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_format.h"
  21 #include "xfs_log_format.h"
  22 #include "xfs_shared.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_bit.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_mount.h"
  27 #include "xfs_inode.h"
  28 #include "xfs_bmap.h"
  29 #include "xfs_bmap_util.h"
  30 #include "xfs_alloc.h"
  31 #include "xfs_quota.h"
  32 #include "xfs_error.h"
  33 #include "xfs_trans.h"
  34 #include "xfs_buf_item.h"
  35 #include "xfs_trans_space.h"
  36 #include "xfs_trans_priv.h"
  37 #include "xfs_qm.h"
  38 #include "xfs_cksum.h"
  39 #include "xfs_trace.h"
  40 #include "xfs_log.h"
  41 #include "xfs_bmap_btree.h"
  42
  43 /*
  44  * Lock order:
  45  *
  46  * ip->i_lock
  47  *   qi->qi_tree_lock
  48  *     dquot->q_qlock (xfs_dqlock() and friends)
  49  *       dquot->q_flush (xfs_dqflock() and friends)
  50  *       qi->qi_lru_lock
  51  *
  52  * If two dquots need to be locked the order is user before group/project,
  53  * otherwise by the lowest id first, see xfs_dqlock2.
  54  */
  55
  56 #ifdef DEBUG
  57 xfs_buftarg_t *xfs_dqerror_target;
  58 int xfs_do_dqerror;
  59 int xfs_dqreq_num;
  60 int xfs_dqerror_mod = 33;
  61 #endif
  62
  63 struct kmem_zone                *xfs_qm_dqtrxzone;
  64 static struct kmem_zone         *xfs_qm_dqzone;
  65
  66 static struct lock_class_key xfs_dquot_group_class;
  67 static struct lock_class_key xfs_dquot_project_class;
  68
  69 /*
  70  * This is called to free all the memory associated with a dquot
  71  */
  72 void
  73 xfs_qm_dqdestroy(
  74         xfs_dquot_t     *dqp)
  75 {
  76         ASSERT(list_empty(&dqp->q_lru));
  77
  78         mutex_destroy(&dqp->q_qlock);
  79         kmem_zone_free(xfs_qm_dqzone, dqp);
  80
  81         XFS_STATS_DEC(xs_qm_dquot);
  82 }
  83
  84 /*
  85  * If default limits are in force, push them into the dquot now.
  86  * We overwrite the dquot limits only if they are zero and this
  87  * is not the root dquot.
  88  */
  89 void
  90 xfs_qm_adjust_dqlimits(
  91         struct xfs_mount        *mp,
  92         struct xfs_dquot        *dq)
  93 {
  94         struct xfs_quotainfo    *q = mp->m_quotainfo;
  95         struct xfs_disk_dquot   *d = &dq->q_core;
  96         int                     prealloc = 0;
  97
  98         ASSERT(d->d_id);
  99
 100         if (q->qi_bsoftlimit && !d->d_blk_softlimit) {
 101                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 102                 prealloc = 1;
 103         }
 104         if (q->qi_bhardlimit && !d->d_blk_hardlimit) {
 105                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 106                 prealloc = 1;
 107         }
 108         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 109                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 110         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 111                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 112         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 113                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 114         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 115                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 116
 117         if (prealloc)
 118                 xfs_dquot_set_prealloc_limits(dq);
 119 }
 120
 121 /*
 122  * Check the limits and timers of a dquot and start or reset timers
 123  * if necessary.
 124  * This gets called even when quota enforcement is OFF, which makes our
 125  * life a little less complicated. (We just don't reject any quota
 126  * reservations in that case, when enforcement is off).
 127  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 128  * enforcement's off.
 129  * In contrast, warnings are a little different in that they don't
 130  * 'automatically' get started when limits get exceeded.  They do
 131  * get reset to zero, however, when we find the count to be under
 132  * the soft limit (they are only ever set non-zero via userspace).
