fs/xfs/quota/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_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_alloc.h"
  27 #include "xfs_quota.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_inode.h"
  31 #include "xfs_bmap.h"
  32 #include "xfs_rtalloc.h"
  33 #include "xfs_error.h"
  34 #include "xfs_itable.h"
  35 #include "xfs_attr.h"
  36 #include "xfs_buf_item.h"
  37 #include "xfs_trans_space.h"
  38 #include "xfs_trans_priv.h"
  39 #include "xfs_qm.h"
  40 #include "xfs_trace.h"
  41
  42
  43 /*
  44    LOCK ORDER
  45
  46    inode lock               (ilock)
  47    dquot hash-chain lock    (hashlock)
  48    xqm dquot freelist lock  (freelistlock
  49    mount's dquot list lock  (mplistlock)
  50    user dquot lock - lock ordering among dquots is based on the uid or gid
  51    group dquot lock - similar to udquots. Between the two dquots, the udquot
  52                       has to be locked first.
  53    pin lock - the dquot lock must be held to take this lock.
  54    flush lock - ditto.
  55 */
  56
  57 #ifdef DEBUG
  58 xfs_buftarg_t *xfs_dqerror_target;
  59 int xfs_do_dqerror;
  60 int xfs_dqreq_num;
  61 int xfs_dqerror_mod = 33;
  62 #endif
  63
  64 static struct lock_class_key xfs_dquot_other_class;
  65
  66 /*
  67  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  68  * we try to reclaim a free dquot if the number of incore dquots are above
  69  * a threshold.
  70  * The only field inside the core that gets initialized at this point
  71  * is the d_id field. The idea is to fill in the entire q_core
  72  * when we read in the on disk dquot.
  73  */
  74 STATIC xfs_dquot_t *
  75 xfs_qm_dqinit(
  76         xfs_mount_t  *mp,
  77         xfs_dqid_t   id,
  78         uint         type)
  79 {
  80         xfs_dquot_t     *dqp;
  81         boolean_t       brandnewdquot;
  82
  83         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  84         dqp->dq_flags = type;
  85         dqp->q_core.d_id = cpu_to_be32(id);
  86         dqp->q_mount = mp;
  87
  88         /*
  89          * No need to re-initialize these if this is a reclaimed dquot.
  90          */
  91         if (brandnewdquot) {
  92                 INIT_LIST_HEAD(&dqp->q_freelist);
  93                 mutex_init(&dqp->q_qlock);
  94                 init_waitqueue_head(&dqp->q_pinwait);
  95
  96                 /*
  97                  * Because we want to use a counting completion, complete
  98                  * the flush completion once to allow a single access to
  99                  * the flush completion without blocking.
 100                  */
 101                 init_completion(&dqp->q_flush);
 102                 complete(&dqp->q_flush);
 103
 104                 trace_xfs_dqinit(dqp);
 105         } else {
 106                 /*
 107                  * Only the q_core portion was zeroed in dqreclaim_one().
 108                  * So, we need to reset others.
 109                  */
 110                 dqp->q_nrefs = 0;
 111                 dqp->q_blkno = 0;
 112                 INIT_LIST_HEAD(&dqp->q_mplist);
 113                 INIT_LIST_HEAD(&dqp->q_hashlist);
 114                 dqp->q_bufoffset = 0;
 115                 dqp->q_fileoffset = 0;
 116                 dqp->q_transp = NULL;
 117                 dqp->q_gdquot = NULL;
 118                 dqp->q_res_bcount = 0;
 119                 dqp->q_res_icount = 0;
 120                 dqp->q_res_rtbcount = 0;
 121                 atomic_set(&dqp->q_pincount, 0);
 122                 dqp->q_hash = NULL;
 123                 ASSERT(list_empty(&dqp->q_freelist));
 124
 125                 trace_xfs_dqreuse(dqp);
 126         }
 127
 128         /*
 129          * In either case we need to make sure group quotas have a different
 130          * lock class than user quotas, to make sure lockdep knows we can
 131          * locks of one of each at the same time.
 132          */
 133         if (!(type & XFS_DQ_USER))
 134                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 135
 136         /*
 137          * log item gets initialized later
 138          */
 139         return (dqp);
 140 }
 141
 142 /*
 143  * This is called to free all the memory associated with a dquot
 144  */
 145 void
 146 xfs_qm_dqdestroy(
 147         xfs_dquot_t     *dqp)
 148 {
 149         ASSERT(list_empty(&dqp->q_freelist));
 150
 151         mutex_destroy(&dqp->q_qlock);
 152         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 153
 154         atomic_dec(&xfs_Gqm->qm_totaldquots);
 155 }
 156
 157 /*
 158  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 159  */
 160 STATIC void
 161 xfs_qm_dqinit_core(
 162         xfs_dqid_t      id,
 163         uint            type,
 164         xfs_dqblk_t     *d)
 165 {
 166         /*
 167          * Caller has zero'd the entire dquot 'chunk' already.
 168          */
 169         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 170         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 171         d->dd_diskdq.d_id = cpu_to_be32(id);
 172         d->dd_diskdq.d_flags = type;
 173 }
 174
 175 /*
 176  * If default limits are in force, push them into the dquot now.
 177  * We overwrite the dquot limits only if they are zero and this
 178  * is not the root dquot.
 179  */
 180 void
 181 xfs_qm_adjust_dqlimits(
 182         xfs_mount_t             *mp,
 183         xfs_disk_dquot_t        *d)
 184 {
 185         xfs_quotainfo_t         *q = mp->m_quotainfo;
 186
 187         ASSERT(d->d_id);
 188
 189         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 190                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 191         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 192                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 193         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 194                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 195         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 196                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 197         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 198                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 199         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 200                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 201 }
 202
 203 /*
 204  * Check the limits and timers of a dquot and start or reset timers
 205  * if necessary.
 206  * This gets called even when quota enforcement is OFF, which makes our
 207  * life a little less complicated. (We just don't reject any quota
 208  * reservations in that case, when enforcement is off).
 209  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 210  * enforcement's off.
