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_trans.h"
  23 #include "xfs_sb.h"
  24 #include "xfs_ag.h"
  25 #include "xfs_alloc.h"
  26 #include "xfs_quota.h"
  27 #include "xfs_mount.h"
  28 #include "xfs_bmap_btree.h"
  29 #include "xfs_inode.h"
  30 #include "xfs_bmap.h"
  31 #include "xfs_rtalloc.h"
  32 #include "xfs_error.h"
  33 #include "xfs_itable.h"
  34 #include "xfs_attr.h"
  35 #include "xfs_buf_item.h"
  36 #include "xfs_trans_space.h"
  37 #include "xfs_trans_priv.h"
  38 #include "xfs_qm.h"
  39 #include "xfs_trace.h"
  40
  41 /*
  42  * Lock order:
  43  *
  44  * ip->i_lock
  45  *   qi->qi_tree_lock
  46  *     dquot->q_qlock (xfs_dqlock() and friends)
  47  *       dquot->q_flush (xfs_dqflock() and friends)
  48  *       qi->qi_lru_lock
  49  *
  50  * If two dquots need to be locked the order is user before group/project,
  51  * otherwise by the lowest id first, see xfs_dqlock2.
  52  */
  53
  54 #ifdef DEBUG
  55 xfs_buftarg_t *xfs_dqerror_target;
  56 int xfs_do_dqerror;
  57 int xfs_dqreq_num;
  58 int xfs_dqerror_mod = 33;
  59 #endif
  60
  61 struct kmem_zone                *xfs_qm_dqtrxzone;
  62 static struct kmem_zone         *xfs_qm_dqzone;
  63
  64 static struct lock_class_key xfs_dquot_other_class;
  65
  66 /*
  67  * This is called to free all the memory associated with a dquot
  68  */
  69 void
  70 xfs_qm_dqdestroy(
  71         xfs_dquot_t     *dqp)
  72 {
  73         ASSERT(list_empty(&dqp->q_lru));
  74
  75         mutex_destroy(&dqp->q_qlock);
  76         kmem_zone_free(xfs_qm_dqzone, dqp);
  77
  78         XFS_STATS_DEC(xs_qm_dquot);
  79 }
  80
  81 /*
  82  * If default limits are in force, push them into the dquot now.
  83  * We overwrite the dquot limits only if they are zero and this
  84  * is not the root dquot.
  85  */
  86 void
  87 xfs_qm_adjust_dqlimits(
  88         struct xfs_mount        *mp,
  89         struct xfs_dquot        *dq)
  90 {
  91         struct xfs_quotainfo    *q = mp->m_quotainfo;
  92         struct xfs_disk_dquot   *d = &dq->q_core;
  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 static void
 253 xfs_dquot_buf_verify(
 254         struct xfs_buf          *bp)
 255 {
 256         struct xfs_mount        *mp = bp->b_target->bt_mount;
 257         struct xfs_dqblk        *d = (struct xfs_dqblk *)bp->b_addr;
 258         struct xfs_disk_dquot   *ddq;
 259         xfs_dqid_t              id = 0;
 260         int                     i;
 261
 262         /*
 263          * On the first read of the buffer, verify that each dquot is valid.
 264          * We don't know what the id of the dquot is supposed to be, just that
 265          * they should be increasing monotonically within the buffer. If the
 266          * first id is corrupt, then it will fail on the second dquot in the
 267          * buffer so corruptions could point to the wrong dquot in this case.
 268          */
 269         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 270                 int     error;
 271
 272                 ddq = &d[i].dd_diskdq;
 273
 274                 if (i == 0)
 275                         id = be32_to_cpu(ddq->d_id);
 276
 277                 error = xfs_qm_dqcheck(mp, ddq, id + i, 0, XFS_QMOPT_DOWARN,
 278                                         "xfs_dquot_read_verify");
 279                 if (error) {
 280                         XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, d);
 281                         xfs_buf_ioerror(bp, EFSCORRUPTED);
 282                         break;
 283                 }
 284         }
 285 }
 286
 287 static void
 288 xfs_dquot_buf_read_verify(
 289         struct xfs_buf  *bp)
 290 {
 291         xfs_dquot_buf_verify(bp);
 292 }
 293
 294 void
 295 xfs_dquot_buf_write_verify(
 296         struct xfs_buf  *bp)
 297 {
 298         xfs_dquot_buf_verify(bp);
 299 }
 300
 301 const struct xfs_buf_ops xfs_dquot_buf_ops = {
 302         .verify_read = xfs_dquot_buf_read_verify,
 303         .verify_write = xfs_dquot_buf_write_verify,
 304 };
 305
 306 /*
 307  * Allocate a block and fill it with dquots.
