarns = 0;
 160                 }
 161         } else {
 162                 if ((!d->d_blk_softlimit ||
 163                      (be64_to_cpu(d->d_bcount) <=
 164                       be64_to_cpu(d->d_blk_softlimit))) &&
 165                     (!d->d_blk_hardlimit ||
 166                     (be64_to_cpu(d->d_bcount) <=
 167                      be64_to_cpu(d->d_blk_hardlimit)))) {
 168                         d->d_btimer = 0;
 169                 }
 170         }
 171
 172         if (!d->d_itimer) {
 173                 if ((d->d_ino_softlimit &&
 174                      (be64_to_cpu(d->d_icount) >
 175                       be64_to_cpu(d->d_ino_softlimit))) ||
 176                     (d->d_ino_hardlimit &&
 177                      (be64_to_cpu(d->d_icount) >
 178                       be64_to_cpu(d->d_ino_hardlimit)))) {
 179                         d->d_itimer = cpu_to_be32(get_seconds() +
 180                                         mp->m_quotainfo->qi_itimelimit);
 181                 } else {
 182                         d->d_iwarns = 0;
 183                 }
 184         } else {
 185                 if ((!d->d_ino_softlimit ||
 186                      (be64_to_cpu(d->d_icount) <=
 187                       be64_to_cpu(d->d_ino_softlimit)))  &&
 188                     (!d->d_ino_hardlimit ||
 189                      (be64_to_cpu(d->d_icount) <=
 190                       be64_to_cpu(d->d_ino_hardlimit)))) {
 191                         d->d_itimer = 0;
 192                 }
 193         }
 194
 195         if (!d->d_rtbtimer) {
 196                 if ((d->d_rtb_softlimit &&
 197                      (be64_to_cpu(d->d_rtbcount) >
 198                       be64_to_cpu(d->d_rtb_softlimit))) ||
 199                     (d->d_rtb_hardlimit &&
 200                      (be64_to_cpu(d->d_rtbcount) >
 201                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 202                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 203                                         mp->m_quotainfo->qi_rtbtimelimit);
 204                 } else {
 205                         d->d_rtbwarns = 0;
 206                 }
 207         } else {
 208                 if ((!d->d_rtb_softlimit ||
 209                      (be64_to_cpu(d->d_rtbcount) <=
 210                       be64_to_cpu(d->d_rtb_softlimit))) &&
 211                     (!d->d_rtb_hardlimit ||
 212                      (be64_to_cpu(d->d_rtbcount) <=
 213                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 214                         d->d_rtbtimer = 0;
 215                 }
 216         }
 217 }
 218
 219 /*
 220  * initialize a buffer full of dquots and log the whole thing
 221  */
 222 STATIC void
 223 xfs_qm_init_dquot_blk(
 224         xfs_trans_t     *tp,
 225         xfs_mount_t     *mp,
 226         xfs_dqid_t      id,
 227         uint            type,
 228         xfs_buf_t       *bp)
 229 {
 230         struct xfs_quotainfo    *q = mp->m_quotainfo;
 231         xfs_dqblk_t     *d;
 232         xfs_dqid_t      curid;
 233         int             i;
 234
 235         ASSERT(tp);
 236         ASSERT(xfs_buf_islocked(bp));
 237
 238         d = bp->b_addr;
 239
 240         /*
 241          * ID of the first dquot in the block - id's are zero based.
 242          */
 243         curid = id - (id % q->qi_dqperchunk);
 244         memset(d, 0, BBTOB(q->qi_dqchunklen));
 245         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 246                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 247                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 248                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 249                 d->dd_diskdq.d_flags = type;
 250                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 251                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
 252                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
 253                                          XFS_DQUOT_CRC_OFF);
 254                 }
 255         }
 256
 257         xfs_trans_dquot_buf(tp, bp,
 258                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 259                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 260                              XFS_BLF_GDQUOT_BUF)));
 261         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 262 }
 263
 264 /*
 265  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
 266  * watermarks correspond to the soft and hard limits by default. If a soft limit
 267  * is not specified, we use 95% of the hard limit.
