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_dir2.h"
  27 #include "xfs_alloc.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_quota.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_btree.h"
  39 #include "xfs_ialloc.h"
  40 #include "xfs_bmap.h"
  41 #include "xfs_rtalloc.h"
  42 #include "xfs_error.h"
  43 #include "xfs_itable.h"
  44 #include "xfs_rw.h"
  45 #include "xfs_attr.h"
  46 #include "xfs_buf_item.h"
  47 #include "xfs_trans_space.h"
  48 #include "xfs_trans_priv.h"
  49 #include "xfs_qm.h"
  50 #include "xfs_trace.h"
  51
  52
  53 /*
  54    LOCK ORDER
  55
  56    inode lock               (ilock)
  57    dquot hash-chain lock    (hashlock)
  58    xqm dquot freelist lock  (freelistlock
  59    mount's dquot list lock  (mplistlock)
  60    user dquot lock - lock ordering among dquots is based on the uid or gid
  61    group dquot lock - similar to udquots. Between the two dquots, the udquot
  62                       has to be locked first.
  63    pin lock - the dquot lock must be held to take this lock.
  64    flush lock - ditto.
  65 */
  66
  67 STATIC void             xfs_qm_dqflush_done(xfs_buf_t *, xfs_dq_logitem_t *);
  68
  69 #ifdef DEBUG
  70 xfs_buftarg_t *xfs_dqerror_target;
  71 int xfs_do_dqerror;
  72 int xfs_dqreq_num;
  73 int xfs_dqerror_mod = 33;
  74 #endif
  75
  76 static struct lock_class_key xfs_dquot_other_class;
  77
  78 /*
  79  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  80  * we try to reclaim a free dquot if the number of incore dquots are above
  81  * a threshold.
  82  * The only field inside the core that gets initialized at this point
  83  * is the d_id field. The idea is to fill in the entire q_core
  84  * when we read in the on disk dquot.
  85  */
  86 STATIC xfs_dquot_t *
  87 xfs_qm_dqinit(
  88         xfs_mount_t  *mp,
  89         xfs_dqid_t   id,
  90         uint         type)
  91 {
  92         xfs_dquot_t     *dqp;
  93         boolean_t       brandnewdquot;
  94
  95         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  96         dqp->dq_flags = type;
  97         dqp->q_core.d_id = cpu_to_be32(id);
  98         dqp->q_mount = mp;
  99
 100         /*
 101          * No need to re-initialize these if this is a reclaimed dquot.
 102          */
 103         if (brandnewdquot) {
 104                 dqp->dq_flnext = dqp->dq_flprev = dqp;
 105                 mutex_init(&dqp->q_qlock);
 106                 init_waitqueue_head(&dqp->q_pinwait);
 107
 108                 /*
 109                  * Because we want to use a counting completion, complete
 110                  * the flush completion once to allow a single access to
 111                  * the flush completion without blocking.
 112                  */
 113                 init_completion(&dqp->q_flush);
 114                 complete(&dqp->q_flush);
 115
 116                 trace_xfs_dqinit(dqp);
 117         } else {
 118                 /*
 119                  * Only the q_core portion was zeroed in dqreclaim_one().
 120                  * So, we need to reset others.
 121                  */
 122                  dqp->q_nrefs = 0;
 123                  dqp->q_blkno = 0;
 124                  dqp->MPL_NEXT = dqp->HL_NEXT = NULL;
 125                  dqp->HL_PREVP = dqp->MPL_PREVP = NULL;
 126                  dqp->q_bufoffset = 0;
 127                  dqp->q_fileoffset = 0;
 128                  dqp->q_transp = NULL;
 129                  dqp->q_gdquot = NULL;
 130                  dqp->q_res_bcount = 0;
 131                  dqp->q_res_icount = 0;
 132                  dqp->q_res_rtbcount = 0;
 133                  atomic_set(&dqp->q_pincount, 0);
 134                  dqp->q_hash = NULL;
 135                  ASSERT(dqp->dq_flnext == dqp->dq_flprev);
 136
 137                 trace_xfs_dqreuse(dqp);
 138         }
 139
 140         /*
 141          * In either case we need to make sure group quotas have a different
 142          * lock class than user quotas, to make sure lockdep knows we can
 143          * locks of one of each at the same time.
 144          */
 145         if (!(type & XFS_DQ_USER))
 146                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 147
 148         /*
 149          * log item gets initialized later
 150          */
 151         return (dqp);
 152 }
 153
 154 /*
 155  * This is called to free all the memory associated with a dquot
 156  */
 157 void
 158 xfs_qm_dqdestroy(
 159         xfs_dquot_t     *dqp)
 160 {
 161         ASSERT(! XFS_DQ_IS_ON_FREELIST(dqp));
 162
 163         mutex_destroy(&dqp->q_qlock);
 164         sv_destroy(&dqp->q_pinwait);
 165         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 166
 167         atomic_dec(&xfs_Gqm->qm_totaldquots);
 168 }
 169
 170 /*
 171  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 172  */
 173 STATIC void
 174 xfs_qm_dqinit_core(
 175         xfs_dqid_t      id,
 176         uint            type,
 177         xfs_dqblk_t     *d)
 178 {
 179         /*
 180          * Caller has zero'd the entire dquot 'chunk' already.
 181          */
 182         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 183         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 184         d->dd_diskdq.d_id = cpu_to_be32(id);
 185         d->dd_diskdq.d_flags = type;
 186 }
 187
 188 /*
 189  * If default limits are in force, push them into the dquot now.
 190  * We overwrite the dquot limits only if they are zero and this
 191  * is not the root dquot.
 192  */
 193 void
 194 xfs_qm_adjust_dqlimits(
 195         xfs_mount_t             *mp,
 196         xfs_disk_dquot_t        *d)
 197 {
 198         xfs_quotainfo_t         *q = mp->m_quotainfo;
 199
 200         ASSERT(d->d_id);
 201
 202         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 203                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 204         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 205                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 206         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 207                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 208         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 209                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 210         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 211                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 212         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 213                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 214 }
 215
 216 /*
 217  * Check the limits and timers of a dquot and start or reset timers
 218  * if necessary.
 219  * This gets called even when quota enforcement is OFF, which makes our
 220  * life a little less complicated. (We just don't reject any quota
 221  * reservations in that case, when enforcement is off).
 222  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 223  * enforcement's off.
