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_acl.h"
  46 #include "xfs_attr.h"
  47 #include "xfs_buf_item.h"
  48 #include "xfs_trans_space.h"
  49 #include "xfs_trans_priv.h"
  50 #include "xfs_qm.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 /*
  77  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  78  * we try to reclaim a free dquot if the number of incore dquots are above
  79  * a threshold.
  80  * The only field inside the core that gets initialized at this point
  81  * is the d_id field. The idea is to fill in the entire q_core
  82  * when we read in the on disk dquot.
  83  */
  84 STATIC xfs_dquot_t *
  85 xfs_qm_dqinit(
  86         xfs_mount_t  *mp,
  87         xfs_dqid_t   id,
  88         uint         type)
  89 {
  90         xfs_dquot_t     *dqp;
  91         boolean_t       brandnewdquot;
  92
  93         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  94         dqp->dq_flags = type;
  95         dqp->q_core.d_id = cpu_to_be32(id);
  96         dqp->q_mount = mp;
  97
  98         /*
  99          * No need to re-initialize these if this is a reclaimed dquot.
 100          */
 101         if (brandnewdquot) {
 102                 dqp->dq_flnext = dqp->dq_flprev = dqp;
 103                 mutex_init(&dqp->q_qlock);
 104                 initnsema(&dqp->q_flock, 1, "fdq");
 105                 sv_init(&dqp->q_pinwait, SV_DEFAULT, "pdq");
 106
 107 #ifdef XFS_DQUOT_TRACE
 108                 dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_SLEEP);
 109                 xfs_dqtrace_entry(dqp, "DQINIT");
 110 #endif
 111         } else {
 112                 /*
 113                  * Only the q_core portion was zeroed in dqreclaim_one().
 114                  * So, we need to reset others.
 115                  */
 116                  dqp->q_nrefs = 0;
 117                  dqp->q_blkno = 0;
 118                  dqp->MPL_NEXT = dqp->HL_NEXT = NULL;
 119                  dqp->HL_PREVP = dqp->MPL_PREVP = NULL;
 120                  dqp->q_bufoffset = 0;
 121                  dqp->q_fileoffset = 0;
 122                  dqp->q_transp = NULL;
 123                  dqp->q_gdquot = NULL;
 124                  dqp->q_res_bcount = 0;
 125                  dqp->q_res_icount = 0;
 126                  dqp->q_res_rtbcount = 0;
 127                  dqp->q_pincount = 0;
 128                  dqp->q_hash = NULL;
 129                  ASSERT(dqp->dq_flnext == dqp->dq_flprev);
 130
 131 #ifdef XFS_DQUOT_TRACE
 132                  ASSERT(dqp->q_trace);
 133                  xfs_dqtrace_entry(dqp, "DQRECLAIMED_INIT");
 134 #endif
 135          }
 136
 137         /*
 138          * log item gets initialized later
 139          */
 140         return (dqp);
 141 }
 142
 143 /*
 144  * This is called to free all the memory associated with a dquot
 145  */
 146 void
 147 xfs_qm_dqdestroy(
 148         xfs_dquot_t     *dqp)
 149 {
 150         ASSERT(! XFS_DQ_IS_ON_FREELIST(dqp));
 151
 152         mutex_destroy(&dqp->q_qlock);
 153         freesema(&dqp->q_flock);
 154         sv_destroy(&dqp->q_pinwait);
 155
 156 #ifdef XFS_DQUOT_TRACE
 157         if (dqp->q_trace)
 158              ktrace_free(dqp->q_trace);
 159         dqp->q_trace = NULL;
 160 #endif
 161         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 162         atomic_dec(&xfs_Gqm->qm_totaldquots);
 163 }
 164
 165 /*
 166  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 167  */
 168 STATIC void
 169 xfs_qm_dqinit_core(
 170         xfs_dqid_t      id,
 171         uint            type,
 172         xfs_dqblk_t     *d)
 173 {
 174         /*
 175          * Caller has zero'd the entire dquot 'chunk' already.
 176          */
 177         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 178         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 179         d->dd_diskdq.d_id = cpu_to_be32(id);
 180         d->dd_diskdq.d_flags = type;
 181 }
 182
 183
 184 #ifdef XFS_DQUOT_TRACE
 185 /*
 186  * Dquot tracing for debugging.
 187  */
 188 /* ARGSUSED */
 189 void
 190 __xfs_dqtrace_entry(
 191         xfs_dquot_t     *dqp,
 192         char            *func,
 193         void            *retaddr,
 194         xfs_inode_t     *ip)
 195 {
 196         xfs_dquot_t     *udqp = NULL;
 197         xfs_ino_t       ino = 0;
 198
 199         ASSERT(dqp->q_trace);
 200         if (ip) {
 201                 ino = ip->i_ino;
 202                 udqp = ip->i_udquot;
 203         }
 204         ktrace_enter(dqp->q_trace,
 205                      (void *)(__psint_t)DQUOT_KTRACE_ENTRY,
 206                      (void *)func,
 207                      (void *)(__psint_t)dqp->q_nrefs,
 208                      (void *)(__psint_t)dqp->dq_flags,
 209                      (void *)(__psint_t)dqp->q_res_bcount,
 210                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_bcount),
 211                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_icount),
 212                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_hardlimit),
 213                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_softlimit),
 214                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_hardlimit),
 215                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_softlimit),
 216                      (void *)(__psint_t)be32_to_cpu(dqp->q_core.d_id),
 217                      (void *)(__psint_t)current_pid(),
 218                      (void *)(__psint_t)ino,
 219                      (void *)(__psint_t)retaddr,
 220                      (void *)(__psint_t)udqp);
 221         return;
 222 }
 223 #endif
 224
 225
 226 /*
 227  * If default limits are in force, push them into the dquot now.
 228  * We overwrite the dquot limits only if they are zero and this
 229  * is not the root dquot.
 230  */
 231 void
 232 xfs_qm_adjust_dqlimits(
 233         xfs_mount_t             *mp,
 234         xfs_disk_dquot_t        *d)
 235 {
 236         xfs_quotainfo_t         *q = mp->m_quotainfo;
 237
 238         ASSERT(d->d_id);
 239
 240         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 241                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 242         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 243                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 244         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 245                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 246         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 247                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 248         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 249                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 250         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 251                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 252 }
 253
 254 /*
 255  * Check the limits and timers of a dquot and start or reset timers
 256  * if necessary.
 257  * This gets called even when quota enforcement is OFF, which makes our
 258  * life a little less complicated. (We just don't reject any quota
 259  * reservations in that case, when enforcement is off).