 211  * In contrast, warnings are a little different in that they don't
 212  * 'automatically' get started when limits get exceeded.  They do
 213  * get reset to zero, however, when we find the count to be under
 214  * the soft limit (they are only ever set non-zero via userspace).
 215  */
 216 void
 217 xfs_qm_adjust_dqtimers(
 218         xfs_mount_t             *mp,
 219         xfs_disk_dquot_t        *d)
 220 {
 221         ASSERT(d->d_id);
 222
 223 #ifdef QUOTADEBUG
 224         if (d->d_blk_hardlimit)
 225                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 226                        be64_to_cpu(d->d_blk_hardlimit));
 227         if (d->d_ino_hardlimit)
 228                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 229                        be64_to_cpu(d->d_ino_hardlimit));
 230         if (d->d_rtb_hardlimit)
 231                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 232                        be64_to_cpu(d->d_rtb_hardlimit));
 233 #endif
 234         if (!d->d_btimer) {
 235                 if ((d->d_blk_softlimit &&
 236                      (be64_to_cpu(d->d_bcount) >=
 237                       be64_to_cpu(d->d_blk_softlimit))) ||
 238                     (d->d_blk_hardlimit &&
 239                      (be64_to_cpu(d->d_bcount) >=
 240                       be64_to_cpu(d->d_blk_hardlimit)))) {
 241                         d->d_btimer = cpu_to_be32(get_seconds() +
 242                                         mp->m_quotainfo->qi_btimelimit);
 243                 } else {
 244                         d->d_bwarns = 0;
 245                 }
 246         } else {
 247                 if ((!d->d_blk_softlimit ||
 248                      (be64_to_cpu(d->d_bcount) <
 249                       be64_to_cpu(d->d_blk_softlimit))) &&
 250                     (!d->d_blk_hardlimit ||
 251                     (be64_to_cpu(d->d_bcount) <
 252                      be64_to_cpu(d->d_blk_hardlimit)))) {
 253                         d->d_btimer = 0;
 254                 }
 255         }
 256
 257         if (!d->d_itimer) {
 258                 if ((d->d_ino_softlimit &&
 259                      (be64_to_cpu(d->d_icount) >=
 260                       be64_to_cpu(d->d_ino_softlimit))) ||
 261                     (d->d_ino_hardlimit &&
 262                      (be64_to_cpu(d->d_icount) >=
 263                       be64_to_cpu(d->d_ino_hardlimit)))) {
 264                         d->d_itimer = cpu_to_be32(get_seconds() +
 265                                         mp->m_quotainfo->qi_itimelimit);
 266                 } else {
 267                         d->d_iwarns = 0;
 268                 }
 269         } else {
 270                 if ((!d->d_ino_softlimit ||
 271                      (be64_to_cpu(d->d_icount) <
 272                       be64_to_cpu(d->d_ino_softlimit)))  &&
 273                     (!d->d_ino_hardlimit ||
 274                      (be64_to_cpu(d->d_icount) <
 275                       be64_to_cpu(d->d_ino_hardlimit)))) {
 276                         d->d_itimer = 0;
 277                 }
 278         }
 279
 280         if (!d->d_rtbtimer) {
 281                 if ((d->d_rtb_softlimit &&
 282                      (be64_to_cpu(d->d_rtbcount) >=
 283                       be64_to_cpu(d->d_rtb_softlimit))) ||
 284                     (d->d_rtb_hardlimit &&
 285                      (be64_to_cpu(d->d_rtbcount) >=
 286                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 287                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 288                                         mp->m_quotainfo->qi_rtbtimelimit);
 289                 } else {
 290                         d->d_rtbwarns = 0;
 291                 }
 292         } else {
 293                 if ((!d->d_rtb_softlimit ||
 294                      (be64_to_cpu(d->d_rtbcount) <
 295                       be64_to_cpu(d->d_rtb_softlimit))) &&
 296                     (!d->d_rtb_hardlimit ||
 297                      (be64_to_cpu(d->d_rtbcount) <
 298                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 299                         d->d_rtbtimer = 0;
 300                 }
 301         }
 302 }
 303
 304 /*
 305  * initialize a buffer full of dquots and log the whole thing
 306  */
 307 STATIC void
 308 xfs_qm_init_dquot_blk(
 309         xfs_trans_t     *tp,
 310         xfs_mount_t     *mp,
 311         xfs_dqid_t      id,
 312         uint            type,
 313         xfs_buf_t       *bp)
 314 {
 315         struct xfs_quotainfo    *q = mp->m_quotainfo;
 316         xfs_dqblk_t     *d;
 317         int             curid, i;
 318
 319         ASSERT(tp);
 320         ASSERT(XFS_BUF_ISBUSY(bp));
 321         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 322
 323         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 324
 325         /*
 326          * ID of the first dquot in the block - id's are zero based.
 327          */
 328         curid = id - (id % q->qi_dqperchunk);
 329         ASSERT(curid >= 0);
 330         memset(d, 0, BBTOB(q->qi_dqchunklen));
 331         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++)
 332                 xfs_qm_dqinit_core(curid, type, d);
 333         xfs_trans_dquot_buf(tp, bp,
 334                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 335                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 336                              XFS_BLF_GDQUOT_BUF)));
 337         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 338 }
 339
 340
 341
 342 /*
 343  * Allocate a block and fill it with dquots.
 344  * This is called when the bmapi finds a hole.
 345  */
 346 STATIC int
 347 xfs_qm_dqalloc(
 348         xfs_trans_t     **tpp,
 349         xfs_mount_t     *mp,
 350         xfs_dquot_t     *dqp,
 351         xfs_inode_t     *quotip,
 352         xfs_fileoff_t   offset_fsb,
 353         xfs_buf_t       **O_bpp)
 354 {
 355         xfs_fsblock_t   firstblock;
 356         xfs_bmap_free_t flist;
 357         xfs_bmbt_irec_t map;
 358         int             nmaps, error, committed;
 359         xfs_buf_t       *bp;
 360         xfs_trans_t     *tp = *tpp;
 361
 362         ASSERT(tp != NULL);
 363
 364         trace_xfs_dqalloc(dqp);
 365
 366         /*
 367          * Initialize the bmap freelist prior to calling bmapi code.