 308  * This is called when the bmapi finds a hole.
 309  */
 310 STATIC int
 311 xfs_qm_dqalloc(
 312         xfs_trans_t     **tpp,
 313         xfs_mount_t     *mp,
 314         xfs_dquot_t     *dqp,
 315         xfs_inode_t     *quotip,
 316         xfs_fileoff_t   offset_fsb,
 317         xfs_buf_t       **O_bpp)
 318 {
 319         xfs_fsblock_t   firstblock;
 320         xfs_bmap_free_t flist;
 321         xfs_bmbt_irec_t map;
 322         int             nmaps, error, committed;
 323         xfs_buf_t       *bp;
 324         xfs_trans_t     *tp = *tpp;
 325
 326         ASSERT(tp != NULL);
 327
 328         trace_xfs_dqalloc(dqp);
 329
 330         /*
 331          * Initialize the bmap freelist prior to calling bmapi code.
 332          */
 333         xfs_bmap_init(&flist, &firstblock);
 334         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 335         /*
 336          * Return if this type of quotas is turned off while we didn't
 337          * have an inode lock
 338          */
 339         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 340                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 341                 return (ESRCH);
 342         }
 343
 344         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 345         nmaps = 1;
 346         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 347                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 348                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 349                                 &map, &nmaps, &flist);
 350         if (error)
 351                 goto error0;
 352         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 353         ASSERT(nmaps == 1);
 354         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 355                (map.br_startblock != HOLESTARTBLOCK));
 356
 357         /*
 358          * Keep track of the blkno to save a lookup later
 359          */
 360         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 361
 362         /* now we can just get the buffer (there's nothing to read yet) */
 363         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 364                                dqp->q_blkno,
 365                                mp->m_quotainfo->qi_dqchunklen,
 366                                0);
 367
 368         error = xfs_buf_geterror(bp);
 369         if (error)
 370                 goto error1;
 371         bp->b_ops = &xfs_dquot_buf_ops;
 372
 373         /*
 374          * Make a chunk of dquots out of this buffer and log
 375          * the entire thing.
 376          */
 377         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 378                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 379
 380         /*
 381          * xfs_bmap_finish() may commit the current transaction and
 382          * start a second transaction if the freelist is not empty.
 383          *
 384          * Since we still want to modify this buffer, we need to
 385          * ensure that the buffer is not released on commit of
 386          * the first transaction and ensure the buffer is added to the
 387          * second transaction.
 388          *
 389          * If there is only one transaction then don't stop the buffer
 390          * from being released when it commits later on.
 391          */
 392
 393         xfs_trans_bhold(tp, bp);
 394
 395         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 396                 goto error1;
 397         }
 398
 399         if (committed) {
 400                 tp = *tpp;
 401                 xfs_trans_bjoin(tp, bp);
 402         } else {
 403                 xfs_trans_bhold_release(tp, bp);
 404         }
 405
 406         *O_bpp = bp;
 407         return 0;
 408
 409       error1:
 410         xfs_bmap_cancel(&flist);
 411       error0:
 412         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 413
 414         return (error);
 415 }
 416 STATIC int
 417 xfs_qm_dqrepair(
 418         struct xfs_mount        *mp,
 419         struct xfs_trans        *tp,
 420         struct xfs_dquot        *dqp,
 421         xfs_dqid_t              firstid,
 422         struct xfs_buf          **bpp)
 423 {
 424         int                     error;
 425         struct xfs_disk_dquot   *ddq;
 426         struct xfs_dqblk        *d;
 427         int                     i;
 428
 429         /*
 430          * Read the buffer without verification so we get the corrupted
 431          * buffer returned to us. make sure we verify it on write, though.
