git.oblomov.eu Git - linux-2.6/blob - 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                 sv_init(&dqp->q_pinwait, SV_DEFAULT, "pdq");
 105
 106                 /*
 107                  * Because we want to use a counting completion, complete
 108                  * the flush completion once to allow a single access to
 109                  * the flush completion without blocking.
 110                  */
 111                 init_completion(&dqp->q_flush);
 112                 complete(&dqp->q_flush);
 113
 114 #ifdef XFS_DQUOT_TRACE
 115                 dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_NOFS);
 116                 xfs_dqtrace_entry(dqp, "DQINIT");
 117 #endif
 118         } else {
 119                 /*
 120                  * Only the q_core portion was zeroed in dqreclaim_one().
 121                  * So, we need to reset others.
 122                  */
 123                  dqp->q_nrefs = 0;
 124                  dqp->q_blkno = 0;
 125                  dqp->MPL_NEXT = dqp->HL_NEXT = NULL;
 126                  dqp->HL_PREVP = dqp->MPL_PREVP = NULL;
 127                  dqp->q_bufoffset = 0;
 128                  dqp->q_fileoffset = 0;
 129                  dqp->q_transp = NULL;
 130                  dqp->q_gdquot = NULL;
 131                  dqp->q_res_bcount = 0;
 132                  dqp->q_res_icount = 0;
 133                  dqp->q_res_rtbcount = 0;
 134                  dqp->q_pincount = 0;
 135                  dqp->q_hash = NULL;
 136                  ASSERT(dqp->dq_flnext == dqp->dq_flprev);
 137
 138 #ifdef XFS_DQUOT_TRACE
 139                  ASSERT(dqp->q_trace);
 140                  xfs_dqtrace_entry(dqp, "DQRECLAIMED_INIT");
 141 #endif
 142          }
 143
 144         /*
 145          * log item gets initialized later
 146          */
 147         return (dqp);
 148 }
 149
 150 /*
 151  * This is called to free all the memory associated with a dquot
 152  */
 153 void
 154 xfs_qm_dqdestroy(
 155         xfs_dquot_t     *dqp)
 156 {
 157         ASSERT(! XFS_DQ_IS_ON_FREELIST(dqp));
 158
 159         mutex_destroy(&dqp->q_qlock);
 160         sv_destroy(&dqp->q_pinwait);
 161
 162 #ifdef XFS_DQUOT_TRACE
 163         if (dqp->q_trace)
 164              ktrace_free(dqp->q_trace);
 165         dqp->q_trace = NULL;
 166 #endif
 167         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 168         atomic_dec(&xfs_Gqm->qm_totaldquots);
 169 }
 170
 171 /*
 172  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 173  */
 174 STATIC void
 175 xfs_qm_dqinit_core(
 176         xfs_dqid_t      id,
 177         uint            type,
 178         xfs_dqblk_t     *d)
 179 {
 180         /*
 181          * Caller has zero'd the entire dquot 'chunk' already.
 182          */
 183         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 184         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 185         d->dd_diskdq.d_id = cpu_to_be32(id);
 186         d->dd_diskdq.d_flags = type;
 187 }
 188
 189
 190 #ifdef XFS_DQUOT_TRACE
 191 /*
 192  * Dquot tracing for debugging.
 193  */
 194 /* ARGSUSED */
 195 void
 196 __xfs_dqtrace_entry(
 197         xfs_dquot_t     *dqp,
 198         char            *func,
 199         void            *retaddr,
 200         xfs_inode_t     *ip)
 201 {
 202         xfs_dquot_t     *udqp = NULL;
 203         xfs_ino_t       ino = 0;
 204
 205         ASSERT(dqp->q_trace);
 206         if (ip) {
 207                 ino = ip->i_ino;
 208                 udqp = ip->i_udquot;
 209         }
 210         ktrace_enter(dqp->q_trace,
 211                      (void *)(__psint_t)DQUOT_KTRACE_ENTRY,
 212                      (void *)func,
 213                      (void *)(__psint_t)dqp->q_nrefs,
 214                      (void *)(__psint_t)dqp->dq_flags,
 215                      (void *)(__psint_t)dqp->q_res_bcount,
 216                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_bcount),
 217                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_icount),
 218                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_hardlimit),
 219                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_softlimit),
 220                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_hardlimit),
 221                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_softlimit),
 222                      (void *)(__psint_t)be32_to_cpu(dqp->q_core.d_id),
 223                      (void *)(__psint_t)current_pid(),
 224                      (void *)(__psint_t)ino,
 225                      (void *)(__psint_t)retaddr,
 226                      (void *)(__psint_t)udqp);
 227         return;
 228 }
 229 #endif
 230
 231
 232 /*
 233  * If default limits are in force, push them into the dquot now.
 234  * We overwrite the dquot limits only if they are zero and this
 235  * is not the root dquot.
 236  */
 237 void
 238 xfs_qm_adjust_dqlimits(
 239         xfs_mount_t             *mp,
 240         xfs_disk_dquot_t        *d)
 241 {
 242         xfs_quotainfo_t         *q = mp->m_quotainfo;
 243
 244         ASSERT(d->d_id);
 245
 246         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 247                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 248         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 249                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 250         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 251                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 252         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 253                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 254         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 255                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 256         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 257                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 258 }
 259
 260 /*
 261  * Check the limits and timers of a dquot and start or reset timers
 262  * if necessary.
 263  * This gets called even when quota enforcement is OFF, which makes our
 264  * life a little less complicated. (We just don't reject any quota
 265  * reservations in that case, when enforcement is off).
 266  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 267  * enforcement's off.
 268  * In contrast, warnings are a little different in that they don't
 269  * 'automatically' get started when limits get exceeded.  They do
 270  * get reset to zero, however, when we find the count to be under
 271  * the soft limit (they are only ever set non-zero via userspace).