 368          */
 369         xfs_bmap_init(&flist, &firstblock);
 370         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 371         /*
 372          * Return if this type of quotas is turned off while we didn't
 373          * have an inode lock
 374          */
 375         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 376                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 377                 return (ESRCH);
 378         }
 379
 380         xfs_trans_ijoin_ref(tp, quotip, XFS_ILOCK_EXCL);
 381         nmaps = 1;
 382         if ((error = xfs_bmapi(tp, quotip,
 383                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 384                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 385                               &firstblock,
 386                               XFS_QM_DQALLOC_SPACE_RES(mp),
 387                               &map, &nmaps, &flist))) {
 388                 goto error0;
 389         }
 390         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 391         ASSERT(nmaps == 1);
 392         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 393                (map.br_startblock != HOLESTARTBLOCK));
 394
 395         /*
 396          * Keep track of the blkno to save a lookup later
 397          */
 398         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 399
 400         /* now we can just get the buffer (there's nothing to read yet) */
 401         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 402                                dqp->q_blkno,
 403                                mp->m_quotainfo->qi_dqchunklen,
 404                                0);
 405         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 406                 goto error1;
 407         /*
 408          * Make a chunk of dquots out of this buffer and log
 409          * the entire thing.
 410          */
 411         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 412                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 413
 414         /*
 415          * xfs_bmap_finish() may commit the current transaction and
 416          * start a second transaction if the freelist is not empty.
 417          *
 418          * Since we still want to modify this buffer, we need to
 419          * ensure that the buffer is not released on commit of
 420          * the first transaction and ensure the buffer is added to the
 421          * second transaction.
 422          *
 423          * If there is only one transaction then don't stop the buffer
 424          * from being released when it commits later on.
 425          */
 426
 427         xfs_trans_bhold(tp, bp);
 428
 429         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 430                 goto error1;
 431         }
 432
 433         if (committed) {
 434                 tp = *tpp;
 435                 xfs_trans_bjoin(tp, bp);
 436         } else {
 437                 xfs_trans_bhold_release(tp, bp);
 438         }
 439
 440         *O_bpp = bp;
 441         return 0;
 442
 443       error1:
 444         xfs_bmap_cancel(&flist);
 445       error0:
 446         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 447
 448         return (error);
 449 }
 450
 451 /*
 452  * Maps a dquot to the buffer containing its on-disk version.
 453  * This returns a ptr to the buffer containing the on-disk dquot
 454  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 455  */
 456 STATIC int
 457 xfs_qm_dqtobp(
 458         xfs_trans_t             **tpp,
 459         xfs_dquot_t             *dqp,
 460         xfs_disk_dquot_t        **O_ddpp,
 461         xfs_buf_t               **O_bpp,
 462         uint                    flags)
 463 {
 464         xfs_bmbt_irec_t map;
 465         int             nmaps = 1, error;
 466         xfs_buf_t       *bp;
 467         xfs_inode_t     *quotip = XFS_DQ_TO_QIP(dqp);
 468         xfs_mount_t     *mp = dqp->q_mount;
 469         xfs_disk_dquot_t *ddq;
 470         xfs_dqid_t      id = be32_to_cpu(dqp->q_core.d_id);
 471         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 472
 473         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 474
 475         xfs_ilock(quotip, XFS_ILOCK_SHARED);
 476         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 477                 /*
 478                  * Return if this type of quotas is turned off while we
 479                  * didn't have the quota inode lock.
 480                  */
 481                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 482                 return ESRCH;
 483         }
 484
 485         /*
 486          * Find the block map; no allocations yet
 487          */
 488         error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 489                           XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 490                           NULL, 0, &map, &nmaps, NULL);
 491
 492         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 493         if (error)
 494                 return error;
 495
 496         ASSERT(nmaps == 1);
 497         ASSERT(map.br_blockcount == 1);
 498
 499         /*
 500          * Offset of dquot in the (fixed sized) dquot chunk.
 501          */
 502         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 503                 sizeof(xfs_dqblk_t);
 504
 505         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 506         if (map.br_startblock == HOLESTARTBLOCK) {
 507                 /*
 508                  * We don't allocate unless we're asked to
 509                  */
 510                 if (!(flags & XFS_QMOPT_DQALLOC))
 511                         return ENOENT;
 512
 513                 ASSERT(tp);
 514                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 515                                         dqp->q_fileoffset, &bp);
 516                 if (error)
 517                         return error;
 518                 tp = *tpp;
 519         } else {
 520                 trace_xfs_dqtobp_read(dqp);
 521
 522                 /*
 523                  * store the blkno etc so that we don't have to do the
 524                  * mapping all the time
 525                  */
 526                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 527
 528                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 529                                            dqp->q_blkno,
 530                                            mp->m_quotainfo->qi_dqchunklen,
 531                                            0, &bp);
 532                 if (error || !bp)
 533                         return XFS_ERROR(error);
 534         }
 535
 536         ASSERT(XFS_BUF_ISBUSY(bp));
 537         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 538
 539         /*
 540          * calculate the location of the dquot inside the buffer.
 541          */
 542         ddq = (struct xfs_disk_dquot *)(XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 543
 544         /*
 545          * A simple sanity check in case we got a corrupted dquot...
 546          */
 547         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 548                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 549                            "dqtobp")) {
 550                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 551                         xfs_trans_brelse(tp, bp);
 552                         return XFS_ERROR(EIO);
 553                 }
 554                 XFS_BUF_BUSY(bp); /* We dirtied this */
 555         }
 556
 557         *O_bpp = bp;
 558         *O_ddpp = ddq;
 559
 560         return (0);
 561 }
 562
 563
 564 /*
 565  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 566  * and release the buffer immediately.
