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_mount.h"
  26 #include "xfs_defer.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 struct kmem_zone                *xfs_qm_dqtrxzone;
  57 static struct kmem_zone         *xfs_qm_dqzone;
  58
  59 static struct lock_class_key xfs_dquot_group_class;
  60 static struct lock_class_key xfs_dquot_project_class;
  61
  62 /*
  63  * This is called to free all the memory associated with a dquot
  64  */
  65 void
  66 xfs_qm_dqdestroy(
  67         xfs_dquot_t     *dqp)
  68 {
  69         ASSERT(list_empty(&dqp->q_lru));
  70
  71         kmem_free(dqp->q_logitem.qli_item.li_lv_shadow);
  72         mutex_destroy(&dqp->q_qlock);
  73
  74         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
  75         kmem_zone_free(xfs_qm_dqzone, dqp);
  76 }
  77
  78 /*
  79  * If default limits are in force, push them into the dquot now.
  80  * We overwrite the dquot limits only if they are zero and this
  81  * is not the root dquot.
  82  */
  83 void
  84 xfs_qm_adjust_dqlimits(
  85         struct xfs_mount        *mp,
  86         struct xfs_dquot        *dq)
  87 {
  88         struct xfs_quotainfo    *q = mp->m_quotainfo;
  89         struct xfs_disk_dquot   *d = &dq->q_core;
  90         struct xfs_def_quota    *defq;
  91         int                     prealloc = 0;
  92
  93         ASSERT(d->d_id);
  94         defq = xfs_get_defquota(dq, q);
  95
  96         if (defq->bsoftlimit && !d->d_blk_softlimit) {
  97                 d->d_blk_softlimit = cpu_to_be64(defq->bsoftlimit);
  98                 prealloc = 1;
  99         }
 100         if (defq->bhardlimit && !d->d_blk_hardlimit) {
 101                 d->d_blk_hardlimit = cpu_to_be64(defq->bhardlimit);
 102                 prealloc = 1;
 103         }
 104         if (defq->isoftlimit && !d->d_ino_softlimit)
 105                 d->d_ino_softlimit = cpu_to_be64(defq->isoftlimit);
 106         if (defq->ihardlimit && !d->d_ino_hardlimit)
 107                 d->d_ino_hardlimit = cpu_to_be64(defq->ihardlimit);
 108         if (defq->rtbsoftlimit && !d->d_rtb_softlimit)
 109                 d->d_rtb_softlimit = cpu_to_be64(defq->rtbsoftlimit);
 110         if (defq->rtbhardlimit && !d->d_rtb_hardlimit)
 111                 d->d_rtb_hardlimit = cpu_to_be64(defq->rtbhardlimit);
 112
 113         if (prealloc)
 114                 xfs_dquot_set_prealloc_limits(dq);
 115 }
 116
 117 /*
 118  * Check the limits and timers of a dquot and start or reset timers
 119  * if necessary.
 120  * This gets called even when quota enforcement is OFF, which makes our
 121  * life a little less complicated. (We just don't reject any quota
 122  * reservations in that case, when enforcement is off).
 123  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 124  * enforcement's off.
 125  * In contrast, warnings are a little different in that they don't
 126  * 'automatically' get started when limits get exceeded.  They do
 127  * get reset to zero, however, when we find the count to be under
 128  * the soft limit (they are only ever set non-zero via userspace).