 272  */
 273 void
 274 xfs_qm_adjust_dqtimers(
 275         xfs_mount_t             *mp,
 276         xfs_disk_dquot_t        *d)
 277 {
 278         ASSERT(d->d_id);
 279
 280 #ifdef QUOTADEBUG
 281         if (d->d_blk_hardlimit)
 282                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 283                        be64_to_cpu(d->d_blk_hardlimit));
 284         if (d->d_ino_hardlimit)
 285                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 286                        be64_to_cpu(d->d_ino_hardlimit));
 287         if (d->d_rtb_hardlimit)
 288                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 289                        be64_to_cpu(d->d_rtb_hardlimit));
 290 #endif
 291         if (!d->d_btimer) {
 292                 if ((d->d_blk_softlimit &&
 293                      (be64_to_cpu(d->d_bcount) >=
 294                       be64_to_cpu(d->d_blk_softlimit))) ||
 295                     (d->d_blk_hardlimit &&
 296                      (be64_to_cpu(d->d_bcount) >=
 297                       be64_to_cpu(d->d_blk_hardlimit)))) {
 298                         d->d_btimer = cpu_to_be32(get_seconds() +
 299                                         XFS_QI_BTIMELIMIT(mp));
 300                 } else {
 301                         d->d_bwarns = 0;
 302                 }
 303         } else {
 304                 if ((!d->d_blk_softlimit ||
 305                      (be64_to_cpu(d->d_bcount) <
 306                       be64_to_cpu(d->d_blk_softlimit))) &&
 307                     (!d->d_blk_hardlimit ||
 308                     (be64_to_cpu(d->d_bcount) <
 309                      be64_to_cpu(d->d_blk_hardlimit)))) {
 310                         d->d_btimer = 0;
 311                 }
 312         }
 313
 314         if (!d->d_itimer) {
 315                 if ((d->d_ino_softlimit &&
 316                      (be64_to_cpu(d->d_icount) >=
 317                       be64_to_cpu(d->d_ino_softlimit))) ||
 318                     (d->d_ino_hardlimit &&
 319                      (be64_to_cpu(d->d_icount) >=
 320                       be64_to_cpu(d->d_ino_hardlimit)))) {
 321                         d->d_itimer = cpu_to_be32(get_seconds() +
 322                                         XFS_QI_ITIMELIMIT(mp));
 323                 } else {
 324                         d->d_iwarns = 0;
 325                 }
 326         } else {
 327                 if ((!d->d_ino_softlimit ||
 328                      (be64_to_cpu(d->d_icount) <
 329                       be64_to_cpu(d->d_ino_softlimit)))  &&
 330                     (!d->d_ino_hardlimit ||
 331                      (be64_to_cpu(d->d_icount) <
 332                       be64_to_cpu(d->d_ino_hardlimit)))) {
 333                         d->d_itimer = 0;
 334                 }
 335         }
 336
 337         if (!d->d_rtbtimer) {
 338                 if ((d->d_rtb_softlimit &&
 339                      (be64_to_cpu(d->d_rtbcount) >=
 340                       be64_to_cpu(d->d_rtb_softlimit))) ||
 341                     (d->d_rtb_hardlimit &&
 342                      (be64_to_cpu(d->d_rtbcount) >=
 343                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 344                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 345                                         XFS_QI_RTBTIMELIMIT(mp));
 346                 } else {
 347                         d->d_rtbwarns = 0;
 348                 }
 349         } else {
 350                 if ((!d->d_rtb_softlimit ||
 351                      (be64_to_cpu(d->d_rtbcount) <
 352                       be64_to_cpu(d->d_rtb_softlimit))) &&
 353                     (!d->d_rtb_hardlimit ||
 354                      (be64_to_cpu(d->d_rtbcount) <
 355                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 356                         d->d_rtbtimer = 0;
 357                 }
 358         }
 359 }
 360
 361 /*
 362  * initialize a buffer full of dquots and log the whole thing
 363  */
 364 STATIC void
 365 xfs_qm_init_dquot_blk(
 366         xfs_trans_t     *tp,
 367         xfs_mount_t     *mp,
 368         xfs_dqid_t      id,
 369         uint            type,
 370         xfs_buf_t       *bp)
 371 {
 372         xfs_dqblk_t     *d;
 373         int             curid, i;
 374
 375         ASSERT(tp);
 376         ASSERT(XFS_BUF_ISBUSY(bp));
 377         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 378
 379         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 380
 381         /*
 382          * ID of the first dquot in the block - id's are zero based.
 383          */
 384         curid = id - (id % XFS_QM_DQPERBLK(mp));
 385         ASSERT(curid >= 0);
 386         memset(d, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)));
 387         for (i = 0; i < XFS_QM_DQPERBLK(mp); i++, d++, curid++)
 388                 xfs_qm_dqinit_core(curid, type, d);
 389         xfs_trans_dquot_buf(tp, bp,
 390                             (type & XFS_DQ_USER ? XFS_BLI_UDQUOT_BUF :
 391                             ((type & XFS_DQ_PROJ) ? XFS_BLI_PDQUOT_BUF :
 392                              XFS_BLI_GDQUOT_BUF)));
 393         xfs_trans_log_buf(tp, bp, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1);
 394 }
 395
 396
 397
 398 /*
 399  * Allocate a block and fill it with dquots.
 400  * This is called when the bmapi finds a hole.
 401  */
 402 STATIC int
 403 xfs_qm_dqalloc(
 404         xfs_trans_t     **tpp,
 405         xfs_mount_t     *mp,
 406         xfs_dquot_t     *dqp,
 407         xfs_inode_t     *quotip,
 408         xfs_fileoff_t   offset_fsb,
 409         xfs_buf_t       **O_bpp)
 410 {
 411         xfs_fsblock_t   firstblock;
 412         xfs_bmap_free_t flist;
 413         xfs_bmbt_irec_t map;
 414         int             nmaps, error, committed;
 415         xfs_buf_t       *bp;
 416         xfs_trans_t     *tp = *tpp;
 417
 418         ASSERT(tp != NULL);
 419         xfs_dqtrace_entry(dqp, "DQALLOC");
 420
 421         /*
 422          * Initialize the bmap freelist prior to calling bmapi code.
 423          */
 424         XFS_BMAP_INIT(&flist, &firstblock);
 425         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 426         /*
 427          * Return if this type of quotas is turned off while we didn't
 428          * have an inode lock
 429          */
 430         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 431                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 432                 return (ESRCH);
 433         }
 434
 435         /*
 436          * xfs_trans_commit normally decrements the vnode ref count
 437          * when it unlocks the inode. Since we want to keep the quota
 438          * inode around, we bump the vnode ref count now.