 224  * In contrast, warnings are a little different in that they don't
 225  * 'automatically' get started when limits get exceeded.  They do
 226  * get reset to zero, however, when we find the count to be under
 227  * the soft limit (they are only ever set non-zero via userspace).
 228  */
 229 void
 230 xfs_qm_adjust_dqtimers(
 231         xfs_mount_t             *mp,
 232         xfs_disk_dquot_t        *d)
 233 {
 234         ASSERT(d->d_id);
 235
 236 #ifdef QUOTADEBUG
 237         if (d->d_blk_hardlimit)
 238                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 239                        be64_to_cpu(d->d_blk_hardlimit));
 240         if (d->d_ino_hardlimit)
 241                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 242                        be64_to_cpu(d->d_ino_hardlimit));
 243         if (d->d_rtb_hardlimit)
 244                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 245                        be64_to_cpu(d->d_rtb_hardlimit));
 246 #endif
 247         if (!d->d_btimer) {
 248                 if ((d->d_blk_softlimit &&
 249                      (be64_to_cpu(d->d_bcount) >=
 250                       be64_to_cpu(d->d_blk_softlimit))) ||
 251                     (d->d_blk_hardlimit &&
 252                      (be64_to_cpu(d->d_bcount) >=
 253                       be64_to_cpu(d->d_blk_hardlimit)))) {
 254                         d->d_btimer = cpu_to_be32(get_seconds() +
 255                                         XFS_QI_BTIMELIMIT(mp));
 256                 } else {
 257                         d->d_bwarns = 0;
 258                 }
 259         } else {
 260                 if ((!d->d_blk_softlimit ||
 261                      (be64_to_cpu(d->d_bcount) <
 262                       be64_to_cpu(d->d_blk_softlimit))) &&
 263                     (!d->d_blk_hardlimit ||
 264                     (be64_to_cpu(d->d_bcount) <
 265                      be64_to_cpu(d->d_blk_hardlimit)))) {
 266                         d->d_btimer = 0;
 267                 }
 268         }
 269
 270         if (!d->d_itimer) {
 271                 if ((d->d_ino_softlimit &&
 272                      (be64_to_cpu(d->d_icount) >=
 273                       be64_to_cpu(d->d_ino_softlimit))) ||
 274                     (d->d_ino_hardlimit &&
 275                      (be64_to_cpu(d->d_icount) >=
 276                       be64_to_cpu(d->d_ino_hardlimit)))) {
 277                         d->d_itimer = cpu_to_be32(get_seconds() +
 278                                         XFS_QI_ITIMELIMIT(mp));
 279                 } else {
 280                         d->d_iwarns = 0;
 281                 }
 282         } else {
 283                 if ((!d->d_ino_softlimit ||
 284                      (be64_to_cpu(d->d_icount) <
 285                       be64_to_cpu(d->d_ino_softlimit)))  &&
 286                     (!d->d_ino_hardlimit ||
 287                      (be64_to_cpu(d->d_icount) <
 288                       be64_to_cpu(d->d_ino_hardlimit)))) {
 289                         d->d_itimer = 0;
 290                 }
 291         }
 292
 293         if (!d->d_rtbtimer) {
 294                 if ((d->d_rtb_softlimit &&
 295                      (be64_to_cpu(d->d_rtbcount) >=
 296                       be64_to_cpu(d->d_rtb_softlimit))) ||
 297                     (d->d_rtb_hardlimit &&
 298                      (be64_to_cpu(d->d_rtbcount) >=
 299                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 300                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 301                                         XFS_QI_RTBTIMELIMIT(mp));
 302                 } else {
 303                         d->d_rtbwarns = 0;
 304                 }
 305         } else {
 306                 if ((!d->d_rtb_softlimit ||
 307                      (be64_to_cpu(d->d_rtbcount) <
 308                       be64_to_cpu(d->d_rtb_softlimit))) &&
 309                     (!d->d_rtb_hardlimit ||
 310                      (be64_to_cpu(d->d_rtbcount) <
 311                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 312                         d->d_rtbtimer = 0;
 313                 }
 314         }
 315 }
 316
 317 /*
 318  * initialize a buffer full of dquots and log the whole thing
 319  */
 320 STATIC void
 321 xfs_qm_init_dquot_blk(
 322         xfs_trans_t     *tp,
 323         xfs_mount_t     *mp,
 324         xfs_dqid_t      id,
 325         uint            type,
 326         xfs_buf_t       *bp)
 327 {
 328         xfs_dqblk_t     *d;
 329         int             curid, i;
 330
 331         ASSERT(tp);
 332         ASSERT(XFS_BUF_ISBUSY(bp));
 333         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 334
 335         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 336
 337         /*
 338          * ID of the first dquot in the block - id's are zero based.
 339          */
 340         curid = id - (id % XFS_QM_DQPERBLK(mp));
 341         ASSERT(curid >= 0);
 342         memset(d, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)));
 343         for (i = 0; i < XFS_QM_DQPERBLK(mp); i++, d++, curid++)
 344                 xfs_qm_dqinit_core(curid, type, d);
 345         xfs_trans_dquot_buf(tp, bp,
 346                             (type & XFS_DQ_USER ? XFS_BLI_UDQUOT_BUF :
 347                             ((type & XFS_DQ_PROJ) ? XFS_BLI_PDQUOT_BUF :
 348                              XFS_BLI_GDQUOT_BUF)));
 349         xfs_trans_log_buf(tp, bp, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1);
 350 }
 351
 352
 353
 354 /*
 355  * Allocate a block and fill it with dquots.
 356  * This is called when the bmapi finds a hole.
 357  */
 358 STATIC int
 359 xfs_qm_dqalloc(
 360         xfs_trans_t     **tpp,
 361         xfs_mount_t     *mp,
 362         xfs_dquot_t     *dqp,
 363         xfs_inode_t     *quotip,
 364         xfs_fileoff_t   offset_fsb,
 365         xfs_buf_t       **O_bpp)
 366 {
 367         xfs_fsblock_t   firstblock;
 368         xfs_bmap_free_t flist;
 369         xfs_bmbt_irec_t map;
 370         int             nmaps, error, committed;
 371         xfs_buf_t       *bp;
 372         xfs_trans_t     *tp = *tpp;
 373
 374         ASSERT(tp != NULL);
 375
 376         trace_xfs_dqalloc(dqp);
 377
 378         /*
 379          * Initialize the bmap freelist prior to calling bmapi code.