 260  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 261  * enforcement's off.
 262  * In contrast, warnings are a little different in that they don't
 263  * 'automatically' get started when limits get exceeded.  They do
 264  * get reset to zero, however, when we find the count to be under
 265  * the soft limit (they are only ever set non-zero via userspace).
 266  */
 267 void
 268 xfs_qm_adjust_dqtimers(
 269         xfs_mount_t             *mp,
 270         xfs_disk_dquot_t        *d)
 271 {
 272         ASSERT(d->d_id);
 273
 274 #ifdef QUOTADEBUG
 275         if (d->d_blk_hardlimit)
 276                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 277                        be64_to_cpu(d->d_blk_hardlimit));
 278         if (d->d_ino_hardlimit)
 279                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 280                        be64_to_cpu(d->d_ino_hardlimit));
 281         if (d->d_rtb_hardlimit)
 282                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 283                        be64_to_cpu(d->d_rtb_hardlimit));
 284 #endif
 285         if (!d->d_btimer) {
 286                 if ((d->d_blk_softlimit &&
 287                      (be64_to_cpu(d->d_bcount) >=
 288                       be64_to_cpu(d->d_blk_softlimit))) ||
 289                     (d->d_blk_hardlimit &&
 290                      (be64_to_cpu(d->d_bcount) >=
 291                       be64_to_cpu(d->d_blk_hardlimit)))) {
 292                         d->d_btimer = cpu_to_be32(get_seconds() +
 293                                         XFS_QI_BTIMELIMIT(mp));
 294                 } else {
 295                         d->d_bwarns = 0;
 296                 }
 297         } else {
 298                 if ((!d->d_blk_softlimit ||
 299                      (be64_to_cpu(d->d_bcount) <
 300                       be64_to_cpu(d->d_blk_softlimit))) &&
 301                     (!d->d_blk_hardlimit ||
 302                     (be64_to_cpu(d->d_bcount) <
 303                      be64_to_cpu(d->d_blk_hardlimit)))) {
 304                         d->d_btimer = 0;
 305                 }
 306         }
 307
 308         if (!d->d_itimer) {
 309                 if ((d->d_ino_softlimit &&
 310                      (be64_to_cpu(d->d_icount) >=
 311                       be64_to_cpu(d->d_ino_softlimit))) ||
 312                     (d->d_ino_hardlimit &&
 313                      (be64_to_cpu(d->d_icount) >=
 314                       be64_to_cpu(d->d_ino_hardlimit)))) {
 315                         d->d_itimer = cpu_to_be32(get_seconds() +
 316                                         XFS_QI_ITIMELIMIT(mp));
 317                 } else {
 318                         d->d_iwarns = 0;
 319                 }
 320         } else {
 321                 if ((!d->d_ino_softlimit ||
 322                      (be64_to_cpu(d->d_icount) <
 323                       be64_to_cpu(d->d_ino_softlimit)))  &&
 324                     (!d->d_ino_hardlimit ||
 325                      (be64_to_cpu(d->d_icount) <
 326                       be64_to_cpu(d->d_ino_hardlimit)))) {
 327                         d->d_itimer = 0;
 328                 }
 329         }
 330
 331         if (!d->d_rtbtimer) {
 332                 if ((d->d_rtb_softlimit &&
 333                      (be64_to_cpu(d->d_rtbcount) >=
 334                       be64_to_cpu(d->d_rtb_softlimit))) ||
 335                     (d->d_rtb_hardlimit &&
 336                      (be64_to_cpu(d->d_rtbcount) >=
 337                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 338                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 339                                         XFS_QI_RTBTIMELIMIT(mp));
 340                 } else {
 341                         d->d_rtbwarns = 0;
 342                 }
 343         } else {
 344                 if ((!d->d_rtb_softlimit ||
 345                      (be64_to_cpu(d->d_rtbcount) <
 346                       be64_to_cpu(d->d_rtb_softlimit))) &&
 347                     (!d->d_rtb_hardlimit ||
 348                      (be64_to_cpu(d->d_rtbcount) <
 349                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 350                         d->d_rtbtimer = 0;
 351                 }
 352         }
 353 }
 354
 355 /*
 356  * initialize a buffer full of dquots and log the whole thing
 357  */
 358 STATIC void
 359 xfs_qm_init_dquot_blk(
 360         xfs_trans_t     *tp,
 361         xfs_mount_t     *mp,
 362         xfs_dqid_t      id,
 363         uint            type,
 364         xfs_buf_t       *bp)
 365 {
 366         xfs_dqblk_t     *d;
 367         int             curid, i;
 368
 369         ASSERT(tp);
 370         ASSERT(XFS_BUF_ISBUSY(bp));
 371         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 372
 373         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 374
 375         /*
 376          * ID of the first dquot in the block - id's are zero based.
 377          */
 378         curid = id - (id % XFS_QM_DQPERBLK(mp));
 379         ASSERT(curid >= 0);
 380         memset(d, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)));
 381         for (i = 0; i < XFS_QM_DQPERBLK(mp); i++, d++, curid++)
 382                 xfs_qm_dqinit_core(curid, type, d);
 383         xfs_trans_dquot_buf(tp, bp,
 384                             (type & XFS_DQ_USER ? XFS_BLI_UDQUOT_BUF :
 385                             ((type & XFS_DQ_PROJ) ? XFS_BLI_PDQUOT_BUF :
 386                              XFS_BLI_GDQUOT_BUF)));
 387         xfs_trans_log_buf(tp, bp, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1);
 388 }
 389
 390
 391
 392 /*
 393  * Allocate a block and fill it with dquots.
 394  * This is called when the bmapi finds a hole.
 395  */
 396 STATIC int
 397 xfs_qm_dqalloc(
 398         xfs_trans_t     **tpp,
 399         xfs_mount_t     *mp,
 400         xfs_dquot_t     *dqp,
 401         xfs_inode_t     *quotip,
 402         xfs_fileoff_t   offset_fsb,
 403         xfs_buf_t       **O_bpp)
 404 {
 405         xfs_fsblock_t   firstblock;
 406         xfs_bmap_free_t flist;
 407         xfs_bmbt_irec_t map;
 408         int             nmaps, error, committed;
 409         xfs_buf_t       *bp;
 410         xfs_trans_t     *tp = *tpp;
 411
 412         ASSERT(tp != NULL);
 413         xfs_dqtrace_entry(dqp, "DQALLOC");
 414
 415         /*
 416          * Initialize the bmap freelist prior to calling bmapi code.
