fs/xfs/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  * Lock order:
  44  *
  45  * ip->i_lock
  46  *   qi->qi_tree_lock
  47  *     dquot->q_qlock (xfs_dqlock() and friends)
  48  *       dquot->q_flush (xfs_dqflock() and friends)
  49  *       qi->qi_lru_lock
  50  *
  51  * If two dquots need to be locked the order is user before group/project,
  52  * otherwise by the lowest id first, see xfs_dqlock2.
  53  */
  54
  55 #ifdef DEBUG
  56 xfs_buftarg_t *xfs_dqerror_target;
  57 int xfs_do_dqerror;
  58 int xfs_dqreq_num;
  59 int xfs_dqerror_mod = 33;
  60 #endif
  61
  62 struct kmem_zone                *xfs_qm_dqtrxzone;
  63 static struct kmem_zone         *xfs_qm_dqzone;
  64
  65 static struct lock_class_key xfs_dquot_other_class;
  66
  67 /*
  68  * This is called to free all the memory associated with a dquot
  69  */
  70 void
  71 xfs_qm_dqdestroy(
  72         xfs_dquot_t     *dqp)
  73 {
  74         ASSERT(list_empty(&dqp->q_lru));
  75
  76         mutex_destroy(&dqp->q_qlock);
  77         kmem_zone_free(xfs_qm_dqzone, dqp);
  78
  79         XFS_STATS_DEC(xs_qm_dquot);
  80 }
  81
  82 /*
  83  * If default limits are in force, push them into the dquot now.
  84  * We overwrite the dquot limits only if they are zero and this
  85  * is not the root dquot.
  86  */
  87 void
  88 xfs_qm_adjust_dqlimits(
  89         xfs_mount_t             *mp,
  90         xfs_disk_dquot_t        *d)
  91 {
  92         xfs_quotainfo_t         *q = mp->m_quotainfo;
  93
  94         ASSERT(d->d_id);
  95
  96         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
  97                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
  98         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
  99                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 100         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 101                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 102         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 103                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 104         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 105                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 106         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 107                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 108 }
 109
 110 /*
 111  * Check the limits and timers of a dquot and start or reset timers
 112  * if necessary.
 113  * This gets called even when quota enforcement is OFF, which makes our
 114  * life a little less complicated. (We just don't reject any quota
 115  * reservations in that case, when enforcement is off).
 116  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 117  * enforcement's off.
 118  * In contrast, warnings are a little different in that they don't
 119  * 'automatically' get started when limits get exceeded.  They do
 120  * get reset to zero, however, when we find the count to be under
 121  * the soft limit (they are only ever set non-zero via userspace).
 122  */
 123 void
 124 xfs_qm_adjust_dqtimers(
 125         xfs_mount_t             *mp,
 126         xfs_disk_dquot_t        *d)
 127 {
 128         ASSERT(d->d_id);
 129
 130 #ifdef DEBUG
 131         if (d->d_blk_hardlimit)
 132                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 133                        be64_to_cpu(d->d_blk_hardlimit));
 134         if (d->d_ino_hardlimit)
 135                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 136                        be64_to_cpu(d->d_ino_hardlimit));
 137         if (d->d_rtb_hardlimit)
 138                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 139                        be64_to_cpu(d->d_rtb_hardlimit));
 140 #endif
 141
 142         if (!d->d_btimer) {
 143                 if ((d->d_blk_softlimit &&
 144                      (be64_to_cpu(d->d_bcount) >
 145                       be64_to_cpu(d->d_blk_softlimit))) ||
 146                     (d->d_blk_hardlimit &&
 147                      (be64_to_cpu(d->d_bcount) >
 148                       be64_to_cpu(d->d_blk_hardlimit)))) {
 149                         d->d_btimer = cpu_to_be32(get_seconds() +
 150                                         mp->m_quotainfo->qi_btimelimit);
 151                 } else {
 152                         d->d_bwarns = 0;
 153                 }
 154         } else {
 155                 if ((!d->d_blk_softlimit ||
 156                      (be64_to_cpu(d->d_bcount) <=
 157                       be64_to_cpu(d->d_blk_softlimit))) &&
 158                     (!d->d_blk_hardlimit ||