 133  */
 134 void
 135 xfs_qm_adjust_dqtimers(
 136         xfs_mount_t             *mp,
 137         xfs_disk_dquot_t        *d)
 138 {
 139         ASSERT(d->d_id);
 140
 141 #ifdef DEBUG
 142         if (d->d_blk_hardlimit)
 143                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 144                        be64_to_cpu(d->d_blk_hardlimit));
 145         if (d->d_ino_hardlimit)
 146                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 147                        be64_to_cpu(d->d_ino_hardlimit));
 148         if (d->d_rtb_hardlimit)
 149                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 150                        be64_to_cpu(d->d_rtb_hardlimit));
 151 #endif
 152
 153         if (!d->d_btimer) {
 154                 if ((d->d_blk_softlimit &&
 155                      (be64_to_cpu(d->d_bcount) >
 156                       be64_to_cpu(d->d_blk_softlimit))) ||
 157                     (d->d_blk_hardlimit &&
 158                      (be64_to_cpu(d->d_bcount) >
 159                       be64_to_cpu(d->d_blk_hardlimit)))) {
 160                         d->d_btimer = cpu_to_be32(get_seconds() +
 161                                         mp->m_quotainfo->qi_btimelimit);
 162                 } else {
 163                         d->d_bwarns = 0;
 164                 }
 165         } else {
 166                 if ((!d->d_blk_softlimit ||
 167                      (be64_to_cpu(d->d_bcount) <=
 168                       be64_to_cpu(d->d_blk_softlimit))) &&
 169                     (!d->d_blk_hardlimit ||
 170                     (be64_to_cpu(d->d_bcount) <=
 171                      be64_to_cpu(d->d_blk_hardlimit)))) {
 172                         d->d_btimer = 0;
 173                 }
 174         }
 175
 176         if (!d->d_itimer) {
 177                 if ((d->d_ino_softlimit &&
 178                      (be64_to_cpu(d->d_icount) >
 179                       be64_to_cpu(d->d_ino_softlimit))) ||
 180                     (d->d_ino_hardlimit &&
 181                      (be64_to_cpu(d->d_icount) >
 182                       be64_to_cpu(d->d_ino_hardlimit)))) {
 183                         d->d_itimer = cpu_to_be32(get_seconds() +
 184                                         mp->m_quotainfo->qi_itimelimit);
 185                 } else {
 186                         d->d_iwarns = 0;
 187                 }
 188         } else {
 189                 if ((!d->d_ino_softlimit ||
 190                      (be64_to_cpu(d->d_icount) <=
 191                       be64_to_cpu(d->d_ino_softlimit)))  &&
 192                     (!d->d_ino_hardlimit ||
 193                      (be64_to_cpu(d->d_icount) <=
 194                       be64_to_cpu(d->d_ino_hardlimit)))) {
 195                         d->d_itimer = 0;
 196                 }
 197         }
 198
 199         if (!d->d_rtbtimer) {
 200                 if ((d->d_rtb_softlimit &&
 201                      (be64_to_cpu(d->d_rtbcount) >
 202                       be64_to_cpu(d->d_rtb_softlimit))) ||
 203                     (d->d_rtb_hardlimit &&
 204                      (be64_to_cpu(d->d_rtbcount) >
 205                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 206                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 207                                         mp->m_quotainfo->qi_rtbtimelimit);
 208                 } else {
 209                         d->d_rtbwarns = 0;
 210                 }
 211         } else {
 212                 if ((!d->d_rtb_softlimit ||
 213                      (be64_to_cpu(d->d_rtbcount) <=
 214                       be64_to_cpu(d->d_rtb_softlimit))) &&
 215                     (!d->d_rtb_hardlimit ||
 216                      (be64_to_cpu(d->d_rtbcount) <=
 217                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 218                         d->d_rtbtimer = 0;
 219                 }
 220         }
 221 }
 222
 223 /*
 224  * initialize a buffer full of dquots and log the whole thing
 225  */
 226 STATIC void
 227 xfs_qm_init_dquot_blk(
 228         xfs_trans_t     *tp,
 229         xfs_mount_t     *mp,
 230         xfs_dqid_t      id,
 231         uint            type,
 232         xfs_buf_t       *bp)
 233 {
 234         struct xfs_quotainfo    *q = mp->m_quotainfo;
 235         xfs_dqblk_t     *d;
 236         int             curid, i;
 237
 238         ASSERT(tp);
 239         ASSERT(xfs_buf_islocked(bp));
 240
 241         d = bp->b_addr;
 242
 243         /*
 244          * ID of the first dquot in the block - id's are zero based.