 417          */
 418         XFS_BMAP_INIT(&flist, &firstblock);
 419         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 420         /*
 421          * Return if this type of quotas is turned off while we didn't
 422          * have an inode lock
 423          */
 424         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 425                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 426                 return (ESRCH);
 427         }
 428
 429         /*
 430          * xfs_trans_commit normally decrements the vnode ref count
 431          * when it unlocks the inode. Since we want to keep the quota
 432          * inode around, we bump the vnode ref count now.
 433          */
 434         VN_HOLD(XFS_ITOV(quotip));
 435
 436         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 437         nmaps = 1;
 438         if ((error = xfs_bmapi(tp, quotip,
 439                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 440                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 441                               &firstblock,
 442                               XFS_QM_DQALLOC_SPACE_RES(mp),
 443                               &map, &nmaps, &flist, NULL))) {
 444                 goto error0;
 445         }
 446         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 447         ASSERT(nmaps == 1);
 448         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 449                (map.br_startblock != HOLESTARTBLOCK));
 450
 451         /*
 452          * Keep track of the blkno to save a lookup later
 453          */
 454         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 455
 456         /* now we can just get the buffer (there's nothing to read yet) */
 457         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 458                                dqp->q_blkno,
 459                                XFS_QI_DQCHUNKLEN(mp),
 460                                0);
 461         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 462                 goto error1;
 463         /*
 464          * Make a chunk of dquots out of this buffer and log
 465          * the entire thing.
 466          */
 467         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 468                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 469
 470         /*
 471          * xfs_bmap_finish() may commit the current transaction and
 472          * start a second transaction if the freelist is not empty.
 473          *
 474          * Since we still want to modify this buffer, we need to
 475          * ensure that the buffer is not released on commit of
 476          * the first transaction and ensure the buffer is added to the
 477          * second transaction.
 478          *
 479          * If there is only one transaction then don't stop the buffer
 480          * from being released when it commits later on.
 481          */
 482
 483         xfs_trans_bhold(tp, bp);
 484
 485         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 486                 goto error1;
 487         }
 488
 489         if (committed) {
 490                 tp = *tpp;
 491                 xfs_trans_bjoin(tp, bp);
 492         } else {
 493                 xfs_trans_bhold_release(tp, bp);
 494         }
 495
 496         *O_bpp = bp;
 497         return 0;
 498
 499       error1:
 500         xfs_bmap_cancel(&flist);
 501       error0:
 502         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 503
 504         return (error);
 505 }
 506
 507 /*
 508  * Maps a dquot to the buffer containing its on-disk version.
 509  * This returns a ptr to the buffer containing the on-disk dquot
 510  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 511  */
 512 STATIC int
 513 xfs_qm_dqtobp(
 514         xfs_trans_t             **tpp,
 515         xfs_dquot_t             *dqp,
 516         xfs_disk_dquot_t        **O_ddpp,
 517         xfs_buf_t               **O_bpp,
 518         uint                    flags)
 519 {
 520         xfs_bmbt_irec_t map;
 521         int             nmaps, error;
 522         xfs_buf_t       *bp;
 523         xfs_inode_t     *quotip;
 524         xfs_mount_t     *mp;
 525         xfs_disk_dquot_t *ddq;
 526         xfs_dqid_t      id;
 527         boolean_t       newdquot;
 528         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 529
 530         mp = dqp->q_mount;
 531         id = be32_to_cpu(dqp->q_core.d_id);
 532         nmaps = 1;
 533         newdquot = B_FALSE;
 534
 535         /*
 536          * If we don't know where the dquot lives, find out.
 537          */
 538         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 539                 /* We use the id as an index */
 540                 dqp->q_fileoffset = (xfs_fileoff_t)id / XFS_QM_DQPERBLK(mp);
 541                 nmaps = 1;
 542                 quotip = XFS_DQ_TO_QIP(dqp);
 543                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 544                 /*
 545                  * Return if this type of quotas is turned off while we didn't
 546                  * have an inode lock
 547                  */
 548                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 549                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 550                         return (ESRCH);
 551                 }
 552                 /*
 553                  * Find the block map; no allocations yet
 554                  */
 555                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 556                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 557                                   XFS_BMAPI_METADATA,
 558                                   NULL, 0, &map, &nmaps, NULL, NULL);
 559
 560                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 561                 if (error)
 562                         return (error);
 563                 ASSERT(nmaps == 1);
 564                 ASSERT(map.br_blockcount == 1);
 565
 566                 /*
 567                  * offset of dquot in the (fixed sized) dquot chunk.
 568                  */
 569                 dqp->q_bufoffset = (id % XFS_QM_DQPERBLK(mp)) *
 570                         sizeof(xfs_dqblk_t);
 571                 if (map.br_startblock == HOLESTARTBLOCK) {
 572                         /*
 573                          * We don't allocate unless we're asked to
 574                          */
 575                         if (!(flags & XFS_QMOPT_DQALLOC))
 576                                 return (ENOENT);
 577
 578                         ASSERT(tp);
 579                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 580                                                 dqp->q_fileoffset, &bp)))
 581                                 return (error);
 582                         tp = *tpp;
 583                         newdquot = B_TRUE;
 584                 } else {
 585                         /*
 586                          * store the blkno etc so that we don't have to do the
 587                          * mapping all the time
 588                          */
 589                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 590                 }
 591         }
 592         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 593         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 594
 595         /*
 596          * Read in the buffer, unless we've just done the allocation
 597          * (in which case we already have the buf).
 598          */
 599         if (! newdquot) {
 600                 xfs_dqtrace_entry(dqp, "DQTOBP READBUF");
 601                 if ((error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 602                                                dqp->q_blkno,
 603                                                XFS_QI_DQCHUNKLEN(mp),
 604                                                0, &bp))) {
 605                         return (error);
 606                 }
 607                 if (error || !bp)
 608                         return XFS_ERROR(error);
 609         }
 610         ASSERT(XFS_BUF_ISBUSY(bp));
 611         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 612
 613         /*
 614          * calculate the location of the dquot inside the buffer.
 615          */
 616         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 617
 618         /*
 619          * A simple sanity check in case we got a corrupted dquot...
 620          */
 621         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 622                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 623                            "dqtobp")) {
 624                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 625                         xfs_trans_brelse(tp, bp);
 626                         return XFS_ERROR(EIO);
 627                 }
 628                 XFS_BUF_BUSY(bp); /* We dirtied this */
 629         }
 630
 631         *O_bpp = bp;
 632         *O_ddpp = ddq;
 633
 634         return (0);
 635 }
 636
 637
 638 /*
 639  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 640  * and release the buffer immediately.