 129  */
 130 void
 131 xfs_qm_adjust_dqtimers(
 132         xfs_mount_t             *mp,
 133         xfs_disk_dquot_t        *d)
 134 {
 135         ASSERT(d->d_id);
 136
 137 #ifdef DEBUG
 138         if (d->d_blk_hardlimit)
 139                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 140                        be64_to_cpu(d->d_blk_hardlimit));
 141         if (d->d_ino_hardlimit)
 142                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 143                        be64_to_cpu(d->d_ino_hardlimit));
 144         if (d->d_rtb_hardlimit)
 145                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 146                        be64_to_cpu(d->d_rtb_hardlimit));
 147 #endif
 148
 149         if (!d->d_btimer) {
 150                 if ((d->d_blk_softlimit &&
 151                      (be64_to_cpu(d->d_bcount) >
 152                       be64_to_cpu(d->d_blk_softlimit))) ||
 153                     (d->d_blk_hardlimit &&
 154                      (be64_to_cpu(d->d_bcount) >
 155                       be64_to_cpu(d->d_blk_hardlimit)))) {
 156                         d->d_btimer = cpu_to_be32(get_seconds() +
 157                                         mp->m_quotainfo->qi_btimelimit);
 158                 } else {
 159                         d->d_bwarns = 0;
 160                 }
 161         } else {
 162                 if ((!d->d_blk_softlimit ||
 163                      (be64_to_cpu(d->d_bcount) <=
 164                       be64_to_cpu(d->d_blk_softlimit))) &&
 165                     (!d->d_blk_hardlimit ||
 166                     (be64_to_cpu(d->d_bcount) <=
 167                      be64_to_cpu(d->d_blk_hardlimit)))) {
 168                         d->d_btimer = 0;
 169                 }
 170         }
 171
 172         if (!d->d_itimer) {
 173                 if ((d->d_ino_softlimit &&
 174                      (be64_to_cpu(d->d_icount) >
 175                       be64_to_cpu(d->d_ino_softlimit))) ||
 176                     (d->d_ino_hardlimit &&
 177                      (be64_to_cpu(d->d_icount) >
 178                       be64_to_cpu(d->d_ino_hardlimit)))) {
 179                         d->d_itimer = cpu_to_be32(get_seconds() +
 180                                         mp->m_quotainfo->qi_itimelimit);
 181                 } else {
 182                         d->d_iwarns = 0;
 183                 }
 184         } else {
 185                 if ((!d->d_ino_softlimit ||
 186                      (be64_to_cpu(d->d_icount) <=
 187                       be64_to_cpu(d->d_ino_softlimit)))  &&
 188                     (!d->d_ino_hardlimit ||
 189                      (be64_to_cpu(d->d_icount) <=
 190                       be64_to_cpu(d->d_ino_hardlimit)))) {
 191                         d->d_itimer = 0;
 192                 }
 193         }
 194
 195         if (!d->d_rtbtimer) {
 196                 if ((d->d_rtb_softlimit &&
 197                      (be64_to_cpu(d->d_rtbcount) >
 198                       be64_to_cpu(d->d_rtb_softlimit))) ||
 199                     (d->d_rtb_hardlimit &&
 200                      (be64_to_cpu(d->d_rtbcount) >
 201                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 202                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 203                                         mp->m_quotainfo->qi_rtbtimelimit);
 204                 } else {
 205                         d->d_rtbwarns = 0;
 206                 }
 207         } else {
 208                 if ((!d->d_rtb_softlimit ||
 209                      (be64_to_cpu(d->d_rtbcount) <=
 210                       be64_to_cpu(d->d_rtb_softlimit))) &&
 211                     (!d->d_rtb_hardlimit ||
 212                      (be64_to_cpu(d->d_rtbcount) <=
 213                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 214                         d->d_rtbtimer = 0;
 215                 }
 216         }
 217 }
 218
 219 /*
 220  * initialize a buffer full of dquots and log the whole thing
 221  */
 222 STATIC void
 223 xfs_qm_init_dquot_blk(
 224         xfs_trans_t     *tp,
 225         xfs_mount_t     *mp,
 226         xfs_dqid_t      id,
 227         uint            type,
 228         xfs_buf_t       *bp)
 229 {
 230         struct xfs_quotainfo    *q = mp->m_quotainfo;
 231         xfs_dqblk_t     *d;
 232         xfs_dqid_t      curid;
 233         int             i;
 234
 235         ASSERT(tp);
 236         ASSERT(xfs_buf_islocked(bp));
 237
 238         d = bp->b_addr;
 239
 240         /*
 241          * ID of the first dquot in the block - id's are zero based.
 242          */
 243         curid = id - (id % q->qi_dqperchunk);
 244         memset(d, 0, BBTOB(q->qi_dqchunklen));
 245         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
 246                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 247                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 248                 d->dd_diskdq.d_id = cpu_to_be32(curid);
 249                 d->dd_diskdq.d_flags = type;
 250                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 251                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
 252                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
 253                                          XFS_DQUOT_CRC_OFF);
 254                 }
 255         }
 256
 257         xfs_trans_dquot_buf(tp, bp,
 258                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
 259                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
 260                              XFS_BLF_GDQUOT_BUF)));
 261         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
 262 }
 263
 264 /*
 265  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
 266  * watermarks correspond to the soft and hard limits by default. If a soft limit
 267  * is not specified, we use 95% of the hard limit.