 245          */
 246         curid = id - (id % q->qi_dqperchunk);
 247         ASSERT(curid >= 0);
 248         memset(d, 0, BBTOB(q->qi_dqchunklen));
 249         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 250                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 251                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 252                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 253                 d->dd_diskdq.d_flags = type;
 254                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 255                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_uuid);
 256                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
 257                                          XFS_DQUOT_CRC_OFF);
 258                 }
 259         }
 260
 261         xfs_trans_dquot_buf(tp, bp,
 262                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 263                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 264                              XFS_BLF_GDQUOT_BUF)));
 265         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 266 }
 267
 268 /*
 269  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
 270  * watermarks correspond to the soft and hard limits by default. If a soft limit
 271  * is not specified, we use 95% of the hard limit.
 272  */
 273 void
 274 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
 275 {
 276         __uint64_t space;
 277
 278         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
 279         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
 280         if (!dqp->q_prealloc_lo_wmark) {
 281                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
 282                 do_div(dqp->q_prealloc_lo_wmark, 100);
 283                 dqp->q_prealloc_lo_wmark *= 95;
 284         }
 285
 286         space = dqp->q_prealloc_hi_wmark;
 287
 288         do_div(space, 100);
 289         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
 290         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
 291         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
 292 }
 293
 294 /*
 295  * Allocate a block and fill it with dquots.
 296  * This is called when the bmapi finds a hole.
 297  */
 298 STATIC int
 299 xfs_qm_dqalloc(
 300         xfs_trans_t     **tpp,
 301         xfs_mount_t     *mp,
 302         xfs_dquot_t     *dqp,
 303         xfs_inode_t     *quotip,
 304         xfs_fileoff_t   offset_fsb,
 305         xfs_buf_t       **O_bpp)
 306 {
 307         xfs_fsblock_t   firstblock;
 308         xfs_bmap_free_t flist;
 309         xfs_bmbt_irec_t map;
 310         int             nmaps, error, committed;
 311         xfs_buf_t       *bp;
 312         xfs_trans_t     *tp = *tpp;
 313
 314         ASSERT(tp != NULL);
 315
 316         trace_xfs_dqalloc(dqp);
 317
 318         /*
 319          * Initialize the bmap freelist prior to calling bmapi code.
 320          */
 321         xfs_bmap_init(&flist, &firstblock);
 322         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 323         /*
 324          * Return if this type of quotas is turned off while we didn't
 325          * have an inode lock
 326          */
 327         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 328                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 329                 return -ESRCH;
 330         }
 331
 332         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 333         nmaps = 1;
 334         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 335                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 336                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 337                                 &map, &nmaps, &flist);
 338         if (error)
 339                 goto error0;
 340         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 341         ASSERT(nmaps == 1);
 342         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 343                (map.br_startblock != HOLESTARTBLOCK));
 344
 345         /*
 346          * Keep track of the blkno to save a lookup later
 347          */
 348         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 349
 350         /* now we can just get the buffer (there's nothing to read yet) */
 351         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 352                                dqp->q_blkno,
 353                                mp->m_quotainfo->qi_dqchunklen,
 354                                0);
 355         if (!bp) {
 356                 error = -ENOMEM;
 357                 goto error1;
 358         }
 359         bp->b_ops = &xfs_dquot_buf_ops;
 360
 361         /*
 362          * Make a chunk of dquots out of this buffer and log
 363          * the entire thing.
 364          */
 365         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 366                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 367
 368         /*
 369          * xfs_bmap_finish() may commit the current transaction and
 370          * start a second transaction if the freelist is not empty.
 371          *
 372          * Since we still want to modify this buffer, we need to
 373          * ensure that the buffer is not released on commit of
 374          * the first transaction and ensure the buffer is added to the
 375          * second transaction.
 376          *
 377          * If there is only one transaction then don't stop the buffer
 378          * from being released when it commits later on.