 432          */
 433         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 434                                    mp->m_quotainfo->qi_dqchunklen,
 435                                    0, bpp, NULL);
 436
 437         if (error) {
 438                 ASSERT(*bpp == NULL);
 439                 return XFS_ERROR(error);
 440         }
 441         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 442
 443         ASSERT(xfs_buf_islocked(*bpp));
 444         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 445
 446         /* Do the actual repair of dquots in this buffer */
 447         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 448                 ddq = &d[i].dd_diskdq;
 449                 error = xfs_qm_dqcheck(mp, ddq, firstid + i,
 450                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 451                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 452                 if (error) {
 453                         /* repair failed, we're screwed */
 454                         xfs_trans_brelse(tp, *bpp);
 455                         return XFS_ERROR(EIO);
 456                 }
 457         }
 458
 459         return 0;
 460 }
 461
 462 /*
 463  * Maps a dquot to the buffer containing its on-disk version.
 464  * This returns a ptr to the buffer containing the on-disk dquot
 465  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 466  */
 467 STATIC int
 468 xfs_qm_dqtobp(
 469         xfs_trans_t             **tpp,
 470         xfs_dquot_t             *dqp,
 471         xfs_disk_dquot_t        **O_ddpp,
 472         xfs_buf_t               **O_bpp,
 473         uint                    flags)
 474 {
 475         xfs_bmbt_irec_t map;
 476         int             nmaps = 1, error;
 477         xfs_buf_t       *bp;
 478         xfs_inode_t     *quotip = XFS_DQ_TO_QIP(dqp);
 479         xfs_mount_t     *mp = dqp->q_mount;
 480         xfs_dqid_t      id = be32_to_cpu(dqp->q_core.d_id);
 481         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 482
 483         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 484
 485         xfs_ilock(quotip, XFS_ILOCK_SHARED);
 486         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 487                 /*
 488                  * Return if this type of quotas is turned off while we
 489                  * didn't have the quota inode lock.
 490                  */
 491                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 492                 return ESRCH;
 493         }
 494
 495         /*
 496          * Find the block map; no allocations yet
 497          */
 498         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 499                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 500
 501         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 502         if (error)
 503                 return error;
 504
 505         ASSERT(nmaps == 1);
 506         ASSERT(map.br_blockcount == 1);
 507
 508         /*
 509          * Offset of dquot in the (fixed sized) dquot chunk.
 510          */
 511         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 512                 sizeof(xfs_dqblk_t);
 513
 514         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 515         if (map.br_startblock == HOLESTARTBLOCK) {
 516                 /*
 517                  * We don't allocate unless we're asked to
 518                  */
 519                 if (!(flags & XFS_QMOPT_DQALLOC))
 520                         return ENOENT;
 521
 522                 ASSERT(tp);
 523                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 524                                         dqp->q_fileoffset, &bp);
 525                 if (error)
 526                         return error;
 527                 tp = *tpp;
 528         } else {
 529                 trace_xfs_dqtobp_read(dqp);
 530
 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                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 538                                            dqp->q_blkno,
 539                                            mp->m_quotainfo->qi_dqchunklen,
 540                                            0, &bp, &xfs_dquot_buf_ops);
 541
 542                 if (error == EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 543                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 544                                                 mp->m_quotainfo->qi_dqperchunk;
 545                         ASSERT(bp == NULL);
 546                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 547                 }
 548
 549                 if (error) {
 550                         ASSERT(bp == NULL);
 551                         return XFS_ERROR(error);
 552                 }
 553         }
 554
 555         ASSERT(xfs_buf_islocked(bp));
 556         *O_bpp = bp;
 557         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 558
 559         return (0);
 560 }
 561
 562
 563 /*
 564  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 565  * and release the buffer immediately.
 566  *
 567  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 568  */
 569 int
 570 xfs_qm_dqread(
 571         struct xfs_mount        *mp,
 572         xfs_dqid_t              id,
 573         uint                    type,
 574         uint                    flags,
 575         struct xfs_dquot        **O_dqpp)
 576 {
 577         struct xfs_dquot        *dqp;
 578         struct xfs_disk_dquot   *ddqp;
 579         struct xfs_buf          *bp;
 580         struct xfs_trans        *tp = NULL;
 581         int                     error;
 582         int                     cancelflags = 0;
 583
 584
 585         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 586
 587         dqp->dq_flags = type;
 588         dqp->q_core.d_id = cpu_to_be32(id);
 589         dqp->q_mount = mp;
 590         INIT_LIST_HEAD(&dqp->q_lru);
 591         mutex_init(&dqp->q_qlock);
 592         init_waitqueue_head(&dqp->q_pinwait);
 593
 594         /*
 595          * Because we want to use a counting completion, complete
 596          * the flush completion once to allow a single access to
 597          * the flush completion without blocking.