 268  */
 269 void
 270 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
 271 {
 272         uint64_t space;
 273
 274         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
 275         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
 276         if (!dqp->q_prealloc_lo_wmark) {
 277                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
 278                 do_div(dqp->q_prealloc_lo_wmark, 100);
 279                 dqp->q_prealloc_lo_wmark *= 95;
 280         }
 281
 282         space = dqp->q_prealloc_hi_wmark;
 283
 284         do_div(space, 100);
 285         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
 286         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
 287         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
 288 }
 289
 290 /*
 291  * Allocate a block and fill it with dquots.
 292  * This is called when the bmapi finds a hole.
 293  */
 294 STATIC int
 295 xfs_qm_dqalloc(
 296         xfs_trans_t     **tpp,
 297         xfs_mount_t     *mp,
 298         xfs_dquot_t     *dqp,
 299         xfs_inode_t     *quotip,
 300         xfs_fileoff_t   offset_fsb,
 301         xfs_buf_t       **O_bpp)
 302 {
 303         xfs_fsblock_t   firstblock;
 304         struct xfs_defer_ops dfops;
 305         xfs_bmbt_irec_t map;
 306         int             nmaps, error;
 307         xfs_buf_t       *bp;
 308         xfs_trans_t     *tp = *tpp;
 309
 310         ASSERT(tp != NULL);
 311
 312         trace_xfs_dqalloc(dqp);
 313
 314         /*
 315          * Initialize the bmap freelist prior to calling bmapi code.
 316          */
 317         xfs_defer_init(&dfops, &firstblock);
 318         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 319         /*
 320          * Return if this type of quotas is turned off while we didn't
 321          * have an inode lock
 322          */
 323         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 324                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 325                 return -ESRCH;
 326         }
 327
 328         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 329         nmaps = 1;
 330         error = xfs_bmapi_write(tp, quotip, offset_fsb,
 331                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
 332                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
 333                                 &map, &nmaps, &dfops);
 334         if (error)
 335                 goto error0;
 336         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 337         ASSERT(nmaps == 1);
 338         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 339                (map.br_startblock != HOLESTARTBLOCK));
 340
 341         /*
 342          * Keep track of the blkno to save a lookup later
 343          */
 344         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 345
 346         /* now we can just get the buffer (there's nothing to read yet) */
 347         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 348                                dqp->q_blkno,
 349                                mp->m_quotainfo->qi_dqchunklen,
 350                                0);
 351         if (!bp) {
 352                 error = -ENOMEM;
 353                 goto error1;
 354         }
 355         bp->b_ops = &xfs_dquot_buf_ops;
 356
 357         /*
 358          * Make a chunk of dquots out of this buffer and log
 359          * the entire thing.
 360          */
 361         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 362                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 363
 364         /*
 365          * xfs_defer_finish() may commit the current transaction and
 366          * start a second transaction if the freelist is not empty.
 367          *
 368          * Since we still want to modify this buffer, we need to
 369          * ensure that the buffer is not released on commit of
 370          * the first transaction and ensure the buffer is added to the
 371          * second transaction.
 372          *
 373          * If there is only one transaction then don't stop the buffer
 374          * from being released when it commits later on.
 375          */
 376
 377         xfs_trans_bhold(tp, bp);
 378
 379         error = xfs_defer_finish(tpp, &dfops);
 380         if (error)
 381                 goto error1;
 382
 383         /* Transaction was committed? */
 384         if (*tpp != tp) {
 385                 tp = *tpp;
 386                 xfs_trans_bjoin(tp, bp);
 387         } else {
 388                 xfs_trans_bhold_release(tp, bp);
 389         }
 390
 391         *O_bpp = bp;
 392         return 0;
 393
 394 error1:
 395         xfs_defer_cancel(&dfops);
 396 error0:
 397         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 398
 399         return error;
 400 }
 401
 402 /*
 403  * Maps a dquot to the buffer containing its on-disk version.
 404  * This returns a ptr to the buffer containing the on-disk dquot
 405  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 406  */
 407 STATIC int
 408 xfs_qm_dqtobp(
 409         xfs_trans_t             **tpp,
 410         xfs_dquot_t             *dqp,
 411         xfs_disk_dquot_t        **O_ddpp,
 412         xfs_buf_t               **O_bpp,
 413         uint                    flags)
 414 {
 415         struct xfs_bmbt_irec    map;
 416         int                     nmaps = 1, error;
 417         struct xfs_buf          *bp;
 418         struct xfs_inode        *quotip;
 419         struct xfs_mount        *mp = dqp->q_mount;
 420         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
 421         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
 422         uint                    lock_mode;
 423
 424         quotip = xfs_quota_inode(dqp->q_mount, dqp->dq_flags);
 425         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
 426
 427         lock_mode = xfs_ilock_data_map_shared(quotip);
 428         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
 429                 /*
 430                  * Return if this type of quotas is turned off while we
 431                  * didn't have the quota inode lock.