 379          */
 380
 381         xfs_trans_bhold(tp, bp);
 382
 383         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 384                 goto error1;
 385         }
 386
 387         if (committed) {
 388                 tp = *tpp;
 389                 xfs_trans_bjoin(tp, bp);
 390         } else {
 391                 xfs_trans_bhold_release(tp, bp);
 392         }
 393
 394         *O_bpp = bp;
 395         return 0;
 396
 397       error1:
 398         xfs_bmap_cancel(&flist);
 399       error0:
 400         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 401
 402         return error;
 403 }
 404
 405 STATIC int
 406 xfs_qm_dqrepair(
 407         struct xfs_mount        *mp,
 408         struct xfs_trans        *tp,
 409         struct xfs_dquot        *dqp,
 410         xfs_dqid_t              firstid,
 411         struct xfs_buf          **bpp)
 412 {
 413         int                     error;
 414         struct xfs_disk_dquot   *ddq;
 415         struct xfs_dqblk        *d;
 416         int                     i;
 417
 418         /*
 419          * Read the buffer without verification so we get the corrupted
 420          * buffer returned to us. make sure we verify it on write, though.
 421          */
 422         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
 423                                    mp->m_quotainfo->qi_dqchunklen,
 424                                    0, bpp, NULL);
 425
 426         if (error) {
 427                 ASSERT(*bpp == NULL);
 428                 return error;
 429         }
 430         (*bpp)->b_ops = &xfs_dquot_buf_ops;
 431
 432         ASSERT(xfs_buf_islocked(*bpp));
 433         d = (struct xfs_dqblk *)(*bpp)->b_addr;
 434
 435         /* Do the actual repair of dquots in this buffer */
 436         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
 437                 ddq = &d[i].dd_diskdq;
 438                 error = xfs_dqcheck(mp, ddq, firstid + i,
 439                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
 440                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
 441                 if (error) {
 442                         /* repair failed, we're screwed */
 443                         xfs_trans_brelse(tp, *bpp);
 444                         return -EIO;
 445                 }
 446         }
 447
 448         return 0;
 449 }
 450
 451 /*
 452  * Maps a dquot to the buffer containing its on-disk version.
 453  * This returns a ptr to the buffer containing the on-disk dquot
 454  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 455  */
 456 STATIC int
 457 xfs_qm_dqtobp(
 458         xfs_trans_t             **tpp,
 459         xfs_dquot_t             *dqp,
 460         xfs_disk_dquot_t        **O_ddpp,
 461         xfs_buf_t               **O_bpp,
 462         uint                    flags)
 463 {
 464         struct xfs_bmbt_irec    map;
 465         int                     nmaps = 1, error;
 466         struct xfs_buf          *bp;
 467         struct xfs_inode        *quotip = xfs_dq_to_quota_inode(dqp);
 468         struct xfs_mount        *mp = dqp->q_mount;
 469         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 470         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 471         uint                    lock_mode;
 472
 473         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 474
 475         lock_mode = xfs_ilock_data_map_shared(quotip);
 476         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 477                 /*
 478                  * Return if this type of quotas is turned off while we
 479                  * didn't have the quota inode lock.
 480                  */
 481                 xfs_iunlock(quotip, lock_mode);
 482                 return -ESRCH;
 483         }
 484
 485         /*
 486          * Find the block map; no allocations yet
 487          */
 488         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 489                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 490
 491         xfs_iunlock(quotip, lock_mode);
 492         if (error)
 493                 return error;
 494
 495         ASSERT(nmaps == 1);
 496         ASSERT(map.br_blockcount == 1);
 497
 498         /*
 499          * Offset of dquot in the (fixed sized) dquot chunk.
 500          */
 501         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 502                 sizeof(xfs_dqblk_t);
 503
 504         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 505         if (map.br_startblock == HOLESTARTBLOCK) {
 506                 /*
 507                  * We don't allocate unless we're asked to
 508                  */
 509                 if (!(flags & XFS_QMOPT_DQALLOC))
 510                         return -ENOENT;
 511
 512                 ASSERT(tp);
 513                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 514                                         dqp->q_fileoffset, &bp);
 515                 if (error)
 516                         return error;
 517                 tp = *tpp;
 518         } else {
 519                 trace_xfs_dqtobp_read(dqp);
 520
 521                 /*
 522                  * store the blkno etc so that we don't have to do the
 523                  * mapping all the time
 524                  */
 525                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 526
 527                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 528                                            dqp->q_blkno,
 529                                            mp->m_quotainfo->qi_dqchunklen,
 530                                            0, &bp, &xfs_dquot_buf_ops);
 531
 532                 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
 533                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
 534                                                 mp->m_quotainfo->qi_dqperchunk;
 535                         ASSERT(bp == NULL);
 536                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
 537                 }
 538
 539                 if (error) {
 540                         ASSERT(bp == NULL);
 541                         return error;
 542                 }
 543         }
 544
 545         ASSERT(xfs_buf_islocked(bp));
 546         *O_bpp = bp;
 547         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 548
 549         return 0;
 550 }
 551
 552
 553 /*
 554  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 555  * and release the buffer immediately.
