fs/xfs/quota/xfs_dquot.c

   1 /*
   2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_alloc.h"
  27 #include "xfs_quota.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_inode.h"
  31 #include "xfs_bmap.h"
  32 #include "xfs_rtalloc.h"
  33 #include "xfs_error.h"
  34 #include "xfs_itable.h"
  35 #include "xfs_attr.h"
  36 #include "xfs_buf_item.h"
  37 #include "xfs_trans_space.h"
  38 #include "xfs_trans_priv.h"
  39 #include "xfs_qm.h"
  40 #include "xfs_trace.h"
  41
  42
  43 /*
  44    LOCK ORDER
  45
  46    inode lock               (ilock)
  47    dquot hash-chain lock    (hashlock)
  48    xqm dquot freelist lock  (freelistlock
  49    mount's dquot list lock  (mplistlock)
  50    user dquot lock - lock ordering among dquots is based on the uid or gid
  51    group dquot lock - similar to udquots. Between the two dquots, the udquot
  52                       has to be locked first.
  53    pin lock - the dquot lock must be held to take this lock.
  54    flush lock - ditto.
  55 */
  56
  57 #ifdef DEBUG
  58 xfs_buftarg_t *xfs_dqerror_target;
  59 int xfs_do_dqerror;
  60 int xfs_dqreq_num;
  61 int xfs_dqerror_mod = 33;
  62 #endif
  63
  64 static struct lock_class_key xfs_dquot_other_class;
  65
  66 /*
  67  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  68  * we try to reclaim a free dquot if the number of incore dquots are above
  69  * a threshold.
  70  * The only field inside the core that gets initialized at this point
  71  * is the d_id field. The idea is to fill in the entire q_core
  72  * when we read in the on disk dquot.
  73  */
  74 STATIC xfs_dquot_t *
  75 xfs_qm_dqinit(
  76         xfs_mount_t  *mp,
  77         xfs_dqid_t   id,
  78         uint         type)
  79 {
  80         xfs_dquot_t     *dqp;
  81         boolean_t       brandnewdquot;
  82
  83         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  84         dqp->dq_flags = type;
  85         dqp->q_core.d_id = cpu_to_be32(id);
  86         dqp->q_mount = mp;
  87
  88         /*
  89          * No need to re-initialize these if this is a reclaimed dquot.
  90          */
  91         if (brandnewdquot) {
  92                 INIT_LIST_HEAD(&dqp->q_freelist);
  93                 mutex_init(&dqp->q_qlock);
  94                 init_waitqueue_head(&dqp->q_pinwait);
  95
  96                 /*
  97                  * Because we want to use a counting completion, complete
  98                  * the flush completion once to allow a single access to
  99                  * the flush completion without blocking.
 100                  */
 101                 init_completion(&dqp->q_flush);
 102                 complete(&dqp->q_flush);
 103
 104                 trace_xfs_dqinit(dqp);
 105         } else {
 106                 /*
 107                  * Only the q_core portion was zeroed in dqreclaim_one().
 108                  * So, we need to reset others.
 109                  */
 110                 dqp->q_nrefs = 0;
 111                 dqp->q_blkno = 0;
 112                 INIT_LIST_HEAD(&dqp->q_mplist);
 113                 INIT_LIST_HEAD(&dqp->q_hashlist);
 114                 dqp->q_bufoffset = 0;
 115                 dqp->q_fileoffset = 0;
 116                 dqp->q_transp = NULL;
 117                 dqp->q_gdquot = NULL;
 118                 dqp->q_res_bcount = 0;
 119                 dqp->q_res_icount = 0;
 120                 dqp->q_res_rtbcount = 0;
 121                 atomic_set(&dqp->q_pincount, 0);
 122                 dqp->q_hash = NULL;
 123                 ASSERT(list_empty(&dqp->q_freelist));
 124
 125                 trace_xfs_dqreuse(dqp);
 126         }
 127
 128         /*
 129          * In either case we need to make sure group quotas have a different
 130          * lock class than user quotas, to make sure lockdep knows we can
 131          * locks of one of each at the same time.
 132          */
 133         if (!(type & XFS_DQ_USER))
 134                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 135
 136         /*
 137          * log item gets initialized later
 138          */
 139         return (dqp);
 140 }
 141
 142 /*
 143  * This is called to free all the memory associated with a dquot
 144  */
 145 void
 146 xfs_qm_dqdestroy(
 147         xfs_dquot_t     *dqp)
 148 {
 149         ASSERT(list_empty(&dqp->q_freelist));
 150
 151         mutex_destroy(&dqp->q_qlock);
 152         sv_destroy(&dqp->q_pinwait);
 153         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 154
 155         atomic_dec(&xfs_Gqm->qm_totaldquots);
 156 }
 157
 158 /*
 159  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 160  */
 161 STATIC void
 162 xfs_qm_dqinit_core(
 163         xfs_dqid_t      id,
 164         uint            type,
 165         xfs_dqblk_t     *d)
 166 {
 167         /*
 168          * Caller has zero'd the entire dquot 'chunk' already.
 169          */
 170         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 171         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 172         d->dd_diskdq.d_id = cpu_to_be32(id);
 173         d->dd_diskdq.d_flags = type;
 174 }
 175
 176 /*
 177  * If default limits are in force, push them into the dquot now.
 178  * We overwrite the dquot limits only if they are zero and this
 179  * is not the root dquot.
 180  */
 181 void
 182 xfs_qm_adjust_dqlimits(
 183         xfs_mount_t             *mp,
 184         xfs_disk_dquot_t        *d)
 185 {
 186         xfs_quotainfo_t         *q = mp->m_quotainfo;
 187
 188         ASSERT(d->d_id);
 189
 190         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 191                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 192         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 193                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 194         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 195                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 196         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 197                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 198         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 199                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 200         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 201                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 202 }
 203
 204 /*
 205  * Check the limits and timers of a dquot and start or reset timers
 206  * if necessary.
 207  * This gets called even when quota enforcement is OFF, which makes our
 208  * life a little less complicated. (We just don't reject any quota
 209  * reservations in that case, when enforcement is off).
