fs/xfs/quota/xfs_dquot.c

   1 /*
   2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_dir2.h"
  27 #include "xfs_alloc.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_quota.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_btree.h"
  39 #include "xfs_ialloc.h"
  40 #include "xfs_bmap.h"
  41 #include "xfs_rtalloc.h"
  42 #include "xfs_error.h"
  43 #include "xfs_itable.h"
  44 #include "xfs_rw.h"
  45 #include "xfs_attr.h"
  46 #include "xfs_buf_item.h"
  47 #include "xfs_trans_space.h"
  48 #include "xfs_trans_priv.h"
  49 #include "xfs_qm.h"
  50
  51
  52 /*
  53    LOCK ORDER
  54
  55    inode lock               (ilock)
  56    dquot hash-chain lock    (hashlock)
  57    xqm dquot freelist lock  (freelistlock
  58    mount's dquot list lock  (mplistlock)
  59    user dquot lock - lock ordering among dquots is based on the uid or gid
  60    group dquot lock - similar to udquots. Between the two dquots, the udquot
  61                       has to be locked first.
  62    pin lock - the dquot lock must be held to take this lock.
  63    flush lock - ditto.
  64 */
  65
  66 STATIC void             xfs_qm_dqflush_done(xfs_buf_t *, xfs_dq_logitem_t *);
  67
  68 #ifdef DEBUG
  69 xfs_buftarg_t *xfs_dqerror_target;
  70 int xfs_do_dqerror;
  71 int xfs_dqreq_num;
  72 int xfs_dqerror_mod = 33;
  73 #endif
  74
  75 static struct lock_class_key xfs_dquot_other_class;
  76
  77 /*
  78  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  79  * we try to reclaim a free dquot if the number of incore dquots are above
  80  * a threshold.
  81  * The only field inside the core that gets initialized at this point
  82  * is the d_id field. The idea is to fill in the entire q_core
  83  * when we read in the on disk dquot.
  84  */
  85 STATIC xfs_dquot_t *
  86 xfs_qm_dqinit(
  87         xfs_mount_t  *mp,
  88         xfs_dqid_t   id,
  89         uint         type)
  90 {
  91         xfs_dquot_t     *dqp;
  92         boolean_t       brandnewdquot;
  93
  94         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  95         dqp->dq_flags = type;
  96         dqp->q_core.d_id = cpu_to_be32(id);
  97         dqp->q_mount = mp;
  98
  99         /*
 100          * No need to re-initialize these if this is a reclaimed dquot.
 101          */
 102         if (brandnewdquot) {
 103                 dqp->dq_flnext = dqp->dq_flprev = dqp;
 104                 mutex_init(&dqp->q_qlock);
 105                 init_waitqueue_head(&dqp->q_pinwait);
 106
 107                 /*
 108                  * Because we want to use a counting completion, complete
 109                  * the flush completion once to allow a single access to
 110                  * the flush completion without blocking.
 111                  */
 112                 init_completion(&dqp->q_flush);
 113                 complete(&dqp->q_flush);
 114
 115 #ifdef XFS_DQUOT_TRACE
 116                 dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_NOFS);
 117                 xfs_dqtrace_entry(dqp, "DQINIT");
 118 #endif
 119         } else {
 120                 /*
 121                  * Only the q_core portion was zeroed in dqreclaim_one().
 122                  * So, we need to reset others.
 123                  */
 124                  dqp->q_nrefs = 0;
 125                  dqp->q_blkno = 0;
 126                  dqp->MPL_NEXT = dqp->HL_NEXT = NULL;
 127                  dqp->HL_PREVP = dqp->MPL_PREVP = NULL;
 128                  dqp->q_bufoffset = 0;
 129                  dqp->q_fileoffset = 0;
 130                  dqp->q_transp = NULL;
 131                  dqp->q_gdquot = NULL;
 132                  dqp->q_res_bcount = 0;
 133                  dqp->q_res_icount = 0;
 134                  dqp->q_res_rtbcount = 0;
 135                  atomic_set(&dqp->q_pincount, 0);
 136                  dqp->q_hash = NULL;
 137                  ASSERT(dqp->dq_flnext == dqp->dq_flprev);
 138
 139 #ifdef XFS_DQUOT_TRACE
 140                  ASSERT(dqp->q_trace);
 141                  xfs_dqtrace_entry(dqp, "DQRECLAIMED_INIT");
 142 #endif
 143         }
 144
 145         /*
 146          * In either case we need to make sure group quotas have a different
 147          * lock class than user quotas, to make sure lockdep knows we can
 148          * locks of one of each at the same time.
 149          */
 150         if (!(type & XFS_DQ_USER))
 151                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 152
 153         /*
 154          * log item gets initialized later
 155          */
 156         return (dqp);
 157 }
 158
 159 /*
 160  * This is called to free all the memory associated with a dquot
 161  */
 162 void
 163 xfs_qm_dqdestroy(
 164         xfs_dquot_t     *dqp)
 165 {
 166         ASSERT(! XFS_DQ_IS_ON_FREELIST(dqp));
 167
 168         mutex_destroy(&dqp->q_qlock);
 169         sv_destroy(&dqp->q_pinwait);
 170
 171 #ifdef XFS_DQUOT_TRACE
 172         if (dqp->q_trace)
 173              ktrace_free(dqp->q_trace);
 174         dqp->q_trace = NULL;
 175 #endif
 176         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 177         atomic_dec(&xfs_Gqm->qm_totaldquots);
 178 }
 179
 180 /*
 181  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 182  */
 183 STATIC void
 184 xfs_qm_dqinit_core(
 185         xfs_dqid_t      id,
 186         uint            type,
 187         xfs_dqblk_t     *d)
 188 {
 189         /*
 190          * Caller has zero'd the entire dquot 'chunk' already.
 191          */
 192         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 193         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 194         d->dd_diskdq.d_id = cpu_to_be32(id);
 195         d->dd_diskdq.d_flags = type;
 196 }
 197
 198
 199 #ifdef XFS_DQUOT_TRACE
 200 /*
 201  * Dquot tracing for debugging.