 641  *
 642  */
 643 /* ARGSUSED */
 644 STATIC int
 645 xfs_qm_dqread(
 646         xfs_trans_t     **tpp,
 647         xfs_dqid_t      id,
 648         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 649         uint            flags)
 650 {
 651         xfs_disk_dquot_t *ddqp;
 652         xfs_buf_t        *bp;
 653         int              error;
 654         xfs_trans_t      *tp;
 655
 656         ASSERT(tpp);
 657
 658         /*
 659          * get a pointer to the on-disk dquot and the buffer containing it
 660          * dqp already knows its own type (GROUP/USER).
 661          */
 662         xfs_dqtrace_entry(dqp, "DQREAD");
 663         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 664                 return (error);
 665         }
 666         tp = *tpp;
 667
 668         /* copy everything from disk dquot to the incore dquot */
 669         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 670         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 671         xfs_qm_dquot_logitem_init(dqp);
 672
 673         /*
 674          * Reservation counters are defined as reservation plus current usage
 675          * to avoid having to add everytime.
 676          */
 677         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 678         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 679         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 680
 681         /* Mark the buf so that this will stay incore a little longer */
 682         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 683
 684         /*
 685          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 686          * So we need to release with xfs_trans_brelse().
 687          * The strategy here is identical to that of inodes; we lock
 688          * the dquot in xfs_qm_dqget() before making it accessible to
 689          * others. This is because dquots, like inodes, need a good level of
 690          * concurrency, and we don't want to take locks on the entire buffers
 691          * for dquot accesses.
 692          * Note also that the dquot buffer may even be dirty at this point, if
 693          * this particular dquot was repaired. We still aren't afraid to
 694          * brelse it because we have the changes incore.
 695          */
 696         ASSERT(XFS_BUF_ISBUSY(bp));
 697         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 698         xfs_trans_brelse(tp, bp);
 699
 700         return (error);
 701 }
 702
 703
 704 /*
 705  * allocate an incore dquot from the kernel heap,
 706  * and fill its core with quota information kept on disk.
 707  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 708  * if it wasn't already allocated.
 709  */
 710 STATIC int
 711 xfs_qm_idtodq(
 712         xfs_mount_t     *mp,
 713         xfs_dqid_t      id,      /* gid or uid, depending on type */
 714         uint            type,    /* UDQUOT or GDQUOT */
 715         uint            flags,   /* DQALLOC, DQREPAIR */
 716         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 717 {
 718         xfs_dquot_t     *dqp;
 719         int             error;
 720         xfs_trans_t     *tp;
 721         int             cancelflags=0;
 722
 723         dqp = xfs_qm_dqinit(mp, id, type);
 724         tp = NULL;
 725         if (flags & XFS_QMOPT_DQALLOC) {
 726                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 727                 if ((error = xfs_trans_reserve(tp,
 728                                        XFS_QM_DQALLOC_SPACE_RES(mp),
 729                                        XFS_WRITE_LOG_RES(mp) +
 730                                               BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1 +
 731                                               128,
 732                                        0,
 733                                        XFS_TRANS_PERM_LOG_RES,
 734                                        XFS_WRITE_LOG_COUNT))) {
 735                         cancelflags = 0;
 736                         goto error0;
 737                 }
 738                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 739         }
 740
 741         /*
 742          * Read it from disk; xfs_dqread() takes care of
 743          * all the necessary initialization of dquot's fields (locks, etc)
 744          */
 745         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 746                 /*
 747                  * This can happen if quotas got turned off (ESRCH),
 748                  * or if the dquot didn't exist on disk and we ask to
 749                  * allocate (ENOENT).
 750                  */
 751                 xfs_dqtrace_entry(dqp, "DQREAD FAIL");
 752                 cancelflags |= XFS_TRANS_ABORT;
 753                 goto error0;
 754         }
 755         if (tp) {
 756                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 757                         goto error1;
 758         }
 759
 760         *O_dqpp = dqp;
 761         return (0);
 762
 763  error0:
 764         ASSERT(error);
 765         if (tp)
 766                 xfs_trans_cancel(tp, cancelflags);
 767  error1:
 768         xfs_qm_dqdestroy(dqp);
 769         *O_dqpp = NULL;
 770         return (error);
 771 }
 772
 773 /*
 774  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 775  * hashtables for user and group dquots; and, these are global tables
 776  * inside the XQM, not per-filesystem tables.
 777  * The hash chain must be locked by caller, and it is left locked
 778  * on return. Returning dquot is locked.
 779  */
 780 STATIC int
 781 xfs_qm_dqlookup(
 782         xfs_mount_t             *mp,
 783         xfs_dqid_t              id,
 784         xfs_dqhash_t            *qh,
 785         xfs_dquot_t             **O_dqpp)
 786 {
 787         xfs_dquot_t             *dqp;
 788         uint                    flist_locked;
 789         xfs_dquot_t             *d;
 790
 791         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 792
 793         flist_locked = B_FALSE;
 794
 795         /*
 796          * Traverse the hashchain looking for a match
 797          */
 798         for (dqp = qh->qh_next; dqp != NULL; dqp = dqp->HL_NEXT) {
 799                 /*
 800                  * We already have the hashlock. We don't need the
 801                  * dqlock to look at the id field of the dquot, since the
 802                  * id can't be modified without the hashlock anyway.
 803                  */
 804                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 805                         xfs_dqtrace_entry(dqp, "DQFOUND BY LOOKUP");
 806                         /*
 807                          * All in core dquots must be on the dqlist of mp
 808                          */
 809                         ASSERT(dqp->MPL_PREVP != NULL);
 810
 811                         xfs_dqlock(dqp);
 812                         if (dqp->q_nrefs == 0) {
 813                                 ASSERT (XFS_DQ_IS_ON_FREELIST(dqp));
 814                                 if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
 815                                         xfs_dqtrace_entry(dqp, "DQLOOKUP: WANT");
 816
 817                                         /*
 818                                          * We may have raced with dqreclaim_one()
 819                                          * (and lost). So, flag that we don't
 820                                          * want the dquot to be reclaimed.