 567  *
 568  */
 569 /* ARGSUSED */
 570 STATIC int
 571 xfs_qm_dqread(
 572         xfs_trans_t     **tpp,
 573         xfs_dqid_t      id,
 574         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 575         uint            flags)
 576 {
 577         xfs_disk_dquot_t *ddqp;
 578         xfs_buf_t        *bp;
 579         int              error;
 580         xfs_trans_t      *tp;
 581
 582         ASSERT(tpp);
 583
 584         trace_xfs_dqread(dqp);
 585
 586         /*
 587          * get a pointer to the on-disk dquot and the buffer containing it
 588          * dqp already knows its own type (GROUP/USER).
 589          */
 590         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 591                 return (error);
 592         }
 593         tp = *tpp;
 594
 595         /* copy everything from disk dquot to the incore dquot */
 596         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 597         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 598         xfs_qm_dquot_logitem_init(dqp);
 599
 600         /*
 601          * Reservation counters are defined as reservation plus current usage
 602          * to avoid having to add everytime.
 603          */
 604         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 605         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 606         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 607
 608         /* Mark the buf so that this will stay incore a little longer */
 609         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 610
 611         /*
 612          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 613          * So we need to release with xfs_trans_brelse().
 614          * The strategy here is identical to that of inodes; we lock
 615          * the dquot in xfs_qm_dqget() before making it accessible to
 616          * others. This is because dquots, like inodes, need a good level of
 617          * concurrency, and we don't want to take locks on the entire buffers
 618          * for dquot accesses.
 619          * Note also that the dquot buffer may even be dirty at this point, if
 620          * this particular dquot was repaired. We still aren't afraid to
 621          * brelse it because we have the changes incore.
 622          */
 623         ASSERT(XFS_BUF_ISBUSY(bp));
 624         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 625         xfs_trans_brelse(tp, bp);
 626
 627         return (error);
 628 }
 629
 630
 631 /*
 632  * allocate an incore dquot from the kernel heap,
 633  * and fill its core with quota information kept on disk.
 634  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 635  * if it wasn't already allocated.
 636  */
 637 STATIC int
 638 xfs_qm_idtodq(
 639         xfs_mount_t     *mp,
 640         xfs_dqid_t      id,      /* gid or uid, depending on type */
 641         uint            type,    /* UDQUOT or GDQUOT */
 642         uint            flags,   /* DQALLOC, DQREPAIR */
 643         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 644 {
 645         xfs_dquot_t     *dqp;
 646         int             error;
 647         xfs_trans_t     *tp;
 648         int             cancelflags=0;
 649
 650         dqp = xfs_qm_dqinit(mp, id, type);
 651         tp = NULL;
 652         if (flags & XFS_QMOPT_DQALLOC) {
 653                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 654                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
 655                                 XFS_WRITE_LOG_RES(mp) +
 656                                 BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 +
 657                                 128,
 658                                 0,
 659                                 XFS_TRANS_PERM_LOG_RES,
 660                                 XFS_WRITE_LOG_COUNT);
 661                 if (error) {
 662                         cancelflags = 0;
 663                         goto error0;
 664                 }
 665                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 666         }
 667
 668         /*
 669          * Read it from disk; xfs_dqread() takes care of
 670          * all the necessary initialization of dquot's fields (locks, etc)
 671          */
 672         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 673                 /*
 674                  * This can happen if quotas got turned off (ESRCH),
 675                  * or if the dquot didn't exist on disk and we ask to
 676                  * allocate (ENOENT).
 677                  */
 678                 trace_xfs_dqread_fail(dqp);
 679                 cancelflags |= XFS_TRANS_ABORT;
 680                 goto error0;
 681         }
 682         if (tp) {
 683                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 684                         goto error1;
 685         }
 686
 687         *O_dqpp = dqp;
 688         return (0);
 689
 690  error0:
 691         ASSERT(error);
 692         if (tp)
 693                 xfs_trans_cancel(tp, cancelflags);
 694  error1:
 695         xfs_qm_dqdestroy(dqp);
 696         *O_dqpp = NULL;
 697         return (error);
 698 }
 699
 700 /*
 701  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 702  * hashtables for user and group dquots; and, these are global tables
 703  * inside the XQM, not per-filesystem tables.
 704  * The hash chain must be locked by caller, and it is left locked
 705  * on return. Returning dquot is locked.
 706  */
 707 STATIC int
 708 xfs_qm_dqlookup(
 709         xfs_mount_t             *mp,
 710         xfs_dqid_t              id,
 711         xfs_dqhash_t            *qh,
 712         xfs_dquot_t             **O_dqpp)
 713 {
 714         xfs_dquot_t             *dqp;
 715         uint                    flist_locked;
 716
 717         ASSERT(mutex_is_locked(&qh->qh_lock));
 718
 719         flist_locked = B_FALSE;
 720
 721         /*
 722          * Traverse the hashchain looking for a match
 723          */
 724         list_for_each_entry(dqp, &qh->qh_list, q_hashlist) {
 725                 /*
 726                  * We already have the hashlock. We don't need the
 727                  * dqlock to look at the id field of the dquot, since the
 728                  * id can't be modified without the hashlock anyway.
 729                  */
 730                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 731                         trace_xfs_dqlookup_found(dqp);
 732
 733                         /*
 734                          * All in core dquots must be on the dqlist of mp
 735                          */
 736                         ASSERT(!list_empty(&dqp->q_mplist));
 737
 738                         xfs_dqlock(dqp);
 739                         if (dqp->q_nrefs == 0) {
 740                                 ASSERT(!list_empty(&dqp->q_freelist));
 741                                 if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
 742                                         trace_xfs_dqlookup_want(dqp);
 743
 744                                         /*
 745                                          * We may have raced with dqreclaim_one()
 746                                          * (and lost). So, flag that we don't
 747                                          * want the dquot to be reclaimed.