 202  */
 203 /* ARGSUSED */
 204 void
 205 __xfs_dqtrace_entry(
 206         xfs_dquot_t     *dqp,
 207         char            *func,
 208         void            *retaddr,
 209         xfs_inode_t     *ip)
 210 {
 211         xfs_dquot_t     *udqp = NULL;
 212         xfs_ino_t       ino = 0;
 213
 214         ASSERT(dqp->q_trace);
 215         if (ip) {
 216                 ino = ip->i_ino;
 217                 udqp = ip->i_udquot;
 218         }
 219         ktrace_enter(dqp->q_trace,
 220                      (void *)(__psint_t)DQUOT_KTRACE_ENTRY,
 221                      (void *)func,
 222                      (void *)(__psint_t)dqp->q_nrefs,
 223                      (void *)(__psint_t)dqp->dq_flags,
 224                      (void *)(__psint_t)dqp->q_res_bcount,
 225                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_bcount),
 226                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_icount),
 227                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_hardlimit),
 228                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_softlimit),
 229                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_hardlimit),
 230                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_softlimit),
 231                      (void *)(__psint_t)be32_to_cpu(dqp->q_core.d_id),
 232                      (void *)(__psint_t)current_pid(),
 233                      (void *)(__psint_t)ino,
 234                      (void *)(__psint_t)retaddr,
 235                      (void *)(__psint_t)udqp);
 236         return;
 237 }
 238 #endif
 239
 240
 241 /*
 242  * If default limits are in force, push them into the dquot now.
 243  * We overwrite the dquot limits only if they are zero and this
 244  * is not the root dquot.
 245  */
 246 void
 247 xfs_qm_adjust_dqlimits(
 248         xfs_mount_t             *mp,
 249         xfs_disk_dquot_t        *d)
 250 {
 251         xfs_quotainfo_t         *q = mp->m_quotainfo;
 252
 253         ASSERT(d->d_id);
 254
 255         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 256                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 257         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 258                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 259         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 260                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 261         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 262                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 263         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 264                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 265         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 266                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 267 }
 268
 269 /*
 270  * Check the limits and timers of a dquot and start or reset timers
 271  * if necessary.
 272  * This gets called even when quota enforcement is OFF, which makes our
 273  * life a little less complicated. (We just don't reject any quota
 274  * reservations in that case, when enforcement is off).
 275  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 276  * enforcement's off.
 277  * In contrast, warnings are a little different in that they don't
 278  * 'automatically' get started when limits get exceeded.  They do
 279  * get reset to zero, however, when we find the count to be under
 280  * the soft limit (they are only ever set non-zero via userspace).
 281  */
 282 void
 283 xfs_qm_adjust_dqtimers(
 284         xfs_mount_t             *mp,
 285         xfs_disk_dquot_t        *d)
 286 {
 287         ASSERT(d->d_id);
 288
 289 #ifdef QUOTADEBUG
 290         if (d->d_blk_hardlimit)
 291                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 292                        be64_to_cpu(d->d_blk_hardlimit));
 293         if (d->d_ino_hardlimit)
 294                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 295                        be64_to_cpu(d->d_ino_hardlimit));
 296         if (d->d_rtb_hardlimit)
 297                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 298                        be64_to_cpu(d->d_rtb_hardlimit));
 299 #endif
 300         if (!d->d_btimer) {
 301                 if ((d->d_blk_softlimit &&
 302                      (be64_to_cpu(d->d_bcount) >=
 303                       be64_to_cpu(d->d_blk_softlimit))) ||
 304                     (d->d_blk_hardlimit &&
 305                      (be64_to_cpu(d->d_bcount) >=
 306                       be64_to_cpu(d->d_blk_hardlimit)))) {
 307                         d->d_btimer = cpu_to_be32(get_seconds() +
 308                                         XFS_QI_BTIMELIMIT(mp));
 309                 } else {
 310                         d->d_bwarns = 0;
 311                 }
 312         } else {
 313                 if ((!d->d_blk_softlimit ||
 314                      (be64_to_cpu(d->d_bcount) <
 315                       be64_to_cpu(d->d_blk_softlimit))) &&
 316                     (!d->d_blk_hardlimit ||
 317                     (be64_to_cpu(d->d_bcount) <
 318                      be64_to_cpu(d->d_blk_hardlimit)))) {
 319                         d->d_btimer = 0;
 320                 }
 321         }
 322
 323         if (!d->d_itimer) {
 324                 if ((d->d_ino_softlimit &&
 325                      (be64_to_cpu(d->d_icount) >=
 326                       be64_to_cpu(d->d_ino_softlimit))) ||
 327                     (d->d_ino_hardlimit &&
 328                      (be64_to_cpu(d->d_icount) >=
 329                       be64_to_cpu(d->d_ino_hardlimit)))) {
 330                         d->d_itimer = cpu_to_be32(get_seconds() +
 331                                         XFS_QI_ITIMELIMIT(mp));
 332                 } else {
 333                         d->d_iwarns = 0;
 334                 }
 335         } else {
 336                 if ((!d->d_ino_softlimit ||
 337                      (be64_to_cpu(d->d_icount) <
 338                       be64_to_cpu(d->d_ino_softlimit)))  &&
 339                     (!d->d_ino_hardlimit ||
 340                      (be64_to_cpu(d->d_icount) <
 341                       be64_to_cpu(d->d_ino_hardlimit)))) {
 342                         d->d_itimer = 0;
 343                 }
 344         }
 345
 346         if (!d->d_rtbtimer) {
 347                 if ((d->d_rtb_softlimit &&
 348                      (be64_to_cpu(d->d_rtbcount) >=
 349                       be64_to_cpu(d->d_rtb_softlimit))) ||
 350                     (d->d_rtb_hardlimit &&
 351                      (be64_to_cpu(d->d_rtbcount) >=
 352                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 353                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 354                                         XFS_QI_RTBTIMELIMIT(mp));
 355                 } else {
 356                         d->d_rtbwarns = 0;
 357                 }
 358         } else {
 359                 if ((!d->d_rtb_softlimit ||
 360                      (be64_to_cpu(d->d_rtbcount) <
 361                       be64_to_cpu(d->d_rtb_softlimit))) &&
 362                     (!d->d_rtb_hardlimit ||
 363                      (be64_to_cpu(d->d_rtbcount) <
 364                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 365                         d->d_rtbtimer = 0;
 366                 }
 367         }
 368 }
 369
 370 /*
 371  * initialize a buffer full of dquots and log the whole thing
 372  */
 373 STATIC void
 374 xfs_qm_init_dquot_blk(
 375         xfs_trans_t     *tp,
 376         xfs_mount_t     *mp,
 377         xfs_dqid_t      id,
 378         uint            type,
 379         xfs_buf_t       *bp)
 380 {
 381         xfs_dqblk_t     *d;
 382         int             curid, i;
 383
 384         ASSERT(tp);
 385         ASSERT(XFS_BUF_ISBUSY(bp));
 386         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 387
 388         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 389
 390         /*
 391          * ID of the first dquot in the block - id's are zero based.
 392          */
 393         curid = id - (id % XFS_QM_DQPERBLK(mp));
 394         ASSERT(curid >= 0);
 395         memset(d, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)));
 396         for (i = 0; i < XFS_QM_DQPERBLK(mp); i++, d++, curid++)
 397                 xfs_qm_dqinit_core(curid, type, d);
 398         xfs_trans_dquot_buf(tp, bp,
 399                             (type & XFS_DQ_USER ? XFS_BLI_UDQUOT_BUF :
 400                             ((type & XFS_DQ_PROJ) ? XFS_BLI_PDQUOT_BUF :
 401                              XFS_BLI_GDQUOT_BUF)));
 402         xfs_trans_log_buf(tp, bp, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1);
 403 }
 404
 405
 406
 407 /*
 408  * Allocate a block and fill it with dquots.