 268  */
 269 void
 270 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
 271 {
 272         uint64_t space;
 273
 274         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
 275         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
 276         if (!dqp->q_prealloc_lo_wmark) {
 277                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
 278                 do_div(dqp->q_prealloc_lo_wmark, 100);
 279                 dqp->q_prealloc_lo_wmark *= 95;
 280         }
 281
 282         space = dqp->q_prealloc_hi_wmark;
 283
 284         do_div(space, 100);
 285         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
 286         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
 287         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
 288 }
 289
 290 /*
 291  * Allocate a block and fill it with dquots.
 292  * This is called when the bmapi finds a hole.
 293  */
 294 STATIC int
 295 xfs_qm_dqalloc(
 296         xfs_trans_t     **tpp,
 297         xfs_mount_t     *mp,
 298         xfs_dquot_t     *dqp,
 299         xfs_inode_t     *quotip,
 300         xfs_fileoff_t   offset_fsb,
 301         xfs_buf_t       **O_bpp)
 302 {
 303         xfs_fsblock_t   firstblock;
 304         struct xfs_defer_ops dfops;
 305         xfs_bmbt_irec_t map;
 306         int             nmaps, error;
 307         xfs_buf_t       *bp;
 308         xfs_trans_t     *tp = *tpp;
 309
 310         ASSERT(tp != NULL);
 311
 312         trace_xfs_dqalloc(dqp);
 313
 314         /*
 315          * Initialize the bmap freelist prior to calling bmapi code.
 316          */
 317         xfs_defer_init(&dfops, &firstblock);
 318         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 319         /*
 320          * Return if this type of quotas is turned off while we didn't
 321          * have an inode lock
 322          */
 323         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 324                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 325                 return -ESRCH;
 326         }
 327
 328         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 329         nmaps = 1;
 330         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 331                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 332                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 333                                 &map, &nmaps, &dfops);
 334         if (error)
 335                 goto error0;
 336         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 337         ASSERT(nmaps == 1);
 338         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 339                (map.br_startblock != HOLESTARTBLOCK));
 340
 341         /*
 342          * Keep track of the blkno to save a lookup later
 343          */
 344         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 345
 346         /* now we can just get the buffer (there's nothing to read yet) */
 347         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 348                                dqp->q_blkno,
 349                                mp->m_quotainfo->qi_dqchunklen,
 350                                0);
 351         if (!bp) {
 352                 error = -ENOMEM;
 353                 goto error1;
 354         }
 355         bp->b_ops = &xfs_dquot_buf_ops;
 356
 357         /*
 358          * Make a chunk of dquots out of this buffer and log
 359          * the entire thing.
 360          */
 361         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 362                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 363
 364         /*
 365          * xfs_defer_finish() may commit the current transaction and
 366          * start a second transaction if the freelist is not empty.
 367          *
 368          * Since we still want to modify this buffer, we need to
 369          * ensure that the buffer is not released on commit of
 370          * the first transaction and ensure the buffer is added to the
 371          * second transaction.
 372          *
 373          * If there is only one transaction then don't stop the buffer
 374          * from being released when it commits later on.
 375          */
 376
 377         xfs_trans_bhold(tp, bp);
 378
 379         error = xfs_defer_finish(tpp, &dfops);
 380         if (error)
 381                 goto error1;
 382
 383         /* Transaction was committed? */
 384         if (*tpp != tp) {
 385                 tp = *tpp;
 386                 xfs_trans_bjoin(tp, bp);
 387         } else {
 388                 xfs_trans_bhold_release(tp, bp);
 389         }
 390
 391         *O_bpp = bp;
 392         return 0;
 393
 394 error1:
 395         xfs_defer_cancel(&dfops);
 396 error0:
 397         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 398
 399         return error;
 400 }
 401
 402 STATIC int
 403 xfs_qm_dqrepair(
 404         struct xfs_mount        *mp,
 405         struct xfs_trans        *tp,
 406         struct xfs_dquot        *dqp,
 407         xfs_dqid_t              firstid,
 408         struct xfs_buf          **bpp)
 409 {
 410         int                     error;
 411         struct xfs_disk_dquot   *ddq;
 412         struct xfs_dqblk        *d;
 413         int                     i;
 414
 415         /*
 416          * Read the buffer without verification so we get the corrupted
 417          * buffer returned to us. make sure we verify it on write, though.