 439          */
 440         IHOLD(quotip);
 441
 442         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 443         nmaps = 1;
 444         if ((error = xfs_bmapi(tp, quotip,
 445                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 446                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 447                               &firstblock,
 448                               XFS_QM_DQALLOC_SPACE_RES(mp),
 449                               &map, &nmaps, &flist, NULL))) {
 450                 goto error0;
 451         }
 452         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 453         ASSERT(nmaps == 1);
 454         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 455                (map.br_startblock != HOLESTARTBLOCK));
 456
 457         /*
 458          * Keep track of the blkno to save a lookup later
 459          */
 460         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 461
 462         /* now we can just get the buffer (there's nothing to read yet) */
 463         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 464                                dqp->q_blkno,
 465                                XFS_QI_DQCHUNKLEN(mp),
 466                                0);
 467         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 468                 goto error1;
 469         /*
 470          * Make a chunk of dquots out of this buffer and log
 471          * the entire thing.
 472          */
 473         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 474                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 475
 476         /*
 477          * xfs_bmap_finish() may commit the current transaction and
 478          * start a second transaction if the freelist is not empty.
 479          *
 480          * Since we still want to modify this buffer, we need to
 481          * ensure that the buffer is not released on commit of
 482          * the first transaction and ensure the buffer is added to the
 483          * second transaction.
 484          *
 485          * If there is only one transaction then don't stop the buffer
 486          * from being released when it commits later on.
 487          */
 488
 489         xfs_trans_bhold(tp, bp);
 490
 491         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 492                 goto error1;
 493         }
 494
 495         if (committed) {
 496                 tp = *tpp;
 497                 xfs_trans_bjoin(tp, bp);
 498         } else {
 499                 xfs_trans_bhold_release(tp, bp);
 500         }
 501
 502         *O_bpp = bp;
 503         return 0;
 504
 505       error1:
 506         xfs_bmap_cancel(&flist);
 507       error0:
 508         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 509
 510         return (error);
 511 }
 512
 513 /*
 514  * Maps a dquot to the buffer containing its on-disk version.
 515  * This returns a ptr to the buffer containing the on-disk dquot
 516  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 517  */
 518 STATIC int
 519 xfs_qm_dqtobp(
 520         xfs_trans_t             **tpp,
 521         xfs_dquot_t             *dqp,
 522         xfs_disk_dquot_t        **O_ddpp,
 523         xfs_buf_t               **O_bpp,
 524         uint                    flags)
 525 {
 526         xfs_bmbt_irec_t map;
 527         int             nmaps, error;
 528         xfs_buf_t       *bp;
 529         xfs_inode_t     *quotip;
 530         xfs_mount_t     *mp;
 531         xfs_disk_dquot_t *ddq;
 532         xfs_dqid_t      id;
 533         boolean_t       newdquot;
 534         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 535
 536         mp = dqp->q_mount;
 537         id = be32_to_cpu(dqp->q_core.d_id);
 538         nmaps = 1;
 539         newdquot = B_FALSE;
 540
 541         /*
 542          * If we don't know where the dquot lives, find out.
 543          */
 544         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 545                 /* We use the id as an index */
 546                 dqp->q_fileoffset = (xfs_fileoff_t)id / XFS_QM_DQPERBLK(mp);
 547                 nmaps = 1;
 548                 quotip = XFS_DQ_TO_QIP(dqp);
 549                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 550                 /*
 551                  * Return if this type of quotas is turned off while we didn't
 552                  * have an inode lock
 553                  */
 554                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 555                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 556                         return (ESRCH);
 557                 }
 558                 /*
 559                  * Find the block map; no allocations yet
 560                  */
 561                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 562                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 563                                   XFS_BMAPI_METADATA,
 564                                   NULL, 0, &map, &nmaps, NULL, NULL);
 565
 566                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 567                 if (error)
 568                         return (error);
 569                 ASSERT(nmaps == 1);
 570                 ASSERT(map.br_blockcount == 1);
 571
 572                 /*
 573                  * offset of dquot in the (fixed sized) dquot chunk.
 574                  */
 575                 dqp->q_bufoffset = (id % XFS_QM_DQPERBLK(mp)) *
 576                         sizeof(xfs_dqblk_t);
 577                 if (map.br_startblock == HOLESTARTBLOCK) {
 578                         /*
 579                          * We don't allocate unless we're asked to
 580                          */
 581                         if (!(flags & XFS_QMOPT_DQALLOC))
 582                                 return (ENOENT);
 583
 584                         ASSERT(tp);
 585                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 586                                                 dqp->q_fileoffset, &bp)))
 587                                 return (error);
 588                         tp = *tpp;
 589                         newdquot = B_TRUE;
 590                 } else {
 591                         /*
 592                          * store the blkno etc so that we don't have to do the
 593                          * mapping all the time
 594                          */
 595                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 596                 }
 597         }
 598         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 599         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 600
 601         /*
 602          * Read in the buffer, unless we've just done the allocation
 603          * (in which case we already have the buf).
 604          */
 605         if (! newdquot) {
 606                 xfs_dqtrace_entry(dqp, "DQTOBP READBUF");
 607                 if ((error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 608                                                dqp->q_blkno,
 609                                                XFS_QI_DQCHUNKLEN(mp),
 610                                                0, &bp))) {
 611                         return (error);
 612                 }
 613                 if (error || !bp)
 614                         return XFS_ERROR(error);
 615         }
 616         ASSERT(XFS_BUF_ISBUSY(bp));
 617         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 618
 619         /*
 620          * calculate the location of the dquot inside the buffer.
 621          */
 622         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 623
 624         /*
 625          * A simple sanity check in case we got a corrupted dquot...
 626          */
 627         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 628                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 629                            "dqtobp")) {
 630                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 631                         xfs_trans_brelse(tp, bp);
 632                         return XFS_ERROR(EIO);
 633                 }
 634                 XFS_BUF_BUSY(bp); /* We dirtied this */
 635         }
 636
 637         *O_bpp = bp;
 638         *O_ddpp = ddq;
 639
 640         return (0);
 641 }
 642
 643
 644 /*
 645  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 646  * and release the buffer immediately.
 647  *
 648  */
 649 /* ARGSUSED */
 650 STATIC int
 651 xfs_qm_dqread(
 652         xfs_trans_t     **tpp,
 653         xfs_dqid_t      id,
 654         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 655         uint            flags)
 656 {
 657         xfs_disk_dquot_t *ddqp;
 658         xfs_buf_t        *bp;
 659         int              error;
 660         xfs_trans_t      *tp;
 661
 662         ASSERT(tpp);
 663
 664         /*
 665          * get a pointer to the on-disk dquot and the buffer containing it
 666          * dqp already knows its own type (GROUP/USER).