 409  * This is called when the bmapi finds a hole.
 410  */
 411 STATIC int
 412 xfs_qm_dqalloc(
 413         xfs_trans_t     **tpp,
 414         xfs_mount_t     *mp,
 415         xfs_dquot_t     *dqp,
 416         xfs_inode_t     *quotip,
 417         xfs_fileoff_t   offset_fsb,
 418         xfs_buf_t       **O_bpp)
 419 {
 420         xfs_fsblock_t   firstblock;
 421         xfs_bmap_free_t flist;
 422         xfs_bmbt_irec_t map;
 423         int             nmaps, error, committed;
 424         xfs_buf_t       *bp;
 425         xfs_trans_t     *tp = *tpp;
 426
 427         ASSERT(tp != NULL);
 428         xfs_dqtrace_entry(dqp, "DQALLOC");
 429
 430         /*
 431          * Initialize the bmap freelist prior to calling bmapi code.
 432          */
 433         xfs_bmap_init(&flist, &firstblock);
 434         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 435         /*
 436          * Return if this type of quotas is turned off while we didn't
 437          * have an inode lock
 438          */
 439         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 440                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 441                 return (ESRCH);
 442         }
 443
 444         /*
 445          * xfs_trans_commit normally decrements the vnode ref count
 446          * when it unlocks the inode. Since we want to keep the quota
 447          * inode around, we bump the vnode ref count now.
 448          */
 449         IHOLD(quotip);
 450
 451         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 452         nmaps = 1;
 453         if ((error = xfs_bmapi(tp, quotip,
 454                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 455                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 456                               &firstblock,
 457                               XFS_QM_DQALLOC_SPACE_RES(mp),
 458                               &map, &nmaps, &flist, NULL))) {
 459                 goto error0;
 460         }
 461         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 462         ASSERT(nmaps == 1);
 463         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 464                (map.br_startblock != HOLESTARTBLOCK));
 465
 466         /*
 467          * Keep track of the blkno to save a lookup later
 468          */
 469         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 470
 471         /* now we can just get the buffer (there's nothing to read yet) */
 472         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 473                                dqp->q_blkno,
 474                                XFS_QI_DQCHUNKLEN(mp),
 475                                0);
 476         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 477                 goto error1;
 478         /*
 479          * Make a chunk of dquots out of this buffer and log
 480          * the entire thing.
 481          */
 482         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 483                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 484
 485         /*
 486          * xfs_bmap_finish() may commit the current transaction and
 487          * start a second transaction if the freelist is not empty.
 488          *
 489          * Since we still want to modify this buffer, we need to
 490          * ensure that the buffer is not released on commit of
 491          * the first transaction and ensure the buffer is added to the
 492          * second transaction.
 493          *
 494          * If there is only one transaction then don't stop the buffer
 495          * from being released when it commits later on.
 496          */
 497
 498         xfs_trans_bhold(tp, bp);
 499
 500         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 501                 goto error1;
 502         }
 503
 504         if (committed) {
 505                 tp = *tpp;
 506                 xfs_trans_bjoin(tp, bp);
 507         } else {
 508                 xfs_trans_bhold_release(tp, bp);
 509         }
 510
 511         *O_bpp = bp;
 512         return 0;
 513
 514       error1:
 515         xfs_bmap_cancel(&flist);
 516       error0:
 517         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 518
 519         return (error);
 520 }
 521
 522 /*
 523  * Maps a dquot to the buffer containing its on-disk version.
 524  * This returns a ptr to the buffer containing the on-disk dquot
 525  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 526  */
 527 STATIC int
 528 xfs_qm_dqtobp(
 529         xfs_trans_t             **tpp,
 530         xfs_dquot_t             *dqp,
 531         xfs_disk_dquot_t        **O_ddpp,
 532         xfs_buf_t               **O_bpp,
 533         uint                    flags)
 534 {
 535         xfs_bmbt_irec_t map;
 536         int             nmaps, error;
 537         xfs_buf_t       *bp;
 538         xfs_inode_t     *quotip;
 539         xfs_mount_t     *mp;
 540         xfs_disk_dquot_t *ddq;
 541         xfs_dqid_t      id;
 542         boolean_t       newdquot;
 543         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 544
 545         mp = dqp->q_mount;
 546         id = be32_to_cpu(dqp->q_core.d_id);
 547         nmaps = 1;
 548         newdquot = B_FALSE;
 549
 550         /*
 551          * If we don't know where the dquot lives, find out.
 552          */
 553         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 554                 /* We use the id as an index */
 555                 dqp->q_fileoffset = (xfs_fileoff_t)id / XFS_QM_DQPERBLK(mp);
 556                 nmaps = 1;
 557                 quotip = XFS_DQ_TO_QIP(dqp);
 558                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 559                 /*
 560                  * Return if this type of quotas is turned off while we didn't
 561                  * have an inode lock
 562                  */
 563                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 564                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 565                         return (ESRCH);
 566                 }
 567                 /*
 568                  * Find the block map; no allocations yet
 569                  */
 570                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 571                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 572                                   XFS_BMAPI_METADATA,
 573                                   NULL, 0, &map, &nmaps, NULL, NULL);
 574
 575                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 576                 if (error)
 577                         return (error);
 578                 ASSERT(nmaps == 1);
 579                 ASSERT(map.br_blockcount == 1);
 580
 581                 /*
 582                  * offset of dquot in the (fixed sized) dquot chunk.
 583                  */
 584                 dqp->q_bufoffset = (id % XFS_QM_DQPERBLK(mp)) *
 585                         sizeof(xfs_dqblk_t);
 586                 if (map.br_startblock == HOLESTARTBLOCK) {
 587                         /*
 588                          * We don't allocate unless we're asked to
 589                          */
 590                         if (!(flags & XFS_QMOPT_DQALLOC))
 591                                 return (ENOENT);
 592
 593                         ASSERT(tp);
 594                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 595                                                 dqp->q_fileoffset, &bp)))
 596                                 return (error);
 597                         tp = *tpp;
 598                         newdquot = B_TRUE;
 599                 } else {
 600                         /*
 601                          * store the blkno etc so that we don't have to do the
 602                          * mapping all the time
 603                          */
 604                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 605                 }
 606         }
 607         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 608         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 609
 610         /*
 611          * Read in the buffer, unless we've just done the allocation
 612          * (in which case we already have the buf).