 432                  */
 433                 xfs_iunlock(quotip, lock_mode);
 434                 return -ESRCH;
 435         }
 436
 437         /*
 438          * Find the block map; no allocations yet
 439          */
 440         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
 441                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
 442
 443         xfs_iunlock(quotip, lock_mode);
 444         if (error)
 445                 return error;
 446
 447         ASSERT(nmaps == 1);
 448         ASSERT(map.br_blockcount == 1);
 449
 450         /*
 451          * Offset of dquot in the (fixed sized) dquot chunk.
 452          */
 453         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
 454                 sizeof(xfs_dqblk_t);
 455
 456         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
 457         if (map.br_startblock == HOLESTARTBLOCK) {
 458                 /*
 459                  * We don't allocate unless we're asked to
 460                  */
 461                 if (!(flags & XFS_QMOPT_DQALLOC))
 462                         return -ENOENT;
 463
 464                 ASSERT(tp);
 465                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 466                                         dqp->q_fileoffset, &bp);
 467                 if (error)
 468                         return error;
 469                 tp = *tpp;
 470         } else {
 471                 trace_xfs_dqtobp_read(dqp);
 472
 473                 /*
 474                  * store the blkno etc so that we don't have to do the
 475                  * mapping all the time
 476                  */
 477                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 478
 479                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 480                                            dqp->q_blkno,
 481                                            mp->m_quotainfo->qi_dqchunklen,
 482                                            0, &bp, &xfs_dquot_buf_ops);
 483                 if (error) {
 484                         ASSERT(bp == NULL);
 485                         return error;
 486                 }
 487         }
 488
 489         ASSERT(xfs_buf_islocked(bp));
 490         *O_bpp = bp;
 491         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
 492
 493         return 0;
 494 }
 495
 496
 497 /*
 498  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 499  * and release the buffer immediately.
 500  *
 501  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
 502  */
 503 int
 504 xfs_qm_dqread(
 505         struct xfs_mount        *mp,
 506         xfs_dqid_t              id,
 507         uint                    type,
 508         uint                    flags,
 509         struct xfs_dquot        **O_dqpp)
 510 {
 511         struct xfs_dquot        *dqp;
 512         struct xfs_disk_dquot   *ddqp;
 513         struct xfs_buf          *bp;
 514         struct xfs_trans        *tp = NULL;
 515         int                     error;
 516
 517         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
 518
 519         dqp->dq_flags = type;
 520         dqp->q_core.d_id = cpu_to_be32(id);
 521         dqp->q_mount = mp;
 522         INIT_LIST_HEAD(&dqp->q_lru);
 523         mutex_init(&dqp->q_qlock);
 524         init_waitqueue_head(&dqp->q_pinwait);
 525
 526         /*
 527          * Because we want to use a counting completion, complete
 528          * the flush completion once to allow a single access to
 529          * the flush completion without blocking.
 530          */
 531         init_completion(&dqp->q_flush);
 532         complete(&dqp->q_flush);
 533
 534         /*
 535          * Make sure group quotas have a different lock class than user
 536          * quotas.
 537          */
 538         switch (type) {
 539         case XFS_DQ_USER:
 540                 /* uses the default lock class */
 541                 break;
 542         case XFS_DQ_GROUP:
 543                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
 544                 break;
 545         case XFS_DQ_PROJ:
 546                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
 547                 break;
 548         default:
 549                 ASSERT(0);
 550                 break;
 551         }
 552
 553         XFS_STATS_INC(mp, xs_qm_dquot);
 554
 555         trace_xfs_dqread(dqp);
 556
 557         if (flags & XFS_QMOPT_DQALLOC) {
 558                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
 559                                 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
 560                 if (error)
 561                         goto error0;
 562         }
 563
 564         /*
 565          * get a pointer to the on-disk dquot and the buffer containing it
 566          * dqp already knows its own type (GROUP/USER).
 567          */
 568         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
 569         if (error) {
 570                 /*
 571                  * This can happen if quotas got turned off (ESRCH),
 572                  * or if the dquot didn't exist on disk and we ask to
 573                  * allocate (ENOENT).
 574                  */
 575                 trace_xfs_dqread_fail(dqp);
 576                 goto error1;
 577         }
 578
 579         /* copy everything from disk dquot to the incore dquot */
 580         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 581         xfs_qm_dquot_logitem_init(dqp);
 582
 583         /*
 584          * Reservation counters are defined as reservation plus current usage
 585          * to avoid having to add every time.