 210  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 211  * enforcement's off.
 212  * In contrast, warnings are a little different in that they don't
 213  * 'automatically' get started when limits get exceeded.  They do
 214  * get reset to zero, however, when we find the count to be under
 215  * the soft limit (they are only ever set non-zero via userspace).
 216  */
 217 void
 218 xfs_qm_adjust_dqtimers(
 219         xfs_mount_t             *mp,
 220         xfs_disk_dquot_t        *d)
 221 {
 222         ASSERT(d->d_id);
 223
 224 #ifdef QUOTADEBUG
 225         if (d->d_blk_hardlimit)
 226                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 227                        be64_to_cpu(d->d_blk_hardlimit));
 228         if (d->d_ino_hardlimit)
 229                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 230                        be64_to_cpu(d->d_ino_hardlimit));
 231         if (d->d_rtb_hardlimit)
 232                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 233                        be64_to_cpu(d->d_rtb_hardlimit));
 234 #endif
 235         if (!d->d_btimer) {
 236                 if ((d->d_blk_softlimit &&
 237                      (be64_to_cpu(d->d_bcount) >=
 238                       be64_to_cpu(d->d_blk_softlimit))) ||
 239                     (d->d_blk_hardlimit &&
 240                      (be64_to_cpu(d->d_bcount) >=
 241                       be64_to_cpu(d->d_blk_hardlimit)))) {
 242                         d->d_btimer = cpu_to_be32(get_seconds() +
 243                                         mp->m_quotainfo->qi_btimelimit);
 244                 } else {
 245                         d->d_bwarns = 0;
 246                 }
 247         } else {
 248                 if ((!d->d_blk_softlimit ||
 249                      (be64_to_cpu(d->d_bcount) <
 250                       be64_to_cpu(d->d_blk_softlimit))) &&
 251                     (!d->d_blk_hardlimit ||
 252                     (be64_to_cpu(d->d_bcount) <
 253                      be64_to_cpu(d->d_blk_hardlimit)))) {
 254                         d->d_btimer = 0;
 255                 }
 256         }
 257
 258         if (!d->d_itimer) {
 259                 if ((d->d_ino_softlimit &&
 260                      (be64_to_cpu(d->d_icount) >=
 261                       be64_to_cpu(d->d_ino_softlimit))) ||
 262                     (d->d_ino_hardlimit &&
 263                      (be64_to_cpu(d->d_icount) >=
 264                       be64_to_cpu(d->d_ino_hardlimit)))) {
 265                         d->d_itimer = cpu_to_be32(get_seconds() +
 266                                         mp->m_quotainfo->qi_itimelimit);
 267                 } else {
 268                         d->d_iwarns = 0;
 269                 }
 270         } else {
 271                 if ((!d->d_ino_softlimit ||
 272                      (be64_to_cpu(d->d_icount) <
 273                       be64_to_cpu(d->d_ino_softlimit)))  &&
 274                     (!d->d_ino_hardlimit ||
 275                      (be64_to_cpu(d->d_icount) <
 276                       be64_to_cpu(d->d_ino_hardlimit)))) {
 277                         d->d_itimer = 0;
 278                 }
 279         }
 280
 281         if (!d->d_rtbtimer) {
 282                 if ((d->d_rtb_softlimit &&
 283                      (be64_to_cpu(d->d_rtbcount) >=
 284                       be64_to_cpu(d->d_rtb_softlimit))) ||
 285                     (d->d_rtb_hardlimit &&
 286                      (be64_to_cpu(d->d_rtbcount) >=
 287                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 288                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 289                                         mp->m_quotainfo->qi_rtbtimelimit);
 290                 } else {
 291                         d->d_rtbwarns = 0;
 292                 }
 293         } else {
 294                 if ((!d->d_rtb_softlimit ||
 295                      (be64_to_cpu(d->d_rtbcount) <
 296                       be64_to_cpu(d->d_rtb_softlimit))) &&
 297                     (!d->d_rtb_hardlimit ||
 298                      (be64_to_cpu(d->d_rtbcount) <
 299                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 300                         d->d_rtbtimer = 0;
 301                 }
 302         }
 303 }
 304
 305 /*
 306  * initialize a buffer full of dquots and log the whole thing
 307  */
 308 STATIC void
 309 xfs_qm_init_dquot_blk(
 310         xfs_trans_t     *tp,
 311         xfs_mount_t     *mp,
 312         xfs_dqid_t      id,
 313         uint            type,
 314         xfs_buf_t       *bp)
 315 {
 316         struct xfs_quotainfo    *q = mp->m_quotainfo;
 317         xfs_dqblk_t     *d;
 318         int             curid, i;
 319
 320         ASSERT(tp);
 321         ASSERT(XFS_BUF_ISBUSY(bp));
 322         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 323
 324         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 325
 326         /*
 327          * ID of the first dquot in the block - id's are zero based.
 328          */
 329         curid = id - (id % q->qi_dqperchunk);
 330         ASSERT(curid >= 0);
 331         memset(d, 0, BBTOB(q->qi_dqchunklen));
 332         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++)
 333                 xfs_qm_dqinit_core(curid, type, d);
 334         xfs_trans_dquot_buf(tp, bp,
 335                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 336                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 337                              XFS_BLF_GDQUOT_BUF)));
 338         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 339 }
 340
 341
 342
 343 /*
 344  * Allocate a block and fill it with dquots.
 345  * This is called when the bmapi finds a hole.
 346  */
 347 STATIC int
 348 xfs_qm_dqalloc(
 349         xfs_trans_t     **tpp,
 350         xfs_mount_t     *mp,
 351         xfs_dquot_t     *dqp,
 352         xfs_inode_t     *quotip,
 353         xfs_fileoff_t   offset_fsb,
 354         xfs_buf_t       **O_bpp)
 355 {
 356         xfs_fsblock_t   firstblock;
 357         xfs_bmap_free_t flist;
 358         xfs_bmbt_irec_t map;
 359         int             nmaps, error, committed;
 360         xfs_buf_t       *bp;
 361         xfs_trans_t     *tp = *tpp;
 362
 363         ASSERT(tp != NULL);
 364
 365         trace_xfs_dqalloc(dqp);
 366
 367         /*
 368          * Initialize the bmap freelist prior to calling bmapi code.