 380          */
 381         xfs_bmap_init(&flist, &firstblock);
 382         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 383         /*
 384          * Return if this type of quotas is turned off while we didn't
 385          * have an inode lock
 386          */
 387         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 388                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 389                 return (ESRCH);
 390         }
 391
 392         /*
 393          * xfs_trans_commit normally decrements the vnode ref count
 394          * when it unlocks the inode. Since we want to keep the quota
 395          * inode around, we bump the vnode ref count now.
 396          */
 397         IHOLD(quotip);
 398
 399         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 400         nmaps = 1;
 401         if ((error = xfs_bmapi(tp, quotip,
 402                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 403                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 404                               &firstblock,
 405                               XFS_QM_DQALLOC_SPACE_RES(mp),
 406                               &map, &nmaps, &flist, NULL))) {
 407                 goto error0;
 408         }
 409         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 410         ASSERT(nmaps == 1);
 411         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 412                (map.br_startblock != HOLESTARTBLOCK));
 413
 414         /*
 415          * Keep track of the blkno to save a lookup later
 416          */
 417         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 418
 419         /* now we can just get the buffer (there's nothing to read yet) */
 420         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 421                                dqp->q_blkno,
 422                                XFS_QI_DQCHUNKLEN(mp),
 423                                0);
 424         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 425                 goto error1;
 426         /*
 427          * Make a chunk of dquots out of this buffer and log
 428          * the entire thing.
 429          */
 430         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 431                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 432
 433         /*
 434          * xfs_bmap_finish() may commit the current transaction and
 435          * start a second transaction if the freelist is not empty.
 436          *
 437          * Since we still want to modify this buffer, we need to
 438          * ensure that the buffer is not released on commit of
 439          * the first transaction and ensure the buffer is added to the
 440          * second transaction.
 441          *
 442          * If there is only one transaction then don't stop the buffer
 443          * from being released when it commits later on.
 444          */
 445
 446         xfs_trans_bhold(tp, bp);
 447
 448         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 449                 goto error1;
 450         }
 451
 452         if (committed) {
 453                 tp = *tpp;
 454                 xfs_trans_bjoin(tp, bp);
 455         } else {
 456                 xfs_trans_bhold_release(tp, bp);
 457         }
 458
 459         *O_bpp = bp;
 460         return 0;
 461
 462       error1:
 463         xfs_bmap_cancel(&flist);
 464       error0:
 465         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 466
 467         return (error);
 468 }
 469
 470 /*
 471  * Maps a dquot to the buffer containing its on-disk version.
 472  * This returns a ptr to the buffer containing the on-disk dquot
 473  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 474  */
 475 STATIC int
 476 xfs_qm_dqtobp(
 477         xfs_trans_t             **tpp,
 478         xfs_dquot_t             *dqp,
 479         xfs_disk_dquot_t        **O_ddpp,
 480         xfs_buf_t               **O_bpp,
 481         uint                    flags)
 482 {
 483         xfs_bmbt_irec_t map;
 484         int             nmaps, error;
 485         xfs_buf_t       *bp;
 486         xfs_inode_t     *quotip;
 487         xfs_mount_t     *mp;
 488         xfs_disk_dquot_t *ddq;
 489         xfs_dqid_t      id;
 490         boolean_t       newdquot;
 491         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 492
 493         mp = dqp->q_mount;
 494         id = be32_to_cpu(dqp->q_core.d_id);
 495         nmaps = 1;
 496         newdquot = B_FALSE;
 497
 498         /*
 499          * If we don't know where the dquot lives, find out.
 500          */
 501         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 502                 /* We use the id as an index */
 503                 dqp->q_fileoffset = (xfs_fileoff_t)id / XFS_QM_DQPERBLK(mp);
 504                 nmaps = 1;
 505                 quotip = XFS_DQ_TO_QIP(dqp);
 506                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 507                 /*
 508                  * Return if this type of quotas is turned off while we didn't
 509                  * have an inode lock
 510                  */
 511                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 512                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 513                         return (ESRCH);
 514                 }
 515                 /*
 516                  * Find the block map; no allocations yet
 517                  */
 518                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 519                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 520                                   XFS_BMAPI_METADATA,
 521                                   NULL, 0, &map, &nmaps, NULL, NULL);
 522
 523                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 524                 if (error)
 525                         return (error);
 526                 ASSERT(nmaps == 1);
 527                 ASSERT(map.br_blockcount == 1);
 528
 529                 /*
 530                  * offset of dquot in the (fixed sized) dquot chunk.
 531                  */
 532                 dqp->q_bufoffset = (id % XFS_QM_DQPERBLK(mp)) *
 533                         sizeof(xfs_dqblk_t);
 534                 if (map.br_startblock == HOLESTARTBLOCK) {
 535                         /*
 536                          * We don't allocate unless we're asked to
 537                          */
 538                         if (!(flags & XFS_QMOPT_DQALLOC))
 539                                 return (ENOENT);
 540
 541                         ASSERT(tp);
 542                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 543                                                 dqp->q_fileoffset, &bp)))
 544                                 return (error);
 545                         tp = *tpp;
 546                         newdquot = B_TRUE;
 547                 } else {
 548                         /*
 549                          * store the blkno etc so that we don't have to do the
 550                          * mapping all the time
 551                          */
 552                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 553                 }
 554         }
 555         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 556         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 557
 558         /*
 559          * Read in the buffer, unless we've just done the allocation
 560          * (in which case we already have the buf).
 561          */
 562         if (! newdquot) {
 563                 trace_xfs_dqtobp_read(dqp);
 564
 565                 if ((error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 566                                                dqp->q_blkno,
 567                                                XFS_QI_DQCHUNKLEN(mp),
 568                                                0, &bp))) {
 569                         return (error);
 570                 }
 571                 if (error || !bp)
 572                         return XFS_ERROR(error);
 573         }
 574         ASSERT(XFS_BUF_ISBUSY(bp));
 575         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 576
 577         /*
 578          * calculate the location of the dquot inside the buffer.
 579          */
 580         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 581
 582         /*
 583          * A simple sanity check in case we got a corrupted dquot...
