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_acl.h"
  46 #include "xfs_attr.h"
  47 #include "xfs_buf_item.h"
  48 #include "xfs_trans_space.h"
  49 #include "xfs_trans_priv.h"
  50 #include "xfs_qm.h"
  51
  52
  53 /*
  54    LOCK ORDER
  55
  56    inode lock               (ilock)
  57    dquot hash-chain lock    (hashlock)
  58    xqm dquot freelist lock  (freelistlock
  59    mount's dquot list lock  (mplistlock)
  60    user dquot lock - lock ordering among dquots is based on the uid or gid
  61    group dquot lock - similar to udquots. Between the two dquots, the udquot
  62                       has to be locked first.
  63    pin lock - the dquot lock must be held to take this lock.
  64    flush lock - ditto.
  65 */
  66
  67 STATIC void             xfs_qm_dqflush_done(xfs_buf_t *, xfs_dq_logitem_t *);
  68
  69 #ifdef DEBUG
  70 xfs_buftarg_t *xfs_dqerror_target;
  71 int xfs_do_dqerror;
  72 int xfs_dqreq_num;
  73 int xfs_dqerror_mod = 33;
  74 #endif
  75
  76 static struct lock_class_key xfs_dquot_other_class;
  77
  78 /*
  79  * Allocate and initialize a dquot. We don't always allocate fresh memory;
  80  * we try to reclaim a free dquot if the number of incore dquots are above
  81  * a threshold.
  82  * The only field inside the core that gets initialized at this point
  83  * is the d_id field. The idea is to fill in the entire q_core
  84  * when we read in the on disk dquot.
  85  */
  86 STATIC xfs_dquot_t *
  87 xfs_qm_dqinit(
  88         xfs_mount_t  *mp,
  89         xfs_dqid_t   id,
  90         uint         type)
  91 {
  92         xfs_dquot_t     *dqp;
  93         boolean_t       brandnewdquot;
  94
  95         brandnewdquot = xfs_qm_dqalloc_incore(&dqp);
  96         dqp->dq_flags = type;
  97         dqp->q_core.d_id = cpu_to_be32(id);
  98         dqp->q_mount = mp;
  99
 100         /*
 101          * No need to re-initialize these if this is a reclaimed dquot.
 102          */
 103         if (brandnewdquot) {
 104                 dqp->dq_flnext = dqp->dq_flprev = dqp;
 105                 mutex_init(&dqp->q_qlock);
 106                 init_waitqueue_head(&dqp->q_pinwait);
 107
 108                 /*
 109                  * Because we want to use a counting completion, complete
 110                  * the flush completion once to allow a single access to
 111                  * the flush completion without blocking.
 112                  */
 113                 init_completion(&dqp->q_flush);
 114                 complete(&dqp->q_flush);
 115
 116 #ifdef XFS_DQUOT_TRACE
 117                 dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_NOFS);
 118                 xfs_dqtrace_entry(dqp, "DQINIT");
 119 #endif
 120         } else {
 121                 /*
 122                  * Only the q_core portion was zeroed in dqreclaim_one().
 123                  * So, we need to reset others.
 124                  */
 125                  dqp->q_nrefs = 0;
 126                  dqp->q_blkno = 0;
 127                  dqp->MPL_NEXT = dqp->HL_NEXT = NULL;
 128                  dqp->HL_PREVP = dqp->MPL_PREVP = NULL;
 129                  dqp->q_bufoffset = 0;
 130                  dqp->q_fileoffset = 0;
 131                  dqp->q_transp = NULL;
 132                  dqp->q_gdquot = NULL;
 133                  dqp->q_res_bcount = 0;
 134                  dqp->q_res_icount = 0;
 135                  dqp->q_res_rtbcount = 0;
 136                  atomic_set(&dqp->q_pincount, 0);
 137                  dqp->q_hash = NULL;
 138                  ASSERT(dqp->dq_flnext == dqp->dq_flprev);
 139
 140 #ifdef XFS_DQUOT_TRACE
 141                  ASSERT(dqp->q_trace);
 142                  xfs_dqtrace_entry(dqp, "DQRECLAIMED_INIT");
 143 #endif
 144         }
 145
 146         /*
 147          * In either case we need to make sure group quotas have a different
 148          * lock class than user quotas, to make sure lockdep knows we can
 149          * locks of one of each at the same time.
 150          */
 151         if (!(type & XFS_DQ_USER))
 152                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
 153
 154         /*
 155          * log item gets initialized later
 156          */
 157         return (dqp);
 158 }
 159
 160 /*
 161  * This is called to free all the memory associated with a dquot
 162  */
 163 void
 164 xfs_qm_dqdestroy(
 165         xfs_dquot_t     *dqp)
 166 {
 167         ASSERT(! XFS_DQ_IS_ON_FREELIST(dqp));
 168
 169         mutex_destroy(&dqp->q_qlock);
 170         sv_destroy(&dqp->q_pinwait);
 171
 172 #ifdef XFS_DQUOT_TRACE
 173         if (dqp->q_trace)
 174              ktrace_free(dqp->q_trace);
 175         dqp->q_trace = NULL;
 176 #endif
 177         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
 178         atomic_dec(&xfs_Gqm->qm_totaldquots);
 179 }
 180
 181 /*
 182  * This is what a 'fresh' dquot inside a dquot chunk looks like on disk.
 183  */
 184 STATIC void
 185 xfs_qm_dqinit_core(
 186         xfs_dqid_t      id,
 187         uint            type,
 188         xfs_dqblk_t     *d)
 189 {
 190         /*
 191          * Caller has zero'd the entire dquot 'chunk' already.
 192          */
 193         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
 194         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
 195         d->dd_diskdq.d_id = cpu_to_be32(id);
 196         d->dd_diskdq.d_flags = type;
 197 }
 198
 199
 200 #ifdef XFS_DQUOT_TRACE
 201 /*
 202  * Dquot tracing for debugging.