 369          */
 370         xfs_bmap_init(&flist, &firstblock);
 371         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 372         /*
 373          * Return if this type of quotas is turned off while we didn't
 374          * have an inode lock
 375          */
 376         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 377                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 378                 return (ESRCH);
 379         }
 380
 381         xfs_trans_ijoin_ref(tp, quotip, XFS_ILOCK_EXCL);
 382         nmaps = 1;
 383         if ((error = xfs_bmapi(tp, quotip,
 384                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 385                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 386                               &firstblock,
 387                               XFS_QM_DQALLOC_SPACE_RES(mp),
 388                               &map, &nmaps, &flist))) {
 389                 goto error0;
 390         }
 391         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 392         ASSERT(nmaps == 1);
 393         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 394                (map.br_startblock != HOLESTARTBLOCK));
 395
 396         /*
 397          * Keep track of the blkno to save a lookup later
 398          */
 399         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 400
 401         /* now we can just get the buffer (there's nothing to read yet) */
 402         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 403                                dqp->q_blkno,
 404                                mp->m_quotainfo->qi_dqchunklen,
 405                                0);
 406         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 407                 goto error1;
 408         /*
 409          * Make a chunk of dquots out of this buffer and log
 410          * the entire thing.
 411          */
 412         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 413                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 414
 415         /*
 416          * xfs_bmap_finish() may commit the current transaction and
 417          * start a second transaction if the freelist is not empty.
 418          *
 419          * Since we still want to modify this buffer, we need to
 420          * ensure that the buffer is not released on commit of
 421          * the first transaction and ensure the buffer is added to the
 422          * second transaction.
 423          *
 424          * If there is only one transaction then don't stop the buffer
 425          * from being released when it commits later on.
 426          */
 427
 428         xfs_trans_bhold(tp, bp);
 429
 430         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 431                 goto error1;
 432         }
 433
 434         if (committed) {
 435                 tp = *tpp;
 436                 xfs_trans_bjoin(tp, bp);
 437         } else {
 438                 xfs_trans_bhold_release(tp, bp);
 439         }
 440
 441         *O_bpp = bp;
 442         return 0;
 443
 444       error1:
 445         xfs_bmap_cancel(&flist);
 446       error0:
 447         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 448
 449         return (error);
 450 }
 451
 452 /*
 453  * Maps a dquot to the buffer containing its on-disk version.
 454  * This returns a ptr to the buffer containing the on-disk dquot
 455  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 456  */
 457 STATIC int
 458 xfs_qm_dqtobp(
 459         xfs_trans_t             **tpp,
 460         xfs_dquot_t             *dqp,
 461         xfs_disk_dquot_t        **O_ddpp,
 462         xfs_buf_t               **O_bpp,
 463         uint                    flags)
 464 {
 465         xfs_bmbt_irec_t map;
 466         int             nmaps, error;
 467         xfs_buf_t       *bp;
 468         xfs_inode_t     *quotip;
 469         xfs_mount_t     *mp;
 470         xfs_disk_dquot_t *ddq;
 471         xfs_dqid_t      id;
 472         boolean_t       newdquot;
 473         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 474
 475         mp = dqp->q_mount;
 476         id = be32_to_cpu(dqp->q_core.d_id);
 477         nmaps = 1;
 478         newdquot = B_FALSE;
 479
 480         /*
 481          * If we don't know where the dquot lives, find out.
 482          */
 483         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 484                 /* We use the id as an index */
 485                 dqp->q_fileoffset = (xfs_fileoff_t)id /
 486                                         mp->m_quotainfo->qi_dqperchunk;
 487                 nmaps = 1;
 488                 quotip = XFS_DQ_TO_QIP(dqp);
 489                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 490                 /*
 491                  * Return if this type of quotas is turned off while we didn't
 492                  * have an inode lock
 493                  */
 494                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 495                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 496                         return (ESRCH);
 497                 }
 498                 /*
 499                  * Find the block map; no allocations yet
 500                  */
 501                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 502                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 503                                   XFS_BMAPI_METADATA,
 504                                   NULL, 0, &map, &nmaps, NULL);
 505
 506                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 507                 if (error)
 508                         return (error);
 509                 ASSERT(nmaps == 1);
 510                 ASSERT(map.br_blockcount == 1);
 511
 512                 /*
 513                  * offset of dquot in the (fixed sized) dquot chunk.
 514                  */
 515                 dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 516                         sizeof(xfs_dqblk_t);
 517                 if (map.br_startblock == HOLESTARTBLOCK) {
 518                         /*
 519                          * We don't allocate unless we're asked to
 520                          */
 521                         if (!(flags & XFS_QMOPT_DQALLOC))
 522                                 return (ENOENT);
 523
 524                         ASSERT(tp);
 525                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 526                                                 dqp->q_fileoffset, &bp)))
 527                                 return (error);
 528                         tp = *tpp;
 529                         newdquot = B_TRUE;
 530                 } else {
 531                         /*
 532                          * store the blkno etc so that we don't have to do the
 533                          * mapping all the time
 534                          */
 535                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 536                 }
 537         }
 538         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 539         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 540
 541         /*
 542          * Read in the buffer, unless we've just done the allocation
 543          * (in which case we already have the buf).
 544          */
 545         if (!newdquot) {
 546                 trace_xfs_dqtobp_read(dqp);
 547
 548                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 549                                            dqp->q_blkno,
 550                                            mp->m_quotainfo->qi_dqchunklen,
 551                                            0, &bp);
 552                 if (error || !bp)
 553                         return XFS_ERROR(error);
 554         }
 555         ASSERT(XFS_BUF_ISBUSY(bp));
 556         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 557
 558         /*
 559          * calculate the location of the dquot inside the buffer.
 560          */
 561         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 562
 563         /*
 564          * A simple sanity check in case we got a corrupted dquot...