 667          */
 668         xfs_dqtrace_entry(dqp, "DQREAD");
 669         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 670                 return (error);
 671         }
 672         tp = *tpp;
 673
 674         /* copy everything from disk dquot to the incore dquot */
 675         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 676         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 677         xfs_qm_dquot_logitem_init(dqp);
 678
 679         /*
 680          * Reservation counters are defined as reservation plus current usage
 681          * to avoid having to add everytime.
 682          */
 683         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 684         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 685         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 686
 687         /* Mark the buf so that this will stay incore a little longer */
 688         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 689
 690         /*
 691          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 692          * So we need to release with xfs_trans_brelse().
 693          * The strategy here is identical to that of inodes; we lock
 694          * the dquot in xfs_qm_dqget() before making it accessible to
 695          * others. This is because dquots, like inodes, need a good level of
 696          * concurrency, and we don't want to take locks on the entire buffers
 697          * for dquot accesses.
 698          * Note also that the dquot buffer may even be dirty at this point, if
 699          * this particular dquot was repaired. We still aren't afraid to
 700          * brelse it because we have the changes incore.
 701          */
 702         ASSERT(XFS_BUF_ISBUSY(bp));
 703         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 704         xfs_trans_brelse(tp, bp);
 705
 706         return (error);
 707 }
 708
 709
 710 /*
 711  * allocate an incore dquot from the kernel heap,
 712  * and fill its core with quota information kept on disk.
 713  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 714  * if it wasn't already allocated.
 715  */
 716 STATIC int
 717 xfs_qm_idtodq(
 718         xfs_mount_t     *mp,
 719         xfs_dqid_t      id,      /* gid or uid, depending on type */
 720         uint            type,    /* UDQUOT or GDQUOT */
 721         uint            flags,   /* DQALLOC, DQREPAIR */
 722         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 723 {
 724         xfs_dquot_t     *dqp;
 725         int             error;
 726         xfs_trans_t     *tp;
 727         int             cancelflags=0;
 728
 729         dqp = xfs_qm_dqinit(mp, id, type);
 730         tp = NULL;
 731         if (flags & XFS_QMOPT_DQALLOC) {
 732                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 733                 if ((error = xfs_trans_reserve(tp,
 734                                        XFS_QM_DQALLOC_SPACE_RES(mp),
 735                                        XFS_WRITE_LOG_RES(mp) +
 736                                               BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1 +
 737                                               128,
 738                                        0,
 739                                        XFS_TRANS_PERM_LOG_RES,
 740                                        XFS_WRITE_LOG_COUNT))) {
 741                         cancelflags = 0;
 742                         goto error0;
 743                 }
 744                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 745         }
 746
 747         /*
 748          * Read it from disk; xfs_dqread() takes care of
 749          * all the necessary initialization of dquot's fields (locks, etc)
 750          */
 751         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 752                 /*
 753                  * This can happen if quotas got turned off (ESRCH),
 754                  * or if the dquot didn't exist on disk and we ask to
 755                  * allocate (ENOENT).
 756                  */
 757                 xfs_dqtrace_entry(dqp, "DQREAD FAIL");
 758                 cancelflags |= XFS_TRANS_ABORT;
 759                 goto error0;
 760         }
 761         if (tp) {
 762                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 763                         goto error1;
 764         }
 765
 766         *O_dqpp = dqp;
 767         return (0);
 768
 769  error0:
 770         ASSERT(error);
 771         if (tp)
 772                 xfs_trans_cancel(tp, cancelflags);
 773  error1:
 774         xfs_qm_dqdestroy(dqp);
 775         *O_dqpp = NULL;
 776         return (error);
 777 }
 778
 779 /*
 780  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 781  * hashtables for user and group dquots; and, these are global tables
 782  * inside the XQM, not per-filesystem tables.
 783  * The hash chain must be locked by caller, and it is left locked
 784  * on return. Returning dquot is locked.
 785  */
 786 STATIC int
 787 xfs_qm_dqlookup(
 788         xfs_mount_t             *mp,
 789         xfs_dqid_t              id,
 790         xfs_dqhash_t            *qh,
 791         xfs_dquot_t             **O_dqpp)
 792 {
 793         xfs_dquot_t             *dqp;
 794         uint                    flist_locked;
 795         xfs_dquot_t             *d;
 796
 797         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 798
 799         flist_locked = B_FALSE;
 800
 801         /*
 802          * Traverse the hashchain looking for a match
 803          */
 804         for (dqp = qh->qh_next; dqp != NULL; dqp = dqp->HL_NEXT) {
 805                 /*
 806                  * We already have the hashlock. We don't need the
 807                  * dqlock to look at the id field of the dquot, since the
 808                  * id can't be modified without the hashlock anyway.
 809                  */
 810                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 811                         xfs_dqtrace_entry(dqp, "DQFOUND BY LOOKUP");
 812                         /*
 813                          * All in core dquots must be on the dqlist of mp
 814                          */
 815                         ASSERT(dqp->MPL_PREVP != NULL);
 816
 817                         xfs_dqlock(dqp);
 818                         if (dqp->q_nrefs == 0) {
 819                                 ASSERT (XFS_DQ_IS_ON_FREELIST(dqp));
 820                                 if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
 821                                         xfs_dqtrace_entry(dqp, "DQLOOKUP: WANT");
 822
 823                                         /*
 824                                          * We may have raced with dqreclaim_one()
 825                                          * (and lost). So, flag that we don't
 826                                          * want the dquot to be reclaimed.