 586          */
 587         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 588         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 589         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 590
 591         /* initialize the dquot speculative prealloc thresholds */
 592         xfs_dquot_set_prealloc_limits(dqp);
 593
 594         /* Mark the buf so that this will stay incore a little longer */
 595         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
 596
 597         /*
 598          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 599          * So we need to release with xfs_trans_brelse().
 600          * The strategy here is identical to that of inodes; we lock
 601          * the dquot in xfs_qm_dqget() before making it accessible to
 602          * others. This is because dquots, like inodes, need a good level of
 603          * concurrency, and we don't want to take locks on the entire buffers
 604          * for dquot accesses.
 605          * Note also that the dquot buffer may even be dirty at this point, if
 606          * this particular dquot was repaired. We still aren't afraid to
 607          * brelse it because we have the changes incore.
 608          */
 609         ASSERT(xfs_buf_islocked(bp));
 610         xfs_trans_brelse(tp, bp);
 611
 612         if (tp) {
 613                 error = xfs_trans_commit(tp);
 614                 if (error)
 615                         goto error0;
 616         }
 617
 618         *O_dqpp = dqp;
 619         return error;
 620
 621 error1:
 622         if (tp)
 623                 xfs_trans_cancel(tp);
 624 error0:
 625         xfs_qm_dqdestroy(dqp);
 626         *O_dqpp = NULL;
 627         return error;
 628 }
 629
 630 /*
 631  * Advance to the next id in the current chunk, or if at the
 632  * end of the chunk, skip ahead to first id in next allocated chunk
 633  * using the SEEK_DATA interface.
 634  */
 635 static int
 636 xfs_dq_get_next_id(
 637         struct xfs_mount        *mp,
 638         uint                    type,
 639         xfs_dqid_t              *id)
 640 {
 641         struct xfs_inode        *quotip = xfs_quota_inode(mp, type);
 642         xfs_dqid_t              next_id = *id + 1; /* simple advance */
 643         uint                    lock_flags;
 644         struct xfs_bmbt_irec    got;
 645         struct xfs_iext_cursor  cur;
 646         xfs_fsblock_t           start;
 647         int                     error = 0;
 648
 649         /* If we'd wrap past the max ID, stop */
 650         if (next_id < *id)
 651                 return -ENOENT;
 652
 653         /* If new ID is within the current chunk, advancing it sufficed */
 654         if (next_id % mp->m_quotainfo->qi_dqperchunk) {
 655                 *id = next_id;
 656                 return 0;
 657         }
 658
 659         /* Nope, next_id is now past the current chunk, so find the next one */
 660         start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
 661
 662         lock_flags = xfs_ilock_data_map_shared(quotip);
 663         if (!(quotip->i_df.if_flags & XFS_IFEXTENTS)) {
 664                 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
 665                 if (error)
 666                         return error;
 667         }
 668
 669         if (xfs_iext_lookup_extent(quotip, &quotip->i_df, start, &cur, &got)) {
 670                 /* contiguous chunk, bump startoff for the id calculation */
 671                 if (got.br_startoff < start)
 672                         got.br_startoff = start;
 673                 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
 674         } else {
 675                 error = -ENOENT;
 676         }
 677
 678         xfs_iunlock(quotip, lock_flags);
 679
 680         return error;
 681 }
 682
 683 /*
 684  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 685  * a locked dquot, doing an allocation (if requested) as needed.
 686  * When both an inode and an id are given, the inode's id takes precedence.
 687  * That is, if the id changes while we don't hold the ilock inside this
 688  * function, the new dquot is returned, not necessarily the one requested
 689  * in the id argument.