 565          */
 566         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 567                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 568                            "dqtobp")) {
 569                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 570                         xfs_trans_brelse(tp, bp);
 571                         return XFS_ERROR(EIO);
 572                 }
 573                 XFS_BUF_BUSY(bp); /* We dirtied this */
 574         }
 575
 576         *O_bpp = bp;
 577         *O_ddpp = ddq;
 578
 579         return (0);
 580 }
 581
 582
 583 /*
 584  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 585  * and release the buffer immediately.
 586  *
 587  */
 588 /* ARGSUSED */
 589 STATIC int
 590 xfs_qm_dqread(
 591         xfs_trans_t     **tpp,
 592         xfs_dqid_t      id,
 593         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 594         uint            flags)
 595 {
 596         xfs_disk_dquot_t *ddqp;
 597         xfs_buf_t        *bp;
 598         int              error;
 599         xfs_trans_t      *tp;
 600
 601         ASSERT(tpp);
 602
 603         trace_xfs_dqread(dqp);
 604
 605         /*
 606          * get a pointer to the on-disk dquot and the buffer containing it
 607          * dqp already knows its own type (GROUP/USER).
 608          */
 609         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 610                 return (error);
 611         }
 612         tp = *tpp;
 613
 614         /* copy everything from disk dquot to the incore dquot */
 615         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 616         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 617         xfs_qm_dquot_logitem_init(dqp);
 618
 619         /*
 620          * Reservation counters are defined as reservation plus current usage
 621          * to avoid having to add everytime.
 622          */
 623         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 624         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 625         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 626
 627         /* Mark the buf so that this will stay incore a little longer */
 628         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 629
 630         /*
 631          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 632          * So we need to release with xfs_trans_brelse().
 633          * The strategy here is identical to that of inodes; we lock
 634          * the dquot in xfs_qm_dqget() before making it accessible to
 635          * others. This is because dquots, like inodes, need a good level of
 636          * concurrency, and we don't want to take locks on the entire buffers
 637          * for dquot accesses.
 638          * Note also that the dquot buffer may even be dirty at this point, if
 639          * this particular dquot was repaired. We still aren't afraid to
 640          * brelse it because we have the changes incore.
 641          */
 642         ASSERT(XFS_BUF_ISBUSY(bp));
 643         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 644         xfs_trans_brelse(tp, bp);
 645
 646         return (error);
 647 }
 648
 649
 650 /*
 651  * allocate an incore dquot from the kernel heap,
 652  * and fill its core with quota information kept on disk.
 653  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 654  * if it wasn't already allocated.
 655  */
 656 STATIC int
 657 xfs_qm_idtodq(
 658         xfs_mount_t     *mp,
 659         xfs_dqid_t      id,      /* gid or uid, depending on type */
 660         uint            type,    /* UDQUOT or GDQUOT */
 661         uint            flags,   /* DQALLOC, DQREPAIR */
 662         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 663 {
 664         xfs_dquot_t     *dqp;
 665         int             error;
 666         xfs_trans_t     *tp;
 667         int             cancelflags=0;
 668
 669         dqp = xfs_qm_dqinit(mp, id, type);
 670         tp = NULL;
 671         if (flags & XFS_QMOPT_DQALLOC) {
 672                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 673                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
 674                                 XFS_WRITE_LOG_RES(mp) +
 675                                 BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 +
 676                                 128,
 677                                 0,
 678                                 XFS_TRANS_PERM_LOG_RES,
 679                                 XFS_WRITE_LOG_COUNT);
 680                 if (error) {
 681                         cancelflags = 0;
 682                         goto error0;
 683                 }
 684                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 685         }
 686
 687         /*
 688          * Read it from disk; xfs_dqread() takes care of
 689          * all the necessary initialization of dquot's fields (locks, etc)
 690          */
 691         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 692                 /*
 693                  * This can happen if quotas got turned off (ESRCH),
 694                  * or if the dquot didn't exist on disk and we ask to
 695                  * allocate (ENOENT).
 696                  */
 697                 trace_xfs_dqread_fail(dqp);
 698                 cancelflags |= XFS_TRANS_ABORT;
 699                 goto error0;
 700         }
 701         if (tp) {
 702                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 703                         goto error1;
 704         }
 705
 706         *O_dqpp = dqp;
 707         return (0);
 708
 709  error0:
 710         ASSERT(error);
 711         if (tp)
 712                 xfs_trans_cancel(tp, cancelflags);
 713  error1:
 714         xfs_qm_dqdestroy(dqp);
 715         *O_dqpp = NULL;
 716         return (error);
 717 }
 718
 719 /*
 720  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 721  * hashtables for user and group dquots; and, these are global tables
 722  * inside the XQM, not per-filesystem tables.
 723  * The hash chain must be locked by caller, and it is left locked
 724  * on return. Returning dquot is locked.
 725  */
 726 STATIC int
 727 xfs_qm_dqlookup(
 728         xfs_mount_t             *mp,
 729         xfs_dqid_t              id,
 730         xfs_dqhash_t            *qh,
 731         xfs_dquot_t             **O_dqpp)
 732 {
 733         xfs_dquot_t             *dqp;
 734         uint                    flist_locked;
 735
 736         ASSERT(mutex_is_locked(&qh->qh_lock));
 737
 738         flist_locked = B_FALSE;
 739
 740         /*
 741          * Traverse the hashchain looking for a match
 742          */
 743         list_for_each_entry(dqp, &qh->qh_list, q_hashlist) {
 744                 /*
 745                  * We already have the hashlock. We don't need the
 746                  * dqlock to look at the id field of the dquot, since the
 747                  * id can't be modified without the hashlock anyway.
 748                  */
 749                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 750                         trace_xfs_dqlookup_found(dqp);
 751
 752                         /*
 753                          * All in core dquots must be on the dqlist of mp
 754                          */
 755                         ASSERT(!list_empty(&dqp->q_mplist));
 756
 757                         xfs_dqlock(dqp);
 758                         if (dqp->q_nrefs == 0) {
 759                                 ASSERT(!list_empty(&dqp->q_freelist));
 760                                 if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
 761                                         trace_xfs_dqlookup_want(dqp);
 762
 763                                         /*
 764                                          * We may have raced with dqreclaim_one()
 765                                          * (and lost). So, flag that we don't
 766                                          * want the dquot to be reclaimed.