 159                     (be64_to_cpu(d->d_bcount) <=
 160                      be64_to_cpu(d->d_blk_hardlimit)))) {
 161                         d->d_btimer = 0;
 162                 }
 163         }
 164
 165         if (!d->d_itimer) {
 166                 if ((d->d_ino_softlimit &&
 167                      (be64_to_cpu(d->d_icount) >
 168                       be64_to_cpu(d->d_ino_softlimit))) ||
 169                     (d->d_ino_hardlimit &&
 170                      (be64_to_cpu(d->d_icount) >
 171                       be64_to_cpu(d->d_ino_hardlimit)))) {
 172                         d->d_itimer = cpu_to_be32(get_seconds() +
 173                                         mp->m_quotainfo->qi_itimelimit);
 174                 } else {
 175                         d->d_iwarns = 0;
 176                 }
 177         } else {
 178                 if ((!d->d_ino_softlimit ||
 179                      (be64_to_cpu(d->d_icount) <=
 180                       be64_to_cpu(d->d_ino_softlimit)))  &&
 181                     (!d->d_ino_hardlimit ||
 182                      (be64_to_cpu(d->d_icount) <=
 183                       be64_to_cpu(d->d_ino_hardlimit)))) {
 184                         d->d_itimer = 0;
 185                 }
 186         }
 187
 188         if (!d->d_rtbtimer) {
 189                 if ((d->d_rtb_softlimit &&
 190                      (be64_to_cpu(d->d_rtbcount) >
 191                       be64_to_cpu(d->d_rtb_softlimit))) ||
 192                     (d->d_rtb_hardlimit &&
 193                      (be64_to_cpu(d->d_rtbcount) >
 194                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 195                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 196                                         mp->m_quotainfo->qi_rtbtimelimit);
 197                 } else {
 198                         d->d_rtbwarns = 0;
 199                 }
 200         } else {
 201                 if ((!d->d_rtb_softlimit ||
 202                      (be64_to_cpu(d->d_rtbcount) <=
 203                       be64_to_cpu(d->d_rtb_softlimit))) &&
 204                     (!d->d_rtb_hardlimit ||
 205                      (be64_to_cpu(d->d_rtbcount) <=
 206                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 207                         d->d_rtbtimer = 0;
 208                 }
 209         }
 210 }
 211
 212 /*
 213  * initialize a buffer full of dquots and log the whole thing
 214  */
 215 STATIC void
 216 xfs_qm_init_dquot_blk(
 217         xfs_trans_t     *tp,
 218         xfs_mount_t     *mp,
 219         xfs_dqid_t      id,
 220         uint            type,
 221         xfs_buf_t       *bp)
 222 {
 223         struct xfs_quotainfo    *q = mp->m_quotainfo;
 224         xfs_dqblk_t     *d;
 225         int             curid, i;
 226
 227         ASSERT(tp);
 228         ASSERT(xfs_buf_islocked(bp));
 229
 230         d = bp->b_addr;
 231
 232         /*
 233          * ID of the first dquot in the block - id's are zero based.
 234          */
 235         curid = id - (id % q->qi_dqperchunk);
 236         ASSERT(curid >= 0);
 237         memset(d, 0, BBTOB(q->qi_dqchunklen));
 238         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 239                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 240                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 241                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 242                 d->dd_diskdq.d_flags = type;
 243         }
 244
 245         xfs_trans_dquot_buf(tp, bp,
 246                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 247                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 248                              XFS_BLF_GDQUOT_BUF)));
 249         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 250 }
 251
 252
 253
 254 /*
 255  * Allocate a block and fill it with dquots.
 256  * This is called when the bmapi finds a hole.
 257  */
 258 STATIC int
 259 xfs_qm_dqalloc(
 260         xfs_trans_t     **tpp,
 261         xfs_mount_t     *mp,
 262         xfs_dquot_t     *dqp,
 263         xfs_inode_t     *quotip,
 264         xfs_fileoff_t   offset_fsb,
 265         xfs_buf_t       **O_bpp)
 266 {
 267         xfs_fsblock_t   firstblock;
 268         xfs_bmap_free_t flist;
 269         xfs_bmbt_irec_t map;
 270         int             nmaps, error, committed;
 271         xfs_buf_t       *bp;
 272         xfs_trans_t     *tp = *tpp;
 273
 274         ASSERT(tp != NULL);
 275
 276         trace_xfs_dqalloc(dqp);
 277
 278         /*
 279          * Initialize the bmap freelist prior to calling bmapi code.