 584          */
 585         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 586                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 587                            "dqtobp")) {
 588                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 589                         xfs_trans_brelse(tp, bp);
 590                         return XFS_ERROR(EIO);
 591                 }
 592                 XFS_BUF_BUSY(bp); /* We dirtied this */
 593         }
 594
 595         *O_bpp = bp;
 596         *O_ddpp = ddq;
 597
 598         return (0);
 599 }
 600
 601
 602 /*
 603  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 604  * and release the buffer immediately.
 605  *
 606  */
 607 /* ARGSUSED */
 608 STATIC int
 609 xfs_qm_dqread(
 610         xfs_trans_t     **tpp,
 611         xfs_dqid_t      id,
 612         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 613         uint            flags)
 614 {
 615         xfs_disk_dquot_t *ddqp;
 616         xfs_buf_t        *bp;
 617         int              error;
 618         xfs_trans_t      *tp;
 619
 620         ASSERT(tpp);
 621
 622         trace_xfs_dqread(dqp);
 623
 624         /*
 625          * get a pointer to the on-disk dquot and the buffer containing it
 626          * dqp already knows its own type (GROUP/USER).
 627          */
 628         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 629                 return (error);
 630         }
 631         tp = *tpp;
 632
 633         /* copy everything from disk dquot to the incore dquot */
 634         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 635         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 636         xfs_qm_dquot_logitem_init(dqp);
 637
 638         /*
 639          * Reservation counters are defined as reservation plus current usage
 640          * to avoid having to add everytime.
 641          */
 642         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 643         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 644         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 645
 646         /* Mark the buf so that this will stay incore a little longer */
 647         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 648
 649         /*
 650          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 651          * So we need to release with xfs_trans_brelse().
 652          * The strategy here is identical to that of inodes; we lock
 653          * the dquot in xfs_qm_dqget() before making it accessible to
 654          * others. This is because dquots, like inodes, need a good level of
 655          * concurrency, and we don't want to take locks on the entire buffers
 656          * for dquot accesses.
 657          * Note also that the dquot buffer may even be dirty at this point, if
 658          * this particular dquot was repaired. We still aren't afraid to
 659          * brelse it because we have the changes incore.
 660          */
 661         ASSERT(XFS_BUF_ISBUSY(bp));
 662         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 663         xfs_trans_brelse(tp, bp);
 664
 665         return (error);
 666 }
 667
 668
 669 /*
 670  * allocate an incore dquot from the kernel heap,
 671  * and fill its core with quota information kept on disk.
 672  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 673  * if it wasn't already allocated.
 674  */
 675 STATIC int
 676 xfs_qm_idtodq(
 677         xfs_mount_t     *mp,
 678         xfs_dqid_t      id,      /* gid or uid, depending on type */
 679         uint            type,    /* UDQUOT or GDQUOT */
 680         uint            flags,   /* DQALLOC, DQREPAIR */
 681         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 682 {
 683         xfs_dquot_t     *dqp;
 684         int             error;
 685         xfs_trans_t     *tp;
 686         int             cancelflags=0;
 687
 688         dqp = xfs_qm_dqinit(mp, id, type);
 689         tp = NULL;
 690         if (flags & XFS_QMOPT_DQALLOC) {
 691                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 692                 if ((error = xfs_trans_reserve(tp,
 693                                        XFS_QM_DQALLOC_SPACE_RES(mp),
 694                                        XFS_WRITE_LOG_RES(mp) +
 695                                               BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1 +
 696                                               128,
 697                                        0,
 698                                        XFS_TRANS_PERM_LOG_RES,
 699                                        XFS_WRITE_LOG_COUNT))) {
 700                         cancelflags = 0;
 701                         goto error0;
 702                 }
 703                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 704         }
 705
 706         /*
 707          * Read it from disk; xfs_dqread() takes care of
 708          * all the necessary initialization of dquot's fields (locks, etc)
 709          */
 710         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 711                 /*
 712                  * This can happen if quotas got turned off (ESRCH),
 713                  * or if the dquot didn't exist on disk and we ask to
 714                  * allocate (ENOENT).
 715                  */
 716                 trace_xfs_dqread_fail(dqp);
 717                 cancelflags |= XFS_TRANS_ABORT;
 718                 goto error0;
 719         }
 720         if (tp) {
 721                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 722                         goto error1;
 723         }
 724
 725         *O_dqpp = dqp;
 726         return (0);
 727
 728  error0:
 729         ASSERT(error);
 730         if (tp)
 731                 xfs_trans_cancel(tp, cancelflags);
 732  error1:
 733         xfs_qm_dqdestroy(dqp);
 734         *O_dqpp = NULL;
 735         return (error);
 736 }
 737
 738 /*
 739  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 740  * hashtables for user and group dquots; and, these are global tables
 741  * inside the XQM, not per-filesystem tables.
 742  * The hash chain must be locked by caller, and it is left locked
 743  * on return. Returning dquot is locked.
 744  */
 745 STATIC int
 746 xfs_qm_dqlookup(
 747         xfs_mount_t             *mp,
 748         xfs_dqid_t              id,
 749         xfs_dqhash_t            *qh,
 750         xfs_dquot_t             **O_dqpp)
 751 {
 752         xfs_dquot_t             *dqp;
 753         uint                    flist_locked;
 754         xfs_dquot_t             *d;
 755
 756         ASSERT(mutex_is_locked(&qh->qh_lock));
 757
 758         flist_locked = B_FALSE;
 759
 760         /*
 761          * Traverse the hashchain looking for a match
 762          */
 763         for (dqp = qh->qh_next; dqp != NULL; dqp = dqp->HL_NEXT) {
 764                 /*
 765                  * We already have the hashlock. We don't need the
 766                  * dqlock to look at the id field of the dquot, since the
 767                  * id can't be modified without the hashlock anyway.
 768                  */
 769                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 770                         trace_xfs_dqlookup_found(dqp);
 771
 772                         /*
 773                          * All in core dquots must be on the dqlist of mp
 774                          */
 775                         ASSERT(dqp->MPL_PREVP != NULL);
 776
 777                         xfs_dqlock(dqp);
 778                         if (dqp->q_nrefs == 0) {
 779                                 ASSERT (XFS_DQ_IS_ON_FREELIST(dqp));
 780                                 if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
 781                                         trace_xfs_dqlookup_want(dqp);
 782
 783                                         /*
 784                                          * We may have raced with dqreclaim_one()
 785                                          * (and lost). So, flag that we don't
 786                                          * want the dquot to be reclaimed.