 767                                          */
 768                                         dqp->dq_flags |= XFS_DQ_WANT;
 769                                         xfs_dqunlock(dqp);
 770                                         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
 771                                         xfs_dqlock(dqp);
 772                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 773                                 }
 774                                 flist_locked = B_TRUE;
 775                         }
 776
 777                         /*
 778                          * id couldn't have changed; we had the hashlock all
 779                          * along
 780                          */
 781                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 782
 783                         if (flist_locked) {
 784                                 if (dqp->q_nrefs != 0) {
 785                                         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 786                                         flist_locked = B_FALSE;
 787                                 } else {
 788                                         /* take it off the freelist */
 789                                         trace_xfs_dqlookup_freelist(dqp);
 790                                         list_del_init(&dqp->q_freelist);
 791                                         xfs_Gqm->qm_dqfrlist_cnt--;
 792                                 }
 793                         }
 794
 795                         XFS_DQHOLD(dqp);
 796
 797                         if (flist_locked)
 798                                 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 799                         /*
 800                          * move the dquot to the front of the hashchain
 801                          */
 802                         ASSERT(mutex_is_locked(&qh->qh_lock));
 803                         list_move(&dqp->q_hashlist, &qh->qh_list);
 804                         trace_xfs_dqlookup_done(dqp);
 805                         *O_dqpp = dqp;
 806                         return 0;
 807                 }
 808         }
 809
 810         *O_dqpp = NULL;
 811         ASSERT(mutex_is_locked(&qh->qh_lock));
 812         return (1);
 813 }
 814
 815 /*
 816  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 817  * a locked dquot, doing an allocation (if requested) as needed.
 818  * When both an inode and an id are given, the inode's id takes precedence.
 819  * That is, if the id changes while we don't hold the ilock inside this
 820  * function, the new dquot is returned, not necessarily the one requested
 821  * in the id argument.
 822  */
 823 int
 824 xfs_qm_dqget(
 825         xfs_mount_t     *mp,
 826         xfs_inode_t     *ip,      /* locked inode (optional) */
 827         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 828         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 829         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 830         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 831 {
 832         xfs_dquot_t     *dqp;
 833         xfs_dqhash_t    *h;
 834         uint            version;
 835         int             error;
 836
 837         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 838         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 839             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 840             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 841                 return (ESRCH);
 842         }
 843         h = XFS_DQ_HASH(mp, id, type);
 844
 845 #ifdef DEBUG
 846         if (xfs_do_dqerror) {
 847                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 848                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 849                         cmn_err(CE_DEBUG, "Returning error in dqget");
 850                         return (EIO);
 851                 }
 852         }
 853 #endif
 854
 855  again:
 856
 857 #ifdef DEBUG
 858         ASSERT(type == XFS_DQ_USER ||
 859                type == XFS_DQ_PROJ ||
 860                type == XFS_DQ_GROUP);
 861         if (ip) {
 862                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 863                 if (type == XFS_DQ_USER)
 864                         ASSERT(ip->i_udquot == NULL);
 865                 else
 866                         ASSERT(ip->i_gdquot == NULL);
 867         }
 868 #endif
 869         mutex_lock(&h->qh_lock);
 870
 871         /*
 872          * Look in the cache (hashtable).
 873          * The chain is kept locked during lookup.
 874          */
 875         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 876                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 877                 /*
 878                  * The dquot was found, moved to the front of the chain,
 879                  * taken off the freelist if it was on it, and locked
 880                  * at this point. Just unlock the hashchain and return.
 881                  */
 882                 ASSERT(*O_dqpp);
 883                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 884                 mutex_unlock(&h->qh_lock);
 885                 trace_xfs_dqget_hit(*O_dqpp);
 886                 return (0);     /* success */
 887         }
 888         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 889
 890         /*
 891          * Dquot cache miss. We don't want to keep the inode lock across
 892          * a (potential) disk read. Also we don't want to deal with the lock
 893          * ordering between quotainode and this inode. OTOH, dropping the inode
 894          * lock here means dealing with a chown that can happen before
 895          * we re-acquire the lock.
 896          */
 897         if (ip)
 898                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 899         /*
 900          * Save the hashchain version stamp, and unlock the chain, so that
 901          * we don't keep the lock across a disk read
 902          */
 903         version = h->qh_version;
 904         mutex_unlock(&h->qh_lock);
 905
 906         /*
 907          * Allocate the dquot on the kernel heap, and read the ondisk
 908          * portion off the disk. Also, do all the necessary initialization
 909          * This can return ENOENT if dquot didn't exist on disk and we didn't
 910          * ask it to allocate; ESRCH if quotas got turned off suddenly.
 911          */
 912         if ((error = xfs_qm_idtodq(mp, id, type,
 913                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
 914                                            XFS_QMOPT_DOWARN),
 915                                   &dqp))) {
 916                 if (ip)
 917                         xfs_ilock(ip, XFS_ILOCK_EXCL);
 918                 return (error);
 919         }
 920
 921         /*
 922          * See if this is mount code calling to look at the overall quota limits
 923          * which are stored in the id == 0 user or group's dquot.
 924          * Since we may not have done a quotacheck by this point, just return
 925          * the dquot without attaching it to any hashtables, lists, etc, or even
 926          * taking a reference.
 927          * The caller must dqdestroy this once done.
 928          */
 929         if (flags & XFS_QMOPT_DQSUSER) {
 930                 ASSERT(id == 0);
 931                 ASSERT(! ip);
 932                 goto dqret;
 933         }
 934
 935         /*
 936          * Dquot lock comes after hashlock in the lock ordering
 937          */
 938         if (ip) {
 939                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 940
 941                 /*
 942                  * A dquot could be attached to this inode by now, since
 943                  * we had dropped the ilock.