 613          */
 614         if (! newdquot) {
 615                 xfs_dqtrace_entry(dqp, "DQTOBP READBUF");
 616                 if ((error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 617                                                dqp->q_blkno,
 618                                                XFS_QI_DQCHUNKLEN(mp),
 619                                                0, &bp))) {
 620                         return (error);
 621                 }
 622                 if (error || !bp)
 623                         return XFS_ERROR(error);
 624         }
 625         ASSERT(XFS_BUF_ISBUSY(bp));
 626         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 627
 628         /*
 629          * calculate the location of the dquot inside the buffer.
 630          */
 631         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 632
 633         /*
 634          * A simple sanity check in case we got a corrupted dquot...
 635          */
 636         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 637                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 638                            "dqtobp")) {
 639                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 640                         xfs_trans_brelse(tp, bp);
 641                         return XFS_ERROR(EIO);
 642                 }
 643                 XFS_BUF_BUSY(bp); /* We dirtied this */
 644         }
 645
 646         *O_bpp = bp;
 647         *O_ddpp = ddq;
 648
 649         return (0);
 650 }
 651
 652
 653 /*
 654  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 655  * and release the buffer immediately.
 656  *
 657  */
 658 /* ARGSUSED */
 659 STATIC int
 660 xfs_qm_dqread(
 661         xfs_trans_t     **tpp,
 662         xfs_dqid_t      id,
 663         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 664         uint            flags)
 665 {
 666         xfs_disk_dquot_t *ddqp;
 667         xfs_buf_t        *bp;
 668         int              error;
 669         xfs_trans_t      *tp;
 670
 671         ASSERT(tpp);
 672
 673         /*
 674          * get a pointer to the on-disk dquot and the buffer containing it
 675          * dqp already knows its own type (GROUP/USER).
 676          */
 677         xfs_dqtrace_entry(dqp, "DQREAD");
 678         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 679                 return (error);
 680         }
 681         tp = *tpp;
 682
 683         /* copy everything from disk dquot to the incore dquot */
 684         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 685         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 686         xfs_qm_dquot_logitem_init(dqp);
 687
 688         /*
 689          * Reservation counters are defined as reservation plus current usage
 690          * to avoid having to add everytime.
 691          */
 692         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 693         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 694         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 695
 696         /* Mark the buf so that this will stay incore a little longer */
 697         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 698
 699         /*
 700          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 701          * So we need to release with xfs_trans_brelse().
 702          * The strategy here is identical to that of inodes; we lock
 703          * the dquot in xfs_qm_dqget() before making it accessible to
 704          * others. This is because dquots, like inodes, need a good level of
 705          * concurrency, and we don't want to take locks on the entire buffers
 706          * for dquot accesses.
 707          * Note also that the dquot buffer may even be dirty at this point, if
 708          * this particular dquot was repaired. We still aren't afraid to
 709          * brelse it because we have the changes incore.
 710          */
 711         ASSERT(XFS_BUF_ISBUSY(bp));
 712         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 713         xfs_trans_brelse(tp, bp);
 714
 715         return (error);
 716 }
 717
 718
 719 /*
 720  * allocate an incore dquot from the kernel heap,
 721  * and fill its core with quota information kept on disk.
 722  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 723  * if it wasn't already allocated.
 724  */
 725 STATIC int
 726 xfs_qm_idtodq(
 727         xfs_mount_t     *mp,
 728         xfs_dqid_t      id,      /* gid or uid, depending on type */
 729         uint            type,    /* UDQUOT or GDQUOT */
 730         uint            flags,   /* DQALLOC, DQREPAIR */
 731         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 732 {
 733         xfs_dquot_t     *dqp;
 734         int             error;
 735         xfs_trans_t     *tp;
 736         int             cancelflags=0;
 737
 738         dqp = xfs_qm_dqinit(mp, id, type);
 739         tp = NULL;
 740         if (flags & XFS_QMOPT_DQALLOC) {
 741                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 742                 if ((error = xfs_trans_reserve(tp,
 743                                        XFS_QM_DQALLOC_SPACE_RES(mp),
 744                                        XFS_WRITE_LOG_RES(mp) +
 745                                               BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1 +
 746                                               128,
 747                                        0,
 748                                        XFS_TRANS_PERM_LOG_RES,
 749                                        XFS_WRITE_LOG_COUNT))) {
 750                         cancelflags = 0;
 751                         goto error0;
 752                 }
 753                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 754         }
 755
 756         /*
 757          * Read it from disk; xfs_dqread() takes care of
 758          * all the necessary initialization of dquot's fields (locks, etc)
 759          */
 760         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 761                 /*
 762                  * This can happen if quotas got turned off (ESRCH),
 763                  * or if the dquot didn't exist on disk and we ask to
 764                  * allocate (ENOENT).
 765                  */
 766                 xfs_dqtrace_entry(dqp, "DQREAD FAIL");
 767                 cancelflags |= XFS_TRANS_ABORT;
 768                 goto error0;
 769         }
 770         if (tp) {
 771                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 772                         goto error1;
 773         }
 774
 775         *O_dqpp = dqp;
 776         return (0);
 777
 778  error0:
 779         ASSERT(error);
 780         if (tp)
 781                 xfs_trans_cancel(tp, cancelflags);
 782  error1:
 783         xfs_qm_dqdestroy(dqp);
 784         *O_dqpp = NULL;
 785         return (error);
 786 }
 787
 788 /*
 789  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 790  * hashtables for user and group dquots; and, these are global tables
 791  * inside the XQM, not per-filesystem tables.
 792  * The hash chain must be locked by caller, and it is left locked
 793  * on return. Returning dquot is locked.
 794  */
 795 STATIC int
 796 xfs_qm_dqlookup(
 797         xfs_mount_t             *mp,
 798         xfs_dqid_t              id,
 799         xfs_dqhash_t            *qh,
 800         xfs_dquot_t             **O_dqpp)
 801 {
 802         xfs_dquot_t             *dqp;
 803         uint                    flist_locked;
 804         xfs_dquot_t             *d;
 805
 806         ASSERT(mutex_is_locked(&qh->qh_lock));
 807
 808         flist_locked = B_FALSE;
 809
 810         /*
 811          * Traverse the hashchain looking for a match
 812          */
 813         for (dqp = qh->qh_next; dqp != NULL; dqp = dqp->HL_NEXT) {
 814                 /*
 815                  * We already have the hashlock. We don't need the
 816                  * dqlock to look at the id field of the dquot, since the
 817                  * id can't be modified without the hashlock anyway.