 203  */
 204 /* ARGSUSED */
 205 void
 206 __xfs_dqtrace_entry(
 207         xfs_dquot_t     *dqp,
 208         char            *func,
 209         void            *retaddr,
 210         xfs_inode_t     *ip)
 211 {
 212         xfs_dquot_t     *udqp = NULL;
 213         xfs_ino_t       ino = 0;
 214
 215         ASSERT(dqp->q_trace);
 216         if (ip) {
 217                 ino = ip->i_ino;
 218                 udqp = ip->i_udquot;
 219         }
 220         ktrace_enter(dqp->q_trace,
 221                      (void *)(__psint_t)DQUOT_KTRACE_ENTRY,
 222                      (void *)func,
 223                      (void *)(__psint_t)dqp->q_nrefs,
 224                      (void *)(__psint_t)dqp->dq_flags,
 225                      (void *)(__psint_t)dqp->q_res_bcount,
 226                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_bcount),
 227                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_icount),
 228                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_hardlimit),
 229                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_blk_softlimit),
 230                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_hardlimit),
 231                      (void *)(__psint_t)be64_to_cpu(dqp->q_core.d_ino_softlimit),
 232                      (void *)(__psint_t)be32_to_cpu(dqp->q_core.d_id),
 233                      (void *)(__psint_t)current_pid(),
 234                      (void *)(__psint_t)ino,
 235                      (void *)(__psint_t)retaddr,
 236                      (void *)(__psint_t)udqp);
 237         return;
 238 }
 239 #endif
 240
 241
 242 /*
 243  * If default limits are in force, push them into the dquot now.
 244  * We overwrite the dquot limits only if they are zero and this
 245  * is not the root dquot.
 246  */
 247 void
 248 xfs_qm_adjust_dqlimits(
 249         xfs_mount_t             *mp,
 250         xfs_disk_dquot_t        *d)
 251 {
 252         xfs_quotainfo_t         *q = mp->m_quotainfo;
 253
 254         ASSERT(d->d_id);
 255
 256         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
 257                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
 258         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
 259                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
 260         if (q->qi_isoftlimit && !d->d_ino_softlimit)
 261                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
 262         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
 263                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
 264         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
 265                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
 266         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
 267                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
 268 }
 269
 270 /*
 271  * Check the limits and timers of a dquot and start or reset timers
 272  * if necessary.
 273  * This gets called even when quota enforcement is OFF, which makes our
 274  * life a little less complicated. (We just don't reject any quota
 275  * reservations in that case, when enforcement is off).
 276  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
 277  * enforcement's off.
 278  * In contrast, warnings are a little different in that they don't
 279  * 'automatically' get started when limits get exceeded.  They do
 280  * get reset to zero, however, when we find the count to be under
 281  * the soft limit (they are only ever set non-zero via userspace).
 282  */
 283 void
 284 xfs_qm_adjust_dqtimers(
 285         xfs_mount_t             *mp,
 286         xfs_disk_dquot_t        *d)
 287 {
 288         ASSERT(d->d_id);
 289
 290 #ifdef QUOTADEBUG
 291         if (d->d_blk_hardlimit)
 292                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
 293                        be64_to_cpu(d->d_blk_hardlimit));
 294         if (d->d_ino_hardlimit)
 295                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
 296                        be64_to_cpu(d->d_ino_hardlimit));
 297         if (d->d_rtb_hardlimit)
 298                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
 299                        be64_to_cpu(d->d_rtb_hardlimit));
 300 #endif
 301         if (!d->d_btimer) {
 302                 if ((d->d_blk_softlimit &&
 303                      (be64_to_cpu(d->d_bcount) >=
 304                       be64_to_cpu(d->d_blk_softlimit))) ||
 305                     (d->d_blk_hardlimit &&
 306                      (be64_to_cpu(d->d_bcount) >=
 307                       be64_to_cpu(d->d_blk_hardlimit)))) {
 308                         d->d_btimer = cpu_to_be32(get_seconds() +
 309                                         XFS_QI_BTIMELIMIT(mp));
 310                 } else {
 311                         d->d_bwarns = 0;
 312                 }
 313         } else {
 314                 if ((!d->d_blk_softlimit ||
 315                      (be64_to_cpu(d->d_bcount) <
 316                       be64_to_cpu(d->d_blk_softlimit))) &&
 317                     (!d->d_blk_hardlimit ||
 318                     (be64_to_cpu(d->d_bcount) <
 319                      be64_to_cpu(d->d_blk_hardlimit)))) {
 320                         d->d_btimer = 0;
 321                 }
 322         }
 323
 324         if (!d->d_itimer) {
 325                 if ((d->d_ino_softlimit &&
 326                      (be64_to_cpu(d->d_icount) >=
 327                       be64_to_cpu(d->d_ino_softlimit))) ||
 328                     (d->d_ino_hardlimit &&
 329                      (be64_to_cpu(d->d_icount) >=
 330                       be64_to_cpu(d->d_ino_hardlimit)))) {
 331                         d->d_itimer = cpu_to_be32(get_seconds() +
 332                                         XFS_QI_ITIMELIMIT(mp));
 333                 } else {
 334                         d->d_iwarns = 0;
 335                 }
 336         } else {
 337                 if ((!d->d_ino_softlimit ||
 338                      (be64_to_cpu(d->d_icount) <
 339                       be64_to_cpu(d->d_ino_softlimit)))  &&
 340                     (!d->d_ino_hardlimit ||
 341                      (be64_to_cpu(d->d_icount) <
 342                       be64_to_cpu(d->d_ino_hardlimit)))) {
 343                         d->d_itimer = 0;
 344                 }
 345         }
 346
 347         if (!d->d_rtbtimer) {
 348                 if ((d->d_rtb_softlimit &&
 349                      (be64_to_cpu(d->d_rtbcount) >=
 350                       be64_to_cpu(d->d_rtb_softlimit))) ||
 351                     (d->d_rtb_hardlimit &&
 352                      (be64_to_cpu(d->d_rtbcount) >=
 353                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 354                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
 355                                         XFS_QI_RTBTIMELIMIT(mp));
 356                 } else {
 357                         d->d_rtbwarns = 0;
 358                 }
 359         } else {
 360                 if ((!d->d_rtb_softlimit ||
 361                      (be64_to_cpu(d->d_rtbcount) <
 362                       be64_to_cpu(d->d_rtb_softlimit))) &&
 363                     (!d->d_rtb_hardlimit ||
 364                      (be64_to_cpu(d->d_rtbcount) <
 365                       be64_to_cpu(d->d_rtb_hardlimit)))) {
 366                         d->d_rtbtimer = 0;
 367                 }
 368         }
 369 }
 370
 371 /*
 372  * initialize a buffer full of dquots and log the whole thing
 373  */
 374 STATIC void
 375 xfs_qm_init_dquot_blk(
 376         xfs_trans_t     *tp,
 377         xfs_mount_t     *mp,
 378         xfs_dqid_t      id,
 379         uint            type,
 380         xfs_buf_t       *bp)
 381 {
 382         xfs_dqblk_t     *d;
 383         int             curid, i;
 384
 385         ASSERT(tp);
 386         ASSERT(XFS_BUF_ISBUSY(bp));
 387         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 388
 389         d = (xfs_dqblk_t *)XFS_BUF_PTR(bp);
 390
 391         /*
 392          * ID of the first dquot in the block - id's are zero based.