 821                                          */
 822                                         dqp->dq_flags |= XFS_DQ_WANT;
 823                                         xfs_dqunlock(dqp);
 824                                         xfs_qm_freelist_lock(xfs_Gqm);
 825                                         xfs_dqlock(dqp);
 826                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 827                                 }
 828                                 flist_locked = B_TRUE;
 829                         }
 830
 831                         /*
 832                          * id couldn't have changed; we had the hashlock all
 833                          * along
 834                          */
 835                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 836
 837                         if (flist_locked) {
 838                                 if (dqp->q_nrefs != 0) {
 839                                         xfs_qm_freelist_unlock(xfs_Gqm);
 840                                         flist_locked = B_FALSE;
 841                                 } else {
 842                                         /*
 843                                          * take it off the freelist
 844                                          */
 845                                         xfs_dqtrace_entry(dqp,
 846                                                         "DQLOOKUP: TAKEOFF FL");
 847                                         XQM_FREELIST_REMOVE(dqp);
 848                                         /* xfs_qm_freelist_print(&(xfs_Gqm->
 849                                                         qm_dqfreelist),
 850                                                         "after removal"); */
 851                                 }
 852                         }
 853
 854                         /*
 855                          * grab a reference
 856                          */
 857                         XFS_DQHOLD(dqp);
 858
 859                         if (flist_locked)
 860                                 xfs_qm_freelist_unlock(xfs_Gqm);
 861                         /*
 862                          * move the dquot to the front of the hashchain
 863                          */
 864                         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 865                         if (dqp->HL_PREVP != &qh->qh_next) {
 866                                 xfs_dqtrace_entry(dqp,
 867                                                   "DQLOOKUP: HASH MOVETOFRONT");
 868                                 if ((d = dqp->HL_NEXT))
 869                                         d->HL_PREVP = dqp->HL_PREVP;
 870                                 *(dqp->HL_PREVP) = d;
 871                                 d = qh->qh_next;
 872                                 d->HL_PREVP = &dqp->HL_NEXT;
 873                                 dqp->HL_NEXT = d;
 874                                 dqp->HL_PREVP = &qh->qh_next;
 875                                 qh->qh_next = dqp;
 876                         }
 877                         xfs_dqtrace_entry(dqp, "LOOKUP END");
 878                         *O_dqpp = dqp;
 879                         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 880                         return (0);
 881                 }
 882         }
 883
 884         *O_dqpp = NULL;
 885         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 886         return (1);
 887 }
 888
 889 /*
 890  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 891  * a locked dquot, doing an allocation (if requested) as needed.
 892  * When both an inode and an id are given, the inode's id takes precedence.
 893  * That is, if the id changes while we don't hold the ilock inside this
 894  * function, the new dquot is returned, not necessarily the one requested
 895  * in the id argument.
 896  */
 897 int
 898 xfs_qm_dqget(
 899         xfs_mount_t     *mp,
 900         xfs_inode_t     *ip,      /* locked inode (optional) */
 901         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 902         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 903         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 904         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 905 {
 906         xfs_dquot_t     *dqp;
 907         xfs_dqhash_t    *h;
 908         uint            version;
 909         int             error;
 910
 911         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 912         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 913             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 914             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 915                 return (ESRCH);
 916         }
 917         h = XFS_DQ_HASH(mp, id, type);
 918
 919 #ifdef DEBUG
 920         if (xfs_do_dqerror) {
 921                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 922                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 923                         cmn_err(CE_DEBUG, "Returning error in dqget");
 924                         return (EIO);
 925                 }
 926         }
 927 #endif
 928
 929  again:
 930
 931 #ifdef DEBUG
 932         ASSERT(type == XFS_DQ_USER ||
 933                type == XFS_DQ_PROJ ||
 934                type == XFS_DQ_GROUP);
 935         if (ip) {
 936                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
 937                 if (type == XFS_DQ_USER)
 938                         ASSERT(ip->i_udquot == NULL);
 939                 else
 940                         ASSERT(ip->i_gdquot == NULL);
 941         }
 942 #endif
 943         XFS_DQ_HASH_LOCK(h);
 944
 945         /*
 946          * Look in the cache (hashtable).
 947          * The chain is kept locked during lookup.
 948          */
 949         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 950                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 951                 /*
 952                  * The dquot was found, moved to the front of the chain,
 953                  * taken off the freelist if it was on it, and locked
 954                  * at this point. Just unlock the hashchain and return.
 955                  */
 956                 ASSERT(*O_dqpp);
 957                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 958                 XFS_DQ_HASH_UNLOCK(h);
 959                 xfs_dqtrace_entry(*O_dqpp, "DQGET DONE (FROM CACHE)");
 960                 return (0);     /* success */
 961         }
 962         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 963
 964         /*
 965          * Dquot cache miss. We don't want to keep the inode lock across
 966          * a (potential) disk read. Also we don't want to deal with the lock
 967          * ordering between quotainode and this inode. OTOH, dropping the inode
 968          * lock here means dealing with a chown that can happen before
 969          * we re-acquire the lock.
 970          */
 971         if (ip)
 972                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 973         /*
 974          * Save the hashchain version stamp, and unlock the chain, so that
 975          * we don't keep the lock across a disk read
 976          */
 977         version = h->qh_version;
 978         XFS_DQ_HASH_UNLOCK(h);
 979
 980         /*
 981          * Allocate the dquot on the kernel heap, and read the ondisk
 982          * portion off the disk. Also, do all the necessary initialization
 983          * This can return ENOENT if dquot didn't exist on disk and we didn't
 984          * ask it to allocate; ESRCH if quotas got turned off suddenly.
 985          */
 986         if ((error = xfs_qm_idtodq(mp, id, type,
 987                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
 988                                            XFS_QMOPT_DOWARN),
 989                                   &dqp))) {
 990                 if (ip)
 991                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 992                 return (error);
 993         }
 994
 995         /*
 996          * See if this is mount code calling to look at the overall quota limits
 997          * which are stored in the id == 0 user or group's dquot.
 998          * Since we may not have done a quotacheck by this point, just return
 999          * the dquot without attaching it to any hashtables, lists, etc, or even
1000          * taking a reference.
1001          * The caller must dqdestroy this once done.
1002          */
1003         if (flags & XFS_QMOPT_DQSUSER) {
1004                 ASSERT(id == 0);
1005                 ASSERT(! ip);
1006                 goto dqret;
1007         }
1008
1009         /*
1010          * Dquot lock comes after hashlock in the lock ordering
1011          */
1012         if (ip) {
1013                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1014                 if (! XFS_IS_DQTYPE_ON(mp, type)) {
1015                         /* inode stays locked on return */
1016                         xfs_qm_dqdestroy(dqp);
1017                         return XFS_ERROR(ESRCH);
1018                 }
1019                 /*
1020                  * A dquot could be attached to this inode by now, since
1021                  * we had dropped the ilock.