 556  *
 557  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 558  */
 559 int
 560 xfs_qm_dqread(
 561         struct xfs_mount        *mp,
 562         xfs_dqid_t              id,
 563         uint                    type,
 564         uint                    flags,
 565         struct xfs_dquot        **O_dqpp)
 566 {
 567         struct xfs_dquot        *dqp;
 568         struct xfs_disk_dquot   *ddqp;
 569         struct xfs_buf          *bp;
 570         struct xfs_trans        *tp = NULL;
 571         int                     error;
 572         int                     cancelflags = 0;
 573
 574
 575         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 576
 577         dqp->dq_flags = type;
 578         dqp->q_core.d_id = cpu_to_be32(id);
 579         dqp->q_mount = mp;
 580         INIT_LIST_HEAD(&dqp->q_lru);
 581         mutex_init(&dqp->q_qlock);
 582         init_waitqueue_head(&dqp->q_pinwait);
 583
 584         /*
 585          * Because we want to use a counting completion, complete
 586          * the flush completion once to allow a single access to
 587          * the flush completion without blocking.
 588          */
 589         init_completion(&dqp->q_flush);
 590         complete(&dqp->q_flush);
 591
 592         /*
 593          * Make sure group quotas have a different lock class than user
 594          * quotas.
 595          */
 596         switch (type) {
 597         case XFS_DQ_USER:
 598                 /* uses the default lock class */
 599                 break;
 600         case XFS_DQ_GROUP:
 601                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 602                 break;
 603         case XFS_DQ_PROJ:
 604                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 605                 break;
 606         default:
 607                 ASSERT(0);
 608                 break;
 609         }
 610
 611         XFS_STATS_INC(xs_qm_dquot);
 612
 613         trace_xfs_dqread(dqp);
 614
 615         if (flags & XFS_QMOPT_DQALLOC) {
 616                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 617                 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_qm_dqalloc,
 618                                           XFS_QM_DQALLOC_SPACE_RES(mp), 0);
 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         /* initialize the dquot speculative prealloc thresholds */
 653         xfs_dquot_set_prealloc_limits(dqp);
 654
 655         /* Mark the buf so that this will stay incore a little longer */
 656         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 657
 658         /*
 659          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 660          * So we need to release with xfs_trans_brelse().
 661          * The strategy here is identical to that of inodes; we lock
 662          * the dquot in xfs_qm_dqget() before making it accessible to
 663          * others. This is because dquots, like inodes, need a good level of
 664          * concurrency, and we don't want to take locks on the entire buffers
 665          * for dquot accesses.
 666          * Note also that the dquot buffer may even be dirty at this point, if
 667          * this particular dquot was repaired. We still aren't afraid to
 668          * brelse it because we have the changes incore.
 669          */
 670         ASSERT(xfs_buf_islocked(bp));
 671         xfs_trans_brelse(tp, bp);
 672
 673         if (tp) {
 674                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 675                 if (error)
 676                         goto error0;
 677         }
 678
 679         *O_dqpp = dqp;
 680         return error;
 681
 682 error1:
 683         if (tp)
 684                 xfs_trans_cancel(tp, cancelflags);
 685 error0:
 686         xfs_qm_dqdestroy(dqp);
 687         *O_dqpp = NULL;
 688         return error;
 689 }
 690
 691 /*
 692  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 693  * a locked dquot, doing an allocation (if requested) as needed.
 694  * When both an inode and an id are given, the inode's id takes precedence.
 695  * That is, if the id changes while we don't hold the ilock inside this
 696  * function, the new dquot is returned, not necessarily the one requested
 697  * in the id argument.