 418          */
 419         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 420                                    mp->m_quotainfo->qi_dqchunklen,
 421                                    0, bpp, NULL);
 422
 423         if (error) {
 424                 ASSERT(*bpp == NULL);
 425                 return error;
 426         }
 427         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 428
 429         ASSERT(xfs_buf_islocked(*bpp));
 430         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 431
 432         /* Do the actual repair of dquots in this buffer */
 433         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 434                 ddq = &d[i].dd_diskdq;
 435                 error = xfs_dqcheck(mp, ddq, firstid + i,
 436                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 437                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 438                 if (error) {
 439                         /* repair failed, we're screwed */
 440                         xfs_trans_brelse(tp, *bpp);
 441                         return -EIO;
 442                 }
 443         }
 444
 445         return 0;
 446 }
 447
 448 /*
 449  * Maps a dquot to the buffer containing its on-disk version.
 450  * This returns a ptr to the buffer containing the on-disk dquot
 451  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 452  */
 453 STATIC int
 454 xfs_qm_dqtobp(
 455         xfs_trans_t             **tpp,
 456         xfs_dquot_t             *dqp,
 457         xfs_disk_dquot_t        **O_ddpp,
 458         xfs_buf_t               **O_bpp,
 459         uint                    flags)
 460 {
 461         struct xfs_bmbt_irec    map;
 462         int                     nmaps = 1, error;
 463         struct xfs_buf          *bp;
 464         struct xfs_inode        *quotip;
 465         struct xfs_mount        *mp = dqp->q_mount;
 466         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 467         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 468         uint                    lock_mode;
 469
 470         quotip = xfs_quota_inode(dqp->q_mount, dqp->dq_flags);
 471         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 472
 473         lock_mode = xfs_ilock_data_map_shared(quotip);
 474         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 475                 /*
 476                  * Return if this type of quotas is turned off while we
 477                  * didn't have the quota inode lock.
 478                  */
 479                 xfs_iunlock(quotip, lock_mode);
 480                 return -ESRCH;
 481         }
 482
 483         /*
 484          * Find the block map; no allocations yet
 485          */
 486         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 487                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 488
 489         xfs_iunlock(quotip, lock_mode);
 490         if (error)
 491                 return error;
 492
 493         ASSERT(nmaps == 1);
 494         ASSERT(map.br_blockcount == 1);
 495
 496         /*
 497          * Offset of dquot in the (fixed sized) dquot chunk.
 498          */
 499         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 500                 sizeof(xfs_dqblk_t);
 501
 502         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 503         if (map.br_startblock == HOLESTARTBLOCK) {
 504                 /*
 505                  * We don't allocate unless we're asked to
 506                  */
 507                 if (!(flags & XFS_QMOPT_DQALLOC))
 508                         return -ENOENT;
 509
 510                 ASSERT(tp);
 511                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 512                                         dqp->q_fileoffset, &bp);
 513                 if (error)
 514                         return error;
 515                 tp = *tpp;
 516         } else {
 517                 trace_xfs_dqtobp_read(dqp);
 518
 519                 /*
 520                  * store the blkno etc so that we don't have to do the
 521                  * mapping all the time
 522                  */
 523                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 524
 525                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 526                                            dqp->q_blkno,
 527                                            mp->m_quotainfo->qi_dqchunklen,
 528                                            0, &bp, &xfs_dquot_buf_ops);
 529
 530                 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 531                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 532                                                 mp->m_quotainfo->qi_dqperchunk;
 533                         ASSERT(bp == NULL);
 534                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 535                 }
 536
 537                 if (error) {
 538                         ASSERT(bp == NULL);
 539                         return error;
 540                 }
 541         }
 542
 543         ASSERT(xfs_buf_islocked(bp));
 544         *O_bpp = bp;
 545         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 546
 547         return 0;
 548 }
 549
 550
 551 /*
 552  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 553  * and release the buffer immediately.