 787                                          */
 788                                         dqp->dq_flags |= XFS_DQ_WANT;
 789                                         xfs_dqunlock(dqp);
 790                                         xfs_qm_freelist_lock(xfs_Gqm);
 791                                         xfs_dqlock(dqp);
 792                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 793                                 }
 794                                 flist_locked = B_TRUE;
 795                         }
 796
 797                         /*
 798                          * id couldn't have changed; we had the hashlock all
 799                          * along
 800                          */
 801                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 802
 803                         if (flist_locked) {
 804                                 if (dqp->q_nrefs != 0) {
 805                                         xfs_qm_freelist_unlock(xfs_Gqm);
 806                                         flist_locked = B_FALSE;
 807                                 } else {
 808                                         /*
 809                                          * take it off the freelist
 810                                          */
 811                                         trace_xfs_dqlookup_freelist(dqp);
 812                                         XQM_FREELIST_REMOVE(dqp);
 813                                         /* xfs_qm_freelist_print(&(xfs_Gqm->
 814                                                         qm_dqfreelist),
 815                                                         "after removal"); */
 816                                 }
 817                         }
 818
 819                         /*
 820                          * grab a reference
 821                          */
 822                         XFS_DQHOLD(dqp);
 823
 824                         if (flist_locked)
 825                                 xfs_qm_freelist_unlock(xfs_Gqm);
 826                         /*
 827                          * move the dquot to the front of the hashchain
 828                          */
 829                         ASSERT(mutex_is_locked(&qh->qh_lock));
 830                         if (dqp->HL_PREVP != &qh->qh_next) {
 831                                 trace_xfs_dqlookup_move(dqp);
 832                                 if ((d = dqp->HL_NEXT))
 833                                         d->HL_PREVP = dqp->HL_PREVP;
 834                                 *(dqp->HL_PREVP) = d;
 835                                 d = qh->qh_next;
 836                                 d->HL_PREVP = &dqp->HL_NEXT;
 837                                 dqp->HL_NEXT = d;
 838                                 dqp->HL_PREVP = &qh->qh_next;
 839                                 qh->qh_next = dqp;
 840                         }
 841                         trace_xfs_dqlookup_done(dqp);
 842                         *O_dqpp = dqp;
 843                         ASSERT(mutex_is_locked(&qh->qh_lock));
 844                         return (0);
 845                 }
 846         }
 847
 848         *O_dqpp = NULL;
 849         ASSERT(mutex_is_locked(&qh->qh_lock));
 850         return (1);
 851 }
 852
 853 /*
 854  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 855  * a locked dquot, doing an allocation (if requested) as needed.
 856  * When both an inode and an id are given, the inode's id takes precedence.
 857  * That is, if the id changes while we don't hold the ilock inside this
 858  * function, the new dquot is returned, not necessarily the one requested
 859  * in the id argument.
 860  */
 861 int
 862 xfs_qm_dqget(
 863         xfs_mount_t     *mp,
 864         xfs_inode_t     *ip,      /* locked inode (optional) */
 865         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 866         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 867         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 868         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 869 {
 870         xfs_dquot_t     *dqp;
 871         xfs_dqhash_t    *h;
 872         uint            version;
 873         int             error;
 874
 875         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 876         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 877             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 878             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 879                 return (ESRCH);
 880         }
 881         h = XFS_DQ_HASH(mp, id, type);
 882
 883 #ifdef DEBUG
 884         if (xfs_do_dqerror) {
 885                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 886                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 887                         cmn_err(CE_DEBUG, "Returning error in dqget");
 888                         return (EIO);
 889                 }
 890         }
 891 #endif
 892
 893  again:
 894
 895 #ifdef DEBUG
 896         ASSERT(type == XFS_DQ_USER ||
 897                type == XFS_DQ_PROJ ||
 898                type == XFS_DQ_GROUP);
 899         if (ip) {
 900                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 901                 if (type == XFS_DQ_USER)
 902                         ASSERT(ip->i_udquot == NULL);
 903                 else
 904                         ASSERT(ip->i_gdquot == NULL);
 905         }
 906 #endif
 907         mutex_lock(&h->qh_lock);
 908
 909         /*
 910          * Look in the cache (hashtable).
 911          * The chain is kept locked during lookup.
 912          */
 913         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 914                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 915                 /*
 916                  * The dquot was found, moved to the front of the chain,
 917                  * taken off the freelist if it was on it, and locked
 918                  * at this point. Just unlock the hashchain and return.
 919                  */
 920                 ASSERT(*O_dqpp);
 921                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 922                 mutex_unlock(&h->qh_lock);
 923                 trace_xfs_dqget_hit(*O_dqpp);
 924                 return (0);     /* success */
 925         }
 926         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 927
 928         /*
 929          * Dquot cache miss. We don't want to keep the inode lock across
 930          * a (potential) disk read. Also we don't want to deal with the lock
 931          * ordering between quotainode and this inode. OTOH, dropping the inode
 932          * lock here means dealing with a chown that can happen before
 933          * we re-acquire the lock.
 934          */
 935         if (ip)
 936                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 937         /*
 938          * Save the hashchain version stamp, and unlock the chain, so that
 939          * we don't keep the lock across a disk read
 940          */
 941         version = h->qh_version;
 942         mutex_unlock(&h->qh_lock);
 943
 944         /*
 945          * Allocate the dquot on the kernel heap, and read the ondisk
 946          * portion off the disk. Also, do all the necessary initialization
 947          * This can return ENOENT if dquot didn't exist on disk and we didn't
 948          * ask it to allocate; ESRCH if quotas got turned off suddenly.
 949          */
 950         if ((error = xfs_qm_idtodq(mp, id, type,
 951                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
 952                                            XFS_QMOPT_DOWARN),
 953                                   &dqp))) {
 954                 if (ip)
 955                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 956                 return (error);
 957         }
 958
 959         /*
 960          * See if this is mount code calling to look at the overall quota limits
 961          * which are stored in the id == 0 user or group's dquot.
 962          * Since we may not have done a quotacheck by this point, just return
 963          * the dquot without attaching it to any hashtables, lists, etc, or even
 964          * taking a reference.
 965          * The caller must dqdestroy this once done.