 944                  */
 945                 if (type == XFS_DQ_USER) {
 946                         if (!XFS_IS_UQUOTA_ON(mp)) {
 947                                 /* inode stays locked on return */
 948                                 xfs_qm_dqdestroy(dqp);
 949                                 return XFS_ERROR(ESRCH);
 950                         }
 951                         if (ip->i_udquot) {
 952                                 xfs_qm_dqdestroy(dqp);
 953                                 dqp = ip->i_udquot;
 954                                 xfs_dqlock(dqp);
 955                                 goto dqret;
 956                         }
 957                 } else {
 958                         if (!XFS_IS_OQUOTA_ON(mp)) {
 959                                 /* inode stays locked on return */
 960                                 xfs_qm_dqdestroy(dqp);
 961                                 return XFS_ERROR(ESRCH);
 962                         }
 963                         if (ip->i_gdquot) {
 964                                 xfs_qm_dqdestroy(dqp);
 965                                 dqp = ip->i_gdquot;
 966                                 xfs_dqlock(dqp);
 967                                 goto dqret;
 968                         }
 969                 }
 970         }
 971
 972         /*
 973          * Hashlock comes after ilock in lock order
 974          */
 975         mutex_lock(&h->qh_lock);
 976         if (version != h->qh_version) {
 977                 xfs_dquot_t *tmpdqp;
 978                 /*
 979                  * Now, see if somebody else put the dquot in the
 980                  * hashtable before us. This can happen because we didn't
 981                  * keep the hashchain lock. We don't have to worry about
 982                  * lock order between the two dquots here since dqp isn't
 983                  * on any findable lists yet.
 984                  */
 985                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
 986                         /*
 987                          * Duplicate found. Just throw away the new dquot
 988                          * and start over.
 989                          */
 990                         xfs_qm_dqput(tmpdqp);
 991                         mutex_unlock(&h->qh_lock);
 992                         xfs_qm_dqdestroy(dqp);
 993                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
 994                         goto again;
 995                 }
 996         }
 997
 998         /*
 999          * Put the dquot at the beginning of the hash-chain and mp's list
1000          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1001          */
1002         ASSERT(mutex_is_locked(&h->qh_lock));
1003         dqp->q_hash = h;
1004         list_add(&dqp->q_hashlist, &h->qh_list);
1005         h->qh_version++;
1006
1007         /*
1008          * Attach this dquot to this filesystem's list of all dquots,
1009          * kept inside the mount structure in m_quotainfo field
1010          */
1011         mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
1012
1013         /*
1014          * We return a locked dquot to the caller, with a reference taken
1015          */
1016         xfs_dqlock(dqp);
1017         dqp->q_nrefs = 1;
1018
1019         list_add(&dqp->q_mplist, &mp->m_quotainfo->qi_dqlist);
1020         mp->m_quotainfo->qi_dquots++;
1021         mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1022         mutex_unlock(&h->qh_lock);
1023  dqret:
1024         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1025         trace_xfs_dqget_miss(dqp);
1026         *O_dqpp = dqp;
1027         return (0);
1028 }
1029
1030
1031 /*
1032  * Release a reference to the dquot (decrement ref-count)
1033  * and unlock it. If there is a group quota attached to this
1034  * dquot, carefully release that too without tripping over
1035  * deadlocks'n'stuff.
1036  */
1037 void
1038 xfs_qm_dqput(
1039         xfs_dquot_t     *dqp)
1040 {
1041         xfs_dquot_t     *gdqp;
1042
1043         ASSERT(dqp->q_nrefs > 0);
1044         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1045
1046         trace_xfs_dqput(dqp);
1047
1048         if (dqp->q_nrefs != 1) {
1049                 dqp->q_nrefs--;
1050                 xfs_dqunlock(dqp);
1051                 return;
1052         }
1053
1054         /*
1055          * drop the dqlock and acquire the freelist and dqlock
1056          * in the right order; but try to get it out-of-order first
1057          */
1058         if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
1059                 trace_xfs_dqput_wait(dqp);
1060                 xfs_dqunlock(dqp);
1061                 mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1062                 xfs_dqlock(dqp);
1063         }
1064
1065         while (1) {
1066                 gdqp = NULL;
1067
1068                 /* We can't depend on nrefs being == 1 here */
1069                 if (--dqp->q_nrefs == 0) {
1070                         trace_xfs_dqput_free(dqp);
1071
1072                         list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
1073                         xfs_Gqm->qm_dqfrlist_cnt++;
1074
1075                         /*
1076                          * If we just added a udquot to the freelist, then
1077                          * we want to release the gdquot reference that
1078                          * it (probably) has. Otherwise it'll keep the
1079                          * gdquot from getting reclaimed.
1080                          */
1081                         if ((gdqp = dqp->q_gdquot)) {
1082                                 /*
1083                                  * Avoid a recursive dqput call
1084                                  */
1085                                 xfs_dqlock(gdqp);
1086                                 dqp->q_gdquot = NULL;
1087                         }
1088                 }
1089                 xfs_dqunlock(dqp);
1090
1091                 /*
1092                  * If we had a group quota inside the user quota as a hint,
1093                  * release it now.
1094                  */
1095                 if (! gdqp)
1096                         break;
1097                 dqp = gdqp;
1098         }
1099         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1100 }
1101
1102 /*
1103  * Release a dquot. Flush it if dirty, then dqput() it.
1104  * dquot must not be locked.
1105  */
1106 void
1107 xfs_qm_dqrele(
1108         xfs_dquot_t     *dqp)
1109 {
1110         if (!dqp)
1111                 return;
1112
1113         trace_xfs_dqrele(dqp);
1114
1115         xfs_dqlock(dqp);
1116         /*
1117          * We don't care to flush it if the dquot is dirty here.
1118          * That will create stutters that we want to avoid.
1119          * Instead we do a delayed write when we try to reclaim
1120          * a dirty dquot. Also xfs_sync will take part of the burden...
1121          */
1122         xfs_qm_dqput(dqp);
1123 }
1124
1125 /*
1126  * This is the dquot flushing I/O completion routine.  It is called
1127  * from interrupt level when the buffer containing the dquot is
1128  * flushed to disk.  It is responsible for removing the dquot logitem
1129  * from the AIL if it has not been re-logged, and unlocking the dquot's
1130  * flush lock. This behavior is very similar to that of inodes..