 393          */
 394         curid = id - (id % XFS_QM_DQPERBLK(mp));
 395         ASSERT(curid >= 0);
 396         memset(d, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)));
 397         for (i = 0; i < XFS_QM_DQPERBLK(mp); i++, d++, curid++)
 398                 xfs_qm_dqinit_core(curid, type, d);
 399         xfs_trans_dquot_buf(tp, bp,
 400                             (type & XFS_DQ_USER ? XFS_BLI_UDQUOT_BUF :
 401                             ((type & XFS_DQ_PROJ) ? XFS_BLI_PDQUOT_BUF :
 402                              XFS_BLI_GDQUOT_BUF)));
 403         xfs_trans_log_buf(tp, bp, 0, BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1);
 404 }
 405
 406
 407
 408 /*
 409  * Allocate a block and fill it with dquots.
 410  * This is called when the bmapi finds a hole.
 411  */
 412 STATIC int
 413 xfs_qm_dqalloc(
 414         xfs_trans_t     **tpp,
 415         xfs_mount_t     *mp,
 416         xfs_dquot_t     *dqp,
 417         xfs_inode_t     *quotip,
 418         xfs_fileoff_t   offset_fsb,
 419         xfs_buf_t       **O_bpp)
 420 {
 421         xfs_fsblock_t   firstblock;
 422         xfs_bmap_free_t flist;
 423         xfs_bmbt_irec_t map;
 424         int             nmaps, error, committed;
 425         xfs_buf_t       *bp;
 426         xfs_trans_t     *tp = *tpp;
 427
 428         ASSERT(tp != NULL);
 429         xfs_dqtrace_entry(dqp, "DQALLOC");
 430
 431         /*
 432          * Initialize the bmap freelist prior to calling bmapi code.
 433          */
 434         xfs_bmap_init(&flist, &firstblock);
 435         xfs_ilock(quotip, XFS_ILOCK_EXCL);
 436         /*
 437          * Return if this type of quotas is turned off while we didn't
 438          * have an inode lock
 439          */
 440         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 441                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 442                 return (ESRCH);
 443         }
 444
 445         /*
 446          * xfs_trans_commit normally decrements the vnode ref count
 447          * when it unlocks the inode. Since we want to keep the quota
 448          * inode around, we bump the vnode ref count now.
 449          */
 450         IHOLD(quotip);
 451
 452         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
 453         nmaps = 1;
 454         if ((error = xfs_bmapi(tp, quotip,
 455                               offset_fsb, XFS_DQUOT_CLUSTER_SIZE_FSB,
 456                               XFS_BMAPI_METADATA | XFS_BMAPI_WRITE,
 457                               &firstblock,
 458                               XFS_QM_DQALLOC_SPACE_RES(mp),
 459                               &map, &nmaps, &flist, NULL))) {
 460                 goto error0;
 461         }
 462         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
 463         ASSERT(nmaps == 1);
 464         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
 465                (map.br_startblock != HOLESTARTBLOCK));
 466
 467         /*
 468          * Keep track of the blkno to save a lookup later
 469          */
 470         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 471
 472         /* now we can just get the buffer (there's nothing to read yet) */
 473         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 474                                dqp->q_blkno,
 475                                XFS_QI_DQCHUNKLEN(mp),
 476                                0);
 477         if (!bp || (error = XFS_BUF_GETERROR(bp)))
 478                 goto error1;
 479         /*
 480          * Make a chunk of dquots out of this buffer and log
 481          * the entire thing.
 482          */
 483         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
 484                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
 485
 486         /*
 487          * xfs_bmap_finish() may commit the current transaction and
 488          * start a second transaction if the freelist is not empty.
 489          *
 490          * Since we still want to modify this buffer, we need to
 491          * ensure that the buffer is not released on commit of
 492          * the first transaction and ensure the buffer is added to the
 493          * second transaction.
 494          *
 495          * If there is only one transaction then don't stop the buffer
 496          * from being released when it commits later on.
 497          */
 498
 499         xfs_trans_bhold(tp, bp);
 500
 501         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
 502                 goto error1;
 503         }
 504
 505         if (committed) {
 506                 tp = *tpp;
 507                 xfs_trans_bjoin(tp, bp);
 508         } else {
 509                 xfs_trans_bhold_release(tp, bp);
 510         }
 511
 512         *O_bpp = bp;
 513         return 0;
 514
 515       error1:
 516         xfs_bmap_cancel(&flist);
 517       error0:
 518         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
 519
 520         return (error);
 521 }
 522
 523 /*
 524  * Maps a dquot to the buffer containing its on-disk version.
 525  * This returns a ptr to the buffer containing the on-disk dquot
 526  * in the bpp param, and a ptr to the on-disk dquot within that buffer
 527  */
 528 STATIC int
 529 xfs_qm_dqtobp(
 530         xfs_trans_t             **tpp,
 531         xfs_dquot_t             *dqp,
 532         xfs_disk_dquot_t        **O_ddpp,
 533         xfs_buf_t               **O_bpp,
 534         uint                    flags)
 535 {
 536         xfs_bmbt_irec_t map;
 537         int             nmaps, error;
 538         xfs_buf_t       *bp;
 539         xfs_inode_t     *quotip;
 540         xfs_mount_t     *mp;
 541         xfs_disk_dquot_t *ddq;
 542         xfs_dqid_t      id;
 543         boolean_t       newdquot;
 544         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
 545
 546         mp = dqp->q_mount;
 547         id = be32_to_cpu(dqp->q_core.d_id);
 548         nmaps = 1;
 549         newdquot = B_FALSE;
 550
 551         /*
 552          * If we don't know where the dquot lives, find out.
 553          */
 554         if (dqp->q_blkno == (xfs_daddr_t) 0) {
 555                 /* We use the id as an index */
 556                 dqp->q_fileoffset = (xfs_fileoff_t)id / XFS_QM_DQPERBLK(mp);
 557                 nmaps = 1;
 558                 quotip = XFS_DQ_TO_QIP(dqp);
 559                 xfs_ilock(quotip, XFS_ILOCK_SHARED);
 560                 /*
 561                  * Return if this type of quotas is turned off while we didn't
 562                  * have an inode lock
 563                  */
 564                 if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
 565                         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 566                         return (ESRCH);
 567                 }
 568                 /*
 569                  * Find the block map; no allocations yet
 570                  */
 571                 error = xfs_bmapi(NULL, quotip, dqp->q_fileoffset,
 572                                   XFS_DQUOT_CLUSTER_SIZE_FSB,
 573                                   XFS_BMAPI_METADATA,
 574                                   NULL, 0, &map, &nmaps, NULL, NULL);
 575
 576                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
 577                 if (error)
 578                         return (error);
 579                 ASSERT(nmaps == 1);
 580                 ASSERT(map.br_blockcount == 1);
 581
 582                 /*
 583                  * offset of dquot in the (fixed sized) dquot chunk.