 966          */
 967         if (flags & XFS_QMOPT_DQSUSER) {
 968                 ASSERT(id == 0);
 969                 ASSERT(! ip);
 970                 goto dqret;
 971         }
 972
 973         /*
 974          * Dquot lock comes after hashlock in the lock ordering
 975          */
 976         if (ip) {
 977                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 978                 if (! XFS_IS_DQTYPE_ON(mp, type)) {
 979                         /* inode stays locked on return */
 980                         xfs_qm_dqdestroy(dqp);
 981                         return XFS_ERROR(ESRCH);
 982                 }
 983                 /*
 984                  * A dquot could be attached to this inode by now, since
 985                  * we had dropped the ilock.
 986                  */
 987                 if (type == XFS_DQ_USER) {
 988                         if (ip->i_udquot) {
 989                                 xfs_qm_dqdestroy(dqp);
 990                                 dqp = ip->i_udquot;
 991                                 xfs_dqlock(dqp);
 992                                 goto dqret;
 993                         }
 994                 } else {
 995                         if (ip->i_gdquot) {
 996                                 xfs_qm_dqdestroy(dqp);
 997                                 dqp = ip->i_gdquot;
 998                                 xfs_dqlock(dqp);
 999                                 goto dqret;
1000                         }
1001                 }
1002         }
1003
1004         /*
1005          * Hashlock comes after ilock in lock order
1006          */
1007         mutex_lock(&h->qh_lock);
1008         if (version != h->qh_version) {
1009                 xfs_dquot_t *tmpdqp;
1010                 /*
1011                  * Now, see if somebody else put the dquot in the
1012                  * hashtable before us. This can happen because we didn't
1013                  * keep the hashchain lock. We don't have to worry about
1014                  * lock order between the two dquots here since dqp isn't
1015                  * on any findable lists yet.
1016                  */
1017                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
1018                         /*
1019                          * Duplicate found. Just throw away the new dquot
1020                          * and start over.
1021                          */
1022                         xfs_qm_dqput(tmpdqp);
1023                         mutex_unlock(&h->qh_lock);
1024                         xfs_qm_dqdestroy(dqp);
1025                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
1026                         goto again;
1027                 }
1028         }
1029
1030         /*
1031          * Put the dquot at the beginning of the hash-chain and mp's list
1032          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1033          */
1034         ASSERT(mutex_is_locked(&h->qh_lock));
1035         dqp->q_hash = h;
1036         XQM_HASHLIST_INSERT(h, dqp);
1037
1038         /*
1039          * Attach this dquot to this filesystem's list of all dquots,
1040          * kept inside the mount structure in m_quotainfo field
1041          */
1042         xfs_qm_mplist_lock(mp);
1043
1044         /*
1045          * We return a locked dquot to the caller, with a reference taken
1046          */
1047         xfs_dqlock(dqp);
1048         dqp->q_nrefs = 1;
1049
1050         XQM_MPLIST_INSERT(&(XFS_QI_MPL_LIST(mp)), dqp);
1051
1052         xfs_qm_mplist_unlock(mp);
1053         mutex_unlock(&h->qh_lock);
1054  dqret:
1055         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1056         trace_xfs_dqget_miss(dqp);
1057         *O_dqpp = dqp;
1058         return (0);
1059 }
1060
1061
1062 /*
1063  * Release a reference to the dquot (decrement ref-count)
1064  * and unlock it. If there is a group quota attached to this
1065  * dquot, carefully release that too without tripping over
1066  * deadlocks'n'stuff.
1067  */
1068 void
1069 xfs_qm_dqput(
1070         xfs_dquot_t     *dqp)
1071 {
1072         xfs_dquot_t     *gdqp;
1073
1074         ASSERT(dqp->q_nrefs > 0);
1075         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1076
1077         trace_xfs_dqput(dqp);
1078
1079         if (dqp->q_nrefs != 1) {
1080                 dqp->q_nrefs--;
1081                 xfs_dqunlock(dqp);
1082                 return;
1083         }
1084
1085         /*
1086          * drop the dqlock and acquire the freelist and dqlock
1087          * in the right order; but try to get it out-of-order first
1088          */
1089         if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
1090                 trace_xfs_dqput_wait(dqp);
1091                 xfs_dqunlock(dqp);
1092                 xfs_qm_freelist_lock(xfs_Gqm);
1093                 xfs_dqlock(dqp);
1094         }
1095
1096         while (1) {
1097                 gdqp = NULL;
1098
1099                 /* We can't depend on nrefs being == 1 here */
1100                 if (--dqp->q_nrefs == 0) {
1101                         trace_xfs_dqput_free(dqp);
1102
1103                         /*
1104                          * insert at end of the freelist.
1105                          */
1106                         XQM_FREELIST_INSERT(&(xfs_Gqm->qm_dqfreelist), dqp);
1107
1108                         /*
1109                          * If we just added a udquot to the freelist, then
1110                          * we want to release the gdquot reference that
1111                          * it (probably) has. Otherwise it'll keep the
1112                          * gdquot from getting reclaimed.
1113                          */
1114                         if ((gdqp = dqp->q_gdquot)) {
1115                                 /*
1116                                  * Avoid a recursive dqput call
1117                                  */
1118                                 xfs_dqlock(gdqp);
1119                                 dqp->q_gdquot = NULL;
1120                         }
1121
1122                         /* xfs_qm_freelist_print(&(xfs_Gqm->qm_dqfreelist),
1123                            "@@@@@++ Free list (after append) @@@@@+");
1124                            */
1125                 }
1126                 xfs_dqunlock(dqp);
1127
1128                 /*
1129                  * If we had a group quota inside the user quota as a hint,
1130                  * release it now.
1131                  */
1132                 if (! gdqp)
1133                         break;
1134                 dqp = gdqp;
1135         }
1136         xfs_qm_freelist_unlock(xfs_Gqm);
1137 }
1138
1139 /*
1140  * Release a dquot. Flush it if dirty, then dqput() it.
1141  * dquot must not be locked.
1142  */
1143 void
1144 xfs_qm_dqrele(
1145         xfs_dquot_t     *dqp)
1146 {
1147         if (!dqp)
1148                 return;
1149
1150         trace_xfs_dqrele(dqp);
1151
1152         xfs_dqlock(dqp);
1153         /*
1154          * We don't care to flush it if the dquot is dirty here.
1155          * That will create stutters that we want to avoid.
1156          * Instead we do a delayed write when we try to reclaim
1157          * a dirty dquot. Also xfs_sync will take part of the burden...