 818                  */
 819                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 820                         xfs_dqtrace_entry(dqp, "DQFOUND BY LOOKUP");
 821                         /*
 822                          * All in core dquots must be on the dqlist of mp
 823                          */
 824                         ASSERT(dqp->MPL_PREVP != NULL);
 825
 826                         xfs_dqlock(dqp);
 827                         if (dqp->q_nrefs == 0) {
 828                                 ASSERT (XFS_DQ_IS_ON_FREELIST(dqp));
 829                                 if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
 830                                         xfs_dqtrace_entry(dqp, "DQLOOKUP: WANT");
 831
 832                                         /*
 833                                          * We may have raced with dqreclaim_one()
 834                                          * (and lost). So, flag that we don't
 835                                          * want the dquot to be reclaimed.
 836                                          */
 837                                         dqp->dq_flags |= XFS_DQ_WANT;
 838                                         xfs_dqunlock(dqp);
 839                                         xfs_qm_freelist_lock(xfs_Gqm);
 840                                         xfs_dqlock(dqp);
 841                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 842                                 }
 843                                 flist_locked = B_TRUE;
 844                         }
 845
 846                         /*
 847                          * id couldn't have changed; we had the hashlock all
 848                          * along
 849                          */
 850                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 851
 852                         if (flist_locked) {
 853                                 if (dqp->q_nrefs != 0) {
 854                                         xfs_qm_freelist_unlock(xfs_Gqm);
 855                                         flist_locked = B_FALSE;
 856                                 } else {
 857                                         /*
 858                                          * take it off the freelist
 859                                          */
 860                                         xfs_dqtrace_entry(dqp,
 861                                                         "DQLOOKUP: TAKEOFF FL");
 862                                         XQM_FREELIST_REMOVE(dqp);
 863                                         /* xfs_qm_freelist_print(&(xfs_Gqm->
 864                                                         qm_dqfreelist),
 865                                                         "after removal"); */
 866                                 }
 867                         }
 868
 869                         /*
 870                          * grab a reference
 871                          */
 872                         XFS_DQHOLD(dqp);
 873
 874                         if (flist_locked)
 875                                 xfs_qm_freelist_unlock(xfs_Gqm);
 876                         /*
 877                          * move the dquot to the front of the hashchain
 878                          */
 879                         ASSERT(mutex_is_locked(&qh->qh_lock));
 880                         if (dqp->HL_PREVP != &qh->qh_next) {
 881                                 xfs_dqtrace_entry(dqp,
 882                                                   "DQLOOKUP: HASH MOVETOFRONT");
 883                                 if ((d = dqp->HL_NEXT))
 884                                         d->HL_PREVP = dqp->HL_PREVP;
 885                                 *(dqp->HL_PREVP) = d;
 886                                 d = qh->qh_next;
 887                                 d->HL_PREVP = &dqp->HL_NEXT;
 888                                 dqp->HL_NEXT = d;
 889                                 dqp->HL_PREVP = &qh->qh_next;
 890                                 qh->qh_next = dqp;
 891                         }
 892                         xfs_dqtrace_entry(dqp, "LOOKUP END");
 893                         *O_dqpp = dqp;
 894                         ASSERT(mutex_is_locked(&qh->qh_lock));
 895                         return (0);
 896                 }
 897         }
 898
 899         *O_dqpp = NULL;
 900         ASSERT(mutex_is_locked(&qh->qh_lock));
 901         return (1);
 902 }
 903
 904 /*
 905  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 906  * a locked dquot, doing an allocation (if requested) as needed.
 907  * When both an inode and an id are given, the inode's id takes precedence.
 908  * That is, if the id changes while we don't hold the ilock inside this
 909  * function, the new dquot is returned, not necessarily the one requested
 910  * in the id argument.
 911  */
 912 int
 913 xfs_qm_dqget(
 914         xfs_mount_t     *mp,
 915         xfs_inode_t     *ip,      /* locked inode (optional) */
 916         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 917         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 918         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 919         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 920 {
 921         xfs_dquot_t     *dqp;
 922         xfs_dqhash_t    *h;
 923         uint            version;
 924         int             error;
 925
 926         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 927         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 928             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 929             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 930                 return (ESRCH);
 931         }
 932         h = XFS_DQ_HASH(mp, id, type);
 933
 934 #ifdef DEBUG
 935         if (xfs_do_dqerror) {
 936                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 937                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 938                         cmn_err(CE_DEBUG, "Returning error in dqget");
 939                         return (EIO);
 940                 }
 941         }
 942 #endif
 943
 944  again:
 945
 946 #ifdef DEBUG
 947         ASSERT(type == XFS_DQ_USER ||
 948                type == XFS_DQ_PROJ ||
 949                type == XFS_DQ_GROUP);
 950         if (ip) {
 951                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 952                 if (type == XFS_DQ_USER)
 953                         ASSERT(ip->i_udquot == NULL);
 954                 else
 955                         ASSERT(ip->i_gdquot == NULL);
 956         }
 957 #endif
 958         mutex_lock(&h->qh_lock);
 959
 960         /*
 961          * Look in the cache (hashtable).
 962          * The chain is kept locked during lookup.
 963          */
 964         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 965                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 966                 /*
 967                  * The dquot was found, moved to the front of the chain,
 968                  * taken off the freelist if it was on it, and locked
 969                  * at this point. Just unlock the hashchain and return.
 970                  */
 971                 ASSERT(*O_dqpp);
 972                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 973                 mutex_unlock(&h->qh_lock);
 974                 xfs_dqtrace_entry(*O_dqpp, "DQGET DONE (FROM CACHE)");
 975                 return (0);     /* success */
 976         }
 977         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 978
 979         /*
 980          * Dquot cache miss. We don't want to keep the inode lock across
 981          * a (potential) disk read. Also we don't want to deal with the lock
 982          * ordering between quotainode and this inode. OTOH, dropping the inode
 983          * lock here means dealing with a chown that can happen before
 984          * we re-acquire the lock.
 985          */
 986         if (ip)
 987                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 988         /*
 989          * Save the hashchain version stamp, and unlock the chain, so that
 990          * we don't keep the lock across a disk read
 991          */
 992         version = h->qh_version;
 993         mutex_unlock(&h->qh_lock);
 994
 995         /*
 996          * Allocate the dquot on the kernel heap, and read the ondisk
 997          * portion off the disk. Also, do all the necessary initialization
 998          * This can return ENOENT if dquot didn't exist on disk and we didn't
 999          * ask it to allocate; ESRCH if quotas got turned off suddenly.
1000          */
1001         if ((error = xfs_qm_idtodq(mp, id, type,
1002                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
1003                                            XFS_QMOPT_DOWARN),
1004                                   &dqp))) {
1005                 if (ip)
1006                         xfs_ilock(ip, XFS_ILOCK_EXCL);
1007                 return (error);
1008         }
1009
1010         /*
1011          * See if this is mount code calling to look at the overall quota limits
1012          * which are stored in the id == 0 user or group's dquot.