 690  */
 691 int
 692 xfs_qm_dqget(
 693         xfs_mount_t     *mp,
 694         xfs_inode_t     *ip,      /* locked inode (optional) */
 695         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 696         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 697         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 698         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 699 {
 700         struct xfs_quotainfo    *qi = mp->m_quotainfo;
 701         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
 702         struct xfs_dquot        *dqp;
 703         int                     error;
 704
 705         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 706         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 707             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 708             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 709                 return -ESRCH;
 710         }
 711
 712         ASSERT(type == XFS_DQ_USER ||
 713                type == XFS_DQ_PROJ ||
 714                type == XFS_DQ_GROUP);
 715         if (ip) {
 716                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 717                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
 718         }
 719
 720 restart:
 721         mutex_lock(&qi->qi_tree_lock);
 722         dqp = radix_tree_lookup(tree, id);
 723         if (dqp) {
 724                 xfs_dqlock(dqp);
 725                 if (dqp->dq_flags & XFS_DQ_FREEING) {
 726                         xfs_dqunlock(dqp);
 727                         mutex_unlock(&qi->qi_tree_lock);
 728                         trace_xfs_dqget_freeing(dqp);
 729                         delay(1);
 730                         goto restart;
 731                 }
 732
 733                 /* uninit / unused quota found in radix tree, keep looking  */
 734                 if (flags & XFS_QMOPT_DQNEXT) {
 735                         if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 736                                 xfs_dqunlock(dqp);
 737                                 mutex_unlock(&qi->qi_tree_lock);
 738                                 error = xfs_dq_get_next_id(mp, type, &id);
 739                                 if (error)
 740                                         return error;
 741                                 goto restart;
 742                         }
 743                 }
 744
 745                 dqp->q_nrefs++;
 746                 mutex_unlock(&qi->qi_tree_lock);
 747
 748                 trace_xfs_dqget_hit(dqp);
 749                 XFS_STATS_INC(mp, xs_qm_dqcachehits);
 750                 *O_dqpp = dqp;
 751                 return 0;
 752         }
 753         mutex_unlock(&qi->qi_tree_lock);
 754         XFS_STATS_INC(mp, xs_qm_dqcachemisses);
 755
 756         /*
 757          * Dquot cache miss. We don't want to keep the inode lock across
 758          * a (potential) disk read. Also we don't want to deal with the lock
 759          * ordering between quotainode and this inode. OTOH, dropping the inode
 760          * lock here means dealing with a chown that can happen before
 761          * we re-acquire the lock.
 762          */
 763         if (ip)
 764                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 765
 766         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
 767
 768         if (ip)
 769                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 770
 771         /* If we are asked to find next active id, keep looking */
 772         if (error == -ENOENT && (flags & XFS_QMOPT_DQNEXT)) {
 773                 error = xfs_dq_get_next_id(mp, type, &id);
 774                 if (!error)
 775                         goto restart;
 776         }
 777
 778         if (error)
 779                 return error;
 780
 781         if (ip) {
 782                 /*
 783                  * A dquot could be attached to this inode by now, since
 784                  * we had dropped the ilock.
 785                  */
 786                 if (xfs_this_quota_on(mp, type)) {
 787                         struct xfs_dquot        *dqp1;
 788
 789                         dqp1 = xfs_inode_dquot(ip, type);
 790                         if (dqp1) {
 791                                 xfs_qm_dqdestroy(dqp);
 792                                 dqp = dqp1;
 793                                 xfs_dqlock(dqp);
 794                                 goto dqret;
 795                         }
 796                 } else {
 797                         /* inode stays locked on return */
 798                         xfs_qm_dqdestroy(dqp);
 799                         return -ESRCH;
 800                 }
 801         }
 802
 803         mutex_lock(&qi->qi_tree_lock);
 804         error = radix_tree_insert(tree, id, dqp);
 805         if (unlikely(error)) {
 806                 WARN_ON(error != -EEXIST);
 807
 808                 /*
 809                  * Duplicate found. Just throw away the new dquot and start
 810                  * over.
 811                  */
 812                 mutex_unlock(&qi->qi_tree_lock);
 813                 trace_xfs_dqget_dup(dqp);
 814                 xfs_qm_dqdestroy(dqp);
 815                 XFS_STATS_INC(mp, xs_qm_dquot_dups);
 816                 goto restart;
 817         }
 818
 819         /*
 820          * We return a locked dquot to the caller, with a reference taken
 821          */
 822         xfs_dqlock(dqp);
 823         dqp->q_nrefs = 1;
 824
 825         qi->qi_dquots++;
 826         mutex_unlock(&qi->qi_tree_lock);
 827
 828         /* If we are asked to find next active id, keep looking */
 829         if (flags & XFS_QMOPT_DQNEXT) {
 830                 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
 831                         xfs_qm_dqput(dqp);
 832                         error = xfs_dq_get_next_id(mp, type, &id);
 833                         if (error)
 834                                 return error;
 835                         goto restart;
 836                 }
 837         }
 838
 839  dqret:
 840         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
 841         trace_xfs_dqget_miss(dqp);
 842         *O_dqpp = dqp;
 843         return 0;
 844 }
 845
 846 /*
 847  * Release a reference to the dquot (decrement ref-count) and unlock it.