 748                                          */
 749                                         dqp->dq_flags |= XFS_DQ_WANT;
 750                                         xfs_dqunlock(dqp);
 751                                         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
 752                                         xfs_dqlock(dqp);
 753                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 754                                 }
 755                                 flist_locked = B_TRUE;
 756                         }
 757
 758                         /*
 759                          * id couldn't have changed; we had the hashlock all
 760                          * along
 761                          */
 762                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 763
 764                         if (flist_locked) {
 765                                 if (dqp->q_nrefs != 0) {
 766                                         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 767                                         flist_locked = B_FALSE;
 768                                 } else {
 769                                         /* take it off the freelist */
 770                                         trace_xfs_dqlookup_freelist(dqp);
 771                                         list_del_init(&dqp->q_freelist);
 772                                         xfs_Gqm->qm_dqfrlist_cnt--;
 773                                 }
 774                         }
 775
 776                         XFS_DQHOLD(dqp);
 777
 778                         if (flist_locked)
 779                                 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 780                         /*
 781                          * move the dquot to the front of the hashchain
 782                          */
 783                         ASSERT(mutex_is_locked(&qh->qh_lock));
 784                         list_move(&dqp->q_hashlist, &qh->qh_list);
 785                         trace_xfs_dqlookup_done(dqp);
 786                         *O_dqpp = dqp;
 787                         return 0;
 788                 }
 789         }
 790
 791         *O_dqpp = NULL;
 792         ASSERT(mutex_is_locked(&qh->qh_lock));
 793         return (1);
 794 }
 795
 796 /*
 797  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 798  * a locked dquot, doing an allocation (if requested) as needed.
 799  * When both an inode and an id are given, the inode's id takes precedence.
 800  * That is, if the id changes while we don't hold the ilock inside this
 801  * function, the new dquot is returned, not necessarily the one requested
 802  * in the id argument.
 803  */
 804 int
 805 xfs_qm_dqget(
 806         xfs_mount_t     *mp,
 807         xfs_inode_t     *ip,      /* locked inode (optional) */
 808         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 809         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 810         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 811         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 812 {
 813         xfs_dquot_t     *dqp;
 814         xfs_dqhash_t    *h;
 815         uint            version;
 816         int             error;
 817
 818         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 819         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 820             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 821             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 822                 return (ESRCH);
 823         }
 824         h = XFS_DQ_HASH(mp, id, type);
 825
 826 #ifdef DEBUG
 827         if (xfs_do_dqerror) {
 828                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 829                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 830                         cmn_err(CE_DEBUG, "Returning error in dqget");
 831                         return (EIO);
 832                 }
 833         }
 834 #endif
 835
 836  again:
 837
 838 #ifdef DEBUG
 839         ASSERT(type == XFS_DQ_USER ||
 840                type == XFS_DQ_PROJ ||
 841                type == XFS_DQ_GROUP);
 842         if (ip) {
 843                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 844                 if (type == XFS_DQ_USER)
 845                         ASSERT(ip->i_udquot == NULL);
 846                 else
 847                         ASSERT(ip->i_gdquot == NULL);
 848         }
 849 #endif
 850         mutex_lock(&h->qh_lock);
 851
 852         /*
 853          * Look in the cache (hashtable).
 854          * The chain is kept locked during lookup.
 855          */
 856         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 857                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 858                 /*
 859                  * The dquot was found, moved to the front of the chain,
 860                  * taken off the freelist if it was on it, and locked
 861                  * at this point. Just unlock the hashchain and return.
 862                  */
 863                 ASSERT(*O_dqpp);
 864                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 865                 mutex_unlock(&h->qh_lock);
 866                 trace_xfs_dqget_hit(*O_dqpp);
 867                 return (0);     /* success */
 868         }
 869         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 870
 871         /*
 872          * Dquot cache miss. We don't want to keep the inode lock across
 873          * a (potential) disk read. Also we don't want to deal with the lock
 874          * ordering between quotainode and this inode. OTOH, dropping the inode
 875          * lock here means dealing with a chown that can happen before
 876          * we re-acquire the lock.
 877          */
 878         if (ip)
 879                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 880         /*
 881          * Save the hashchain version stamp, and unlock the chain, so that
 882          * we don't keep the lock across a disk read
 883          */
 884         version = h->qh_version;
 885         mutex_unlock(&h->qh_lock);
 886
 887         /*
 888          * Allocate the dquot on the kernel heap, and read the ondisk
 889          * portion off the disk. Also, do all the necessary initialization
 890          * This can return ENOENT if dquot didn't exist on disk and we didn't
 891          * ask it to allocate; ESRCH if quotas got turned off suddenly.
 892          */
 893         if ((error = xfs_qm_idtodq(mp, id, type,
 894                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
 895                                            XFS_QMOPT_DOWARN),
 896                                   &dqp))) {
 897                 if (ip)
 898                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 899                 return (error);
 900         }
 901
 902         /*
 903          * See if this is mount code calling to look at the overall quota limits
 904          * which are stored in the id == 0 user or group's dquot.
 905          * Since we may not have done a quotacheck by this point, just return
 906          * the dquot without attaching it to any hashtables, lists, etc, or even
 907          * taking a reference.
 908          * The caller must dqdestroy this once done.
 909          */
 910         if (flags & XFS_QMOPT_DQSUSER) {
 911                 ASSERT(id == 0);
 912                 ASSERT(! ip);
 913                 goto dqret;
 914         }
 915
 916         /*
 917          * Dquot lock comes after hashlock in the lock ordering
 918          */
 919         if (ip) {
 920                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 921
 922                 /*
 923                  * A dquot could be attached to this inode by now, since
 924                  * we had dropped the ilock.