 598          */
 599         init_completion(&dqp->q_flush);
 600         complete(&dqp->q_flush);
 601
 602         /*
 603          * Make sure group quotas have a different lock class than user
 604          * quotas.
 605          */
 606         if (!(type & XFS_DQ_USER))
 607                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 608
 609         XFS_STATS_INC(xs_qm_dquot);
 610
 611         trace_xfs_dqread(dqp);
 612
 613         if (flags & XFS_QMOPT_DQALLOC) {
 614                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 615                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
 616                                           XFS_QM_DQALLOC_LOG_RES(mp), 0,
 617                                           XFS_TRANS_PERM_LOG_RES,
 618                                           XFS_WRITE_LOG_COUNT);
 619                 if (error)
 620                         goto error1;
 621                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 622         }
 623
 624         /*
 625          * get a pointer to the on-disk dquot and the buffer containing it
 626          * dqp already knows its own type (GROUP/USER).
 627          */
 628         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 629         if (error) {
 630                 /*
 631                  * This can happen if quotas got turned off (ESRCH),
 632                  * or if the dquot didn't exist on disk and we ask to
 633                  * allocate (ENOENT).
 634                  */
 635                 trace_xfs_dqread_fail(dqp);
 636                 cancelflags |= XFS_TRANS_ABORT;
 637                 goto error1;
 638         }
 639
 640         /* copy everything from disk dquot to the incore dquot */
 641         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 642         xfs_qm_dquot_logitem_init(dqp);
 643
 644         /*
 645          * Reservation counters are defined as reservation plus current usage
 646          * to avoid having to add every time.
 647          */
 648         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 649         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 650         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 651
 652         /* Mark the buf so that this will stay incore a little longer */
 653         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 654
 655         /*
 656          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 657          * So we need to release with xfs_trans_brelse().
 658          * The strategy here is identical to that of inodes; we lock
 659          * the dquot in xfs_qm_dqget() before making it accessible to
 660          * others. This is because dquots, like inodes, need a good level of
 661          * concurrency, and we don't want to take locks on the entire buffers
 662          * for dquot accesses.
 663          * Note also that the dquot buffer may even be dirty at this point, if
 664          * this particular dquot was repaired. We still aren't afraid to
 665          * brelse it because we have the changes incore.
 666          */
 667         ASSERT(xfs_buf_islocked(bp));
 668         xfs_trans_brelse(tp, bp);
 669
 670         if (tp) {
 671                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 672                 if (error)
 673                         goto error0;
 674         }
 675
 676         *O_dqpp = dqp;
 677         return error;
 678
 679 error1:
 680         if (tp)
 681                 xfs_trans_cancel(tp, cancelflags);
 682 error0:
 683         xfs_qm_dqdestroy(dqp);
 684         *O_dqpp = NULL;
 685         return error;
 686 }
 687
 688 /*
 689  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 690  * a locked dquot, doing an allocation (if requested) as needed.
 691  * When both an inode and an id are given, the inode's id takes precedence.
 692  * That is, if the id changes while we don't hold the ilock inside this
 693  * function, the new dquot is returned, not necessarily the one requested
 694  * in the id argument.