 698  */
 699 int
 700 xfs_qm_dqget(
 701         xfs_mount_t     *mp,
 702         xfs_inode_t     *ip,      /* locked inode (optional) */
 703         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 704         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 705         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 706         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 707 {
 708         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 709         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 710         struct xfs_dquot        *dqp;
 711         int                     error;
 712
 713         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 714         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 715             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 716             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 717                 return -ESRCH;
 718         }
 719
 720 #ifdef DEBUG
 721         if (xfs_do_dqerror) {
 722                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 723                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 724                         xfs_debug(mp, "Returning error in dqget");
 725                         return -EIO;
 726                 }
 727         }
 728
 729         ASSERT(type == XFS_DQ_USER ||
 730                type == XFS_DQ_PROJ ||
 731                type == XFS_DQ_GROUP);
 732         if (ip) {
 733                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 734                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 735         }
 736 #endif
 737
 738 restart:
 739         mutex_lock(&qi->qi_tree_lock);
 740         dqp = radix_tree_lookup(tree, id);
 741         if (dqp) {
 742                 xfs_dqlock(dqp);
 743                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 744                         xfs_dqunlock(dqp);
 745                         mutex_unlock(&qi->qi_tree_lock);
 746                         trace_xfs_dqget_freeing(dqp);
 747                         delay(1);
 748                         goto restart;
 749                 }
 750
 751                 dqp->q_nrefs++;
 752                 mutex_unlock(&qi->qi_tree_lock);
 753
 754                 trace_xfs_dqget_hit(dqp);
 755                 XFS_STATS_INC(xs_qm_dqcachehits);
 756                 *O_dqpp = dqp;
 757                 return 0;
 758         }
 759         mutex_unlock(&qi->qi_tree_lock);
 760         XFS_STATS_INC(xs_qm_dqcachemisses);
 761
 762         /*
 763          * Dquot cache miss. We don't want to keep the inode lock across
 764          * a (potential) disk read. Also we don't want to deal with the lock
 765          * ordering between quotainode and this inode. OTOH, dropping the inode
 766          * lock here means dealing with a chown that can happen before
 767          * we re-acquire the lock.
 768          */
 769         if (ip)
 770                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 771
 772         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 773
 774         if (ip)
 775                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 776
 777         if (error)
 778                 return error;
 779
 780         if (ip) {
 781                 /*
 782                  * A dquot could be attached to this inode by now, since
 783                  * we had dropped the ilock.
 784                  */
 785                 if (xfs_this_quota_on(mp, type)) {
 786                         struct xfs_dquot        *dqp1;
 787
 788                         dqp1 = xfs_inode_dquot(ip, type);
 789                         if (dqp1) {
 790                                 xfs_qm_dqdestroy(dqp);
 791                                 dqp = dqp1;
 792                                 xfs_dqlock(dqp);
 793                                 goto dqret;
 794                         }
 795                 } else {
 796                         /* inode stays locked on return */
 797                         xfs_qm_dqdestroy(dqp);
 798                         return -ESRCH;
 799                 }
 800         }
 801
 802         mutex_lock(&qi->qi_tree_lock);
 803         error = radix_tree_insert(tree, id, dqp);
 804         if (unlikely(error)) {
 805                 WARN_ON(error != -EEXIST);
 806
 807                 /*
 808                  * Duplicate found. Just throw away the new dquot and start
 809                  * over.
 810                  */
 811                 mutex_unlock(&qi->qi_tree_lock);
 812                 trace_xfs_dqget_dup(dqp);
 813                 xfs_qm_dqdestroy(dqp);
 814                 XFS_STATS_INC(xs_qm_dquot_dups);
 815                 goto restart;
 816         }
 817
 818         /*
 819          * We return a locked dquot to the caller, with a reference taken
 820          */
 821         xfs_dqlock(dqp);
 822         dqp->q_nrefs = 1;
 823
 824         qi->qi_dquots++;
 825         mutex_unlock(&qi->qi_tree_lock);
 826
 827  dqret:
 828         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 829         trace_xfs_dqget_miss(dqp);
 830         *O_dqpp = dqp;
 831         return 0;
 832 }
 833
 834 /*
 835  * Release a reference to the dquot (decrement ref-count) and unlock it.
 836  *
 837  * If there is a group quota attached to this dquot, carefully release that
 838  * too without tripping over deadlocks'n'stuff.