 280          */
 281         xfs_bmap_init(&flist, &firstblock);
 282         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 283         /*
 284          * Return if this type of quotas is turned off while we didn't
 285          * have an inode lock
 286          */
 287         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 288                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 289                 return (ESRCH);
 290         }
 291
 292         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 293         nmaps = 1;
 294         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 295                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 296                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 297                                 &map, &nmaps, &flist);
 298         if (error)
 299                 goto error0;
 300         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 301         ASSERT(nmaps == 1);
 302         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 303                (map.br_startblock != HOLESTARTBLOCK));
 304
 305         /*
 306          * Keep track of the blkno to save a lookup later
 307          */
 308         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 309
 310         /* now we can just get the buffer (there's nothing to read yet) */
 311         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 312                                dqp->q_blkno,
 313                                mp->m_quotainfo->qi_dqchunklen,
 314                                0);
 315
 316         error = xfs_buf_geterror(bp);
 317         if (error)
 318                 goto error1;
 319
 320         /*
 321          * Make a chunk of dquots out of this buffer and log
 322          * the entire thing.
 323          */
 324         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 325                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 326
 327         /*
 328          * xfs_bmap_finish() may commit the current transaction and
 329          * start a second transaction if the freelist is not empty.
 330          *
 331          * Since we still want to modify this buffer, we need to
 332          * ensure that the buffer is not released on commit of
 333          * the first transaction and ensure the buffer is added to the
 334          * second transaction.
 335          *
 336          * If there is only one transaction then don't stop the buffer
 337          * from being released when it commits later on.
 338          */
 339
 340         xfs_trans_bhold(tp, bp);
 341
 342         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 343                 goto error1;
 344         }
 345
 346         if (committed) {
 347                 tp = *tpp;
 348                 xfs_trans_bjoin(tp, bp);
 349         } else {
 350                 xfs_trans_bhold_release(tp, bp);
 351         }
 352
 353         *O_bpp = bp;
 354         return 0;
 355
 356       error1:
 357         xfs_bmap_cancel(&flist);
 358       error0:
 359         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 360
 361         return (error);
 362 }
 363
 364 /*
 365  * Maps a dquot to the buffer containing its on-disk version.
 366  * This returns a ptr to the buffer containing the on-disk dquot
 367  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 368  */
 369 STATIC int
 370 xfs_qm_dqtobp(
 371         xfs_trans_t             **tpp,
 372         xfs_dquot_t             *dqp,
 373         xfs_disk_dquot_t        **O_ddpp,
 374         xfs_buf_t               **O_bpp,
 375         uint                    flags)
 376 {
 377         xfs_bmbt_irec_t map;
 378         int             nmaps = 1, error;
 379         xfs_buf_t       *bp;
 380         xfs_inode_t     *quotip = XFS_DQ_TO_QIP(dqp);
 381         xfs_mount_t     *mp = dqp->q_mount;
 382         xfs_disk_dquot_t *ddq;
 383         xfs_dqid_t      id = be32_to_cpu(dqp->q_core.d_id);
 384         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 385
 386         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 387
 388         xfs_ilock(quotip, XFS_ILOCK_SHARED);
 389         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 390                 /*
 391                  * Return if this type of quotas is turned off while we
 392                  * didn't have the quota inode lock.
 393                  */
 394                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 395                 return ESRCH;
 396         }
 397
 398         /*
 399          * Find the block map; no allocations yet
 400          */
 401         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 402                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 403
 404         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 405         if (error)
 406                 return error;
 407
 408         ASSERT(nmaps == 1);
 409         ASSERT(map.br_blockcount == 1);
 410
 411         /*
 412          * Offset of dquot in the (fixed sized) dquot chunk.