 584                  */
 585                 dqp->q_bufoffset = (id % XFS_QM_DQPERBLK(mp)) *
 586                         sizeof(xfs_dqblk_t);
 587                 if (map.br_startblock == HOLESTARTBLOCK) {
 588                         /*
 589                          * We don't allocate unless we're asked to
 590                          */
 591                         if (!(flags & XFS_QMOPT_DQALLOC))
 592                                 return (ENOENT);
 593
 594                         ASSERT(tp);
 595                         if ((error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
 596                                                 dqp->q_fileoffset, &bp)))
 597                                 return (error);
 598                         tp = *tpp;
 599                         newdquot = B_TRUE;
 600                 } else {
 601                         /*
 602                          * store the blkno etc so that we don't have to do the
 603                          * mapping all the time
 604                          */
 605                         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
 606                 }
 607         }
 608         ASSERT(dqp->q_blkno != DELAYSTARTBLOCK);
 609         ASSERT(dqp->q_blkno != HOLESTARTBLOCK);
 610
 611         /*
 612          * Read in the buffer, unless we've just done the allocation
 613          * (in which case we already have the buf).
 614          */
 615         if (! newdquot) {
 616                 xfs_dqtrace_entry(dqp, "DQTOBP READBUF");
 617                 if ((error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
 618                                                dqp->q_blkno,
 619                                                XFS_QI_DQCHUNKLEN(mp),
 620                                                0, &bp))) {
 621                         return (error);
 622                 }
 623                 if (error || !bp)
 624                         return XFS_ERROR(error);
 625         }
 626         ASSERT(XFS_BUF_ISBUSY(bp));
 627         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 628
 629         /*
 630          * calculate the location of the dquot inside the buffer.
 631          */
 632         ddq = (xfs_disk_dquot_t *)((char *)XFS_BUF_PTR(bp) + dqp->q_bufoffset);
 633
 634         /*
 635          * A simple sanity check in case we got a corrupted dquot...
 636          */
 637         if (xfs_qm_dqcheck(ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
 638                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
 639                            "dqtobp")) {
 640                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
 641                         xfs_trans_brelse(tp, bp);
 642                         return XFS_ERROR(EIO);
 643                 }
 644                 XFS_BUF_BUSY(bp); /* We dirtied this */
 645         }
 646
 647         *O_bpp = bp;
 648         *O_ddpp = ddq;
 649
 650         return (0);
 651 }
 652
 653
 654 /*
 655  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
 656  * and release the buffer immediately.
 657  *
 658  */
 659 /* ARGSUSED */
 660 STATIC int
 661 xfs_qm_dqread(
 662         xfs_trans_t     **tpp,
 663         xfs_dqid_t      id,
 664         xfs_dquot_t     *dqp,   /* dquot to get filled in */
 665         uint            flags)
 666 {
 667         xfs_disk_dquot_t *ddqp;
 668         xfs_buf_t        *bp;
 669         int              error;
 670         xfs_trans_t      *tp;
 671
 672         ASSERT(tpp);
 673
 674         /*
 675          * get a pointer to the on-disk dquot and the buffer containing it
 676          * dqp already knows its own type (GROUP/USER).
 677          */
 678         xfs_dqtrace_entry(dqp, "DQREAD");
 679         if ((error = xfs_qm_dqtobp(tpp, dqp, &ddqp, &bp, flags))) {
 680                 return (error);
 681         }
 682         tp = *tpp;
 683
 684         /* copy everything from disk dquot to the incore dquot */
 685         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
 686         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 687         xfs_qm_dquot_logitem_init(dqp);
 688
 689         /*
 690          * Reservation counters are defined as reservation plus current usage
 691          * to avoid having to add everytime.
 692          */
 693         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
 694         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
 695         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
 696
 697         /* Mark the buf so that this will stay incore a little longer */
 698         XFS_BUF_SET_VTYPE_REF(bp, B_FS_DQUOT, XFS_DQUOT_REF);
 699
 700         /*
 701          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
 702          * So we need to release with xfs_trans_brelse().
 703          * The strategy here is identical to that of inodes; we lock
 704          * the dquot in xfs_qm_dqget() before making it accessible to
 705          * others. This is because dquots, like inodes, need a good level of
 706          * concurrency, and we don't want to take locks on the entire buffers
 707          * for dquot accesses.
 708          * Note also that the dquot buffer may even be dirty at this point, if
 709          * this particular dquot was repaired. We still aren't afraid to
 710          * brelse it because we have the changes incore.
 711          */
 712         ASSERT(XFS_BUF_ISBUSY(bp));
 713         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
 714         xfs_trans_brelse(tp, bp);
 715
 716         return (error);
 717 }
 718
 719
 720 /*
 721  * allocate an incore dquot from the kernel heap,
 722  * and fill its core with quota information kept on disk.
 723  * If XFS_QMOPT_DQALLOC is set, it'll allocate a dquot on disk
 724  * if it wasn't already allocated.
 725  */
 726 STATIC int
 727 xfs_qm_idtodq(
 728         xfs_mount_t     *mp,
 729         xfs_dqid_t      id,      /* gid or uid, depending on type */
 730         uint            type,    /* UDQUOT or GDQUOT */
 731         uint            flags,   /* DQALLOC, DQREPAIR */
 732         xfs_dquot_t     **O_dqpp)/* OUT : incore dquot, not locked */
 733 {
 734         xfs_dquot_t     *dqp;
 735         int             error;
 736         xfs_trans_t     *tp;
 737         int             cancelflags=0;
 738
 739         dqp = xfs_qm_dqinit(mp, id, type);
 740         tp = NULL;
 741         if (flags & XFS_QMOPT_DQALLOC) {
 742                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
 743                 if ((error = xfs_trans_reserve(tp,
 744                                        XFS_QM_DQALLOC_SPACE_RES(mp),
 745                                        XFS_WRITE_LOG_RES(mp) +
 746                                               BBTOB(XFS_QI_DQCHUNKLEN(mp)) - 1 +
 747                                               128,
 748                                        0,
 749                                        XFS_TRANS_PERM_LOG_RES,
 750                                        XFS_WRITE_LOG_COUNT))) {
 751                         cancelflags = 0;
 752                         goto error0;
 753                 }
 754                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
 755         }
 756
 757         /*
 758          * Read it from disk; xfs_dqread() takes care of
 759          * all the necessary initialization of dquot's fields (locks, etc)
 760          */
 761         if ((error = xfs_qm_dqread(&tp, id, dqp, flags))) {
 762                 /*
 763                  * This can happen if quotas got turned off (ESRCH),
 764                  * or if the dquot didn't exist on disk and we ask to
 765                  * allocate (ENOENT).