1013          * Since we may not have done a quotacheck by this point, just return
1014          * the dquot without attaching it to any hashtables, lists, etc, or even
1015          * taking a reference.
1016          * The caller must dqdestroy this once done.
1017          */
1018         if (flags & XFS_QMOPT_DQSUSER) {
1019                 ASSERT(id == 0);
1020                 ASSERT(! ip);
1021                 goto dqret;
1022         }
1023
1024         /*
1025          * Dquot lock comes after hashlock in the lock ordering
1026          */
1027         if (ip) {
1028                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1029                 if (! XFS_IS_DQTYPE_ON(mp, type)) {
1030                         /* inode stays locked on return */
1031                         xfs_qm_dqdestroy(dqp);
1032                         return XFS_ERROR(ESRCH);
1033                 }
1034                 /*
1035                  * A dquot could be attached to this inode by now, since
1036                  * we had dropped the ilock.
1037                  */
1038                 if (type == XFS_DQ_USER) {
1039                         if (ip->i_udquot) {
1040                                 xfs_qm_dqdestroy(dqp);
1041                                 dqp = ip->i_udquot;
1042                                 xfs_dqlock(dqp);
1043                                 goto dqret;
1044                         }
1045                 } else {
1046                         if (ip->i_gdquot) {
1047                                 xfs_qm_dqdestroy(dqp);
1048                                 dqp = ip->i_gdquot;
1049                                 xfs_dqlock(dqp);
1050                                 goto dqret;
1051                         }
1052                 }
1053         }
1054
1055         /*
1056          * Hashlock comes after ilock in lock order
1057          */
1058         mutex_lock(&h->qh_lock);
1059         if (version != h->qh_version) {
1060                 xfs_dquot_t *tmpdqp;
1061                 /*
1062                  * Now, see if somebody else put the dquot in the
1063                  * hashtable before us. This can happen because we didn't
1064                  * keep the hashchain lock. We don't have to worry about
1065                  * lock order between the two dquots here since dqp isn't
1066                  * on any findable lists yet.
1067                  */
1068                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
1069                         /*
1070                          * Duplicate found. Just throw away the new dquot
1071                          * and start over.
1072                          */
1073                         xfs_qm_dqput(tmpdqp);
1074                         mutex_unlock(&h->qh_lock);
1075                         xfs_qm_dqdestroy(dqp);
1076                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
1077                         goto again;
1078                 }
1079         }
1080
1081         /*
1082          * Put the dquot at the beginning of the hash-chain and mp's list
1083          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1084          */
1085         ASSERT(mutex_is_locked(&h->qh_lock));
1086         dqp->q_hash = h;
1087         XQM_HASHLIST_INSERT(h, dqp);
1088
1089         /*
1090          * Attach this dquot to this filesystem's list of all dquots,
1091          * kept inside the mount structure in m_quotainfo field
1092          */
1093         xfs_qm_mplist_lock(mp);
1094
1095         /*
1096          * We return a locked dquot to the caller, with a reference taken
1097          */
1098         xfs_dqlock(dqp);
1099         dqp->q_nrefs = 1;
1100
1101         XQM_MPLIST_INSERT(&(XFS_QI_MPL_LIST(mp)), dqp);
1102
1103         xfs_qm_mplist_unlock(mp);
1104         mutex_unlock(&h->qh_lock);
1105  dqret:
1106         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1107         xfs_dqtrace_entry(dqp, "DQGET DONE");
1108         *O_dqpp = dqp;
1109         return (0);
1110 }
1111
1112
1113 /*
1114  * Release a reference to the dquot (decrement ref-count)
1115  * and unlock it. If there is a group quota attached to this
1116  * dquot, carefully release that too without tripping over
1117  * deadlocks'n'stuff.
1118  */
1119 void
1120 xfs_qm_dqput(
1121         xfs_dquot_t     *dqp)
1122 {
1123         xfs_dquot_t     *gdqp;
1124
1125         ASSERT(dqp->q_nrefs > 0);
1126         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1127         xfs_dqtrace_entry(dqp, "DQPUT");
1128
1129         if (dqp->q_nrefs != 1) {
1130                 dqp->q_nrefs--;
1131                 xfs_dqunlock(dqp);
1132                 return;
1133         }
1134
1135         /*
1136          * drop the dqlock and acquire the freelist and dqlock
1137          * in the right order; but try to get it out-of-order first
1138          */
1139         if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
1140                 xfs_dqtrace_entry(dqp, "DQPUT: FLLOCK-WAIT");
1141                 xfs_dqunlock(dqp);
1142                 xfs_qm_freelist_lock(xfs_Gqm);
1143                 xfs_dqlock(dqp);
1144         }
1145
1146         while (1) {
1147                 gdqp = NULL;
1148
1149                 /* We can't depend on nrefs being == 1 here */
1150                 if (--dqp->q_nrefs == 0) {
1151                         xfs_dqtrace_entry(dqp, "DQPUT: ON FREELIST");
1152                         /*
1153                          * insert at end of the freelist.
1154                          */
1155                         XQM_FREELIST_INSERT(&(xfs_Gqm->qm_dqfreelist), dqp);
1156
1157                         /*
1158                          * If we just added a udquot to the freelist, then
1159                          * we want to release the gdquot reference that
1160                          * it (probably) has. Otherwise it'll keep the
1161                          * gdquot from getting reclaimed.
1162                          */
1163                         if ((gdqp = dqp->q_gdquot)) {
1164                                 /*
1165                                  * Avoid a recursive dqput call
1166                                  */
1167                                 xfs_dqlock(gdqp);
1168                                 dqp->q_gdquot = NULL;
1169                         }
1170
1171                         /* xfs_qm_freelist_print(&(xfs_Gqm->qm_dqfreelist),
1172                            "@@@@@++ Free list (after append) @@@@@+");
1173                            */
1174                 }
1175                 xfs_dqunlock(dqp);
1176
1177                 /*
1178                  * If we had a group quota inside the user quota as a hint,
1179                  * release it now.
1180                  */
1181                 if (! gdqp)
1182                         break;
1183                 dqp = gdqp;
1184         }
1185         xfs_qm_freelist_unlock(xfs_Gqm);
1186 }
1187
1188 /*
1189  * Release a dquot. Flush it if dirty, then dqput() it.
1190  * dquot must not be locked.
1191  */
1192 void
1193 xfs_qm_dqrele(
1194         xfs_dquot_t     *dqp)
1195 {
1196         if (!dqp)
1197                 return;
1198
1199         xfs_dqtrace_entry(dqp, "DQRELE");
1200
1201         xfs_dqlock(dqp);
1202         /*
1203          * We don't care to flush it if the dquot is dirty here.
1204          * That will create stutters that we want to avoid.
1205          * Instead we do a delayed write when we try to reclaim
1206          * a dirty dquot. Also xfs_sync will take part of the burden...