 413          */
 414         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 415                 sizeof(xfs_dqblk_t);
 416
 417         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 418         if (map.br_startblock == HOLESTARTBLOCK) {
 419                 /*
 420                  * We don't allocate unless we're asked to
 421                  */
 422                 if (!(flags & XFS_QMOPT_DQALLOC))
 423                         return ENOENT;
 424
 425                 ASSERT(tp);
 426                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 427                                         dqp->q_fileoffset, &bp);
 428                 if (error)
 429                         return error;
 430                 tp = *tpp;
 431         } else {
 432                 trace_xfs_dqtobp_read(dqp);
 433
 434                 /*
 435                  * store the blkno etc so that we don't have to do the
 436                  * mapping all the time
 437                  */
 438                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 439
 440                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 441                                            dqp->q_blkno,
 442                                            mp->m_quotainfo->qi_dqchunklen,
 443                                            0, &bp);
 444                 if (error || !bp)
 445                         return XFS_ERROR(error);
 446         }
 447
 448         ASSERT(xfs_buf_islocked(bp));
 449
 450         /*
 451          * calculate the location of the dquot inside the buffer.
 452          */
 453         ddq = bp->b_addr + dqp->q_bufoffset;
 454
 455         /*
 456          * A simple sanity check in case we got a corrupted dquot...
 457          */
 458         error = xfs_qm_dqcheck(mp, ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 459                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 460                            "dqtobp");
 461         if (error) {
 462                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 463                         xfs_trans_brelse(tp, bp);
 464                         return XFS_ERROR(EIO);
 465                 }
 466         }
 467
 468         *O_bpp = bp;
 469         *O_ddpp = ddq;
 470
 471         return (0);
 472 }
 473
 474
 475 /*
 476  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 477  * and release the buffer immediately.
 478  *
 479  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 480  */
 481 int
 482 xfs_qm_dqread(
 483         struct xfs_mount        *mp,
 484         xfs_dqid_t              id,
 485         uint                    type,
 486         uint                    flags,
 487         struct xfs_dquot        **O_dqpp)
 488 {
 489         struct xfs_dquot        *dqp;
 490         struct xfs_disk_dquot   *ddqp;
 491         struct xfs_buf          *bp;
 492         struct xfs_trans        *tp = NULL;
 493         int                     error;
 494         int                     cancelflags = 0;
 495
 496
 497         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 498
 499         dqp->dq_flags = type;
 500         dqp->q_core.d_id = cpu_to_be32(id);
 501         dqp->q_mount = mp;
 502         INIT_LIST_HEAD(&dqp->q_lru);
 503         mutex_init(&dqp->q_qlock);
 504         init_waitqueue_head(&dqp->q_pinwait);
 505
 506         /*
 507          * Because we want to use a counting completion, complete
 508          * the flush completion once to allow a single access to
 509          * the flush completion without blocking.
 510          */
 511         init_completion(&dqp->q_flush);
 512         complete(&dqp->q_flush);
 513
 514         /*
 515          * Make sure group quotas have a different lock class than user
 516          * quotas.
 517          */
 518         if (!(type & XFS_DQ_USER))
 519                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 520
 521         XFS_STATS_INC(xs_qm_dquot);
 522
 523         trace_xfs_dqread(dqp);
 524
 525         if (flags & XFS_QMOPT_DQALLOC) {
 526                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 527                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
 528                                 XFS_WRITE_LOG_RES(mp) +
 529                                 /*
 530                                  * Round the chunklen up to the next multiple
 531                                  * of 128 (buf log item chunk size)).
 532                                  */
 533                                 BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 + 128,
 534                                 0,
 535                                 XFS_TRANS_PERM_LOG_RES,
 536                                 XFS_WRITE_LOG_COUNT);
 537                 if (error)
 538                         goto error1;
 539                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 540         }
 541
 542         /*
 543          * get a pointer to the on-disk dquot and the buffer containing it
 544          * dqp already knows its own type (GROUP/USER).
 545          */
 546         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 547         if (error) {
 548                 /*
 549                  * This can happen if quotas got turned off (ESRCH),
 550                  * or if the dquot didn't exist on disk and we ask to
 551                  * allocate (ENOENT).
 552                  */
 553                 trace_xfs_dqread_fail(dqp);
 554                 cancelflags |= XFS_TRANS_ABORT;
 555                 goto error1;
 556         }
 557
 558         /* copy everything from disk dquot to the incore dquot */
 559         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 560         xfs_qm_dquot_logitem_init(dqp);
 561
 562         /*
 563          * Reservation counters are defined as reservation plus current usage
 564          * to avoid having to add every time.