1131  */
1132 STATIC void
1133 xfs_qm_dqflush_done(
1134         struct xfs_buf          *bp,
1135         struct xfs_log_item     *lip)
1136 {
1137         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
1138         xfs_dquot_t             *dqp = qip->qli_dquot;
1139         struct xfs_ail          *ailp = lip->li_ailp;
1140
1141         /*
1142          * We only want to pull the item from the AIL if its
1143          * location in the log has not changed since we started the flush.
1144          * Thus, we only bother if the dquot's lsn has
1145          * not changed. First we check the lsn outside the lock
1146          * since it's cheaper, and then we recheck while
1147          * holding the lock before removing the dquot from the AIL.
1148          */
1149         if ((lip->li_flags & XFS_LI_IN_AIL) &&
1150             lip->li_lsn == qip->qli_flush_lsn) {
1151
1152                 /* xfs_trans_ail_delete() drops the AIL lock. */
1153                 spin_lock(&ailp->xa_lock);
1154                 if (lip->li_lsn == qip->qli_flush_lsn)
1155                         xfs_trans_ail_delete(ailp, lip);
1156                 else
1157                         spin_unlock(&ailp->xa_lock);
1158         }
1159
1160         /*
1161          * Release the dq's flush lock since we're done with it.
1162          */
1163         xfs_dqfunlock(dqp);
1164 }
1165
1166 /*
1167  * Write a modified dquot to disk.
1168  * The dquot must be locked and the flush lock too taken by caller.
1169  * The flush lock will not be unlocked until the dquot reaches the disk,
1170  * but the dquot is free to be unlocked and modified by the caller
1171  * in the interim. Dquot is still locked on return. This behavior is
1172  * identical to that of inodes.
1173  */
1174 int
1175 xfs_qm_dqflush(
1176         xfs_dquot_t             *dqp,
1177         uint                    flags)
1178 {
1179         xfs_mount_t             *mp;
1180         xfs_buf_t               *bp;
1181         xfs_disk_dquot_t        *ddqp;
1182         int                     error;
1183
1184         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1185         ASSERT(!completion_done(&dqp->q_flush));
1186         trace_xfs_dqflush(dqp);
1187
1188         /*
1189          * If not dirty, or it's pinned and we are not supposed to
1190          * block, nada.
1191          */
1192         if (!XFS_DQ_IS_DIRTY(dqp) ||
1193             (!(flags & SYNC_WAIT) && atomic_read(&dqp->q_pincount) > 0)) {
1194                 xfs_dqfunlock(dqp);
1195                 return 0;
1196         }
1197         xfs_qm_dqunpin_wait(dqp);
1198
1199         /*
1200          * This may have been unpinned because the filesystem is shutting
1201          * down forcibly. If that's the case we must not write this dquot
1202          * to disk, because the log record didn't make it to disk!
1203          */
1204         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1205                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1206                 xfs_dqfunlock(dqp);
1207                 return XFS_ERROR(EIO);
1208         }
1209
1210         /*
1211          * Get the buffer containing the on-disk dquot
1212          * We don't need a transaction envelope because we know that the
1213          * the ondisk-dquot has already been allocated for.
1214          */
1215         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1216                 ASSERT(error != ENOENT);
1217                 /*
1218                  * Quotas could have gotten turned off (ESRCH)
1219                  */
1220                 xfs_dqfunlock(dqp);
1221                 return (error);
1222         }
1223
1224         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1225                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1226                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1227                 return XFS_ERROR(EIO);
1228         }
1229
1230         /* This is the only portion of data that needs to persist */
1231         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1232
1233         /*
1234          * Clear the dirty field and remember the flush lsn for later use.
1235          */
1236         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1237         mp = dqp->q_mount;
1238
1239         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1240                                         &dqp->q_logitem.qli_item.li_lsn);
1241
1242         /*
1243          * Attach an iodone routine so that we can remove this dquot from the
1244          * AIL and release the flush lock once the dquot is synced to disk.
1245          */
1246         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1247                                   &dqp->q_logitem.qli_item);
1248
1249         /*
1250          * If the buffer is pinned then push on the log so we won't
1251          * get stuck waiting in the write for too long.
1252          */
1253         if (XFS_BUF_ISPINNED(bp)) {
1254                 trace_xfs_dqflush_force(dqp);
1255                 xfs_log_force(mp, 0);
1256         }
1257
1258         if (flags & SYNC_WAIT)
1259                 error = xfs_bwrite(mp, bp);
1260         else
1261                 xfs_bdwrite(mp, bp);
1262
1263         trace_xfs_dqflush_done(dqp);
1264
1265         /*
1266          * dqp is still locked, but caller is free to unlock it now.
1267          */
1268         return error;
1269
1270 }
1271
1272 int
1273 xfs_qm_dqlock_nowait(
1274         xfs_dquot_t *dqp)
1275 {
1276         return mutex_trylock(&dqp->q_qlock);
1277 }
1278
1279 void
1280 xfs_dqlock(
1281         xfs_dquot_t *dqp)
1282 {
1283         mutex_lock(&dqp->q_qlock);
1284 }
1285
1286 void
1287 xfs_dqunlock(
1288         xfs_dquot_t *dqp)
1289 {
1290         mutex_unlock(&(dqp->q_qlock));
1291         if (dqp->q_logitem.qli_dquot == dqp) {
1292                 /* Once was dqp->q_mount, but might just have been cleared */
1293                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1294                                         (xfs_log_item_t*)&(dqp->q_logitem));
1295         }
1296 }
1297
1298
1299 void
1300 xfs_dqunlock_nonotify(
1301         xfs_dquot_t *dqp)
1302 {
1303         mutex_unlock(&(dqp->q_qlock));
1304 }
1305
1306 /*
1307  * Lock two xfs_dquot structures.
1308  *
1309  * To avoid deadlocks we always lock the quota structure with
1310  * the lowerd id first.