 925                  */
 926                 if (type == XFS_DQ_USER) {
 927                         if (!XFS_IS_UQUOTA_ON(mp)) {
 928                                 /* inode stays locked on return */
 929                                 xfs_qm_dqdestroy(dqp);
 930                                 return XFS_ERROR(ESRCH);
 931                         }
 932                         if (ip->i_udquot) {
 933                                 xfs_qm_dqdestroy(dqp);
 934                                 dqp = ip->i_udquot;
 935                                 xfs_dqlock(dqp);
 936                                 goto dqret;
 937                         }
 938                 } else {
 939                         if (!XFS_IS_OQUOTA_ON(mp)) {
 940                                 /* inode stays locked on return */
 941                                 xfs_qm_dqdestroy(dqp);
 942                                 return XFS_ERROR(ESRCH);
 943                         }
 944                         if (ip->i_gdquot) {
 945                                 xfs_qm_dqdestroy(dqp);
 946                                 dqp = ip->i_gdquot;
 947                                 xfs_dqlock(dqp);
 948                                 goto dqret;
 949                         }
 950                 }
 951         }
 952
 953         /*
 954          * Hashlock comes after ilock in lock order
 955          */
 956         mutex_lock(&h->qh_lock);
 957         if (version != h->qh_version) {
 958                 xfs_dquot_t *tmpdqp;
 959                 /*
 960                  * Now, see if somebody else put the dquot in the
 961                  * hashtable before us. This can happen because we didn't
 962                  * keep the hashchain lock. We don't have to worry about
 963                  * lock order between the two dquots here since dqp isn't
 964                  * on any findable lists yet.
 965                  */
 966                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
 967                         /*
 968                          * Duplicate found. Just throw away the new dquot
 969                          * and start over.
 970                          */
 971                         xfs_qm_dqput(tmpdqp);
 972                         mutex_unlock(&h->qh_lock);
 973                         xfs_qm_dqdestroy(dqp);
 974                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
 975                         goto again;
 976                 }
 977         }
 978
 979         /*
 980          * Put the dquot at the beginning of the hash-chain and mp's list
 981          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
 982          */
 983         ASSERT(mutex_is_locked(&h->qh_lock));
 984         dqp->q_hash = h;
 985         list_add(&dqp->q_hashlist, &h->qh_list);
 986         h->qh_version++;
 987
 988         /*
 989          * Attach this dquot to this filesystem's list of all dquots,
 990          * kept inside the mount structure in m_quotainfo field
 991          */
 992         mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
 993
 994         /*
 995          * We return a locked dquot to the caller, with a reference taken
 996          */
 997         xfs_dqlock(dqp);
 998         dqp->q_nrefs = 1;
 999
1000         list_add(&dqp->q_mplist, &mp->m_quotainfo->qi_dqlist);
1001         mp->m_quotainfo->qi_dquots++;
1002         mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1003         mutex_unlock(&h->qh_lock);
1004  dqret:
1005         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1006         trace_xfs_dqget_miss(dqp);
1007         *O_dqpp = dqp;
1008         return (0);
1009 }
1010
1011
1012 /*
1013  * Release a reference to the dquot (decrement ref-count)
1014  * and unlock it. If there is a group quota attached to this
1015  * dquot, carefully release that too without tripping over
1016  * deadlocks'n'stuff.
1017  */
1018 void
1019 xfs_qm_dqput(
1020         xfs_dquot_t     *dqp)
1021 {
1022         xfs_dquot_t     *gdqp;
1023
1024         ASSERT(dqp->q_nrefs > 0);
1025         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1026
1027         trace_xfs_dqput(dqp);
1028
1029         if (dqp->q_nrefs != 1) {
1030                 dqp->q_nrefs--;
1031                 xfs_dqunlock(dqp);
1032                 return;
1033         }
1034
1035         /*
1036          * drop the dqlock and acquire the freelist and dqlock
1037          * in the right order; but try to get it out-of-order first
1038          */
1039         if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
1040                 trace_xfs_dqput_wait(dqp);
1041                 xfs_dqunlock(dqp);
1042                 mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1043                 xfs_dqlock(dqp);
1044         }
1045
1046         while (1) {
1047                 gdqp = NULL;
1048
1049                 /* We can't depend on nrefs being == 1 here */
1050                 if (--dqp->q_nrefs == 0) {
1051                         trace_xfs_dqput_free(dqp);
1052
1053                         list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
1054                         xfs_Gqm->qm_dqfrlist_cnt++;
1055
1056                         /*
1057                          * If we just added a udquot to the freelist, then
1058                          * we want to release the gdquot reference that
1059                          * it (probably) has. Otherwise it'll keep the
1060                          * gdquot from getting reclaimed.
1061                          */
1062                         if ((gdqp = dqp->q_gdquot)) {
1063                                 /*
1064                                  * Avoid a recursive dqput call
1065                                  */
1066                                 xfs_dqlock(gdqp);
1067                                 dqp->q_gdquot = NULL;
1068                         }
1069                 }
1070                 xfs_dqunlock(dqp);
1071
1072                 /*
1073                  * If we had a group quota inside the user quota as a hint,
1074                  * release it now.
1075                  */
1076                 if (! gdqp)
1077                         break;
1078                 dqp = gdqp;
1079         }
1080         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1081 }
1082
1083 /*
1084  * Release a dquot. Flush it if dirty, then dqput() it.
1085  * dquot must not be locked.
1086  */
1087 void
1088 xfs_qm_dqrele(
1089         xfs_dquot_t     *dqp)
1090 {
1091         if (!dqp)
1092                 return;
1093
1094         trace_xfs_dqrele(dqp);
1095
1096         xfs_dqlock(dqp);
1097         /*
1098          * We don't care to flush it if the dquot is dirty here.
1099          * That will create stutters that we want to avoid.
1100          * Instead we do a delayed write when we try to reclaim
1101          * a dirty dquot. Also xfs_sync will take part of the burden...
1102          */
1103         xfs_qm_dqput(dqp);
1104 }
1105
1106 /*
1107  * This is the dquot flushing I/O completion routine.  It is called
1108  * from interrupt level when the buffer containing the dquot is
1109  * flushed to disk.  It is responsible for removing the dquot logitem
1110  * from the AIL if it has not been re-logged, and unlocking the dquot's
1111  * flush lock. This behavior is very similar to that of inodes..
1112  */
1113 STATIC void
1114 xfs_qm_dqflush_done(
1115         struct xfs_buf          *bp,
1116         struct xfs_log_item     *lip)
1117 {
1118         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
1119         xfs_dquot_t             *dqp = qip->qli_dquot;
1120         struct xfs_ail          *ailp = lip->li_ailp;
1121
1122         /*
1123          * We only want to pull the item from the AIL if its
1124          * location in the log has not changed since we started the flush.