 695  */
 696 int
 697 xfs_qm_dqget(
 698         xfs_mount_t     *mp,
 699         xfs_inode_t     *ip,      /* locked inode (optional) */
 700         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 701         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 702         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 703         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 704 {
 705         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 706         struct radix_tree_root *tree = XFS_DQUOT_TREE(qi, type);
 707         struct xfs_dquot        *dqp;
 708         int                     error;
 709
 710         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 711         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 712             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 713             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 714                 return (ESRCH);
 715         }
 716
 717 #ifdef DEBUG
 718         if (xfs_do_dqerror) {
 719                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 720                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 721                         xfs_debug(mp, "Returning error in dqget");
 722                         return (EIO);
 723                 }
 724         }
 725
 726         ASSERT(type == XFS_DQ_USER ||
 727                type == XFS_DQ_PROJ ||
 728                type == XFS_DQ_GROUP);
 729         if (ip) {
 730                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 731                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 732         }
 733 #endif
 734
 735 restart:
 736         mutex_lock(&qi->qi_tree_lock);
 737         dqp = radix_tree_lookup(tree, id);
 738         if (dqp) {
 739                 xfs_dqlock(dqp);
 740                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 741                         xfs_dqunlock(dqp);
 742                         mutex_unlock(&qi->qi_tree_lock);
 743                         trace_xfs_dqget_freeing(dqp);
 744                         delay(1);
 745                         goto restart;
 746                 }
 747
 748                 dqp->q_nrefs++;
 749                 mutex_unlock(&qi->qi_tree_lock);
 750
 751                 trace_xfs_dqget_hit(dqp);
 752                 XFS_STATS_INC(xs_qm_dqcachehits);
 753                 *O_dqpp = dqp;
 754                 return 0;
 755         }
 756         mutex_unlock(&qi->qi_tree_lock);
 757         XFS_STATS_INC(xs_qm_dqcachemisses);
 758
 759         /*
 760          * Dquot cache miss. We don't want to keep the inode lock across
 761          * a (potential) disk read. Also we don't want to deal with the lock
 762          * ordering between quotainode and this inode. OTOH, dropping the inode
 763          * lock here means dealing with a chown that can happen before
 764          * we re-acquire the lock.
 765          */
 766         if (ip)
 767                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 768
 769         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 770
 771         if (ip)
 772                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 773
 774         if (error)
 775                 return error;
 776
 777         if (ip) {
 778                 /*
 779                  * A dquot could be attached to this inode by now, since
 780                  * we had dropped the ilock.
 781                  */
 782                 if (xfs_this_quota_on(mp, type)) {
 783                         struct xfs_dquot        *dqp1;
 784
 785                         dqp1 = xfs_inode_dquot(ip, type);
 786                         if (dqp1) {
 787                                 xfs_qm_dqdestroy(dqp);
 788                                 dqp = dqp1;
 789                                 xfs_dqlock(dqp);
 790                                 goto dqret;
 791                         }
 792                 } else {
 793                         /* inode stays locked on return */
 794                         xfs_qm_dqdestroy(dqp);
 795                         return XFS_ERROR(ESRCH);
 796                 }
 797         }
 798
 799         mutex_lock(&qi->qi_tree_lock);
 800         error = -radix_tree_insert(tree, id, dqp);
 801         if (unlikely(error)) {
 802                 WARN_ON(error != EEXIST);
 803
 804                 /*
 805                  * Duplicate found. Just throw away the new dquot and start
 806                  * over.
 807                  */
 808                 mutex_unlock(&qi->qi_tree_lock);
 809                 trace_xfs_dqget_dup(dqp);
 810                 xfs_qm_dqdestroy(dqp);
 811                 XFS_STATS_INC(xs_qm_dquot_dups);
 812                 goto restart;
 813         }
 814
 815         /*
 816          * We return a locked dquot to the caller, with a reference taken
 817          */
 818         xfs_dqlock(dqp);
 819         dqp->q_nrefs = 1;
 820
 821         qi->qi_dquots++;
 822         mutex_unlock(&qi->qi_tree_lock);
 823
 824  dqret:
 825         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 826         trace_xfs_dqget_miss(dqp);
 827         *O_dqpp = dqp;
 828         return (0);
 829 }
 830
 831
 832 STATIC void
 833 xfs_qm_dqput_final(
 834         struct xfs_dquot        *dqp)
 835 {
 836         struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 837         struct xfs_dquot        *gdqp;
 838
 839         trace_xfs_dqput_free(dqp);
 840
 841         mutex_lock(&qi->qi_lru_lock);
 842         if (list_empty(&dqp->q_lru)) {
 843                 list_add_tail(&dqp->q_lru, &qi->qi_lru_list);
 844                 qi->qi_lru_count++;
 845                 XFS_STATS_INC(xs_qm_dquot_unused);
 846         }
 847         mutex_unlock(&qi->qi_lru_lock);
 848
 849         /*
 850          * If we just added a udquot to the freelist, then we want to release
 851          * the gdquot reference that it (probably) has. Otherwise it'll keep
 852          * the gdquot from getting reclaimed.