 554  *
 555  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 556  */
 557 int
 558 xfs_qm_dqread(
 559         struct xfs_mount        *mp,
 560         xfs_dqid_t              id,
 561         uint                    type,
 562         uint                    flags,
 563         struct xfs_dquot        **O_dqpp)
 564 {
 565         struct xfs_dquot        *dqp;
 566         struct xfs_disk_dquot   *ddqp;
 567         struct xfs_buf          *bp;
 568         struct xfs_trans        *tp = NULL;
 569         int                     error;
 570
 571         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 572
 573         dqp->dq_flags = type;
 574         dqp->q_core.d_id = cpu_to_be32(id);
 575         dqp->q_mount = mp;
 576         INIT_LIST_HEAD(&dqp->q_lru);
 577         mutex_init(&dqp->q_qlock);
 578         init_waitqueue_head(&dqp->q_pinwait);
 579
 580         /*
 581          * Because we want to use a counting completion, complete
 582          * the flush completion once to allow a single access to
 583          * the flush completion without blocking.
 584          */
 585         init_completion(&dqp->q_flush);
 586         complete(&dqp->q_flush);
 587
 588         /*
 589          * Make sure group quotas have a different lock class than user
 590          * quotas.
 591          */
 592         switch (type) {
 593         case XFS_DQ_USER:
 594                 /* uses the default lock class */
 595                 break;
 596         case XFS_DQ_GROUP:
 597                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 598                 break;
 599         case XFS_DQ_PROJ:
 600                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 601                 break;
 602         default:
 603                 ASSERT(0);
 604                 break;
 605         }
 606
 607         XFS_STATS_INC(mp, xs_qm_dquot);
 608
 609         trace_xfs_dqread(dqp);
 610
 611         if (flags & XFS_QMOPT_DQALLOC) {
 612                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
 613                                 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
 614                 if (error)
 615                         goto error0;
 616         }
 617
 618         /*
 619          * get a pointer to the on-disk dquot and the buffer containing it
 620          * dqp already knows its own type (GROUP/USER).
 621          */
 622         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 623         if (error) {
 624                 /*
 625                  * This can happen if quotas got turned off (ESRCH),
 626                  * or if the dquot didn't exist on disk and we ask to
 627                  * allocate (ENOENT).
 628                  */
 629                 trace_xfs_dqread_fail(dqp);
 630                 goto error1;
 631         }
 632
 633         /* copy everything from disk dquot to the incore dquot */
 634         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 635         xfs_qm_dquot_logitem_init(dqp);
 636
 637         /*
 638          * Reservation counters are defined as reservation plus current usage
 639          * to avoid having to add every time.
 640          */
 641         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 642         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 643         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 644
 645         /* initialize the dquot speculative prealloc thresholds */
 646         xfs_dquot_set_prealloc_limits(dqp);
 647
 648         /* Mark the buf so that this will stay incore a little longer */
 649         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 650
 651         /*
 652          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 653          * So we need to release with xfs_trans_brelse().
 654          * The strategy here is identical to that of inodes; we lock
 655          * the dquot in xfs_qm_dqget() before making it accessible to
 656          * others. This is because dquots, like inodes, need a good level of
 657          * concurrency, and we don't want to take locks on the entire buffers
 658          * for dquot accesses.
 659          * Note also that the dquot buffer may even be dirty at this point, if
 660          * this particular dquot was repaired. We still aren't afraid to
 661          * brelse it because we have the changes incore.