1207          */
1208         xfs_qm_dqput(dqp);
1209 }
1210
1211
1212 /*
1213  * Write a modified dquot to disk.
1214  * The dquot must be locked and the flush lock too taken by caller.
1215  * The flush lock will not be unlocked until the dquot reaches the disk,
1216  * but the dquot is free to be unlocked and modified by the caller
1217  * in the interim. Dquot is still locked on return. This behavior is
1218  * identical to that of inodes.
1219  */
1220 int
1221 xfs_qm_dqflush(
1222         xfs_dquot_t             *dqp,
1223         uint                    flags)
1224 {
1225         xfs_mount_t             *mp;
1226         xfs_buf_t               *bp;
1227         xfs_disk_dquot_t        *ddqp;
1228         int                     error;
1229
1230         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1231         ASSERT(!completion_done(&dqp->q_flush));
1232         xfs_dqtrace_entry(dqp, "DQFLUSH");
1233
1234         /*
1235          * If not dirty, or it's pinned and we are not supposed to
1236          * block, nada.
1237          */
1238         if (!XFS_DQ_IS_DIRTY(dqp) ||
1239             (!(flags & XFS_QMOPT_SYNC) && atomic_read(&dqp->q_pincount) > 0)) {
1240                 xfs_dqfunlock(dqp);
1241                 return 0;
1242         }
1243         xfs_qm_dqunpin_wait(dqp);
1244
1245         /*
1246          * This may have been unpinned because the filesystem is shutting
1247          * down forcibly. If that's the case we must not write this dquot
1248          * to disk, because the log record didn't make it to disk!
1249          */
1250         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1251                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1252                 xfs_dqfunlock(dqp);
1253                 return XFS_ERROR(EIO);
1254         }
1255
1256         /*
1257          * Get the buffer containing the on-disk dquot
1258          * We don't need a transaction envelope because we know that the
1259          * the ondisk-dquot has already been allocated for.
1260          */
1261         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1262                 xfs_dqtrace_entry(dqp, "DQTOBP FAIL");
1263                 ASSERT(error != ENOENT);
1264                 /*
1265                  * Quotas could have gotten turned off (ESRCH)
1266                  */
1267                 xfs_dqfunlock(dqp);
1268                 return (error);
1269         }
1270
1271         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1272                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1273                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1274                 return XFS_ERROR(EIO);
1275         }
1276
1277         /* This is the only portion of data that needs to persist */
1278         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1279
1280         /*
1281          * Clear the dirty field and remember the flush lsn for later use.
1282          */
1283         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1284         mp = dqp->q_mount;
1285
1286         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1287                                         &dqp->q_logitem.qli_item.li_lsn);
1288
1289         /*
1290          * Attach an iodone routine so that we can remove this dquot from the
1291          * AIL and release the flush lock once the dquot is synced to disk.
1292          */
1293         xfs_buf_attach_iodone(bp, (void(*)(xfs_buf_t *, xfs_log_item_t *))
1294                               xfs_qm_dqflush_done, &(dqp->q_logitem.qli_item));
1295         /*
1296          * If the buffer is pinned then push on the log so we won't
1297          * get stuck waiting in the write for too long.
1298          */
1299         if (XFS_BUF_ISPINNED(bp)) {
1300                 xfs_dqtrace_entry(dqp, "DQFLUSH LOG FORCE");
1301                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
1302         }
1303
1304         if (flags & XFS_QMOPT_DELWRI) {
1305                 xfs_bdwrite(mp, bp);
1306         } else if (flags & XFS_QMOPT_ASYNC) {
1307                 error = xfs_bawrite(mp, bp);
1308         } else {
1309                 error = xfs_bwrite(mp, bp);
1310         }
1311         xfs_dqtrace_entry(dqp, "DQFLUSH END");
1312         /*
1313          * dqp is still locked, but caller is free to unlock it now.
1314          */
1315         return (error);
1316
1317 }
1318
1319 /*
1320  * This is the dquot flushing I/O completion routine.  It is called
1321  * from interrupt level when the buffer containing the dquot is
1322  * flushed to disk.  It is responsible for removing the dquot logitem
1323  * from the AIL if it has not been re-logged, and unlocking the dquot's
1324  * flush lock. This behavior is very similar to that of inodes..
1325  */
1326 /*ARGSUSED*/
1327 STATIC void
1328 xfs_qm_dqflush_done(
1329         xfs_buf_t               *bp,
1330         xfs_dq_logitem_t        *qip)
1331 {
1332         xfs_dquot_t             *dqp;
1333         struct xfs_ail          *ailp;
1334
1335         dqp = qip->qli_dquot;
1336         ailp = qip->qli_item.li_ailp;
1337
1338         /*
1339          * We only want to pull the item from the AIL if its
1340          * location in the log has not changed since we started the flush.
1341          * Thus, we only bother if the dquot's lsn has
1342          * not changed. First we check the lsn outside the lock
1343          * since it's cheaper, and then we recheck while
1344          * holding the lock before removing the dquot from the AIL.
1345          */
1346         if ((qip->qli_item.li_flags & XFS_LI_IN_AIL) &&
1347             qip->qli_item.li_lsn == qip->qli_flush_lsn) {
1348
1349                 /* xfs_trans_ail_delete() drops the AIL lock. */
1350                 spin_lock(&ailp->xa_lock);
1351                 if (qip->qli_item.li_lsn == qip->qli_flush_lsn)
1352                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)qip);
1353                 else
1354                         spin_unlock(&ailp->xa_lock);
1355         }
1356
1357         /*
1358          * Release the dq's flush lock since we're done with it.
1359          */
1360         xfs_dqfunlock(dqp);
1361 }
1362
1363 int
1364 xfs_qm_dqlock_nowait(
1365         xfs_dquot_t *dqp)
1366 {
1367         return mutex_trylock(&dqp->q_qlock);
1368 }
1369
1370 void
1371 xfs_dqlock(
1372         xfs_dquot_t *dqp)
1373 {
1374         mutex_lock(&dqp->q_qlock);
1375 }
1376
1377 void
1378 xfs_dqunlock(
1379         xfs_dquot_t *dqp)
1380 {
1381         mutex_unlock(&(dqp->q_qlock));
1382         if (dqp->q_logitem.qli_dquot == dqp) {
1383                 /* Once was dqp->q_mount, but might just have been cleared */
1384                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1385                                         (xfs_log_item_t*)&(dqp->q_logitem));
1386         }
1387 }
1388
1389
1390 void
1391 xfs_dqunlock_nonotify(
1392         xfs_dquot_t *dqp)
1393 {
1394         mutex_unlock(&(dqp->q_qlock));
1395 }
1396
1397 /*
1398  * Lock two xfs_dquot structures.