1022                  */
1023                 if (type == XFS_DQ_USER) {
1024                         if (ip->i_udquot) {
1025                                 xfs_qm_dqdestroy(dqp);
1026                                 dqp = ip->i_udquot;
1027                                 xfs_dqlock(dqp);
1028                                 goto dqret;
1029                         }
1030                 } else {
1031                         if (ip->i_gdquot) {
1032                                 xfs_qm_dqdestroy(dqp);
1033                                 dqp = ip->i_gdquot;
1034                                 xfs_dqlock(dqp);
1035                                 goto dqret;
1036                         }
1037                 }
1038         }
1039
1040         /*
1041          * Hashlock comes after ilock in lock order
1042          */
1043         XFS_DQ_HASH_LOCK(h);
1044         if (version != h->qh_version) {
1045                 xfs_dquot_t *tmpdqp;
1046                 /*
1047                  * Now, see if somebody else put the dquot in the
1048                  * hashtable before us. This can happen because we didn't
1049                  * keep the hashchain lock. We don't have to worry about
1050                  * lock order between the two dquots here since dqp isn't
1051                  * on any findable lists yet.
1052                  */
1053                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
1054                         /*
1055                          * Duplicate found. Just throw away the new dquot
1056                          * and start over.
1057                          */
1058                         xfs_qm_dqput(tmpdqp);
1059                         XFS_DQ_HASH_UNLOCK(h);
1060                         xfs_qm_dqdestroy(dqp);
1061                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
1062                         goto again;
1063                 }
1064         }
1065
1066         /*
1067          * Put the dquot at the beginning of the hash-chain and mp's list
1068          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1069          */
1070         ASSERT(XFS_DQ_IS_HASH_LOCKED(h));
1071         dqp->q_hash = h;
1072         XQM_HASHLIST_INSERT(h, dqp);
1073
1074         /*
1075          * Attach this dquot to this filesystem's list of all dquots,
1076          * kept inside the mount structure in m_quotainfo field
1077          */
1078         xfs_qm_mplist_lock(mp);
1079
1080         /*
1081          * We return a locked dquot to the caller, with a reference taken
1082          */
1083         xfs_dqlock(dqp);
1084         dqp->q_nrefs = 1;
1085
1086         XQM_MPLIST_INSERT(&(XFS_QI_MPL_LIST(mp)), dqp);
1087
1088         xfs_qm_mplist_unlock(mp);
1089         XFS_DQ_HASH_UNLOCK(h);
1090  dqret:
1091         ASSERT((ip == NULL) || XFS_ISLOCKED_INODE_EXCL(ip));
1092         xfs_dqtrace_entry(dqp, "DQGET DONE");
1093         *O_dqpp = dqp;
1094         return (0);
1095 }
1096
1097
1098 /*
1099  * Release a reference to the dquot (decrement ref-count)
1100  * and unlock it. If there is a group quota attached to this
1101  * dquot, carefully release that too without tripping over
1102  * deadlocks'n'stuff.
1103  */
1104 void
1105 xfs_qm_dqput(
1106         xfs_dquot_t     *dqp)
1107 {
1108         xfs_dquot_t     *gdqp;
1109
1110         ASSERT(dqp->q_nrefs > 0);
1111         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1112         xfs_dqtrace_entry(dqp, "DQPUT");
1113
1114         if (dqp->q_nrefs != 1) {
1115                 dqp->q_nrefs--;
1116                 xfs_dqunlock(dqp);
1117                 return;
1118         }
1119
1120         /*
1121          * drop the dqlock and acquire the freelist and dqlock
1122          * in the right order; but try to get it out-of-order first
1123          */
1124         if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
1125                 xfs_dqtrace_entry(dqp, "DQPUT: FLLOCK-WAIT");
1126                 xfs_dqunlock(dqp);
1127                 xfs_qm_freelist_lock(xfs_Gqm);
1128                 xfs_dqlock(dqp);
1129         }
1130
1131         while (1) {
1132                 gdqp = NULL;
1133
1134                 /* We can't depend on nrefs being == 1 here */
1135                 if (--dqp->q_nrefs == 0) {
1136                         xfs_dqtrace_entry(dqp, "DQPUT: ON FREELIST");
1137                         /*
1138                          * insert at end of the freelist.
1139                          */
1140                         XQM_FREELIST_INSERT(&(xfs_Gqm->qm_dqfreelist), dqp);
1141
1142                         /*
1143                          * If we just added a udquot to the freelist, then
1144                          * we want to release the gdquot reference that
1145                          * it (probably) has. Otherwise it'll keep the
1146                          * gdquot from getting reclaimed.
1147                          */
1148                         if ((gdqp = dqp->q_gdquot)) {
1149                                 /*
1150                                  * Avoid a recursive dqput call
1151                                  */
1152                                 xfs_dqlock(gdqp);
1153                                 dqp->q_gdquot = NULL;
1154                         }
1155
1156                         /* xfs_qm_freelist_print(&(xfs_Gqm->qm_dqfreelist),
1157                            "@@@@@++ Free list (after append) @@@@@+");
1158                            */
1159                 }
1160                 xfs_dqunlock(dqp);
1161
1162                 /*
1163                  * If we had a group quota inside the user quota as a hint,
1164                  * release it now.
1165                  */
1166                 if (! gdqp)
1167                         break;
1168                 dqp = gdqp;
1169         }
1170         xfs_qm_freelist_unlock(xfs_Gqm);
1171 }
1172
1173 /*
1174  * Release a dquot. Flush it if dirty, then dqput() it.
1175  * dquot must not be locked.
1176  */
1177 void
1178 xfs_qm_dqrele(
1179         xfs_dquot_t     *dqp)
1180 {
1181         ASSERT(dqp);
1182         xfs_dqtrace_entry(dqp, "DQRELE");
1183
1184         xfs_dqlock(dqp);
1185         /*
1186          * We don't care to flush it if the dquot is dirty here.
1187          * That will create stutters that we want to avoid.
1188          * Instead we do a delayed write when we try to reclaim
1189          * a dirty dquot. Also xfs_sync will take part of the burden...
1190          */
1191         xfs_qm_dqput(dqp);
1192 }
1193
1194
1195 /*
1196  * Write a modified dquot to disk.
1197  * The dquot must be locked and the flush lock too taken by caller.
1198  * The flush lock will not be unlocked until the dquot reaches the disk,
1199  * but the dquot is free to be unlocked and modified by the caller
1200  * in the interim. Dquot is still locked on return. This behavior is
1201  * identical to that of inodes.