1125          * Thus, we only bother if the dquot's lsn has
1126          * not changed. First we check the lsn outside the lock
1127          * since it's cheaper, and then we recheck while
1128          * holding the lock before removing the dquot from the AIL.
1129          */
1130         if ((lip->li_flags & XFS_LI_IN_AIL) &&
1131             lip->li_lsn == qip->qli_flush_lsn) {
1132
1133                 /* xfs_trans_ail_delete() drops the AIL lock. */
1134                 spin_lock(&ailp->xa_lock);
1135                 if (lip->li_lsn == qip->qli_flush_lsn)
1136                         xfs_trans_ail_delete(ailp, lip);
1137                 else
1138                         spin_unlock(&ailp->xa_lock);
1139         }
1140
1141         /*
1142          * Release the dq's flush lock since we're done with it.
1143          */
1144         xfs_dqfunlock(dqp);
1145 }
1146
1147 /*
1148  * Write a modified dquot to disk.
1149  * The dquot must be locked and the flush lock too taken by caller.
1150  * The flush lock will not be unlocked until the dquot reaches the disk,
1151  * but the dquot is free to be unlocked and modified by the caller
1152  * in the interim. Dquot is still locked on return. This behavior is
1153  * identical to that of inodes.
1154  */
1155 int
1156 xfs_qm_dqflush(
1157         xfs_dquot_t             *dqp,
1158         uint                    flags)
1159 {
1160         struct xfs_mount        *mp = dqp->q_mount;
1161         struct xfs_buf          *bp;
1162         struct xfs_disk_dquot   *ddqp;
1163         int                     error;
1164
1165         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1166         ASSERT(!completion_done(&dqp->q_flush));
1167
1168         trace_xfs_dqflush(dqp);
1169
1170         /*
1171          * If not dirty, or it's pinned and we are not supposed to block, nada.
1172          */
1173         if (!XFS_DQ_IS_DIRTY(dqp) ||
1174             (!(flags & SYNC_WAIT) && atomic_read(&dqp->q_pincount) > 0)) {
1175                 xfs_dqfunlock(dqp);
1176                 return 0;
1177         }
1178         xfs_qm_dqunpin_wait(dqp);
1179
1180         /*
1181          * This may have been unpinned because the filesystem is shutting
1182          * down forcibly. If that's the case we must not write this dquot
1183          * to disk, because the log record didn't make it to disk!
1184          */
1185         if (XFS_FORCED_SHUTDOWN(mp)) {
1186                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1187                 xfs_dqfunlock(dqp);
1188                 return XFS_ERROR(EIO);
1189         }
1190
1191         /*
1192          * Get the buffer containing the on-disk dquot
1193          */
1194         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1195                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1196         if (error) {
1197                 ASSERT(error != ENOENT);
1198                 xfs_dqfunlock(dqp);
1199                 return error;
1200         }
1201
1202         /*
1203          * Calculate the location of the dquot inside the buffer.
1204          */
1205         ddqp = (struct xfs_disk_dquot *)(XFS_BUF_PTR(bp) + dqp->q_bufoffset);
1206
1207         /*
1208          * A simple sanity check in case we got a corrupted dquot..
1209          */
1210         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1211                            XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1212                 xfs_buf_relse(bp);
1213                 xfs_dqfunlock(dqp);
1214                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1215                 return XFS_ERROR(EIO);
1216         }
1217
1218         /* This is the only portion of data that needs to persist */
1219         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1220
1221         /*
1222          * Clear the dirty field and remember the flush lsn for later use.
1223          */
1224         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1225
1226         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1227                                         &dqp->q_logitem.qli_item.li_lsn);
1228
1229         /*
1230          * Attach an iodone routine so that we can remove this dquot from the
1231          * AIL and release the flush lock once the dquot is synced to disk.
1232          */
1233         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1234                                   &dqp->q_logitem.qli_item);
1235
1236         /*
1237          * If the buffer is pinned then push on the log so we won't
1238          * get stuck waiting in the write for too long.
1239          */
1240         if (XFS_BUF_ISPINNED(bp)) {
1241                 trace_xfs_dqflush_force(dqp);
1242                 xfs_log_force(mp, 0);
1243         }
1244
1245         if (flags & SYNC_WAIT)
1246                 error = xfs_bwrite(mp, bp);
1247         else
1248                 xfs_bdwrite(mp, bp);
1249
1250         trace_xfs_dqflush_done(dqp);
1251
1252         /*
1253          * dqp is still locked, but caller is free to unlock it now.
1254          */
1255         return error;
1256
1257 }
1258
1259 int
1260 xfs_qm_dqlock_nowait(
1261         xfs_dquot_t *dqp)
1262 {
1263         return mutex_trylock(&dqp->q_qlock);
1264 }
1265
1266 void
1267 xfs_dqlock(
1268         xfs_dquot_t *dqp)
1269 {
1270         mutex_lock(&dqp->q_qlock);
1271 }
1272
1273 void
1274 xfs_dqunlock(
1275         xfs_dquot_t *dqp)
1276 {
1277         mutex_unlock(&(dqp->q_qlock));
1278         if (dqp->q_logitem.qli_dquot == dqp) {
1279                 /* Once was dqp->q_mount, but might just have been cleared */
1280                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1281                                         (xfs_log_item_t*)&(dqp->q_logitem));
1282         }
1283 }
1284
1285
1286 void
1287 xfs_dqunlock_nonotify(
1288         xfs_dquot_t *dqp)
1289 {
1290         mutex_unlock(&(dqp->q_qlock));
1291 }
1292
1293 /*
1294  * Lock two xfs_dquot structures.
1295  *
1296  * To avoid deadlocks we always lock the quota structure with
1297  * the lowerd id first.