 766                  */
 767                 xfs_dqtrace_entry(dqp, "DQREAD FAIL");
 768                 cancelflags |= XFS_TRANS_ABORT;
 769                 goto error0;
 770         }
 771         if (tp) {
 772                 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES)))
 773                         goto error1;
 774         }
 775
 776         *O_dqpp = dqp;
 777         return (0);
 778
 779  error0:
 780         ASSERT(error);
 781         if (tp)
 782                 xfs_trans_cancel(tp, cancelflags);
 783  error1:
 784         xfs_qm_dqdestroy(dqp);
 785         *O_dqpp = NULL;
 786         return (error);
 787 }
 788
 789 /*
 790  * Lookup a dquot in the incore dquot hashtable. We keep two separate
 791  * hashtables for user and group dquots; and, these are global tables
 792  * inside the XQM, not per-filesystem tables.
 793  * The hash chain must be locked by caller, and it is left locked
 794  * on return. Returning dquot is locked.
 795  */
 796 STATIC int
 797 xfs_qm_dqlookup(
 798         xfs_mount_t             *mp,
 799         xfs_dqid_t              id,
 800         xfs_dqhash_t            *qh,
 801         xfs_dquot_t             **O_dqpp)
 802 {
 803         xfs_dquot_t             *dqp;
 804         uint                    flist_locked;
 805         xfs_dquot_t             *d;
 806
 807         ASSERT(mutex_is_locked(&qh->qh_lock));
 808
 809         flist_locked = B_FALSE;
 810
 811         /*
 812          * Traverse the hashchain looking for a match
 813          */
 814         for (dqp = qh->qh_next; dqp != NULL; dqp = dqp->HL_NEXT) {
 815                 /*
 816                  * We already have the hashlock. We don't need the
 817                  * dqlock to look at the id field of the dquot, since the
 818                  * id can't be modified without the hashlock anyway.
 819                  */
 820                 if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
 821                         xfs_dqtrace_entry(dqp, "DQFOUND BY LOOKUP");
 822                         /*
 823                          * All in core dquots must be on the dqlist of mp
 824                          */
 825                         ASSERT(dqp->MPL_PREVP != NULL);
 826
 827                         xfs_dqlock(dqp);
 828                         if (dqp->q_nrefs == 0) {
 829                                 ASSERT (XFS_DQ_IS_ON_FREELIST(dqp));
 830                                 if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
 831                                         xfs_dqtrace_entry(dqp, "DQLOOKUP: WANT");
 832
 833                                         /*
 834                                          * We may have raced with dqreclaim_one()
 835                                          * (and lost). So, flag that we don't
 836                                          * want the dquot to be reclaimed.
 837                                          */
 838                                         dqp->dq_flags |= XFS_DQ_WANT;
 839                                         xfs_dqunlock(dqp);
 840                                         xfs_qm_freelist_lock(xfs_Gqm);
 841                                         xfs_dqlock(dqp);
 842                                         dqp->dq_flags &= ~(XFS_DQ_WANT);
 843                                 }
 844                                 flist_locked = B_TRUE;
 845                         }
 846
 847                         /*
 848                          * id couldn't have changed; we had the hashlock all
 849                          * along
 850                          */
 851                         ASSERT(be32_to_cpu(dqp->q_core.d_id) == id);
 852
 853                         if (flist_locked) {
 854                                 if (dqp->q_nrefs != 0) {
 855                                         xfs_qm_freelist_unlock(xfs_Gqm);
 856                                         flist_locked = B_FALSE;
 857                                 } else {
 858                                         /*
 859                                          * take it off the freelist
 860                                          */
 861                                         xfs_dqtrace_entry(dqp,
 862                                                         "DQLOOKUP: TAKEOFF FL");
 863                                         XQM_FREELIST_REMOVE(dqp);
 864                                         /* xfs_qm_freelist_print(&(xfs_Gqm->
 865                                                         qm_dqfreelist),
 866                                                         "after removal"); */
 867                                 }
 868                         }
 869
 870                         /*
 871                          * grab a reference
 872                          */
 873                         XFS_DQHOLD(dqp);
 874
 875                         if (flist_locked)
 876                                 xfs_qm_freelist_unlock(xfs_Gqm);
 877                         /*
 878                          * move the dquot to the front of the hashchain
 879                          */
 880                         ASSERT(mutex_is_locked(&qh->qh_lock));
 881                         if (dqp->HL_PREVP != &qh->qh_next) {
 882                                 xfs_dqtrace_entry(dqp,
 883                                                   "DQLOOKUP: HASH MOVETOFRONT");
 884                                 if ((d = dqp->HL_NEXT))
 885                                         d->HL_PREVP = dqp->HL_PREVP;
 886                                 *(dqp->HL_PREVP) = d;
 887                                 d = qh->qh_next;
 888                                 d->HL_PREVP = &dqp->HL_NEXT;
 889                                 dqp->HL_NEXT = d;
 890                                 dqp->HL_PREVP = &qh->qh_next;
 891                                 qh->qh_next = dqp;
 892                         }
 893                         xfs_dqtrace_entry(dqp, "LOOKUP END");
 894                         *O_dqpp = dqp;
 895                         ASSERT(mutex_is_locked(&qh->qh_lock));
 896                         return (0);
 897                 }
 898         }
 899
 900         *O_dqpp = NULL;
 901         ASSERT(mutex_is_locked(&qh->qh_lock));
 902         return (1);
 903 }
 904
 905 /*
 906  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
 907  * a locked dquot, doing an allocation (if requested) as needed.
 908  * When both an inode and an id are given, the inode's id takes precedence.
 909  * That is, if the id changes while we don't hold the ilock inside this
 910  * function, the new dquot is returned, not necessarily the one requested
 911  * in the id argument.