1202  */
1203 int
1204 xfs_qm_dqflush(
1205         xfs_dquot_t             *dqp,
1206         uint                    flags)
1207 {
1208         xfs_mount_t             *mp;
1209         xfs_buf_t               *bp;
1210         xfs_disk_dquot_t        *ddqp;
1211         int                     error;
1212         SPLDECL(s);
1213
1214         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1215         ASSERT(XFS_DQ_IS_FLUSH_LOCKED(dqp));
1216         xfs_dqtrace_entry(dqp, "DQFLUSH");
1217
1218         /*
1219          * If not dirty, nada.
1220          */
1221         if (!XFS_DQ_IS_DIRTY(dqp)) {
1222                 xfs_dqfunlock(dqp);
1223                 return (0);
1224         }
1225
1226         /*
1227          * Cant flush a pinned dquot. Wait for it.
1228          */
1229         xfs_qm_dqunpin_wait(dqp);
1230
1231         /*
1232          * This may have been unpinned because the filesystem is shutting
1233          * down forcibly. If that's the case we must not write this dquot
1234          * to disk, because the log record didn't make it to disk!
1235          */
1236         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1237                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1238                 xfs_dqfunlock(dqp);
1239                 return XFS_ERROR(EIO);
1240         }
1241
1242         /*
1243          * Get the buffer containing the on-disk dquot
1244          * We don't need a transaction envelope because we know that the
1245          * the ondisk-dquot has already been allocated for.
1246          */
1247         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1248                 xfs_dqtrace_entry(dqp, "DQTOBP FAIL");
1249                 ASSERT(error != ENOENT);
1250                 /*
1251                  * Quotas could have gotten turned off (ESRCH)
1252                  */
1253                 xfs_dqfunlock(dqp);
1254                 return (error);
1255         }
1256
1257         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1258                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1259                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1260                 return XFS_ERROR(EIO);
1261         }
1262
1263         /* This is the only portion of data that needs to persist */
1264         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1265
1266         /*
1267          * Clear the dirty field and remember the flush lsn for later use.
1268          */
1269         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1270         mp = dqp->q_mount;
1271
1272         /* lsn is 64 bits */
1273         AIL_LOCK(mp, s);
1274         dqp->q_logitem.qli_flush_lsn = dqp->q_logitem.qli_item.li_lsn;
1275         AIL_UNLOCK(mp, s);
1276
1277         /*
1278          * Attach an iodone routine so that we can remove this dquot from the
1279          * AIL and release the flush lock once the dquot is synced to disk.
1280          */
1281         xfs_buf_attach_iodone(bp, (void(*)(xfs_buf_t *, xfs_log_item_t *))
1282                               xfs_qm_dqflush_done, &(dqp->q_logitem.qli_item));
1283         /*
1284          * If the buffer is pinned then push on the log so we won't
1285          * get stuck waiting in the write for too long.
1286          */
1287         if (XFS_BUF_ISPINNED(bp)) {
1288                 xfs_dqtrace_entry(dqp, "DQFLUSH LOG FORCE");
1289                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
1290         }
1291
1292         if (flags & XFS_QMOPT_DELWRI) {
1293                 xfs_bdwrite(mp, bp);
1294         } else if (flags & XFS_QMOPT_ASYNC) {
1295                 xfs_bawrite(mp, bp);
1296         } else {
1297                 error = xfs_bwrite(mp, bp);
1298         }
1299         xfs_dqtrace_entry(dqp, "DQFLUSH END");
1300         /*
1301          * dqp is still locked, but caller is free to unlock it now.
1302          */
1303         return (error);
1304
1305 }
1306
1307 /*
1308  * This is the dquot flushing I/O completion routine.  It is called
1309  * from interrupt level when the buffer containing the dquot is
1310  * flushed to disk.  It is responsible for removing the dquot logitem
1311  * from the AIL if it has not been re-logged, and unlocking the dquot's
1312  * flush lock. This behavior is very similar to that of inodes..
1313  */
1314 /*ARGSUSED*/
1315 STATIC void
1316 xfs_qm_dqflush_done(
1317         xfs_buf_t               *bp,
1318         xfs_dq_logitem_t        *qip)
1319 {
1320         xfs_dquot_t             *dqp;
1321         SPLDECL(s);
1322
1323         dqp = qip->qli_dquot;
1324
1325         /*
1326          * We only want to pull the item from the AIL if its
1327          * location in the log has not changed since we started the flush.
1328          * Thus, we only bother if the dquot's lsn has
1329          * not changed. First we check the lsn outside the lock
1330          * since it's cheaper, and then we recheck while
1331          * holding the lock before removing the dquot from the AIL.
1332          */
1333         if ((qip->qli_item.li_flags & XFS_LI_IN_AIL) &&
1334             qip->qli_item.li_lsn == qip->qli_flush_lsn) {
1335
1336                 AIL_LOCK(dqp->q_mount, s);
1337                 /*
1338                  * xfs_trans_delete_ail() drops the AIL lock.
1339                  */
1340                 if (qip->qli_item.li_lsn == qip->qli_flush_lsn)
1341                         xfs_trans_delete_ail(dqp->q_mount,
1342                                              (xfs_log_item_t*)qip, s);
1343                 else
1344                         AIL_UNLOCK(dqp->q_mount, s);
1345         }
1346
1347         /*
1348          * Release the dq's flush lock since we're done with it.
1349          */
1350         xfs_dqfunlock(dqp);
1351 }
1352
1353
1354 int
1355 xfs_qm_dqflock_nowait(
1356         xfs_dquot_t *dqp)
1357 {
1358         int locked;
1359
1360         locked = cpsema(&((dqp)->q_flock));
1361
1362         /* XXX ifdef these out */
1363         if (locked)
1364                 (dqp)->dq_flags |= XFS_DQ_FLOCKED;
1365         return (locked);
1366 }
1367
1368
1369 int
1370 xfs_qm_dqlock_nowait(
1371         xfs_dquot_t *dqp)
1372 {
1373         return (mutex_trylock(&((dqp)->q_qlock)));
1374 }
1375
1376 void
1377 xfs_dqlock(
1378         xfs_dquot_t *dqp)
1379 {
1380         mutex_lock(&(dqp->q_qlock));
1381 }
1382
1383 void
1384 xfs_dqunlock(
1385         xfs_dquot_t *dqp)
1386 {
1387         mutex_unlock(&(dqp->q_qlock));
1388         if (dqp->q_logitem.qli_dquot == dqp) {
1389                 /* Once was dqp->q_mount, but might just have been cleared */
1390                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_mountp,
1391                                         (xfs_log_item_t*)&(dqp->q_logitem));
1392         }
1393 }
1394
1395
1396 void
1397 xfs_dqunlock_nonotify(
1398         xfs_dquot_t *dqp)
1399 {
1400         mutex_unlock(&(dqp->q_qlock));
1401 }
1402
1403 void
1404 xfs_dqlock2(
1405         xfs_dquot_t     *d1,
1406         xfs_dquot_t     *d2)
1407 {
1408         if (d1 && d2) {
1409                 ASSERT(d1 != d2);
1410                 if (be32_to_cpu(d1->q_core.d_id) >
1411                     be32_to_cpu(d2->q_core.d_id)) {
1412                         xfs_dqlock(d2);
1413                         xfs_dqlock(d1);
1414                 } else {
1415                         xfs_dqlock(d1);
1416                         xfs_dqlock(d2);
1417                 }
1418         } else {
1419                 if (d1) {
1420                         xfs_dqlock(d1);
1421                 } else if (d2) {
1422                         xfs_dqlock(d2);
1423                 }
1424         }
1425 }
1426
1427
1428 /*
1429  * Take a dquot out of the mount's dqlist as well as the hashlist.