 827                                          */
 828                                         dqp->dq_flags |= XFS_DQ_WANT;
 829                                         xfs_dqunlock(dqp);
 830                                         xfs_qm_freelist_lock(xfs_Gqm);
 831                                         xfs_dqlock(dqp);
 832                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 833                                 }
 834                                 flist_locked = B_TRUE;
 835                         }
 836
 837                         /*
 838                          * id couldn't have changed; we had the hashlock all
 839                          * along
 840                          */
 841                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 842
 843                         if (flist_locked) {
 844                                 if (dqp->q_nrefs != 0) {
 845                                         xfs_qm_freelist_unlock(xfs_Gqm);
 846                                         flist_locked = B_FALSE;
 847                                 } else {
 848                                         /*
 849                                          * take it off the freelist
 850                                          */
 851                                         xfs_dqtrace_entry(dqp,
 852                                                         "DQLOOKUP: TAKEOFF FL");
 853                                         XQM_FREELIST_REMOVE(dqp);
 854                                         /* xfs_qm_freelist_print(&(xfs_Gqm->
 855                                                         qm_dqfreelist),
 856                                                         "after removal"); */
 857                                 }
 858                         }
 859
 860                         /*
 861                          * grab a reference
 862                          */
 863                         XFS_DQHOLD(dqp);
 864
 865                         if (flist_locked)
 866                                 xfs_qm_freelist_unlock(xfs_Gqm);
 867                         /*
 868                          * move the dquot to the front of the hashchain
 869                          */
 870                         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 871                         if (dqp->HL_PREVP != &qh->qh_next) {
 872                                 xfs_dqtrace_entry(dqp,
 873                                                   "DQLOOKUP: HASH MOVETOFRONT");
 874                                 if ((d = dqp->HL_NEXT))
 875                                         d->HL_PREVP = dqp->HL_PREVP;
 876                                 *(dqp->HL_PREVP) = d;
 877                                 d = qh->qh_next;
 878                                 d->HL_PREVP = &dqp->HL_NEXT;
 879                                 dqp->HL_NEXT = d;
 880                                 dqp->HL_PREVP = &qh->qh_next;
 881                                 qh->qh_next = dqp;
 882                         }
 883                         xfs_dqtrace_entry(dqp, "LOOKUP END");
 884                         *O_dqpp = dqp;
 885                         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 886                         return (0);
 887                 }
 888         }
 889
 890         *O_dqpp = NULL;
 891         ASSERT(XFS_DQ_IS_HASH_LOCKED(qh));
 892         return (1);
 893 }
 894
 895 /*
 896  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 897  * a locked dquot, doing an allocation (if requested) as needed.
 898  * When both an inode and an id are given, the inode's id takes precedence.
 899  * That is, if the id changes while we don't hold the ilock inside this
 900  * function, the new dquot is returned, not necessarily the one requested
 901  * in the id argument.
 902  */
 903 int
 904 xfs_qm_dqget(
 905         xfs_mount_t     *mp,
 906         xfs_inode_t     *ip,      /* locked inode (optional) */
 907         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 908         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 909         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 910         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 911 {
 912         xfs_dquot_t     *dqp;
 913         xfs_dqhash_t    *h;
 914         uint            version;
 915         int             error;
 916
 917         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 918         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 919             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 920             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 921                 return (ESRCH);
 922         }
 923         h = XFS_DQ_HASH(mp, id, type);
 924
 925 #ifdef DEBUG
 926         if (xfs_do_dqerror) {
 927                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 928                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 929                         cmn_err(CE_DEBUG, "Returning error in dqget");
 930                         return (EIO);
 931                 }
 932         }
 933 #endif
 934
 935  again:
 936
 937 #ifdef DEBUG
 938         ASSERT(type == XFS_DQ_USER ||
 939                type == XFS_DQ_PROJ ||
 940                type == XFS_DQ_GROUP);
 941         if (ip) {
 942                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 943                 if (type == XFS_DQ_USER)
 944                         ASSERT(ip->i_udquot == NULL);
 945                 else
 946                         ASSERT(ip->i_gdquot == NULL);
 947         }
 948 #endif
 949         XFS_DQ_HASH_LOCK(h);
 950
 951         /*
 952          * Look in the cache (hashtable).
 953          * The chain is kept locked during lookup.
 954          */
 955         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 956                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 957                 /*
 958                  * The dquot was found, moved to the front of the chain,
 959                  * taken off the freelist if it was on it, and locked
 960                  * at this point. Just unlock the hashchain and return.
 961                  */
 962                 ASSERT(*O_dqpp);
 963                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 964                 XFS_DQ_HASH_UNLOCK(h);
 965                 xfs_dqtrace_entry(*O_dqpp, "DQGET DONE (FROM CACHE)");
 966                 return (0);     /* success */
 967         }
 968         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 969
 970         /*
 971          * Dquot cache miss. We don't want to keep the inode lock across
 972          * a (potential) disk read. Also we don't want to deal with the lock
 973          * ordering between quotainode and this inode. OTOH, dropping the inode
 974          * lock here means dealing with a chown that can happen before
 975          * we re-acquire the lock.
 976          */
 977         if (ip)
 978                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 979         /*
 980          * Save the hashchain version stamp, and unlock the chain, so that
 981          * we don't keep the lock across a disk read
 982          */
 983         version = h->qh_version;
 984         XFS_DQ_HASH_UNLOCK(h);
 985
 986         /*
 987          * Allocate the dquot on the kernel heap, and read the ondisk
 988          * portion off the disk. Also, do all the necessary initialization
 989          * This can return ENOENT if dquot didn't exist on disk and we didn't
 990          * ask it to allocate; ESRCH if quotas got turned off suddenly.
 991          */
 992         if ((error = xfs_qm_idtodq(mp, id, type,
 993                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
 994                                            XFS_QMOPT_DOWARN),
 995                                   &dqp))) {
 996                 if (ip)
 997                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 998                 return (error);
 999         }
1000
1001         /*
1002          * See if this is mount code calling to look at the overall quota limits
1003          * which are stored in the id == 0 user or group's dquot.
1004          * Since we may not have done a quotacheck by this point, just return
1005          * the dquot without attaching it to any hashtables, lists, etc, or even
1006          * taking a reference.
1007          * The caller must dqdestroy this once done.
1008          */
1009         if (flags & XFS_QMOPT_DQSUSER) {
1010                 ASSERT(id == 0);
1011                 ASSERT(! ip);
1012                 goto dqret;
1013         }
1014
1015         /*
1016          * Dquot lock comes after hashlock in the lock ordering
1017          */
1018         if (ip) {
1019                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1020                 if (! XFS_IS_DQTYPE_ON(mp, type)) {
1021                         /* inode stays locked on return */
1022                         xfs_qm_dqdestroy(dqp);
1023                         return XFS_ERROR(ESRCH);
1024                 }
1025                 /*
1026                  * A dquot could be attached to this inode by now, since
1027                  * we had dropped the ilock.