 848  *
 849  * If there is a group quota attached to this dquot, carefully release that
 850  * too without tripping over deadlocks'n'stuff.
 851  */
 852 void
 853 xfs_qm_dqput(
 854         struct xfs_dquot        *dqp)
 855 {
 856         ASSERT(dqp->q_nrefs > 0);
 857         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 858
 859         trace_xfs_dqput(dqp);
 860
 861         if (--dqp->q_nrefs == 0) {
 862                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
 863                 trace_xfs_dqput_free(dqp);
 864
 865                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
 866                         XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
 867         }
 868         xfs_dqunlock(dqp);
 869 }
 870
 871 /*
 872  * Release a dquot. Flush it if dirty, then dqput() it.
 873  * dquot must not be locked.
 874  */
 875 void
 876 xfs_qm_dqrele(
 877         xfs_dquot_t     *dqp)
 878 {
 879         if (!dqp)
 880                 return;
 881
 882         trace_xfs_dqrele(dqp);
 883
 884         xfs_dqlock(dqp);
 885         /*
 886          * We don't care to flush it if the dquot is dirty here.
 887          * That will create stutters that we want to avoid.
 888          * Instead we do a delayed write when we try to reclaim
 889          * a dirty dquot. Also xfs_sync will take part of the burden...
 890          */
 891         xfs_qm_dqput(dqp);
 892 }
 893
 894 /*
 895  * This is the dquot flushing I/O completion routine.  It is called
 896  * from interrupt level when the buffer containing the dquot is
 897  * flushed to disk.  It is responsible for removing the dquot logitem
 898  * from the AIL if it has not been re-logged, and unlocking the dquot's
 899  * flush lock. This behavior is very similar to that of inodes..
 900  */
 901 STATIC void
 902 xfs_qm_dqflush_done(
 903         struct xfs_buf          *bp,
 904         struct xfs_log_item     *lip)
 905 {
 906         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
 907         xfs_dquot_t             *dqp = qip->qli_dquot;
 908         struct xfs_ail          *ailp = lip->li_ailp;
 909
 910         /*
 911          * We only want to pull the item from the AIL if its
 912          * location in the log has not changed since we started the flush.
 913          * Thus, we only bother if the dquot's lsn has
 914          * not changed. First we check the lsn outside the lock
 915          * since it's cheaper, and then we recheck while
 916          * holding the lock before removing the dquot from the AIL.
 917          */
 918         if ((lip->li_flags & XFS_LI_IN_AIL) &&
 919             ((lip->li_lsn == qip->qli_flush_lsn) ||
 920              (lip->li_flags & XFS_LI_FAILED))) {
 921
 922                 /* xfs_trans_ail_delete() drops the AIL lock. */
 923                 spin_lock(&ailp->xa_lock);
 924                 if (lip->li_lsn == qip->qli_flush_lsn) {
 925                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
 926                 } else {
 927                         /*
 928                          * Clear the failed state since we are about to drop the
 929                          * flush lock
 930                          */
 931                         if (lip->li_flags & XFS_LI_FAILED)
 932                                 xfs_clear_li_failed(lip);
 933                         spin_unlock(&ailp->xa_lock);
 934                 }
 935         }
 936
 937         /*
 938          * Release the dq's flush lock since we're done with it.
 939          */
 940         xfs_dqfunlock(dqp);
 941 }
 942
 943 /*
 944  * Write a modified dquot to disk.
 945  * The dquot must be locked and the flush lock too taken by caller.
 946  * The flush lock will not be unlocked until the dquot reaches the disk,
 947  * but the dquot is free to be unlocked and modified by the caller
 948  * in the interim. Dquot is still locked on return. This behavior is
 949  * identical to that of inodes.
 950  */
 951 int
 952 xfs_qm_dqflush(
 953         struct xfs_dquot        *dqp,
 954         struct xfs_buf          **bpp)
 955 {
 956         struct xfs_mount        *mp = dqp->q_mount;
 957         struct xfs_buf          *bp;
 958         struct xfs_disk_dquot   *ddqp;
 959         xfs_failaddr_t          fa;
 960         int                     error;
 961
 962         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 963         ASSERT(!completion_done(&dqp->q_flush));
 964
 965         trace_xfs_dqflush(dqp);
 966
 967         *bpp = NULL;
 968
 969         xfs_qm_dqunpin_wait(dqp);
 970
 971         /*
 972          * This may have been unpinned because the filesystem is shutting
 973          * down forcibly. If that's the case we must not write this dquot
 974          * to disk, because the log record didn't make it to disk.