 853          */
 854         gdqp = dqp->q_gdquot;
 855         if (gdqp) {
 856                 xfs_dqlock(gdqp);
 857                 dqp->q_gdquot = NULL;
 858         }
 859         xfs_dqunlock(dqp);
 860
 861         /*
 862          * If we had a group quota hint, release it now.
 863          */
 864         if (gdqp)
 865                 xfs_qm_dqput(gdqp);
 866 }
 867
 868 /*
 869  * Release a reference to the dquot (decrement ref-count) and unlock it.
 870  *
 871  * If there is a group quota attached to this dquot, carefully release that
 872  * too without tripping over deadlocks'n'stuff.
 873  */
 874 void
 875 xfs_qm_dqput(
 876         struct xfs_dquot        *dqp)
 877 {
 878         ASSERT(dqp->q_nrefs > 0);
 879         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 880
 881         trace_xfs_dqput(dqp);
 882
 883         if (--dqp->q_nrefs > 0)
 884                 xfs_dqunlock(dqp);
 885         else
 886                 xfs_qm_dqput_final(dqp);
 887 }
 888
 889 /*
 890  * Release a dquot. Flush it if dirty, then dqput() it.
 891  * dquot must not be locked.
 892  */
 893 void
 894 xfs_qm_dqrele(
 895         xfs_dquot_t     *dqp)
 896 {
 897         if (!dqp)
 898                 return;
 899
 900         trace_xfs_dqrele(dqp);
 901
 902         xfs_dqlock(dqp);
 903         /*
 904          * We don't care to flush it if the dquot is dirty here.
 905          * That will create stutters that we want to avoid.
 906          * Instead we do a delayed write when we try to reclaim
 907          * a dirty dquot. Also xfs_sync will take part of the burden...
 908          */
 909         xfs_qm_dqput(dqp);
 910 }
 911
 912 /*
 913  * This is the dquot flushing I/O completion routine.  It is called
 914  * from interrupt level when the buffer containing the dquot is
 915  * flushed to disk.  It is responsible for removing the dquot logitem
 916  * from the AIL if it has not been re-logged, and unlocking the dquot's
 917  * flush lock. This behavior is very similar to that of inodes..
 918  */
 919 STATIC void
 920 xfs_qm_dqflush_done(
 921         struct xfs_buf          *bp,
 922         struct xfs_log_item     *lip)
 923 {
 924         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 925         xfs_dquot_t             *dqp = qip->qli_dquot;
 926         struct xfs_ail          *ailp = lip->li_ailp;
 927
 928         /*
 929          * We only want to pull the item from the AIL if its
 930          * location in the log has not changed since we started the flush.
 931          * Thus, we only bother if the dquot's lsn has
 932          * not changed. First we check the lsn outside the lock
 933          * since it's cheaper, and then we recheck while
 934          * holding the lock before removing the dquot from the AIL.
 935          */
 936         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 937             lip->li_lsn == qip->qli_flush_lsn) {
 938
 939                 /* xfs_trans_ail_delete() drops the AIL lock. */
 940                 spin_lock(&ailp->xa_lock);
 941                 if (lip->li_lsn == qip->qli_flush_lsn)
 942                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
 943                 else
 944                         spin_unlock(&ailp->xa_lock);
 945         }
 946
 947         /*
 948          * Release the dq's flush lock since we're done with it.
 949          */
 950         xfs_dqfunlock(dqp);
 951 }
 952
 953 /*
 954  * Write a modified dquot to disk.
 955  * The dquot must be locked and the flush lock too taken by caller.
 956  * The flush lock will not be unlocked until the dquot reaches the disk,
 957  * but the dquot is free to be unlocked and modified by the caller
 958  * in the interim. Dquot is still locked on return. This behavior is
 959  * identical to that of inodes.
 960  */
 961 int
 962 xfs_qm_dqflush(
 963         struct xfs_dquot        *dqp,
 964         struct xfs_buf          **bpp)
 965 {
 966         struct xfs_mount        *mp = dqp->q_mount;
 967         struct xfs_buf          *bp;
 968         struct xfs_disk_dquot   *ddqp;
 969         int                     error;
 970
 971         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 972         ASSERT(!completion_done(&dqp->q_flush));
 973
 974         trace_xfs_dqflush(dqp);
 975
 976         *bpp = NULL;
 977
 978         xfs_qm_dqunpin_wait(dqp);
 979
 980         /*
 981          * This may have been unpinned because the filesystem is shutting
 982          * down forcibly. If that's the case we must not write this dquot
 983          * to disk, because the log record didn't make it to disk.