1158          */
1159         xfs_qm_dqput(dqp);
1160 }
1161
1162
1163 /*
1164  * Write a modified dquot to disk.
1165  * The dquot must be locked and the flush lock too taken by caller.
1166  * The flush lock will not be unlocked until the dquot reaches the disk,
1167  * but the dquot is free to be unlocked and modified by the caller
1168  * in the interim. Dquot is still locked on return. This behavior is
1169  * identical to that of inodes.
1170  */
1171 int
1172 xfs_qm_dqflush(
1173         xfs_dquot_t             *dqp,
1174         uint                    flags)
1175 {
1176         xfs_mount_t             *mp;
1177         xfs_buf_t               *bp;
1178         xfs_disk_dquot_t        *ddqp;
1179         int                     error;
1180
1181         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1182         ASSERT(!completion_done(&dqp->q_flush));
1183         trace_xfs_dqflush(dqp);
1184
1185         /*
1186          * If not dirty, or it's pinned and we are not supposed to
1187          * block, nada.
1188          */
1189         if (!XFS_DQ_IS_DIRTY(dqp) ||
1190             (!(flags & SYNC_WAIT) && atomic_read(&dqp->q_pincount) > 0)) {
1191                 xfs_dqfunlock(dqp);
1192                 return 0;
1193         }
1194         xfs_qm_dqunpin_wait(dqp);
1195
1196         /*
1197          * This may have been unpinned because the filesystem is shutting
1198          * down forcibly. If that's the case we must not write this dquot
1199          * to disk, because the log record didn't make it to disk!
1200          */
1201         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1202                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1203                 xfs_dqfunlock(dqp);
1204                 return XFS_ERROR(EIO);
1205         }
1206
1207         /*
1208          * Get the buffer containing the on-disk dquot
1209          * We don't need a transaction envelope because we know that the
1210          * the ondisk-dquot has already been allocated for.
1211          */
1212         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1213                 ASSERT(error != ENOENT);
1214                 /*
1215                  * Quotas could have gotten turned off (ESRCH)
1216                  */
1217                 xfs_dqfunlock(dqp);
1218                 return (error);
1219         }
1220
1221         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1222                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1223                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1224                 return XFS_ERROR(EIO);
1225         }
1226
1227         /* This is the only portion of data that needs to persist */
1228         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1229
1230         /*
1231          * Clear the dirty field and remember the flush lsn for later use.
1232          */
1233         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1234         mp = dqp->q_mount;
1235
1236         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1237                                         &dqp->q_logitem.qli_item.li_lsn);
1238
1239         /*
1240          * Attach an iodone routine so that we can remove this dquot from the
1241          * AIL and release the flush lock once the dquot is synced to disk.
1242          */
1243         xfs_buf_attach_iodone(bp, (void(*)(xfs_buf_t *, xfs_log_item_t *))
1244                               xfs_qm_dqflush_done, &(dqp->q_logitem.qli_item));
1245         /*
1246          * If the buffer is pinned then push on the log so we won't
1247          * get stuck waiting in the write for too long.
1248          */
1249         if (XFS_BUF_ISPINNED(bp)) {
1250                 trace_xfs_dqflush_force(dqp);
1251                 xfs_log_force(mp, 0);
1252         }
1253
1254         if (flags & SYNC_WAIT)
1255                 error = xfs_bwrite(mp, bp);
1256         else
1257                 xfs_bdwrite(mp, bp);
1258
1259         trace_xfs_dqflush_done(dqp);
1260
1261         /*
1262          * dqp is still locked, but caller is free to unlock it now.
1263          */
1264         return error;
1265
1266 }
1267
1268 /*
1269  * This is the dquot flushing I/O completion routine.  It is called
1270  * from interrupt level when the buffer containing the dquot is
1271  * flushed to disk.  It is responsible for removing the dquot logitem
1272  * from the AIL if it has not been re-logged, and unlocking the dquot's
1273  * flush lock. This behavior is very similar to that of inodes..
1274  */
1275 /*ARGSUSED*/
1276 STATIC void
1277 xfs_qm_dqflush_done(
1278         xfs_buf_t               *bp,
1279         xfs_dq_logitem_t        *qip)
1280 {
1281         xfs_dquot_t             *dqp;
1282         struct xfs_ail          *ailp;
1283
1284         dqp = qip->qli_dquot;
1285         ailp = qip->qli_item.li_ailp;
1286
1287         /*
1288          * We only want to pull the item from the AIL if its
1289          * location in the log has not changed since we started the flush.
1290          * Thus, we only bother if the dquot's lsn has
1291          * not changed. First we check the lsn outside the lock
1292          * since it's cheaper, and then we recheck while
1293          * holding the lock before removing the dquot from the AIL.
1294          */
1295         if ((qip->qli_item.li_flags & XFS_LI_IN_AIL) &&
1296             qip->qli_item.li_lsn == qip->qli_flush_lsn) {
1297
1298                 /* xfs_trans_ail_delete() drops the AIL lock. */
1299                 spin_lock(&ailp->xa_lock);
1300                 if (qip->qli_item.li_lsn == qip->qli_flush_lsn)
1301                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)qip);
1302                 else
1303                         spin_unlock(&ailp->xa_lock);
1304         }
1305
1306         /*
1307          * Release the dq's flush lock since we're done with it.
1308          */
1309         xfs_dqfunlock(dqp);
1310 }
1311
1312 int
1313 xfs_qm_dqlock_nowait(
1314         xfs_dquot_t *dqp)
1315 {
1316         return mutex_trylock(&dqp->q_qlock);
1317 }
1318
1319 void
1320 xfs_dqlock(
1321         xfs_dquot_t *dqp)
1322 {
1323         mutex_lock(&dqp->q_qlock);
1324 }
1325
1326 void
1327 xfs_dqunlock(
1328         xfs_dquot_t *dqp)
1329 {
1330         mutex_unlock(&(dqp->q_qlock));
1331         if (dqp->q_logitem.qli_dquot == dqp) {
1332                 /* Once was dqp->q_mount, but might just have been cleared */
1333                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1334                                         (xfs_log_item_t*)&(dqp->q_logitem));
1335         }
1336 }
1337
1338
1339 void
1340 xfs_dqunlock_nonotify(
1341         xfs_dquot_t *dqp)
1342 {
1343         mutex_unlock(&(dqp->q_qlock));
1344 }
1345
1346 /*
1347  * Lock two xfs_dquot structures.