 975          *
 976          * We also have to remove the log item from the AIL in this case,
 977          * as we wait for an emptry AIL as part of the unmount process.
 978          */
 979         if (XFS_FORCED_SHUTDOWN(mp)) {
 980                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
 981                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
 982
 983                 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
 984
 985                 error = -EIO;
 986                 goto out_unlock;
 987         }
 988
 989         /*
 990          * Get the buffer containing the on-disk dquot
 991          */
 992         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
 993                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
 994                                    &xfs_dquot_buf_ops);
 995         if (error)
 996                 goto out_unlock;
 997
 998         /*
 999          * Calculate the location of the dquot inside the buffer.
1000          */
1001         ddqp = bp->b_addr + dqp->q_bufoffset;
1002
1003         /*
1004          * A simple sanity check in case we got a corrupted dquot..
1005          */
1006         fa = xfs_dquot_verify(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0, 0);
1007         if (fa) {
1008                 xfs_alert(mp, "corrupt dquot ID 0x%x in memory at %pS",
1009                                 be32_to_cpu(ddqp->d_id), fa);
1010                 xfs_buf_relse(bp);
1011                 xfs_dqfunlock(dqp);
1012                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1013                 return -EIO;
1014         }
1015
1016         /* This is the only portion of data that needs to persist */
1017         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1018
1019         /*
1020          * Clear the dirty field and remember the flush lsn for later use.
1021          */
1022         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1023
1024         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1025                                         &dqp->q_logitem.qli_item.li_lsn);
1026
1027         /*
1028          * copy the lsn into the on-disk dquot now while we have the in memory
1029          * dquot here. This can't be done later in the write verifier as we
1030          * can't get access to the log item at that point in time.
1031          *
1032          * We also calculate the CRC here so that the on-disk dquot in the
1033          * buffer always has a valid CRC. This ensures there is no possibility
1034          * of a dquot without an up-to-date CRC getting to disk.
1035          */
1036         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1037                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1038
1039                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1040                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1041                                  XFS_DQUOT_CRC_OFF);
1042         }
1043
1044         /*
1045          * Attach an iodone routine so that we can remove this dquot from the
1046          * AIL and release the flush lock once the dquot is synced to disk.
1047          */
1048         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1049                                   &dqp->q_logitem.qli_item);
1050
1051         /*
1052          * If the buffer is pinned then push on the log so we won't
1053          * get stuck waiting in the write for too long.
1054          */
1055         if (xfs_buf_ispinned(bp)) {
1056                 trace_xfs_dqflush_force(dqp);
1057                 xfs_log_force(mp, 0);
1058         }
1059
1060         trace_xfs_dqflush_done(dqp);
1061         *bpp = bp;
1062         return 0;
1063
1064 out_unlock:
1065         xfs_dqfunlock(dqp);
1066         return -EIO;
1067 }
1068
1069 /*
1070  * Lock two xfs_dquot structures.
1071  *
1072  * To avoid deadlocks we always lock the quota structure with
1073  * the lowerd id first.
1074  */
1075 void
1076 xfs_dqlock2(
1077         xfs_dquot_t     *d1,
1078         xfs_dquot_t     *d2)
1079 {
1080         if (d1 && d2) {
1081                 ASSERT(d1 != d2);
1082                 if (be32_to_cpu(d1->q_core.d_id) >
1083                     be32_to_cpu(d2->q_core.d_id)) {
1084                         mutex_lock(&d2->q_qlock);
1085                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1086                 } else {
1087                         mutex_lock(&d1->q_qlock);
1088                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1089                 }
1090         } else if (d1) {
1091                 mutex_lock(&d1->q_qlock);
1092         } else if (d2) {
1093                 mutex_lock(&d2->q_qlock);
1094         }
1095 }
1096
1097 int __init
1098 xfs_qm_init(void)
1099 {
1100         xfs_qm_dqzone =
1101                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1102         if (!xfs_qm_dqzone)
1103                 goto out;
1104
1105         xfs_qm_dqtrxzone =
1106                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1107         if (!xfs_qm_dqtrxzone)
1108                 goto out_free_dqzone;
1109
1110         return 0;
1111
1112 out_free_dqzone:
1113         kmem_zone_destroy(xfs_qm_dqzone);
1114 out:
1115         return -ENOMEM;
1116 }
1117
1118 void
1119 xfs_qm_exit(void)
1120 {
1121         kmem_zone_destroy(xfs_qm_dqtrxzone);
1122         kmem_zone_destroy(xfs_qm_dqzone);
1123 }