 912  */
 913 int
 914 xfs_qm_dqget(
 915         xfs_mount_t     *mp,
 916         xfs_inode_t     *ip,      /* locked inode (optional) */
 917         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
 918         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
 919         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
 920         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
 921 {
 922         xfs_dquot_t     *dqp;
 923         xfs_dqhash_t    *h;
 924         uint            version;
 925         int             error;
 926
 927         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 928         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
 929             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
 930             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
 931                 return (ESRCH);
 932         }
 933         h = XFS_DQ_HASH(mp, id, type);
 934
 935 #ifdef DEBUG
 936         if (xfs_do_dqerror) {
 937                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
 938                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
 939                         cmn_err(CE_DEBUG, "Returning error in dqget");
 940                         return (EIO);
 941                 }
 942         }
 943 #endif
 944
 945  again:
 946
 947 #ifdef DEBUG
 948         ASSERT(type == XFS_DQ_USER ||
 949                type == XFS_DQ_PROJ ||
 950                type == XFS_DQ_GROUP);
 951         if (ip) {
 952                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 953                 if (type == XFS_DQ_USER)
 954                         ASSERT(ip->i_udquot == NULL);
 955                 else
 956                         ASSERT(ip->i_gdquot == NULL);
 957         }
 958 #endif
 959         mutex_lock(&h->qh_lock);
 960
 961         /*
 962          * Look in the cache (hashtable).
 963          * The chain is kept locked during lookup.
 964          */
 965         if (xfs_qm_dqlookup(mp, id, h, O_dqpp) == 0) {
 966                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
 967                 /*
 968                  * The dquot was found, moved to the front of the chain,
 969                  * taken off the freelist if it was on it, and locked
 970                  * at this point. Just unlock the hashchain and return.
 971                  */
 972                 ASSERT(*O_dqpp);
 973                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
 974                 mutex_unlock(&h->qh_lock);
 975                 xfs_dqtrace_entry(*O_dqpp, "DQGET DONE (FROM CACHE)");
 976                 return (0);     /* success */
 977         }
 978         XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
 979
 980         /*
 981          * Dquot cache miss. We don't want to keep the inode lock across
 982          * a (potential) disk read. Also we don't want to deal with the lock
 983          * ordering between quotainode and this inode. OTOH, dropping the inode
 984          * lock here means dealing with a chown that can happen before
 985          * we re-acquire the lock.
 986          */
 987         if (ip)
 988                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 989         /*
 990          * Save the hashchain version stamp, and unlock the chain, so that
 991          * we don't keep the lock across a disk read
 992          */
 993         version = h->qh_version;
 994         mutex_unlock(&h->qh_lock);
 995
 996         /*
 997          * Allocate the dquot on the kernel heap, and read the ondisk
 998          * portion off the disk. Also, do all the necessary initialization
 999          * This can return ENOENT if dquot didn't exist on disk and we didn't
1000          * ask it to allocate; ESRCH if quotas got turned off suddenly.
1001          */
1002         if ((error = xfs_qm_idtodq(mp, id, type,
1003                                   flags & (XFS_QMOPT_DQALLOC|XFS_QMOPT_DQREPAIR|
1004                                            XFS_QMOPT_DOWARN),
1005                                   &dqp))) {
1006                 if (ip)
1007                         xfs_ilock(ip, XFS_ILOCK_EXCL);
1008                 return (error);
1009         }
1010
1011         /*
1012          * See if this is mount code calling to look at the overall quota limits
1013          * which are stored in the id == 0 user or group's dquot.
1014          * Since we may not have done a quotacheck by this point, just return
1015          * the dquot without attaching it to any hashtables, lists, etc, or even
1016          * taking a reference.
1017          * The caller must dqdestroy this once done.
1018          */
1019         if (flags & XFS_QMOPT_DQSUSER) {
1020                 ASSERT(id == 0);
1021                 ASSERT(! ip);
1022                 goto dqret;
1023         }
1024
1025         /*
1026          * Dquot lock comes after hashlock in the lock ordering
1027          */
1028         if (ip) {
1029                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1030                 if (! XFS_IS_DQTYPE_ON(mp, type)) {
1031                         /* inode stays locked on return */
1032                         xfs_qm_dqdestroy(dqp);
1033                         return XFS_ERROR(ESRCH);
1034                 }
1035                 /*
1036                  * A dquot could be attached to this inode by now, since
1037                  * we had dropped the ilock.
1038                  */
1039                 if (type == XFS_DQ_USER) {
1040                         if (ip->i_udquot) {
1041                                 xfs_qm_dqdestroy(dqp);
1042                                 dqp = ip->i_udquot;
1043                                 xfs_dqlock(dqp);
1044                                 goto dqret;
1045                         }
1046                 } else {
1047                         if (ip->i_gdquot) {
1048                                 xfs_qm_dqdestroy(dqp);
1049                                 dqp = ip->i_gdquot;
1050                                 xfs_dqlock(dqp);
1051                                 goto dqret;
1052                         }
1053                 }
1054         }
1055
1056         /*
1057          * Hashlock comes after ilock in lock order
1058          */
1059         mutex_lock(&h->qh_lock);
1060         if (version != h->qh_version) {
1061                 xfs_dquot_t *tmpdqp;
1062                 /*
1063                  * Now, see if somebody else put the dquot in the
1064                  * hashtable before us. This can happen because we didn't
1065                  * keep the hashchain lock. We don't have to worry about
1066                  * lock order between the two dquots here since dqp isn't
1067                  * on any findable lists yet.
1068                  */
1069                 if (xfs_qm_dqlookup(mp, id, h, &tmpdqp) == 0) {
1070                         /*
1071                          * Duplicate found. Just throw away the new dquot
1072                          * and start over.
1073                          */
1074                         xfs_qm_dqput(tmpdqp);
1075                         mutex_unlock(&h->qh_lock);
1076                         xfs_qm_dqdestroy(dqp);
1077                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
1078                         goto again;
1079                 }
1080         }
1081
1082         /*
1083          * Put the dquot at the beginning of the hash-chain and mp's list
1084          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
1085          */
1086         ASSERT(mutex_is_locked(&h->qh_lock));
1087         dqp->q_hash = h;
1088         XQM_HASHLIST_INSERT(h, dqp);
1089
1090         /*
1091          * Attach this dquot to this filesystem's list of all dquots,
1092          * kept inside the mount structure in m_quotainfo field
1093          */
1094         xfs_qm_mplist_lock(mp);
1095
1096         /*
1097          * We return a locked dquot to the caller, with a reference taken
1098          */
1099         xfs_dqlock(dqp);
1100         dqp->q_nrefs = 1;
1101
1102         XQM_MPLIST_INSERT(&(XFS_QI_MPL_LIST(mp)), dqp);
1103
1104         xfs_qm_mplist_unlock(mp);
1105         mutex_unlock(&h->qh_lock);
1106  dqret:
1107         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
1108         xfs_dqtrace_entry(dqp, "DQGET DONE");
1109         *O_dqpp = dqp;
1110         return (0);
1111 }
1112
1113
1114 /*
1115  * Release a reference to the dquot (decrement ref-count)
1116  * and unlock it. If there is a group quota attached to this
1117  * dquot, carefully release that too without tripping over
1118  * deadlocks'n'stuff.