 565          */
 566         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 567         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 568         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 569
 570         /* Mark the buf so that this will stay incore a little longer */
 571         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 572
 573         /*
 574          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 575          * So we need to release with xfs_trans_brelse().
 576          * The strategy here is identical to that of inodes; we lock
 577          * the dquot in xfs_qm_dqget() before making it accessible to
 578          * others. This is because dquots, like inodes, need a good level of
 579          * concurrency, and we don't want to take locks on the entire buffers
 580          * for dquot accesses.
 581          * Note also that the dquot buffer may even be dirty at this point, if
 582          * this particular dquot was repaired. We still aren't afraid to
 583          * brelse it because we have the changes incore.
 584          */
 585         ASSERT(xfs_buf_islocked(bp));
 586         xfs_trans_brelse(tp, bp);
 587
 588         if (tp) {
 589                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 590                 if (error)
 591                         goto error0;
 592         }
 593
 594         *O_dqpp = dqp;
 595         return error;
 596
 597 error1:
 598         if (tp)
 599                 xfs_trans_cancel(tp, cancelflags);
 600 error0:
 601         xfs_qm_dqdestroy(dqp);
 602         *O_dqpp = NULL;
 603         return error;
 604 }
 605
 606 /*
 607  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 608  * a locked dquot, doing an allocation (if requested) as needed.
 609  * When both an inode and an id are given, the inode's id takes precedence.
 610  * That is, if the id changes while we don't hold the ilock inside this
 611  * function, the new dquot is returned, not necessarily the one requested
 612  * in the id argument.
 613  */
 614 int
 615 xfs_qm_dqget(
 616         xfs_mount_t     *mp,
 617         xfs_inode_t     *ip,      /* locked inode (optional) */
 618         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 619         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 620         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 621         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 622 {
 623         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 624         struct radix_tree_root *tree = XFS_DQUOT_TREE(qi, type);
 625         struct xfs_dquot        *dqp;
 626         int                     error;
 627
 628         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 629         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 630             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 631             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 632                 return (ESRCH);
 633         }
 634
 635 #ifdef DEBUG
 636         if (xfs_do_dqerror) {
 637                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 638                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 639                         xfs_debug(mp, "Returning error in dqget");
 640                         return (EIO);
 641                 }
 642         }
 643
 644         ASSERT(type == XFS_DQ_USER ||
 645                type == XFS_DQ_PROJ ||
 646                type == XFS_DQ_GROUP);
 647         if (ip) {
 648                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 649                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 650         }
 651 #endif
 652
 653 restart:
 654         mutex_lock(&qi->qi_tree_lock);
 655         dqp = radix_tree_lookup(tree, id);
 656         if (dqp) {
 657                 xfs_dqlock(dqp);
 658                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 659                         xfs_dqunlock(dqp);
 660                         mutex_unlock(&qi->qi_tree_lock);
 661                         trace_xfs_dqget_freeing(dqp);
 662                         delay(1);
 663                         goto restart;
 664                 }
 665
 666                 dqp->q_nrefs++;
 667                 mutex_unlock(&qi->qi_tree_lock);
 668
 669                 trace_xfs_dqget_hit(dqp);
 670                 XFS_STATS_INC(xs_qm_dqcachehits);
 671                 *O_dqpp = dqp;
 672                 return 0;
 673         }
 674         mutex_unlock(&qi->qi_tree_lock);
 675         XFS_STATS_INC(xs_qm_dqcachemisses);
 676
 677         /*
 678          * Dquot cache miss. We don't want to keep the inode lock across
 679          * a (potential) disk read. Also we don't want to deal with the lock
 680          * ordering between quotainode and this inode. OTOH, dropping the inode
 681          * lock here means dealing with a chown that can happen before
 682          * we re-acquire the lock.
 683          */
 684         if (ip)
 685                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 686
 687         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 688
 689         if (ip)
 690                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 691
 692         if (error)
 693                 return error;
 694
 695         if (ip) {
 696                 /*
 697                  * A dquot could be attached to this inode by now, since
 698                  * we had dropped the ilock.