1399  *
1400  * To avoid deadlocks we always lock the quota structure with
1401  * the lowerd id first.
1402  */
1403 void
1404 xfs_dqlock2(
1405         xfs_dquot_t     *d1,
1406         xfs_dquot_t     *d2)
1407 {
1408         if (d1 && d2) {
1409                 ASSERT(d1 != d2);
1410                 if (be32_to_cpu(d1->q_core.d_id) >
1411                     be32_to_cpu(d2->q_core.d_id)) {
1412                         mutex_lock(&d2->q_qlock);
1413                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1414                 } else {
1415                         mutex_lock(&d1->q_qlock);
1416                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1417                 }
1418         } else if (d1) {
1419                 mutex_lock(&d1->q_qlock);
1420         } else if (d2) {
1421                 mutex_lock(&d2->q_qlock);
1422         }
1423 }
1424
1425
1426 /*
1427  * Take a dquot out of the mount's dqlist as well as the hashlist.
1428  * This is called via unmount as well as quotaoff, and the purge
1429  * will always succeed unless there are soft (temp) references
1430  * outstanding.
1431  *
1432  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1433  * that we're returning! XXXsup - not cool.
1434  */
1435 /* ARGSUSED */
1436 int
1437 xfs_qm_dqpurge(
1438         xfs_dquot_t     *dqp)
1439 {
1440         xfs_dqhash_t    *thishash;
1441         xfs_mount_t     *mp = dqp->q_mount;
1442
1443         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
1444         ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1445
1446         xfs_dqlock(dqp);
1447         /*
1448          * We really can't afford to purge a dquot that is
1449          * referenced, because these are hard refs.
1450          * It shouldn't happen in general because we went thru _all_ inodes in
1451          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1452          * However it is possible that we have dquots with temporary
1453          * references that are not attached to an inode. e.g. see xfs_setattr().
1454          */
1455         if (dqp->q_nrefs != 0) {
1456                 xfs_dqunlock(dqp);
1457                 mutex_unlock(&dqp->q_hash->qh_lock);
1458                 return (1);
1459         }
1460
1461         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1462
1463         /*
1464          * If we're turning off quotas, we have to make sure that, for
1465          * example, we don't delete quota disk blocks while dquots are
1466          * in the process of getting written to those disk blocks.
1467          * This dquot might well be on AIL, and we can't leave it there
1468          * if we're turning off quotas. Basically, we need this flush
1469          * lock, and are willing to block on it.
1470          */
1471         if (!xfs_dqflock_nowait(dqp)) {
1472                 /*
1473                  * Block on the flush lock after nudging dquot buffer,
1474                  * if it is incore.
1475                  */
1476                 xfs_qm_dqflock_pushbuf_wait(dqp);
1477         }
1478
1479         /*
1480          * XXXIf we're turning this type of quotas off, we don't care
1481          * about the dirty metadata sitting in this dquot. OTOH, if
1482          * we're unmounting, we do care, so we flush it and wait.
1483          */
1484         if (XFS_DQ_IS_DIRTY(dqp)) {
1485                 int     error;
1486                 xfs_dqtrace_entry(dqp, "DQPURGE ->DQFLUSH: DQDIRTY");
1487                 /* dqflush unlocks dqflock */
1488                 /*
1489                  * Given that dqpurge is a very rare occurrence, it is OK
1490                  * that we're holding the hashlist and mplist locks
1491                  * across the disk write. But, ... XXXsup
1492                  *
1493                  * We don't care about getting disk errors here. We need
1494                  * to purge this dquot anyway, so we go ahead regardless.
1495                  */
1496                 error = xfs_qm_dqflush(dqp, XFS_QMOPT_SYNC);
1497                 if (error)
1498                         xfs_fs_cmn_err(CE_WARN, mp,
1499                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1500                 xfs_dqflock(dqp);
1501         }
1502         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1503         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1504                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1505
1506         thishash = dqp->q_hash;
1507         XQM_HASHLIST_REMOVE(thishash, dqp);
1508         XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(mp)), dqp);
1509         /*
1510          * XXX Move this to the front of the freelist, if we can get the
1511          * freelist lock.
1512          */
1513         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1514
1515         dqp->q_mount = NULL;
1516         dqp->q_hash = NULL;
1517         dqp->dq_flags = XFS_DQ_INACTIVE;
1518         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1519         xfs_dqfunlock(dqp);
1520         xfs_dqunlock(dqp);
1521         mutex_unlock(&thishash->qh_lock);
1522         return (0);
1523 }
1524
1525
1526 #ifdef QUOTADEBUG
1527 void
1528 xfs_qm_dqprint(xfs_dquot_t *dqp)
1529 {
1530         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1531         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1532                 (int)be32_to_cpu(dqp->q_core.d_id));
1533         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1534         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1535         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1536         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1537         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1538                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1539                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1540         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1541                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1542                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1543         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1544                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1545                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1546         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1547                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1548                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1549         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1550                 be64_to_cpu(dqp->q_core.d_bcount),
1551                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1552         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1553                 be64_to_cpu(dqp->q_core.d_icount),
1554                 (int)be64_to_cpu(dqp->q_core.d_icount));
1555         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1556                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1557         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1558                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1559         cmn_err(CE_DEBUG, "---------------------------");
1560 }
1561 #endif
1562
1563 /*
1564  * Give the buffer a little push if it is incore and
1565  * wait on the flush lock.
1566  */
1567 void
1568 xfs_qm_dqflock_pushbuf_wait(
1569         xfs_dquot_t     *dqp)
1570 {
1571         xfs_buf_t       *bp;
1572
1573         /*
1574          * Check to see if the dquot has been flushed delayed
1575          * write.  If so, grab its buffer and send it
1576          * out immediately.  We'll be able to acquire
1577          * the flush lock when the I/O completes.
1578          */
1579         bp = xfs_incore(dqp->q_mount->m_ddev_targp, dqp->q_blkno,
1580                     XFS_QI_DQCHUNKLEN(dqp->q_mount),
1581                     XFS_INCORE_TRYLOCK);
1582         if (bp != NULL) {
1583                 if (XFS_BUF_ISDELAYWRITE(bp)) {
1584                         int     error;
1585                         if (XFS_BUF_ISPINNED(bp)) {
1586                                 xfs_log_force(dqp->q_mount,
1587                                               (xfs_lsn_t)0,
1588                                               XFS_LOG_FORCE);
1589                         }
1590                         error = xfs_bawrite(dqp->q_mount, bp);
1591                         if (error)
1592                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
1593                                         "xfs_qm_dqflock_pushbuf_wait: "
1594                                         "pushbuf error %d on dqp %p, bp %p",
1595                                         error, dqp, bp);
1596                 } else {
1597                         xfs_buf_relse(bp);
1598                 }
1599         }
1600         xfs_dqflock(dqp);
1601 }