1119  */
1120 void
1121 xfs_qm_dqput(
1122         xfs_dquot_t     *dqp)
1123 {
1124         xfs_dquot_t     *gdqp;
1125
1126         ASSERT(dqp->q_nrefs > 0);
1127         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1128         xfs_dqtrace_entry(dqp, "DQPUT");
1129
1130         if (dqp->q_nrefs != 1) {
1131                 dqp->q_nrefs--;
1132                 xfs_dqunlock(dqp);
1133                 return;
1134         }
1135
1136         /*
1137          * drop the dqlock and acquire the freelist and dqlock
1138          * in the right order; but try to get it out-of-order first
1139          */
1140         if (! xfs_qm_freelist_lock_nowait(xfs_Gqm)) {
1141                 xfs_dqtrace_entry(dqp, "DQPUT: FLLOCK-WAIT");
1142                 xfs_dqunlock(dqp);
1143                 xfs_qm_freelist_lock(xfs_Gqm);
1144                 xfs_dqlock(dqp);
1145         }
1146
1147         while (1) {
1148                 gdqp = NULL;
1149
1150                 /* We can't depend on nrefs being == 1 here */
1151                 if (--dqp->q_nrefs == 0) {
1152                         xfs_dqtrace_entry(dqp, "DQPUT: ON FREELIST");
1153                         /*
1154                          * insert at end of the freelist.
1155                          */
1156                         XQM_FREELIST_INSERT(&(xfs_Gqm->qm_dqfreelist), dqp);
1157
1158                         /*
1159                          * If we just added a udquot to the freelist, then
1160                          * we want to release the gdquot reference that
1161                          * it (probably) has. Otherwise it'll keep the
1162                          * gdquot from getting reclaimed.
1163                          */
1164                         if ((gdqp = dqp->q_gdquot)) {
1165                                 /*
1166                                  * Avoid a recursive dqput call
1167                                  */
1168                                 xfs_dqlock(gdqp);
1169                                 dqp->q_gdquot = NULL;
1170                         }
1171
1172                         /* xfs_qm_freelist_print(&(xfs_Gqm->qm_dqfreelist),
1173                            "@@@@@++ Free list (after append) @@@@@+");
1174                            */
1175                 }
1176                 xfs_dqunlock(dqp);
1177
1178                 /*
1179                  * If we had a group quota inside the user quota as a hint,
1180                  * release it now.
1181                  */
1182                 if (! gdqp)
1183                         break;
1184                 dqp = gdqp;
1185         }
1186         xfs_qm_freelist_unlock(xfs_Gqm);
1187 }
1188
1189 /*
1190  * Release a dquot. Flush it if dirty, then dqput() it.
1191  * dquot must not be locked.
1192  */
1193 void
1194 xfs_qm_dqrele(
1195         xfs_dquot_t     *dqp)
1196 {
1197         if (!dqp)
1198                 return;
1199
1200         xfs_dqtrace_entry(dqp, "DQRELE");
1201
1202         xfs_dqlock(dqp);
1203         /*
1204          * We don't care to flush it if the dquot is dirty here.
1205          * That will create stutters that we want to avoid.
1206          * Instead we do a delayed write when we try to reclaim
1207          * a dirty dquot. Also xfs_sync will take part of the burden...
1208          */
1209         xfs_qm_dqput(dqp);
1210 }
1211
1212
1213 /*
1214  * Write a modified dquot to disk.
1215  * The dquot must be locked and the flush lock too taken by caller.
1216  * The flush lock will not be unlocked until the dquot reaches the disk,
1217  * but the dquot is free to be unlocked and modified by the caller
1218  * in the interim. Dquot is still locked on return. This behavior is
1219  * identical to that of inodes.
1220  */
1221 int
1222 xfs_qm_dqflush(
1223         xfs_dquot_t             *dqp,
1224         uint                    flags)
1225 {
1226         xfs_mount_t             *mp;
1227         xfs_buf_t               *bp;
1228         xfs_disk_dquot_t        *ddqp;
1229         int                     error;
1230
1231         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1232         ASSERT(!completion_done(&dqp->q_flush));
1233         xfs_dqtrace_entry(dqp, "DQFLUSH");
1234
1235         /*
1236          * If not dirty, or it's pinned and we are not supposed to
1237          * block, nada.
1238          */
1239         if (!XFS_DQ_IS_DIRTY(dqp) ||
1240             (!(flags & XFS_QMOPT_SYNC) && atomic_read(&dqp->q_pincount) > 0)) {
1241                 xfs_dqfunlock(dqp);
1242                 return 0;
1243         }
1244         xfs_qm_dqunpin_wait(dqp);
1245
1246         /*
1247          * This may have been unpinned because the filesystem is shutting
1248          * down forcibly. If that's the case we must not write this dquot
1249          * to disk, because the log record didn't make it to disk!
1250          */
1251         if (XFS_FORCED_SHUTDOWN(dqp->q_mount)) {
1252                 dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1253                 xfs_dqfunlock(dqp);
1254                 return XFS_ERROR(EIO);
1255         }
1256
1257         /*
1258          * Get the buffer containing the on-disk dquot
1259          * We don't need a transaction envelope because we know that the
1260          * the ondisk-dquot has already been allocated for.
1261          */
1262         if ((error = xfs_qm_dqtobp(NULL, dqp, &ddqp, &bp, XFS_QMOPT_DOWARN))) {
1263                 xfs_dqtrace_entry(dqp, "DQTOBP FAIL");
1264                 ASSERT(error != ENOENT);
1265                 /*
1266                  * Quotas could have gotten turned off (ESRCH)
1267                  */
1268                 xfs_dqfunlock(dqp);
1269                 return (error);
1270         }
1271
1272         if (xfs_qm_dqcheck(&dqp->q_core, be32_to_cpu(ddqp->d_id),
1273                            0, XFS_QMOPT_DOWARN, "dqflush (incore copy)")) {
1274                 xfs_force_shutdown(dqp->q_mount, SHUTDOWN_CORRUPT_INCORE);
1275                 return XFS_ERROR(EIO);
1276         }
1277
1278         /* This is the only portion of data that needs to persist */
1279         memcpy(ddqp, &(dqp->q_core), sizeof(xfs_disk_dquot_t));
1280
1281         /*
1282          * Clear the dirty field and remember the flush lsn for later use.