1298  */
1299 void
1300 xfs_dqlock2(
1301         xfs_dquot_t     *d1,
1302         xfs_dquot_t     *d2)
1303 {
1304         if (d1 && d2) {
1305                 ASSERT(d1 != d2);
1306                 if (be32_to_cpu(d1->q_core.d_id) >
1307                     be32_to_cpu(d2->q_core.d_id)) {
1308                         mutex_lock(&d2->q_qlock);
1309                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1310                 } else {
1311                         mutex_lock(&d1->q_qlock);
1312                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1313                 }
1314         } else if (d1) {
1315                 mutex_lock(&d1->q_qlock);
1316         } else if (d2) {
1317                 mutex_lock(&d2->q_qlock);
1318         }
1319 }
1320
1321
1322 /*
1323  * Take a dquot out of the mount's dqlist as well as the hashlist.
1324  * This is called via unmount as well as quotaoff, and the purge
1325  * will always succeed unless there are soft (temp) references
1326  * outstanding.
1327  *
1328  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1329  * that we're returning! XXXsup - not cool.
1330  */
1331 /* ARGSUSED */
1332 int
1333 xfs_qm_dqpurge(
1334         xfs_dquot_t     *dqp)
1335 {
1336         xfs_dqhash_t    *qh = dqp->q_hash;
1337         xfs_mount_t     *mp = dqp->q_mount;
1338
1339         ASSERT(mutex_is_locked(&mp->m_quotainfo->qi_dqlist_lock));
1340         ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1341
1342         xfs_dqlock(dqp);
1343         /*
1344          * We really can't afford to purge a dquot that is
1345          * referenced, because these are hard refs.
1346          * It shouldn't happen in general because we went thru _all_ inodes in
1347          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1348          * However it is possible that we have dquots with temporary
1349          * references that are not attached to an inode. e.g. see xfs_setattr().
1350          */
1351         if (dqp->q_nrefs != 0) {
1352                 xfs_dqunlock(dqp);
1353                 mutex_unlock(&dqp->q_hash->qh_lock);
1354                 return (1);
1355         }
1356
1357         ASSERT(!list_empty(&dqp->q_freelist));
1358
1359         /*
1360          * If we're turning off quotas, we have to make sure that, for
1361          * example, we don't delete quota disk blocks while dquots are
1362          * in the process of getting written to those disk blocks.
1363          * This dquot might well be on AIL, and we can't leave it there
1364          * if we're turning off quotas. Basically, we need this flush
1365          * lock, and are willing to block on it.
1366          */
1367         if (!xfs_dqflock_nowait(dqp)) {
1368                 /*
1369                  * Block on the flush lock after nudging dquot buffer,
1370                  * if it is incore.
1371                  */
1372                 xfs_qm_dqflock_pushbuf_wait(dqp);
1373         }
1374
1375         /*
1376          * XXXIf we're turning this type of quotas off, we don't care
1377          * about the dirty metadata sitting in this dquot. OTOH, if
1378          * we're unmounting, we do care, so we flush it and wait.
1379          */
1380         if (XFS_DQ_IS_DIRTY(dqp)) {
1381                 int     error;
1382
1383                 /* dqflush unlocks dqflock */
1384                 /*
1385                  * Given that dqpurge is a very rare occurrence, it is OK
1386                  * that we're holding the hashlist and mplist locks
1387                  * across the disk write. But, ... XXXsup
1388                  *
1389                  * We don't care about getting disk errors here. We need
1390                  * to purge this dquot anyway, so we go ahead regardless.
1391                  */
1392                 error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1393                 if (error)
1394                         xfs_fs_cmn_err(CE_WARN, mp,
1395                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1396                 xfs_dqflock(dqp);
1397         }
1398         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1399         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1400                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1401
1402         list_del_init(&dqp->q_hashlist);
1403         qh->qh_version++;
1404         list_del_init(&dqp->q_mplist);
1405         mp->m_quotainfo->qi_dqreclaims++;
1406         mp->m_quotainfo->qi_dquots--;
1407         /*
1408          * XXX Move this to the front of the freelist, if we can get the
1409          * freelist lock.
1410          */
1411         ASSERT(!list_empty(&dqp->q_freelist));
1412
1413         dqp->q_mount = NULL;
1414         dqp->q_hash = NULL;
1415         dqp->dq_flags = XFS_DQ_INACTIVE;
1416         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1417         xfs_dqfunlock(dqp);
1418         xfs_dqunlock(dqp);
1419         mutex_unlock(&qh->qh_lock);
1420         return (0);
1421 }
1422
1423
1424 #ifdef QUOTADEBUG
1425 void
1426 xfs_qm_dqprint(xfs_dquot_t *dqp)
1427 {
1428         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1429         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1430                 (int)be32_to_cpu(dqp->q_core.d_id));
1431         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1432         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1433         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1434         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1435         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1436                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1437                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1438         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1439                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1440                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1441         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1442                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1443                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1444         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1445                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1446                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1447         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1448                 be64_to_cpu(dqp->q_core.d_bcount),
1449                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1450         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1451                 be64_to_cpu(dqp->q_core.d_icount),
1452                 (int)be64_to_cpu(dqp->q_core.d_icount));
1453         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1454                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1455         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1456                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1457         cmn_err(CE_DEBUG, "---------------------------");
1458 }
1459 #endif
1460
1461 /*
1462  * Give the buffer a little push if it is incore and
1463  * wait on the flush lock.
1464  */
1465 void
1466 xfs_qm_dqflock_pushbuf_wait(
1467         xfs_dquot_t     *dqp)
1468 {
1469         xfs_mount_t     *mp = dqp->q_mount;
1470         xfs_buf_t       *bp;
1471
1472         /*
1473          * Check to see if the dquot has been flushed delayed
1474          * write.  If so, grab its buffer and send it
1475          * out immediately.  We'll be able to acquire
1476          * the flush lock when the I/O completes.
1477          */
1478         bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1479                         mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1480         if (!bp)
1481                 goto out_lock;
1482
1483         if (XFS_BUF_ISDELAYWRITE(bp)) {
1484                 if (XFS_BUF_ISPINNED(bp))
1485                         xfs_log_force(mp, 0);
1486                 xfs_buf_delwri_promote(bp);
1487                 wake_up_process(bp->b_target->bt_task);
1488         }
1489         xfs_buf_relse(bp);
1490 out_lock:
1491         xfs_dqflock(dqp);
1492 }