 699                  */
 700                 if (xfs_this_quota_on(mp, type)) {
 701                         struct xfs_dquot        *dqp1;
 702
 703                         dqp1 = xfs_inode_dquot(ip, type);
 704                         if (dqp1) {
 705                                 xfs_qm_dqdestroy(dqp);
 706                                 dqp = dqp1;
 707                                 xfs_dqlock(dqp);
 708                                 goto dqret;
 709                         }
 710                 } else {
 711                         /* inode stays locked on return */
 712                         xfs_qm_dqdestroy(dqp);
 713                         return XFS_ERROR(ESRCH);
 714                 }
 715         }
 716
 717         mutex_lock(&qi->qi_tree_lock);
 718         error = -radix_tree_insert(tree, id, dqp);
 719         if (unlikely(error)) {
 720                 WARN_ON(error != EEXIST);
 721
 722                 /*
 723                  * Duplicate found. Just throw away the new dquot and start
 724                  * over.
 725                  */
 726                 mutex_unlock(&qi->qi_tree_lock);
 727                 trace_xfs_dqget_dup(dqp);
 728                 xfs_qm_dqdestroy(dqp);
 729                 XFS_STATS_INC(xs_qm_dquot_dups);
 730                 goto restart;
 731         }
 732
 733         /*
 734          * We return a locked dquot to the caller, with a reference taken
 735          */
 736         xfs_dqlock(dqp);
 737         dqp->q_nrefs = 1;
 738
 739         qi->qi_dquots++;
 740         mutex_unlock(&qi->qi_tree_lock);
 741
 742  dqret:
 743         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 744         trace_xfs_dqget_miss(dqp);
 745         *O_dqpp = dqp;
 746         return (0);
 747 }
 748
 749
 750 STATIC void
 751 xfs_qm_dqput_final(
 752         struct xfs_dquot        *dqp)
 753 {
 754         struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 755         struct xfs_dquot        *gdqp;
 756
 757         trace_xfs_dqput_free(dqp);
 758
 759         mutex_lock(&qi->qi_lru_lock);
 760         if (list_empty(&dqp->q_lru)) {
 761                 list_add_tail(&dqp->q_lru, &qi->qi_lru_list);
 762                 qi->qi_lru_count++;
 763                 XFS_STATS_INC(xs_qm_dquot_unused);
 764         }
 765         mutex_unlock(&qi->qi_lru_lock);
 766
 767         /*
 768          * If we just added a udquot to the freelist, then we want to release
 769          * the gdquot reference that it (probably) has. Otherwise it'll keep
 770          * the gdquot from getting reclaimed.
 771          */
 772         gdqp = dqp->q_gdquot;
 773         if (gdqp) {
 774                 xfs_dqlock(gdqp);
 775                 dqp->q_gdquot = NULL;
 776         }
 777         xfs_dqunlock(dqp);
 778
 779         /*
 780          * If we had a group quota hint, release it now.
 781          */
 782         if (gdqp)
 783                 xfs_qm_dqput(gdqp);
 784 }
 785
 786 /*
 787  * Release a reference to the dquot (decrement ref-count) and unlock it.
 788  *
 789  * If there is a group quota attached to this dquot, carefully release that
 790  * too without tripping over deadlocks'n'stuff.
 791  */
 792 void
 793 xfs_qm_dqput(
 794         struct xfs_dquot        *dqp)
 795 {
 796         ASSERT(dqp->q_nrefs > 0);
 797         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 798
 799         trace_xfs_dqput(dqp);
 800
 801         if (--dqp->q_nrefs > 0)
 802                 xfs_dqunlock(dqp);
 803         else
 804                 xfs_qm_dqput_final(dqp);
 805 }
 806
 807 /*
 808  * Release a dquot. Flush it if dirty, then dqput() it.
 809  * dquot must not be locked.
 810  */
 811 void
 812 xfs_qm_dqrele(
 813         xfs_dquot_t     *dqp)
 814 {
 815         if (!dqp)
 816                 return;
 817
 818         trace_xfs_dqrele(dqp);
 819
 820         xfs_dqlock(dqp);
 821         /*
 822          * We don't care to flush it if the dquot is dirty here.
 823          * That will create stutters that we want to avoid.