1028                  */
1029                 if (type == XFS_DQ_USER) {
1030                         if (ip->i_udquot) {
1031                                 xfs_qm_dqdestroy(dqp);
1032                                 dqp = ip->i_udquot;
1033                                 xfs_dqlock(dqp);
1034                                 goto dqret;
1035                         }
1036                 } else {
1037                         if (ip->i_gdquot) {
1038                                 xfs_qm_dqdestroy(dqp);
1039                                 dqp = ip->i_gdquot;
1040                                 xfs_dqlock(dqp);
1041                                 goto dqret;
1042                         }
1043                 }
1044         }
1045
1046         /*
1047          * Hashlock comes after ilock in lock order
1048          */
1049         XFS_DQ_HASH_LOCK(h);
1050         if (version != h->qh_version) {
1051                 xfs_dquot_t *tmpdqp;
1052                 /*
1053                  * Now, see if somebody else put the dquot in the
1054                  * hashtable before us. This can happen because we didn't
1055                  * keep the hashchain lock. We don't have to worry about
1056                  * lock order between the two dquots here since dqp isn't
1057                  * on any findable lists yet.
1058                  */
1059                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
1060                         /*
1061                          * Duplicate found. Just throw away the new dquot
1062                          * and start over.
1063                          */
1064                         xfs_qm_dqput(tmpdqp);
1065                         XFS_DQ_HASH_UNLOCK(h);
1066                         xfs_qm_dqdestroy(dqp);
1067                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
1068                         goto again;
1069                 }
1070         }
1071
1072         /*
1073          * Put the dquot at the beginning of the hash-chain and mp's list
1074          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1075          */
1076         ASSERT(XFS_DQ_IS_HASH_LOCKED(h));
1077         dqp->q_hash = h;
1078         XQM_HASHLIST_INSERT(h, dqp);
1079
1080         /*
1081          * Attach this dquot to this filesystem's list of all dquots,
1082          * kept inside the mount structure in m_quotainfo field
1083          */
1084         xfs_qm_mplist_lock(mp);
1085
1086         /*
1087          * We return a locked dquot to the caller, with a reference taken
1088          */
1089         xfs_dqlock(dqp);
1090         dqp->q_nrefs = 1;
1091
1092         XQM_MPLIST_INSERT(&(XFS_QI_MPL_LIST(mp)), dqp);
1093
1094         xfs_qm_mplist_unlock(mp);
1095         XFS_DQ_HASH_UNLOCK(h);
1096  dqret:
1097         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1098         xfs_dqtrace_entry(dqp, "DQGET DONE");
1099         *O_dqpp = dqp;
1100         return (0);
1101 }
1102
1103
1104 /*
1105  * Release a reference to the dquot (decrement ref-count)
1106  * and unlock it. If there is a group quota attached to this
1107  * dquot, carefully release that too without tripping over
1108  * deadlocks'n'stuff.
1109  */
1110 void
1111 xfs_qm_dqput(
1112         xfs_dquot_t     *dqp)
1113 {
1114         xfs_dquot_t     *gdqp;
1115
1116         ASSERT(dqp->q_nrefs > 0);
1117         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1118         xfs_dqtrace_entry(dqp, "DQPUT");
1119
1120         if (dqp->q_nrefs != 1) {
1121                 dqp->q_nrefs--;
1122                 xfs_dqunlock(dqp);
1123                 return;
1124         }
1125
1126         /*
1127          * drop the dqlock and acquire the freelist and dqlock
1128          * in the right order; but try to get it out-of-order first
1129          */
1130         if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
1131                 xfs_dqtrace_entry(dqp, "DQPUT: FLLOCK-WAIT");
1132                 xfs_dqunlock(dqp);
1133                 xfs_qm_freelist_lock(xfs_Gqm);
1134                 xfs_dqlock(dqp);
1135         }
1136
1137         while (1) {
1138                 gdqp = NULL;
1139
1140                 /* We can't depend on nrefs being == 1 here */
1141                 if (--dqp->q_nrefs == 0) {
1142                         xfs_dqtrace_entry(dqp, "DQPUT: ON FREELIST");
1143                         /*
1144                          * insert at end of the freelist.
1145                          */
1146                         XQM_FREELIST_INSERT(&(xfs_Gqm->qm_dqfreelist), dqp);
1147
1148                         /*
1149                          * If we just added a udquot to the freelist, then
1150                          * we want to release the gdquot reference that
1151                          * it (probably) has. Otherwise it'll keep the
1152                          * gdquot from getting reclaimed.
1153                          */
1154                         if ((gdqp = dqp->q_gdquot)) {
1155                                 /*
1156                                  * Avoid a recursive dqput call
1157                                  */
1158                                 xfs_dqlock(gdqp);
1159                                 dqp->q_gdquot = NULL;
1160                         }
1161
1162                         /* xfs_qm_freelist_print(&(xfs_Gqm->qm_dqfreelist),
1163                            "@@@@@++ Free list (after append) @@@@@+");
1164                            */
1165                 }
1166                 xfs_dqunlock(dqp);
1167
1168                 /*
1169                  * If we had a group quota inside the user quota as a hint,
1170                  * release it now.
1171                  */
1172                 if (! gdqp)
1173                         break;
1174                 dqp = gdqp;
1175         }
1176         xfs_qm_freelist_unlock(xfs_Gqm);
1177 }
1178
1179 /*
1180  * Release a dquot. Flush it if dirty, then dqput() it.
1181  * dquot must not be locked.
1182  */
1183 void
1184 xfs_qm_dqrele(
1185         xfs_dquot_t     *dqp)
1186 {
1187         ASSERT(dqp);
1188         xfs_dqtrace_entry(dqp, "DQRELE");
1189
1190         xfs_dqlock(dqp);
1191         /*
1192          * We don't care to flush it if the dquot is dirty here.
1193          * That will create stutters that we want to avoid.
1194          * Instead we do a delayed write when we try to reclaim
1195          * a dirty dquot. Also xfs_sync will take part of the burden...
1196          */
1197         xfs_qm_dqput(dqp);
1198 }
1199
1200
1201 /*
1202  * Write a modified dquot to disk.
1203  * The dquot must be locked and the flush lock too taken by caller.
1204  * The flush lock will not be unlocked until the dquot reaches the disk,
1205  * but the dquot is free to be unlocked and modified by the caller
1206  * in the interim. Dquot is still locked on return. This behavior is
1207  * identical to that of inodes.
1208  */
1209 int
1210 xfs_qm_dqflush(
1211         xfs_dquot_t             *dqp,
1212         uint                    flags)
1213 {
1214         xfs_mount_t             *mp;
1215         xfs_buf_t               *bp;
1216         xfs_disk_dquot_t        *ddqp;
1217         int                     error;
1218
1219         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1220         ASSERT(!completion_done(&dqp->q_flush));
1221         xfs_dqtrace_entry(dqp, "DQFLUSH");
1222
1223         /*
1224          * If not dirty, nada.