1311  */
1312 void
1313 xfs_dqlock2(
1314         xfs_dquot_t     *d1,
1315         xfs_dquot_t     *d2)
1316 {
1317         if (d1 && d2) {
1318                 ASSERT(d1 != d2);
1319                 if (be32_to_cpu(d1->q_core.d_id) >
1320                     be32_to_cpu(d2->q_core.d_id)) {
1321                         mutex_lock(&d2->q_qlock);
1322                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1323                 } else {
1324                         mutex_lock(&d1->q_qlock);
1325                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1326                 }
1327         } else if (d1) {
1328                 mutex_lock(&d1->q_qlock);
1329         } else if (d2) {
1330                 mutex_lock(&d2->q_qlock);
1331         }
1332 }
1333
1334
1335 /*
1336  * Take a dquot out of the mount's dqlist as well as the hashlist.
1337  * This is called via unmount as well as quotaoff, and the purge
1338  * will always succeed unless there are soft (temp) references
1339  * outstanding.
1340  *
1341  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1342  * that we're returning! XXXsup - not cool.
1343  */
1344 /* ARGSUSED */
1345 int
1346 xfs_qm_dqpurge(
1347         xfs_dquot_t     *dqp)
1348 {
1349         xfs_dqhash_t    *qh = dqp->q_hash;
1350         xfs_mount_t     *mp = dqp->q_mount;
1351
1352         ASSERT(mutex_is_locked(&mp->m_quotainfo->qi_dqlist_lock));
1353         ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1354
1355         xfs_dqlock(dqp);
1356         /*
1357          * We really can't afford to purge a dquot that is
1358          * referenced, because these are hard refs.
1359          * It shouldn't happen in general because we went thru _all_ inodes in
1360          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1361          * However it is possible that we have dquots with temporary
1362          * references that are not attached to an inode. e.g. see xfs_setattr().
1363          */
1364         if (dqp->q_nrefs != 0) {
1365                 xfs_dqunlock(dqp);
1366                 mutex_unlock(&dqp->q_hash->qh_lock);
1367                 return (1);
1368         }
1369
1370         ASSERT(!list_empty(&dqp->q_freelist));
1371
1372         /*
1373          * If we're turning off quotas, we have to make sure that, for
1374          * example, we don't delete quota disk blocks while dquots are
1375          * in the process of getting written to those disk blocks.
1376          * This dquot might well be on AIL, and we can't leave it there
1377          * if we're turning off quotas. Basically, we need this flush
1378          * lock, and are willing to block on it.
1379          */
1380         if (!xfs_dqflock_nowait(dqp)) {
1381                 /*
1382                  * Block on the flush lock after nudging dquot buffer,
1383                  * if it is incore.
1384                  */
1385                 xfs_qm_dqflock_pushbuf_wait(dqp);
1386         }
1387
1388         /*
1389          * XXXIf we're turning this type of quotas off, we don't care
1390          * about the dirty metadata sitting in this dquot. OTOH, if
1391          * we're unmounting, we do care, so we flush it and wait.
1392          */
1393         if (XFS_DQ_IS_DIRTY(dqp)) {
1394                 int     error;
1395
1396                 /* dqflush unlocks dqflock */
1397                 /*
1398                  * Given that dqpurge is a very rare occurrence, it is OK
1399                  * that we're holding the hashlist and mplist locks
1400                  * across the disk write. But, ... XXXsup
1401                  *
1402                  * We don't care about getting disk errors here. We need
1403                  * to purge this dquot anyway, so we go ahead regardless.
1404                  */
1405                 error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1406                 if (error)
1407                         xfs_fs_cmn_err(CE_WARN, mp,
1408                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1409                 xfs_dqflock(dqp);
1410         }
1411         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1412         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1413                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1414
1415         list_del_init(&dqp->q_hashlist);
1416         qh->qh_version++;
1417         list_del_init(&dqp->q_mplist);
1418         mp->m_quotainfo->qi_dqreclaims++;
1419         mp->m_quotainfo->qi_dquots--;
1420         /*
1421          * XXX Move this to the front of the freelist, if we can get the
1422          * freelist lock.
1423          */
1424         ASSERT(!list_empty(&dqp->q_freelist));
1425
1426         dqp->q_mount = NULL;
1427         dqp->q_hash = NULL;
1428         dqp->dq_flags = XFS_DQ_INACTIVE;
1429         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1430         xfs_dqfunlock(dqp);
1431         xfs_dqunlock(dqp);
1432         mutex_unlock(&qh->qh_lock);
1433         return (0);
1434 }
1435
1436
1437 #ifdef QUOTADEBUG
1438 void
1439 xfs_qm_dqprint(xfs_dquot_t *dqp)
1440 {
1441         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1442         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1443                 (int)be32_to_cpu(dqp->q_core.d_id));
1444         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1445         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1446         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1447         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1448         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1449                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1450                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1451         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1452                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1453                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1454         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1455                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1456                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1457         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1458                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1459                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1460         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1461                 be64_to_cpu(dqp->q_core.d_bcount),
1462                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1463         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1464                 be64_to_cpu(dqp->q_core.d_icount),
1465                 (int)be64_to_cpu(dqp->q_core.d_icount));
1466         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1467                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1468         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1469                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1470         cmn_err(CE_DEBUG, "---------------------------");
1471 }
1472 #endif
1473
1474 /*
1475  * Give the buffer a little push if it is incore and
1476  * wait on the flush lock.
1477  */
1478 void
1479 xfs_qm_dqflock_pushbuf_wait(
1480         xfs_dquot_t     *dqp)
1481 {
1482         xfs_mount_t     *mp = dqp->q_mount;
1483         xfs_buf_t       *bp;
1484
1485         /*
1486          * Check to see if the dquot has been flushed delayed
1487          * write.  If so, grab its buffer and send it
1488          * out immediately.  We'll be able to acquire
1489          * the flush lock when the I/O completes.
1490          */
1491         bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1492                         mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1493         if (!bp)
1494                 goto out_lock;
1495
1496         if (XFS_BUF_ISDELAYWRITE(bp)) {
1497                 if (XFS_BUF_ISPINNED(bp))
1498                         xfs_log_force(mp, 0);
1499                 xfs_buf_delwri_promote(bp);
1500                 wake_up_process(bp->b_target->bt_task);
1501         }
1502         xfs_buf_relse(bp);
1503 out_lock:
1504         xfs_dqflock(dqp);
1505 }