 662          */
 663         ASSERT(xfs_buf_islocked(bp));
 664         xfs_trans_brelse(tp, bp);
 665
 666         if (tp) {
 667                 error = xfs_trans_commit(tp);
 668                 if (error)
 669                         goto error0;
 670         }
 671
 672         *O_dqpp = dqp;
 673         return error;
 674
 675 error1:
 676         if (tp)
 677                 xfs_trans_cancel(tp);
 678 error0:
 679         xfs_qm_dqdestroy(dqp);
 680         *O_dqpp = NULL;
 681         return error;
 682 }
 683
 684 /*
 685  * Advance to the next id in the current chunk, or if at the
 686  * end of the chunk, skip ahead to first id in next allocated chunk
 687  * using the SEEK_DATA interface.
 688  */
 689 static int
 690 xfs_dq_get_next_id(
 691         struct xfs_mount        *mp,
 692         uint                    type,
 693         xfs_dqid_t              *id)
 694 {
 695         struct xfs_inode        *quotip = xfs_quota_inode(mp, type);
 696         xfs_dqid_t              next_id = *id + 1; /* simple advance */
 697         uint                    lock_flags;
 698         struct xfs_bmbt_irec    got;
 699         struct xfs_iext_cursor  cur;
 700         xfs_fsblock_t           start;
 701         int                     error = 0;
 702
 703         /* If we'd wrap past the max ID, stop */
 704         if (next_id < *id)
 705                 return -ENOENT;
 706
 707         /* If new ID is within the current chunk, advancing it sufficed */
 708         if (next_id % mp->m_quotainfo->qi_dqperchunk) {
 709                 *id = next_id;
 710                 return 0;
 711         }
 712
 713         /* Nope, next_id is now past the current chunk, so find the next one */
 714         start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
 715
 716         lock_flags = xfs_ilock_data_map_shared(quotip);
 717         if (!(quotip->i_df.if_flags & XFS_IFEXTENTS)) {
 718                 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
 719                 if (error)
 720                         return error;
 721         }
 722
 723         if (xfs_iext_lookup_extent(quotip, &quotip->i_df, start, &cur, &got)) {
 724                 /* contiguous chunk, bump startoff for the id calculation */
 725                 if (got.br_startoff < start)
 726                         got.br_startoff = start;
 727                 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
 728         } else {
 729                 error = -ENOENT;
 730         }
 731
 732         xfs_iunlock(quotip, lock_flags);
 733
 734         return error;
 735 }
 736
 737 /*
 738  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 739  * a locked dquot, doing an allocation (if requested) as needed.
 740  * When both an inode and an id are given, the inode's id takes precedence.
 741  * That is, if the id changes while we don't hold the ilock inside this
 742  * function, the new dquot is returned, not necessarily the one requested
 743  * in the id argument.
 744  */
 745 int
 746 xfs_qm_dqget(
 747         xfs_mount_t     *mp,
 748         xfs_inode_t     *ip,      /* locked inode (optional) */
 749         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 750         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 751         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 752         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 753 {
 754         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 755         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 756         struct xfs_dquot        *dqp;
 757         int                     error;
 758
 759         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 760         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 761             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 762             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 763                 return -ESRCH;
 764         }
 765
 766         ASSERT(type == XFS_DQ_USER ||
 767                type == XFS_DQ_PROJ ||
 768                type == XFS_DQ_GROUP);
 769         if (ip) {
 770                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 771                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 772         }
 773
 774 restart:
 775         mutex_lock(&qi->qi_tree_lock);
 776         dqp = radix_tree_lookup(tree, id);
 777         if (dqp) {
 778                 xfs_dqlock(dqp);
 779                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 780                         xfs_dqunlock(dqp);
 781                         mutex_unlock(&qi->qi_tree_lock);
 782                         trace_xfs_dqget_freeing(dqp);
 783                         delay(1);
 784                         goto restart;
 785                 }
 786
 787                 /* uninit / unused quota found in radix tree, keep looking  */
 788                 if (flags & XFS_QMOPT_DQNEXT) {
 789                         if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 790                                 xfs_dqunlock(dqp);
 791                                 mutex_unlock(&qi->qi_tree_lock);
 792                                 error = xfs_dq_get_next_id(mp, type, &id);
 793                                 if (error)
 794                                         return error;
 795                                 goto restart;
 796                         }
 797                 }
 798
 799                 dqp->q_nrefs++;
 800                 mutex_unlock(&qi->qi_tree_lock);
 801
 802                 trace_xfs_dqget_hit(dqp);
 803                 XFS_STATS_INC(mp, xs_qm_dqcachehits);
 804                 *O_dqpp = dqp;
 805                 return 0;
 806         }
 807         mutex_unlock(&qi->qi_tree_lock);
 808         XFS_STATS_INC(mp, xs_qm_dqcachemisses);
 809
 810         /*
 811          * Dquot cache miss. We don't want to keep the inode lock across
 812          * a (potential) disk read. Also we don't want to deal with the lock
 813          * ordering between quotainode and this inode. OTOH, dropping the inode
 814          * lock here means dealing with a chown that can happen before
 815          * we re-acquire the lock.