1348  *
1349  * To avoid deadlocks we always lock the quota structure with
1350  * the lowerd id first.
1351  */
1352 void
1353 xfs_dqlock2(
1354         xfs_dquot_t     *d1,
1355         xfs_dquot_t     *d2)
1356 {
1357         if (d1 && d2) {
1358                 ASSERT(d1 != d2);
1359                 if (be32_to_cpu(d1->q_core.d_id) >
1360                     be32_to_cpu(d2->q_core.d_id)) {
1361                         mutex_lock(&d2->q_qlock);
1362                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1363                 } else {
1364                         mutex_lock(&d1->q_qlock);
1365                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1366                 }
1367         } else if (d1) {
1368                 mutex_lock(&d1->q_qlock);
1369         } else if (d2) {
1370                 mutex_lock(&d2->q_qlock);
1371         }
1372 }
1373
1374
1375 /*
1376  * Take a dquot out of the mount's dqlist as well as the hashlist.
1377  * This is called via unmount as well as quotaoff, and the purge
1378  * will always succeed unless there are soft (temp) references
1379  * outstanding.
1380  *
1381  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1382  * that we're returning! XXXsup - not cool.
1383  */
1384 /* ARGSUSED */
1385 int
1386 xfs_qm_dqpurge(
1387         xfs_dquot_t     *dqp)
1388 {
1389         xfs_dqhash_t    *thishash;
1390         xfs_mount_t     *mp = dqp->q_mount;
1391
1392         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
1393         ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1394
1395         xfs_dqlock(dqp);
1396         /*
1397          * We really can't afford to purge a dquot that is
1398          * referenced, because these are hard refs.
1399          * It shouldn't happen in general because we went thru _all_ inodes in
1400          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1401          * However it is possible that we have dquots with temporary
1402          * references that are not attached to an inode. e.g. see xfs_setattr().
1403          */
1404         if (dqp->q_nrefs != 0) {
1405                 xfs_dqunlock(dqp);
1406                 mutex_unlock(&dqp->q_hash->qh_lock);
1407                 return (1);
1408         }
1409
1410         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1411
1412         /*
1413          * If we're turning off quotas, we have to make sure that, for
1414          * example, we don't delete quota disk blocks while dquots are
1415          * in the process of getting written to those disk blocks.
1416          * This dquot might well be on AIL, and we can't leave it there
1417          * if we're turning off quotas. Basically, we need this flush
1418          * lock, and are willing to block on it.
1419          */
1420         if (!xfs_dqflock_nowait(dqp)) {
1421                 /*
1422                  * Block on the flush lock after nudging dquot buffer,
1423                  * if it is incore.
1424                  */
1425                 xfs_qm_dqflock_pushbuf_wait(dqp);
1426         }
1427
1428         /*
1429          * XXXIf we're turning this type of quotas off, we don't care
1430          * about the dirty metadata sitting in this dquot. OTOH, if
1431          * we're unmounting, we do care, so we flush it and wait.
1432          */
1433         if (XFS_DQ_IS_DIRTY(dqp)) {
1434                 int     error;
1435
1436                 /* dqflush unlocks dqflock */
1437                 /*
1438                  * Given that dqpurge is a very rare occurrence, it is OK
1439                  * that we're holding the hashlist and mplist locks
1440                  * across the disk write. But, ... XXXsup
1441                  *
1442                  * We don't care about getting disk errors here. We need
1443                  * to purge this dquot anyway, so we go ahead regardless.
1444                  */
1445                 error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1446                 if (error)
1447                         xfs_fs_cmn_err(CE_WARN, mp,
1448                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1449                 xfs_dqflock(dqp);
1450         }
1451         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1452         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1453                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1454
1455         thishash = dqp->q_hash;
1456         XQM_HASHLIST_REMOVE(thishash, dqp);
1457         XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(mp)), dqp);
1458         /*
1459          * XXX Move this to the front of the freelist, if we can get the
1460          * freelist lock.
1461          */
1462         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1463
1464         dqp->q_mount = NULL;
1465         dqp->q_hash = NULL;
1466         dqp->dq_flags = XFS_DQ_INACTIVE;
1467         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1468         xfs_dqfunlock(dqp);
1469         xfs_dqunlock(dqp);
1470         mutex_unlock(&thishash->qh_lock);
1471         return (0);
1472 }
1473
1474
1475 #ifdef QUOTADEBUG
1476 void
1477 xfs_qm_dqprint(xfs_dquot_t *dqp)
1478 {
1479         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1480         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1481                 (int)be32_to_cpu(dqp->q_core.d_id));
1482         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1483         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1484         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1485         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1486         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1487                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1488                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1489         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1490                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1491                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1492         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1493                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1494                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1495         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1496                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1497                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1498         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1499                 be64_to_cpu(dqp->q_core.d_bcount),
1500                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1501         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1502                 be64_to_cpu(dqp->q_core.d_icount),
1503                 (int)be64_to_cpu(dqp->q_core.d_icount));
1504         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1505                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1506         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1507                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1508         cmn_err(CE_DEBUG, "---------------------------");
1509 }
1510 #endif
1511
1512 /*
1513  * Give the buffer a little push if it is incore and
1514  * wait on the flush lock.
1515  */
1516 void
1517 xfs_qm_dqflock_pushbuf_wait(
1518         xfs_dquot_t     *dqp)
1519 {
1520         xfs_buf_t       *bp;
1521
1522         /*
1523          * Check to see if the dquot has been flushed delayed
1524          * write.  If so, grab its buffer and send it
1525          * out immediately.  We'll be able to acquire
1526          * the flush lock when the I/O completes.
1527          */
1528         bp = xfs_incore(dqp->q_mount->m_ddev_targp, dqp->q_blkno,
1529                     XFS_QI_DQCHUNKLEN(dqp->q_mount), XBF_TRYLOCK);
1530         if (!bp)
1531                 goto out_lock;
1532
1533         if (XFS_BUF_ISDELAYWRITE(bp)) {
1534                 if (XFS_BUF_ISPINNED(bp))
1535                         xfs_log_force(dqp->q_mount, 0);
1536                 xfs_buf_delwri_promote(bp);
1537                 wake_up_process(bp->b_target->bt_task);
1538         }
1539         xfs_buf_relse(bp);
1540 out_lock:
1541         xfs_dqflock(dqp);
1542 }