 824          * Instead we do a delayed write when we try to reclaim
 825          * a dirty dquot. Also xfs_sync will take part of the burden...
 826          */
 827         xfs_qm_dqput(dqp);
 828 }
 829
 830 /*
 831  * This is the dquot flushing I/O completion routine.  It is called
 832  * from interrupt level when the buffer containing the dquot is
 833  * flushed to disk.  It is responsible for removing the dquot logitem
 834  * from the AIL if it has not been re-logged, and unlocking the dquot's
 835  * flush lock. This behavior is very similar to that of inodes..
 836  */
 837 STATIC void
 838 xfs_qm_dqflush_done(
 839         struct xfs_buf          *bp,
 840         struct xfs_log_item     *lip)
 841 {
 842         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 843         xfs_dquot_t             *dqp = qip->qli_dquot;
 844         struct xfs_ail          *ailp = lip->li_ailp;
 845
 846         /*
 847          * We only want to pull the item from the AIL if its
 848          * location in the log has not changed since we started the flush.
 849          * Thus, we only bother if the dquot's lsn has
 850          * not changed. First we check the lsn outside the lock
 851          * since it's cheaper, and then we recheck while
 852          * holding the lock before removing the dquot from the AIL.
 853          */
 854         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 855             lip->li_lsn == qip->qli_flush_lsn) {
 856
 857                 /* xfs_trans_ail_delete() drops the AIL lock. */
 858                 spin_lock(&ailp->xa_lock);
 859                 if (lip->li_lsn == qip->qli_flush_lsn)
 860                         xfs_trans_ail_delete(ailp, lip);
 861                 else
 862                         spin_unlock(&ailp->xa_lock);
 863         }
 864
 865         /*
 866          * Release the dq's flush lock since we're done with it.
 867          */
 868         xfs_dqfunlock(dqp);
 869 }
 870
 871 /*
 872  * Write a modified dquot to disk.
 873  * The dquot must be locked and the flush lock too taken by caller.
 874  * The flush lock will not be unlocked until the dquot reaches the disk,
 875  * but the dquot is free to be unlocked and modified by the caller
 876  * in the interim. Dquot is still locked on return. This behavior is
 877  * identical to that of inodes.
 878  */
 879 int
 880 xfs_qm_dqflush(
 881         xfs_dquot_t             *dqp,
 882         uint                    flags)
 883 {
 884         struct xfs_mount        *mp = dqp->q_mount;
 885         struct xfs_buf          *bp;
 886         struct xfs_disk_dquot   *ddqp;
 887         int                     error;
 888
 889         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 890         ASSERT(!completion_done(&dqp->q_flush));
 891
 892         trace_xfs_dqflush(dqp);
 893
 894         /*
 895          * If not dirty, or it's pinned and we are not supposed to block, nada.
 896          */
 897         if (!XFS_DQ_IS_DIRTY(dqp) ||
 898             ((flags & SYNC_TRYLOCK) && atomic_read(&dqp->q_pincount) > 0)) {
 899                 xfs_dqfunlock(dqp);
 900                 return 0;
 901         }
 902         xfs_qm_dqunpin_wait(dqp);
 903
 904         /*
 905          * This may have been unpinned because the filesystem is shutting
 906          * down forcibly. If that's the case we must not write this dquot
 907          * to disk, because the log record didn't make it to disk!
 908          */
 909         if (XFS_FORCED_SHUTDOWN(mp)) {
 910                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
 911                 xfs_dqfunlock(dqp);
 912                 return XFS_ERROR(EIO);
 913         }
 914
 915         /*
 916          * Get the buffer containing the on-disk dquot
 917          */
 918         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
 919                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp);
 920         if (error) {
 921                 ASSERT(error != ENOENT);
 922                 xfs_dqfunlock(dqp);
 923                 return error;
 924         }
 925
 926         /*
 927          * Calculate the location of the dquot inside the buffer.
 928          */
 929         ddqp = bp->b_addr + dqp->q_bufoffset;
 930
 931         /*
 932          * A simple sanity check in case we got a corrupted dquot..