 816          */
 817         if (ip)
 818                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 819
 820         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 821
 822         if (ip)
 823                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 824
 825         /* If we are asked to find next active id, keep looking */
 826         if (error == -ENOENT && (flags & XFS_QMOPT_DQNEXT)) {
 827                 error = xfs_dq_get_next_id(mp, type, &id);
 828                 if (!error)
 829                         goto restart;
 830         }
 831
 832         if (error)
 833                 return error;
 834
 835         if (ip) {
 836                 /*
 837                  * A dquot could be attached to this inode by now, since
 838                  * we had dropped the ilock.
 839                  */
 840                 if (xfs_this_quota_on(mp, type)) {
 841                         struct xfs_dquot        *dqp1;
 842
 843                         dqp1 = xfs_inode_dquot(ip, type);
 844                         if (dqp1) {
 845                                 xfs_qm_dqdestroy(dqp);
 846                                 dqp = dqp1;
 847                                 xfs_dqlock(dqp);
 848                                 goto dqret;
 849                         }
 850                 } else {
 851                         /* inode stays locked on return */
 852                         xfs_qm_dqdestroy(dqp);
 853                         return -ESRCH;
 854                 }
 855         }
 856
 857         mutex_lock(&qi->qi_tree_lock);
 858         error = radix_tree_insert(tree, id, dqp);
 859         if (unlikely(error)) {
 860                 WARN_ON(error != -EEXIST);
 861
 862                 /*
 863                  * Duplicate found. Just throw away the new dquot and start
 864                  * over.
 865                  */
 866                 mutex_unlock(&qi->qi_tree_lock);
 867                 trace_xfs_dqget_dup(dqp);
 868                 xfs_qm_dqdestroy(dqp);
 869                 XFS_STATS_INC(mp, xs_qm_dquot_dups);
 870                 goto restart;
 871         }
 872
 873         /*
 874          * We return a locked dquot to the caller, with a reference taken
 875          */
 876         xfs_dqlock(dqp);
 877         dqp->q_nrefs = 1;
 878
 879         qi->qi_dquots++;
 880         mutex_unlock(&qi->qi_tree_lock);
 881
 882         /* If we are asked to find next active id, keep looking */
 883         if (flags & XFS_QMOPT_DQNEXT) {
 884                 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 885                         xfs_qm_dqput(dqp);
 886                         error = xfs_dq_get_next_id(mp, type, &id);
 887                         if (error)
 888                                 return error;
 889                         goto restart;
 890                 }
 891         }
 892
 893  dqret:
 894         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 895         trace_xfs_dqget_miss(dqp);
 896         *O_dqpp = dqp;
 897         return 0;
 898 }
 899
 900 /*
 901  * Release a reference to the dquot (decrement ref-count) and unlock it.
 902  *
 903  * If there is a group quota attached to this dquot, carefully release that
 904  * too without tripping over deadlocks'n'stuff.