1430  * This is called via unmount as well as quotaoff, and the purge
1431  * will always succeed unless there are soft (temp) references
1432  * outstanding.
1433  *
1434  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1435  * that we're returning! XXXsup - not cool.
1436  */
1437 /* ARGSUSED */
1438 int
1439 xfs_qm_dqpurge(
1440         xfs_dquot_t     *dqp,
1441         uint            flags)
1442 {
1443         xfs_dqhash_t    *thishash;
1444         xfs_mount_t     *mp;
1445
1446         mp = dqp->q_mount;
1447
1448         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
1449         ASSERT(XFS_DQ_IS_HASH_LOCKED(dqp->q_hash));
1450
1451         xfs_dqlock(dqp);
1452         /*
1453          * We really can't afford to purge a dquot that is
1454          * referenced, because these are hard refs.
1455          * It shouldn't happen in general because we went thru _all_ inodes in
1456          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1457          * However it is possible that we have dquots with temporary
1458          * references that are not attached to an inode. e.g. see xfs_setattr().
1459          */
1460         if (dqp->q_nrefs != 0) {
1461                 xfs_dqunlock(dqp);
1462                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
1463                 return (1);
1464         }
1465
1466         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1467
1468         /*
1469          * If we're turning off quotas, we have to make sure that, for
1470          * example, we don't delete quota disk blocks while dquots are
1471          * in the process of getting written to those disk blocks.
1472          * This dquot might well be on AIL, and we can't leave it there
1473          * if we're turning off quotas. Basically, we need this flush
1474          * lock, and are willing to block on it.
1475          */
1476         if (! xfs_qm_dqflock_nowait(dqp)) {
1477                 /*
1478                  * Block on the flush lock after nudging dquot buffer,
1479                  * if it is incore.
1480                  */
1481                 xfs_qm_dqflock_pushbuf_wait(dqp);
1482         }
1483
1484         /*
1485          * XXXIf we're turning this type of quotas off, we don't care
1486          * about the dirty metadata sitting in this dquot. OTOH, if
1487          * we're unmounting, we do care, so we flush it and wait.
1488          */
1489         if (XFS_DQ_IS_DIRTY(dqp)) {
1490                 xfs_dqtrace_entry(dqp, "DQPURGE ->DQFLUSH: DQDIRTY");
1491                 /* dqflush unlocks dqflock */
1492                 /*
1493                  * Given that dqpurge is a very rare occurrence, it is OK
1494                  * that we're holding the hashlist and mplist locks
1495                  * across the disk write. But, ... XXXsup
1496                  *
1497                  * We don't care about getting disk errors here. We need
1498                  * to purge this dquot anyway, so we go ahead regardless.
1499                  */
1500                 (void) xfs_qm_dqflush(dqp, XFS_QMOPT_SYNC);
1501                 xfs_dqflock(dqp);
1502         }
1503         ASSERT(dqp->q_pincount == 0);
1504         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1505                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1506
1507         thishash = dqp->q_hash;
1508         XQM_HASHLIST_REMOVE(thishash, dqp);
1509         XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(mp)), dqp);
1510         /*
1511          * XXX Move this to the front of the freelist, if we can get the
1512          * freelist lock.
1513          */
1514         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1515
1516         dqp->q_mount = NULL;
1517         dqp->q_hash = NULL;
1518         dqp->dq_flags = XFS_DQ_INACTIVE;
1519         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1520         xfs_dqfunlock(dqp);
1521         xfs_dqunlock(dqp);
1522         XFS_DQ_HASH_UNLOCK(thishash);
1523         return (0);
1524 }
1525
1526
1527 #ifdef QUOTADEBUG
1528 void
1529 xfs_qm_dqprint(xfs_dquot_t *dqp)
1530 {
1531         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1532         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1533                 (int)be32_to_cpu(dqp->q_core.d_id));
1534         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1535         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1536         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1537         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1538         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1539                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1540                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1541         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1542                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1543                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1544         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1545                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1546                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1547         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1548                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1549                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1550         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1551                 be64_to_cpu(dqp->q_core.d_bcount),
1552                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1553         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1554                 be64_to_cpu(dqp->q_core.d_icount),
1555                 (int)be64_to_cpu(dqp->q_core.d_icount));
1556         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1557                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1558         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1559                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1560         cmn_err(CE_DEBUG, "---------------------------");
1561 }
1562 #endif
1563
1564 /*
1565  * Give the buffer a little push if it is incore and
1566  * wait on the flush lock.
1567  */
1568 void
1569 xfs_qm_dqflock_pushbuf_wait(
1570         xfs_dquot_t     *dqp)
1571 {
1572         xfs_buf_t       *bp;
1573
1574         /*
1575          * Check to see if the dquot has been flushed delayed
1576          * write.  If so, grab its buffer and send it
1577          * out immediately.  We'll be able to acquire
1578          * the flush lock when the I/O completes.
1579          */
1580         bp = xfs_incore(dqp->q_mount->m_ddev_targp, dqp->q_blkno,
1581                     XFS_QI_DQCHUNKLEN(dqp->q_mount),
1582                     XFS_INCORE_TRYLOCK);
1583         if (bp != NULL) {
1584                 if (XFS_BUF_ISDELAYWRITE(bp)) {
1585                         if (XFS_BUF_ISPINNED(bp)) {
1586                                 xfs_log_force(dqp->q_mount,
1587                                               (xfs_lsn_t)0,
1588                                               XFS_LOG_FORCE);
1589                         }
1590                         xfs_bawrite(dqp->q_mount, bp);
1591                 } else {
1592                         xfs_buf_relse(bp);
1593                 }
1594         }
1595         xfs_dqflock(dqp);
1596 }