1225          */
1226         if (!XFS_DQ_IS_DIRTY(dqp)) {
1227                 xfs_dqfunlock(dqp);
1228                 return (0);
1229         }
1230
1231         /*
1232          * Cant flush a pinned dquot. Wait for it.
1233          */
1234         xfs_qm_dqunpin_wait(dqp);
1235
1236         /*
1237          * This may have been unpinned because the filesystem is shutting
1238          * down forcibly. If that's the case we must not write this dquot
1239          * to disk, because the log record didn't make it to disk!
1240          */
1241         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1242                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1243                 xfs_dqfunlock(dqp);
1244                 return XFS_ERROR(EIO);
1245         }
1246
1247         /*
1248          * Get the buffer containing the on-disk dquot
1249          * We don't need a transaction envelope because we know that the
1250          * the ondisk-dquot has already been allocated for.
1251          */
1252         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1253                 xfs_dqtrace_entry(dqp, "DQTOBP FAIL");
1254                 ASSERT(error != ENOENT);
1255                 /*
1256                  * Quotas could have gotten turned off (ESRCH)
1257                  */
1258                 xfs_dqfunlock(dqp);
1259                 return (error);
1260         }
1261
1262         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1263                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1264                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1265                 return XFS_ERROR(EIO);
1266         }
1267
1268         /* This is the only portion of data that needs to persist */
1269         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1270
1271         /*
1272          * Clear the dirty field and remember the flush lsn for later use.
1273          */
1274         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1275         mp = dqp->q_mount;
1276
1277         /* lsn is 64 bits */
1278         spin_lock(&mp->m_ail_lock);
1279         dqp->q_logitem.qli_flush_lsn = dqp->q_logitem.qli_item.li_lsn;
1280         spin_unlock(&mp->m_ail_lock);
1281
1282         /*
1283          * Attach an iodone routine so that we can remove this dquot from the
1284          * AIL and release the flush lock once the dquot is synced to disk.
1285          */
1286         xfs_buf_attach_iodone(bp, (void(*)(xfs_buf_t *, xfs_log_item_t *))
1287                               xfs_qm_dqflush_done, &(dqp->q_logitem.qli_item));
1288         /*
1289          * If the buffer is pinned then push on the log so we won't
1290          * get stuck waiting in the write for too long.
1291          */
1292         if (XFS_BUF_ISPINNED(bp)) {
1293                 xfs_dqtrace_entry(dqp, "DQFLUSH LOG FORCE");
1294                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
1295         }
1296
1297         if (flags & XFS_QMOPT_DELWRI) {
1298                 xfs_bdwrite(mp, bp);
1299         } else if (flags & XFS_QMOPT_ASYNC) {
1300                 error = xfs_bawrite(mp, bp);
1301         } else {
1302                 error = xfs_bwrite(mp, bp);
1303         }
1304         xfs_dqtrace_entry(dqp, "DQFLUSH END");
1305         /*
1306          * dqp is still locked, but caller is free to unlock it now.
1307          */
1308         return (error);
1309
1310 }
1311
1312 /*
1313  * This is the dquot flushing I/O completion routine.  It is called
1314  * from interrupt level when the buffer containing the dquot is
1315  * flushed to disk.  It is responsible for removing the dquot logitem
1316  * from the AIL if it has not been re-logged, and unlocking the dquot's
1317  * flush lock. This behavior is very similar to that of inodes..
1318  */
1319 /*ARGSUSED*/
1320 STATIC void
1321 xfs_qm_dqflush_done(
1322         xfs_buf_t               *bp,
1323         xfs_dq_logitem_t        *qip)
1324 {
1325         xfs_dquot_t             *dqp;
1326
1327         dqp = qip->qli_dquot;
1328
1329         /*
1330          * We only want to pull the item from the AIL if its
1331          * location in the log has not changed since we started the flush.
1332          * Thus, we only bother if the dquot's lsn has
1333          * not changed. First we check the lsn outside the lock
1334          * since it's cheaper, and then we recheck while
1335          * holding the lock before removing the dquot from the AIL.
1336          */
1337         if ((qip->qli_item.li_flags & XFS_LI_IN_AIL) &&
1338             qip->qli_item.li_lsn == qip->qli_flush_lsn) {
1339
1340                 spin_lock(&dqp->q_mount->m_ail_lock);
1341                 /*
1342                  * xfs_trans_delete_ail() drops the AIL lock.
1343                  */
1344                 if (qip->qli_item.li_lsn == qip->qli_flush_lsn)
1345                         xfs_trans_delete_ail(dqp->q_mount,
1346                                              (xfs_log_item_t*)qip);
1347                 else
1348                         spin_unlock(&dqp->q_mount->m_ail_lock);
1349         }
1350
1351         /*
1352          * Release the dq's flush lock since we're done with it.
1353          */
1354         xfs_dqfunlock(dqp);
1355 }
1356
1357 int
1358 xfs_qm_dqlock_nowait(
1359         xfs_dquot_t *dqp)
1360 {
1361         return mutex_trylock(&dqp->q_qlock);
1362 }
1363
1364 void
1365 xfs_dqlock(
1366         xfs_dquot_t *dqp)
1367 {
1368         mutex_lock(&dqp->q_qlock);
1369 }
1370
1371 void
1372 xfs_dqunlock(
1373         xfs_dquot_t *dqp)
1374 {
1375         mutex_unlock(&(dqp->q_qlock));
1376         if (dqp->q_logitem.qli_dquot == dqp) {
1377                 /* Once was dqp->q_mount, but might just have been cleared */
1378                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_mountp,
1379                                         (xfs_log_item_t*)&(dqp->q_logitem));
1380         }
1381 }
1382
1383
1384 void
1385 xfs_dqunlock_nonotify(
1386         xfs_dquot_t *dqp)
1387 {
1388         mutex_unlock(&(dqp->q_qlock));
1389 }
1390
1391 void
1392 xfs_dqlock2(
1393         xfs_dquot_t     *d1,
1394         xfs_dquot_t     *d2)
1395 {
1396         if (d1 && d2) {
1397                 ASSERT(d1 != d2);
1398                 if (be32_to_cpu(d1->q_core.d_id) >
1399                     be32_to_cpu(d2->q_core.d_id)) {
1400                         xfs_dqlock(d2);
1401                         xfs_dqlock(d1);
1402                 } else {
1403                         xfs_dqlock(d1);
1404                         xfs_dqlock(d2);
1405                 }
1406         } else {
1407                 if (d1) {
1408                         xfs_dqlock(d1);
1409                 } else if (d2) {
1410                         xfs_dqlock(d2);
1411                 }
1412         }
1413 }
1414
1415
1416 /*
1417  * Take a dquot out of the mount's dqlist as well as the hashlist.