 905  */
 906 void
 907 xfs_qm_dqput(
 908         struct xfs_dquot        *dqp)
 909 {
 910         ASSERT(dqp->q_nrefs > 0);
 911         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 912
 913         trace_xfs_dqput(dqp);
 914
 915         if (--dqp->q_nrefs == 0) {
 916                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 917                 trace_xfs_dqput_free(dqp);
 918
 919                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 920                         XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
 921         }
 922         xfs_dqunlock(dqp);
 923 }
 924
 925 /*
 926  * Release a dquot. Flush it if dirty, then dqput() it.
 927  * dquot must not be locked.
 928  */
 929 void
 930 xfs_qm_dqrele(
 931         xfs_dquot_t     *dqp)
 932 {
 933         if (!dqp)
 934                 return;
 935
 936         trace_xfs_dqrele(dqp);
 937
 938         xfs_dqlock(dqp);
 939         /*
 940          * We don't care to flush it if the dquot is dirty here.
 941          * That will create stutters that we want to avoid.
 942          * Instead we do a delayed write when we try to reclaim
 943          * a dirty dquot. Also xfs_sync will take part of the burden...
 944          */
 945         xfs_qm_dqput(dqp);
 946 }
 947
 948 /*
 949  * This is the dquot flushing I/O completion routine.  It is called
 950  * from interrupt level when the buffer containing the dquot is
 951  * flushed to disk.  It is responsible for removing the dquot logitem
 952  * from the AIL if it has not been re-logged, and unlocking the dquot's
 953  * flush lock. This behavior is very similar to that of inodes..
 954  */
 955 STATIC void
 956 xfs_qm_dqflush_done(
 957         struct xfs_buf          *bp,
 958         struct xfs_log_item     *lip)
 959 {
 960         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 961         xfs_dquot_t             *dqp = qip->qli_dquot;
 962         struct xfs_ail          *ailp = lip->li_ailp;
 963
 964         /*
 965          * We only want to pull the item from the AIL if its
 966          * location in the log has not changed since we started the flush.
 967          * Thus, we only bother if the dquot's lsn has
 968          * not changed. First we check the lsn outside the lock
 969          * since it's cheaper, and then we recheck while
 970          * holding the lock before removing the dquot from the AIL.
 971          */
 972         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 973             lip->li_lsn == qip->qli_flush_lsn) {
 974
 975                 /* xfs_trans_ail_delete() drops the AIL lock. */
 976                 spin_lock(&ailp->xa_lock);
 977                 if (lip->li_lsn == qip->qli_flush_lsn)
 978                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
 979                 else
 980                         spin_unlock(&ailp->xa_lock);
 981         }
 982
 983         /*
 984          * Release the dq's flush lock since we're done with it.
 985          */
 986         xfs_dqfunlock(dqp);
 987 }
 988
 989 /*
 990  * Write a modified dquot to disk.
 991  * The dquot must be locked and the flush lock too taken by caller.
 992  * The flush lock will not be unlocked until the dquot reaches the disk,
 993  * but the dquot is free to be unlocked and modified by the caller
 994  * in the interim. Dquot is still locked on return. This behavior is
 995  * identical to that of inodes.
 996  */
 997 int
 998 xfs_qm_dqflush(
 999         struct xfs_dquot        *dqp,
1000         struct xfs_buf          **bpp)
1001 {
1002         struct xfs_mount        *mp = dqp->q_mount;
1003         struct xfs_buf          *bp;
1004         struct xfs_disk_dquot   *ddqp;
1005         int                     error;
1006
1007         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1008         ASSERT(!completion_done(&dqp->q_flush));
1009
1010         trace_xfs_dqflush(dqp);
1011
1012         *bpp = NULL;
1013
1014         xfs_qm_dqunpin_wait(dqp);
1015
1016         /*
1017          * This may have been unpinned because the filesystem is shutting
1018          * down forcibly. If that's the case we must not write this dquot
1019          * to disk, because the log record didn't make it to disk.
1020          *
1021          * We also have to remove the log item from the AIL in this case,
1022          * as we wait for an emptry AIL as part of the unmount process.