 933          */
 934         error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
 935                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
 936         if (error) {
 937                 xfs_buf_relse(bp);
 938                 xfs_dqfunlock(dqp);
 939                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 940                 return XFS_ERROR(EIO);
 941         }
 942
 943         /* This is the only portion of data that needs to persist */
 944         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
 945
 946         /*
 947          * Clear the dirty field and remember the flush lsn for later use.
 948          */
 949         dqp->dq_flags &= ~XFS_DQ_DIRTY;
 950
 951         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
 952                                         &dqp->q_logitem.qli_item.li_lsn);
 953
 954         /*
 955          * Attach an iodone routine so that we can remove this dquot from the
 956          * AIL and release the flush lock once the dquot is synced to disk.
 957          */
 958         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
 959                                   &dqp->q_logitem.qli_item);
 960
 961         /*
 962          * If the buffer is pinned then push on the log so we won't
 963          * get stuck waiting in the write for too long.
 964          */
 965         if (xfs_buf_ispinned(bp)) {
 966                 trace_xfs_dqflush_force(dqp);
 967                 xfs_log_force(mp, 0);
 968         }
 969
 970         if (flags & SYNC_WAIT)
 971                 error = xfs_bwrite(bp);
 972         else
 973                 xfs_buf_delwri_queue(bp);
 974
 975         xfs_buf_relse(bp);
 976
 977         trace_xfs_dqflush_done(dqp);
 978
 979         /*
 980          * dqp is still locked, but caller is free to unlock it now.
 981          */
 982         return error;
 983
 984 }
 985
 986 /*
 987  * Lock two xfs_dquot structures.
 988  *
 989  * To avoid deadlocks we always lock the quota structure with
 990  * the lowerd id first.
 991  */
 992 void
 993 xfs_dqlock2(
 994         xfs_dquot_t     *d1,
 995         xfs_dquot_t     *d2)
 996 {
 997         if (d1 && d2) {
 998                 ASSERT(d1 != d2);
 999                 if (be32_to_cpu(d1->q_core.d_id) >
1000                     be32_to_cpu(d2->q_core.d_id)) {
1001                         mutex_lock(&d2->q_qlock);
1002                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1003                 } else {
1004                         mutex_lock(&d1->q_qlock);
1005                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1006                 }
1007         } else if (d1) {
1008                 mutex_lock(&d1->q_qlock);
1009         } else if (d2) {
1010                 mutex_lock(&d2->q_qlock);
1011         }
1012 }
1013
1014 /*
1015  * Give the buffer a little push if it is incore and
1016  * wait on the flush lock.
1017  */
1018 void
1019 xfs_dqflock_pushbuf_wait(
1020         xfs_dquot_t     *dqp)
1021 {
1022         xfs_mount_t     *mp = dqp->q_mount;
1023         xfs_buf_t       *bp;
1024
1025         /*
1026          * Check to see if the dquot has been flushed delayed
1027          * write.  If so, grab its buffer and send it
1028          * out immediately.  We'll be able to acquire
1029          * the flush lock when the I/O completes.
1030          */
1031         bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1032                         mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1033         if (!bp)
1034                 goto out_lock;
1035
1036         if (XFS_BUF_ISDELAYWRITE(bp)) {
1037                 if (xfs_buf_ispinned(bp))
1038                         xfs_log_force(mp, 0);
1039                 xfs_buf_delwri_promote(bp);
1040                 wake_up_process(bp->b_target->bt_task);
1041         }
1042         xfs_buf_relse(bp);
1043 out_lock:
1044         xfs_dqflock(dqp);
1045 }
1046
1047 int __init
1048 xfs_qm_init(void)
1049 {
1050         xfs_qm_dqzone =
1051                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1052         if (!xfs_qm_dqzone)
1053                 goto out;
1054
1055         xfs_qm_dqtrxzone =
1056                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1057         if (!xfs_qm_dqtrxzone)
1058                 goto out_free_dqzone;
1059
1060         return 0;
1061
1062 out_free_dqzone:
1063         kmem_zone_destroy(xfs_qm_dqzone);
1064 out:
1065         return -ENOMEM;
1066 }
1067
1068 void
1069 xfs_qm_exit(void)
1070 {
1071         kmem_zone_destroy(xfs_qm_dqtrxzone);
1072         kmem_zone_destroy(xfs_qm_dqzone);
1073 }