 839  */
 840 void
 841 xfs_qm_dqput(
 842         struct xfs_dquot        *dqp)
 843 {
 844         ASSERT(dqp->q_nrefs > 0);
 845         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 846
 847         trace_xfs_dqput(dqp);
 848
 849         if (--dqp->q_nrefs == 0) {
 850                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 851                 trace_xfs_dqput_free(dqp);
 852
 853                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 854                         XFS_STATS_INC(xs_qm_dquot_unused);
 855         }
 856         xfs_dqunlock(dqp);
 857 }
 858
 859 /*
 860  * Release a dquot. Flush it if dirty, then dqput() it.
 861  * dquot must not be locked.
 862  */
 863 void
 864 xfs_qm_dqrele(
 865         xfs_dquot_t     *dqp)
 866 {
 867         if (!dqp)
 868                 return;
 869
 870         trace_xfs_dqrele(dqp);
 871
 872         xfs_dqlock(dqp);
 873         /*
 874          * We don't care to flush it if the dquot is dirty here.
 875          * That will create stutters that we want to avoid.
 876          * Instead we do a delayed write when we try to reclaim
 877          * a dirty dquot. Also xfs_sync will take part of the burden...
 878          */
 879         xfs_qm_dqput(dqp);
 880 }
 881
 882 /*
 883  * This is the dquot flushing I/O completion routine.  It is called
 884  * from interrupt level when the buffer containing the dquot is
 885  * flushed to disk.  It is responsible for removing the dquot logitem
 886  * from the AIL if it has not been re-logged, and unlocking the dquot's
 887  * flush lock. This behavior is very similar to that of inodes..
 888  */
 889 STATIC void
 890 xfs_qm_dqflush_done(
 891         struct xfs_buf          *bp,
 892         struct xfs_log_item     *lip)
 893 {
 894         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 895         xfs_dquot_t             *dqp = qip->qli_dquot;
 896         struct xfs_ail          *ailp = lip->li_ailp;
 897
 898         /*
 899          * We only want to pull the item from the AIL if its
 900          * location in the log has not changed since we started the flush.
 901          * Thus, we only bother if the dquot's lsn has
 902          * not changed. First we check the lsn outside the lock
 903          * since it's cheaper, and then we recheck while
 904          * holding the lock before removing the dquot from the AIL.
 905          */
 906         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 907             lip->li_lsn == qip->qli_flush_lsn) {
 908
 909                 /* xfs_trans_ail_delete() drops the AIL lock. */
 910                 spin_lock(&ailp->xa_lock);
 911                 if (lip->li_lsn == qip->qli_flush_lsn)
 912                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
 913                 else
 914                         spin_unlock(&ailp->xa_lock);
 915         }
 916
 917         /*
 918          * Release the dq's flush lock since we're done with it.
 919          */
 920         xfs_dqfunlock(dqp);
 921 }
 922
 923 /*
 924  * Write a modified dquot to disk.
 925  * The dquot must be locked and the flush lock too taken by caller.
 926  * The flush lock will not be unlocked until the dquot reaches the disk,
 927  * but the dquot is free to be unlocked and modified by the caller
 928  * in the interim. Dquot is still locked on return. This behavior is
 929  * identical to that of inodes.
 930  */
 931 int
 932 xfs_qm_dqflush(
 933         struct xfs_dquot        *dqp,
 934         struct xfs_buf          **bpp)
 935 {
 936         struct xfs_mount        *mp = dqp->q_mount;
 937         struct xfs_buf          *bp;
 938         struct xfs_disk_dquot   *ddqp;
 939         int                     error;
 940
 941         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 942         ASSERT(!completion_done(&dqp->q_flush));
 943
 944         trace_xfs_dqflush(dqp);
 945
 946         *bpp = NULL;
 947
 948         xfs_qm_dqunpin_wait(dqp);
 949
 950         /*
 951          * This may have been unpinned because the filesystem is shutting
 952          * down forcibly. If that's the case we must not write this dquot
 953          * to disk, because the log record didn't make it to disk.
 954          *
 955          * We also have to remove the log item from the AIL in this case,
 956          * as we wait for an emptry AIL as part of the unmount process.