1023          */
1024         if (XFS_FORCED_SHUTDOWN(mp)) {
1025                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
1026                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1027
1028                 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
1029
1030                 error = -EIO;
1031                 goto out_unlock;
1032         }
1033
1034         /*
1035          * Get the buffer containing the on-disk dquot
1036          */
1037         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1038                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
1039                                    &xfs_dquot_buf_ops);
1040         if (error)
1041                 goto out_unlock;
1042
1043         /*
1044          * Calculate the location of the dquot inside the buffer.
1045          */
1046         ddqp = bp->b_addr + dqp->q_bufoffset;
1047
1048         /*
1049          * A simple sanity check in case we got a corrupted dquot..
1050          */
1051         error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1052                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1053         if (error) {
1054                 xfs_buf_relse(bp);
1055                 xfs_dqfunlock(dqp);
1056                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1057                 return -EIO;
1058         }
1059
1060         /* This is the only portion of data that needs to persist */
1061         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1062
1063         /*
1064          * Clear the dirty field and remember the flush lsn for later use.
1065          */
1066         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1067
1068         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1069                                         &dqp->q_logitem.qli_item.li_lsn);
1070
1071         /*
1072          * copy the lsn into the on-disk dquot now while we have the in memory
1073          * dquot here. This can't be done later in the write verifier as we
1074          * can't get access to the log item at that point in time.
1075          *
1076          * We also calculate the CRC here so that the on-disk dquot in the
1077          * buffer always has a valid CRC. This ensures there is no possibility
1078          * of a dquot without an up-to-date CRC getting to disk.
1079          */
1080         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1081                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1082
1083                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1084                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1085                                  XFS_DQUOT_CRC_OFF);
1086         }
1087
1088         /*
1089          * Attach an iodone routine so that we can remove this dquot from the
1090          * AIL and release the flush lock once the dquot is synced to disk.
1091          */
1092         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1093                                   &dqp->q_logitem.qli_item);
1094
1095         /*
1096          * If the buffer is pinned then push on the log so we won't
1097          * get stuck waiting in the write for too long.
1098          */
1099         if (xfs_buf_ispinned(bp)) {
1100                 trace_xfs_dqflush_force(dqp);
1101                 xfs_log_force(mp, 0);
1102         }
1103
1104         trace_xfs_dqflush_done(dqp);
1105         *bpp = bp;
1106         return 0;
1107
1108 out_unlock:
1109         xfs_dqfunlock(dqp);
1110         return -EIO;
1111 }
1112
1113 /*
1114  * Lock two xfs_dquot structures.
1115  *
1116  * To avoid deadlocks we always lock the quota structure with
1117  * the lowerd id first.
1118  */
1119 void
1120 xfs_dqlock2(
1121         xfs_dquot_t     *d1,
1122         xfs_dquot_t     *d2)
1123 {
1124         if (d1 && d2) {
1125                 ASSERT(d1 != d2);
1126                 if (be32_to_cpu(d1->q_core.d_id) >
1127                     be32_to_cpu(d2->q_core.d_id)) {
1128                         mutex_lock(&d2->q_qlock);
1129                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1130                 } else {
1131                         mutex_lock(&d1->q_qlock);
1132                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1133                 }
1134         } else if (d1) {
1135                 mutex_lock(&d1->q_qlock);
1136         } else if (d2) {
1137                 mutex_lock(&d2->q_qlock);
1138         }
1139 }
1140
1141 int __init
1142 xfs_qm_init(void)
1143 {
1144         xfs_qm_dqzone =
1145                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1146         if (!xfs_qm_dqzone)
1147                 goto out;
1148
1149         xfs_qm_dqtrxzone =
1150                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1151         if (!xfs_qm_dqtrxzone)
1152                 goto out_free_dqzone;
1153
1154         return 0;
1155
1156 out_free_dqzone:
1157         kmem_zone_destroy(xfs_qm_dqzone);
1158 out:
1159         return -ENOMEM;
1160 }
1161
1162 void
1163 xfs_qm_exit(void)
1164 {
1165         kmem_zone_destroy(xfs_qm_dqtrxzone);
1166         kmem_zone_destroy(xfs_qm_dqzone);
1167 }