 984          *
 985          * We also have to remove the log item from the AIL in this case,
 986          * as we wait for an emptry AIL as part of the unmount process.
 987          */
 988         if (XFS_FORCED_SHUTDOWN(mp)) {
 989                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
 990                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
 991
 992                 spin_lock(&mp->m_ail->xa_lock);
 993                 if (lip->li_flags & XFS_LI_IN_AIL)
 994                         xfs_trans_ail_delete(mp->m_ail, lip,
 995                                              SHUTDOWN_CORRUPT_INCORE);
 996                 else
 997                         spin_unlock(&mp->m_ail->xa_lock);
 998                 error = XFS_ERROR(EIO);
 999                 goto out_unlock;
1000         }
1001
1002         /*
1003          * Get the buffer containing the on-disk dquot
1004          */
1005         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1006                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp, NULL);
1007         if (error)
1008                 goto out_unlock;
1009
1010         /*
1011          * Calculate the location of the dquot inside the buffer.
1012          */
1013         ddqp = bp->b_addr + dqp->q_bufoffset;
1014
1015         /*
1016          * A simple sanity check in case we got a corrupted dquot..
1017          */
1018         error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1019                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1020         if (error) {
1021                 xfs_buf_relse(bp);
1022                 xfs_dqfunlock(dqp);
1023                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1024                 return XFS_ERROR(EIO);
1025         }
1026
1027         /* This is the only portion of data that needs to persist */
1028         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1029
1030         /*
1031          * Clear the dirty field and remember the flush lsn for later use.
1032          */
1033         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1034
1035         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1036                                         &dqp->q_logitem.qli_item.li_lsn);
1037
1038         /*
1039          * Attach an iodone routine so that we can remove this dquot from the
1040          * AIL and release the flush lock once the dquot is synced to disk.
1041          */
1042         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1043                                   &dqp->q_logitem.qli_item);
1044
1045         /*
1046          * If the buffer is pinned then push on the log so we won't
1047          * get stuck waiting in the write for too long.
1048          */
1049         if (xfs_buf_ispinned(bp)) {
1050                 trace_xfs_dqflush_force(dqp);
1051                 xfs_log_force(mp, 0);
1052         }
1053
1054         trace_xfs_dqflush_done(dqp);
1055         *bpp = bp;
1056         return 0;
1057
1058 out_unlock:
1059         xfs_dqfunlock(dqp);
1060         return XFS_ERROR(EIO);
1061 }
1062
1063 /*
1064  * Lock two xfs_dquot structures.
1065  *
1066  * To avoid deadlocks we always lock the quota structure with
1067  * the lowerd id first.
1068  */
1069 void
1070 xfs_dqlock2(
1071         xfs_dquot_t     *d1,
1072         xfs_dquot_t     *d2)
1073 {
1074         if (d1 && d2) {
1075                 ASSERT(d1 != d2);
1076                 if (be32_to_cpu(d1->q_core.d_id) >
1077                     be32_to_cpu(d2->q_core.d_id)) {
1078                         mutex_lock(&d2->q_qlock);
1079                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1080                 } else {
1081                         mutex_lock(&d1->q_qlock);
1082                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1083                 }
1084         } else if (d1) {
1085                 mutex_lock(&d1->q_qlock);
1086         } else if (d2) {
1087                 mutex_lock(&d2->q_qlock);
1088         }
1089 }
1090
1091 int __init
1092 xfs_qm_init(void)
1093 {
1094         xfs_qm_dqzone =
1095                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1096         if (!xfs_qm_dqzone)
1097                 goto out;
1098
1099         xfs_qm_dqtrxzone =
1100                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1101         if (!xfs_qm_dqtrxzone)
1102                 goto out_free_dqzone;
1103
1104         return 0;
1105
1106 out_free_dqzone:
1107         kmem_zone_destroy(xfs_qm_dqzone);
1108 out:
1109         return -ENOMEM;
1110 }
1111
1112 void
1113 xfs_qm_exit(void)
1114 {
1115         kmem_zone_destroy(xfs_qm_dqtrxzone);
1116         kmem_zone_destroy(xfs_qm_dqzone);
1117 }