 957          */
 958         if (XFS_FORCED_SHUTDOWN(mp)) {
 959                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
 960                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
 961
 962                 spin_lock(&mp->m_ail->xa_lock);
 963                 if (lip->li_flags & XFS_LI_IN_AIL)
 964                         xfs_trans_ail_delete(mp->m_ail, lip,
 965                                              SHUTDOWN_CORRUPT_INCORE);
 966                 else
 967                         spin_unlock(&mp->m_ail->xa_lock);
 968                 error = -EIO;
 969                 goto out_unlock;
 970         }
 971
 972         /*
 973          * Get the buffer containing the on-disk dquot
 974          */
 975         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
 976                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
 977                                    &xfs_dquot_buf_ops);
 978         if (error)
 979                 goto out_unlock;
 980
 981         /*
 982          * Calculate the location of the dquot inside the buffer.
 983          */
 984         ddqp = bp->b_addr + dqp->q_bufoffset;
 985
 986         /*
 987          * A simple sanity check in case we got a corrupted dquot..
 988          */
 989         error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
 990                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
 991         if (error) {
 992                 xfs_buf_relse(bp);
 993                 xfs_dqfunlock(dqp);
 994                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 995                 return -EIO;
 996         }
 997
 998         /* This is the only portion of data that needs to persist */
 999         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1000
1001         /*
1002          * Clear the dirty field and remember the flush lsn for later use.
1003          */
1004         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1005
1006         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1007                                         &dqp->q_logitem.qli_item.li_lsn);
1008
1009         /*
1010          * copy the lsn into the on-disk dquot now while we have the in memory
1011          * dquot here. This can't be done later in the write verifier as we
1012          * can't get access to the log item at that point in time.
1013          *
1014          * We also calculate the CRC here so that the on-disk dquot in the
1015          * buffer always has a valid CRC. This ensures there is no possibility
1016          * of a dquot without an up-to-date CRC getting to disk.
1017          */
1018         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1019                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1020
1021                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1022                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1023                                  XFS_DQUOT_CRC_OFF);
1024         }
1025
1026         /*
1027          * Attach an iodone routine so that we can remove this dquot from the
1028          * AIL and release the flush lock once the dquot is synced to disk.
1029          */
1030         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1031                                   &dqp->q_logitem.qli_item);
1032
1033         /*
1034          * If the buffer is pinned then push on the log so we won't
1035          * get stuck waiting in the write for too long.
1036          */
1037         if (xfs_buf_ispinned(bp)) {
1038                 trace_xfs_dqflush_force(dqp);
1039                 xfs_log_force(mp, 0);
1040         }
1041
1042         trace_xfs_dqflush_done(dqp);
1043         *bpp = bp;
1044         return 0;
1045
1046 out_unlock:
1047         xfs_dqfunlock(dqp);
1048         return -EIO;
1049 }
1050
1051 /*
1052  * Lock two xfs_dquot structures.
1053  *
1054  * To avoid deadlocks we always lock the quota structure with
1055  * the lowerd id first.
1056  */
1057 void
1058 xfs_dqlock2(
1059         xfs_dquot_t     *d1,
1060         xfs_dquot_t     *d2)
1061 {
1062         if (d1 && d2) {
1063                 ASSERT(d1 != d2);
1064                 if (be32_to_cpu(d1->q_core.d_id) >
1065                     be32_to_cpu(d2->q_core.d_id)) {
1066                         mutex_lock(&d2->q_qlock);
1067                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1068                 } else {
1069                         mutex_lock(&d1->q_qlock);
1070                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1071                 }
1072         } else if (d1) {
1073                 mutex_lock(&d1->q_qlock);
1074         } else if (d2) {
1075                 mutex_lock(&d2->q_qlock);
1076         }
1077 }
1078
1079 int __init
1080 xfs_qm_init(void)
1081 {
1082         xfs_qm_dqzone =
1083                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1084         if (!xfs_qm_dqzone)
1085                 goto out;
1086
1087         xfs_qm_dqtrxzone =
1088                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1089         if (!xfs_qm_dqtrxzone)
1090                 goto out_free_dqzone;
1091
1092         return 0;
1093
1094 out_free_dqzone:
1095         kmem_zone_destroy(xfs_qm_dqzone);
1096 out:
1097         return -ENOMEM;
1098 }
1099
1100 void
1101 xfs_qm_exit(void)
1102 {
1103         kmem_zone_destroy(xfs_qm_dqtrxzone);
1104         kmem_zone_destroy(xfs_qm_dqzone);
1105 }