1418  * This is called via unmount as well as quotaoff, and the purge
1419  * will always succeed unless there are soft (temp) references
1420  * outstanding.
1421  *
1422  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1423  * that we're returning! XXXsup - not cool.
1424  */
1425 /* ARGSUSED */
1426 int
1427 xfs_qm_dqpurge(
1428         xfs_dquot_t     *dqp)
1429 {
1430         xfs_dqhash_t    *thishash;
1431         xfs_mount_t     *mp = dqp->q_mount;
1432
1433         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
1434         ASSERT(XFS_DQ_IS_HASH_LOCKED(dqp->q_hash));
1435
1436         xfs_dqlock(dqp);
1437         /*
1438          * We really can't afford to purge a dquot that is
1439          * referenced, because these are hard refs.
1440          * It shouldn't happen in general because we went thru _all_ inodes in
1441          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1442          * However it is possible that we have dquots with temporary
1443          * references that are not attached to an inode. e.g. see xfs_setattr().
1444          */
1445         if (dqp->q_nrefs != 0) {
1446                 xfs_dqunlock(dqp);
1447                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
1448                 return (1);
1449         }
1450
1451         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1452
1453         /*
1454          * If we're turning off quotas, we have to make sure that, for
1455          * example, we don't delete quota disk blocks while dquots are
1456          * in the process of getting written to those disk blocks.
1457          * This dquot might well be on AIL, and we can't leave it there
1458          * if we're turning off quotas. Basically, we need this flush
1459          * lock, and are willing to block on it.
1460          */
1461         if (!xfs_dqflock_nowait(dqp)) {
1462                 /*
1463                  * Block on the flush lock after nudging dquot buffer,
1464                  * if it is incore.
1465                  */
1466                 xfs_qm_dqflock_pushbuf_wait(dqp);
1467         }
1468
1469         /*
1470          * XXXIf we're turning this type of quotas off, we don't care
1471          * about the dirty metadata sitting in this dquot. OTOH, if
1472          * we're unmounting, we do care, so we flush it and wait.
1473          */
1474         if (XFS_DQ_IS_DIRTY(dqp)) {
1475                 int     error;
1476                 xfs_dqtrace_entry(dqp, "DQPURGE ->DQFLUSH: DQDIRTY");
1477                 /* dqflush unlocks dqflock */
1478                 /*
1479                  * Given that dqpurge is a very rare occurrence, it is OK
1480                  * that we're holding the hashlist and mplist locks
1481                  * across the disk write. But, ... XXXsup
1482                  *
1483                  * We don't care about getting disk errors here. We need
1484                  * to purge this dquot anyway, so we go ahead regardless.
1485                  */
1486                 error = xfs_qm_dqflush(dqp, XFS_QMOPT_SYNC);
1487                 if (error)
1488                         xfs_fs_cmn_err(CE_WARN, mp,
1489                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1490                 xfs_dqflock(dqp);
1491         }
1492         ASSERT(dqp->q_pincount == 0);
1493         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1494                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1495
1496         thishash = dqp->q_hash;
1497         XQM_HASHLIST_REMOVE(thishash, dqp);
1498         XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(mp)), dqp);
1499         /*
1500          * XXX Move this to the front of the freelist, if we can get the
1501          * freelist lock.
1502          */
1503         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1504
1505         dqp->q_mount = NULL;
1506         dqp->q_hash = NULL;
1507         dqp->dq_flags = XFS_DQ_INACTIVE;
1508         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1509         xfs_dqfunlock(dqp);
1510         xfs_dqunlock(dqp);
1511         XFS_DQ_HASH_UNLOCK(thishash);
1512         return (0);
1513 }
1514
1515
1516 #ifdef QUOTADEBUG
1517 void
1518 xfs_qm_dqprint(xfs_dquot_t *dqp)
1519 {
1520         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1521         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1522                 (int)be32_to_cpu(dqp->q_core.d_id));
1523         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1524         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1525         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1526         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1527         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1528                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1529                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1530         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1531                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1532                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1533         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1534                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1535                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1536         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1537                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1538                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1539         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1540                 be64_to_cpu(dqp->q_core.d_bcount),
1541                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1542         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1543                 be64_to_cpu(dqp->q_core.d_icount),
1544                 (int)be64_to_cpu(dqp->q_core.d_icount));
1545         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1546                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1547         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1548                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1549         cmn_err(CE_DEBUG, "---------------------------");
1550 }
1551 #endif
1552
1553 /*
1554  * Give the buffer a little push if it is incore and
1555  * wait on the flush lock.
1556  */
1557 void
1558 xfs_qm_dqflock_pushbuf_wait(
1559         xfs_dquot_t     *dqp)
1560 {
1561         xfs_buf_t       *bp;
1562
1563         /*
1564          * Check to see if the dquot has been flushed delayed
1565          * write.  If so, grab its buffer and send it
1566          * out immediately.  We'll be able to acquire
1567          * the flush lock when the I/O completes.
1568          */
1569         bp = xfs_incore(dqp->q_mount->m_ddev_targp, dqp->q_blkno,
1570                     XFS_QI_DQCHUNKLEN(dqp->q_mount),
1571                     XFS_INCORE_TRYLOCK);
1572         if (bp != NULL) {
1573                 if (XFS_BUF_ISDELAYWRITE(bp)) {
1574                         int     error;
1575                         if (XFS_BUF_ISPINNED(bp)) {
1576                                 xfs_log_force(dqp->q_mount,
1577                                               (xfs_lsn_t)0,
1578                                               XFS_LOG_FORCE);
1579                         }
1580                         error = xfs_bawrite(dqp->q_mount, bp);
1581                         if (error)
1582                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
1583                                         "xfs_qm_dqflock_pushbuf_wait: "
1584                                         "pushbuf error %d on dqp %p, bp %p",
1585                                         error, dqp, bp);
1586                 } else {
1587                         xfs_buf_relse(bp);
1588                 }
1589         }
1590         xfs_dqflock(dqp);
1591 }