1283          */
1284         dqp->dq_flags &= ~(XFS_DQ_DIRTY);
1285         mp = dqp->q_mount;
1286
1287         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1288                                         &dqp->q_logitem.qli_item.li_lsn);
1289
1290         /*
1291          * Attach an iodone routine so that we can remove this dquot from the
1292          * AIL and release the flush lock once the dquot is synced to disk.
1293          */
1294         xfs_buf_attach_iodone(bp, (void(*)(xfs_buf_t *, xfs_log_item_t *))
1295                               xfs_qm_dqflush_done, &(dqp->q_logitem.qli_item));
1296         /*
1297          * If the buffer is pinned then push on the log so we won't
1298          * get stuck waiting in the write for too long.
1299          */
1300         if (XFS_BUF_ISPINNED(bp)) {
1301                 xfs_dqtrace_entry(dqp, "DQFLUSH LOG FORCE");
1302                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
1303         }
1304
1305         if (flags & XFS_QMOPT_DELWRI) {
1306                 xfs_bdwrite(mp, bp);
1307         } else if (flags & XFS_QMOPT_ASYNC) {
1308                 error = xfs_bawrite(mp, bp);
1309         } else {
1310                 error = xfs_bwrite(mp, bp);
1311         }
1312         xfs_dqtrace_entry(dqp, "DQFLUSH END");
1313         /*
1314          * dqp is still locked, but caller is free to unlock it now.
1315          */
1316         return (error);
1317
1318 }
1319
1320 /*
1321  * This is the dquot flushing I/O completion routine.  It is called
1322  * from interrupt level when the buffer containing the dquot is
1323  * flushed to disk.  It is responsible for removing the dquot logitem
1324  * from the AIL if it has not been re-logged, and unlocking the dquot's
1325  * flush lock. This behavior is very similar to that of inodes..
1326  */
1327 /*ARGSUSED*/
1328 STATIC void
1329 xfs_qm_dqflush_done(
1330         xfs_buf_t               *bp,
1331         xfs_dq_logitem_t        *qip)
1332 {
1333         xfs_dquot_t             *dqp;
1334         struct xfs_ail          *ailp;
1335
1336         dqp = qip->qli_dquot;
1337         ailp = qip->qli_item.li_ailp;
1338
1339         /*
1340          * We only want to pull the item from the AIL if its
1341          * location in the log has not changed since we started the flush.
1342          * Thus, we only bother if the dquot's lsn has
1343          * not changed. First we check the lsn outside the lock
1344          * since it's cheaper, and then we recheck while
1345          * holding the lock before removing the dquot from the AIL.
1346          */
1347         if ((qip->qli_item.li_flags & XFS_LI_IN_AIL) &&
1348             qip->qli_item.li_lsn == qip->qli_flush_lsn) {
1349
1350                 /* xfs_trans_ail_delete() drops the AIL lock. */
1351                 spin_lock(&ailp->xa_lock);
1352                 if (qip->qli_item.li_lsn == qip->qli_flush_lsn)
1353                         xfs_trans_ail_delete(ailp, (xfs_log_item_t*)qip);
1354                 else
1355                         spin_unlock(&ailp->xa_lock);
1356         }
1357
1358         /*
1359          * Release the dq's flush lock since we're done with it.
1360          */
1361         xfs_dqfunlock(dqp);
1362 }
1363
1364 int
1365 xfs_qm_dqlock_nowait(
1366         xfs_dquot_t *dqp)
1367 {
1368         return mutex_trylock(&dqp->q_qlock);
1369 }
1370
1371 void
1372 xfs_dqlock(
1373         xfs_dquot_t *dqp)
1374 {
1375         mutex_lock(&dqp->q_qlock);
1376 }
1377
1378 void
1379 xfs_dqunlock(
1380         xfs_dquot_t *dqp)
1381 {
1382         mutex_unlock(&(dqp->q_qlock));
1383         if (dqp->q_logitem.qli_dquot == dqp) {
1384                 /* Once was dqp->q_mount, but might just have been cleared */
1385                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1386                                         (xfs_log_item_t*)&(dqp->q_logitem));
1387         }
1388 }
1389
1390
1391 void
1392 xfs_dqunlock_nonotify(
1393         xfs_dquot_t *dqp)
1394 {
1395         mutex_unlock(&(dqp->q_qlock));
1396 }
1397
1398 /*
1399  * Lock two xfs_dquot structures.
1400  *
1401  * To avoid deadlocks we always lock the quota structure with
1402  * the lowerd id first.
1403  */
1404 void
1405 xfs_dqlock2(
1406         xfs_dquot_t     *d1,
1407         xfs_dquot_t     *d2)
1408 {
1409         if (d1 && d2) {
1410                 ASSERT(d1 != d2);
1411                 if (be32_to_cpu(d1->q_core.d_id) >
1412                     be32_to_cpu(d2->q_core.d_id)) {
1413                         mutex_lock(&d2->q_qlock);
1414                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1415                 } else {
1416                         mutex_lock(&d1->q_qlock);
1417                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1418                 }
1419         } else if (d1) {
1420                 mutex_lock(&d1->q_qlock);
1421         } else if (d2) {
1422                 mutex_lock(&d2->q_qlock);
1423         }
1424 }
1425
1426
1427 /*
1428  * Take a dquot out of the mount's dqlist as well as the hashlist.
1429  * This is called via unmount as well as quotaoff, and the purge
1430  * will always succeed unless there are soft (temp) references
1431  * outstanding.
1432  *
1433  * This returns 0 if it was purged, 1 if it wasn't. It's not an error code
1434  * that we're returning! XXXsup - not cool.
1435  */
1436 /* ARGSUSED */
1437 int
1438 xfs_qm_dqpurge(
1439         xfs_dquot_t     *dqp)
1440 {
1441         xfs_dqhash_t    *thishash;
1442         xfs_mount_t     *mp = dqp->q_mount;
1443
1444         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
1445         ASSERT(mutex_is_locked(&dqp->q_hash->qh_lock));
1446
1447         xfs_dqlock(dqp);
1448         /*
1449          * We really can't afford to purge a dquot that is
1450          * referenced, because these are hard refs.
1451          * It shouldn't happen in general because we went thru _all_ inodes in
1452          * dqrele_all_inodes before calling this and didn't let the mountlock go.
1453          * However it is possible that we have dquots with temporary
1454          * references that are not attached to an inode. e.g. see xfs_setattr().
1455          */
1456         if (dqp->q_nrefs != 0) {
1457                 xfs_dqunlock(dqp);
1458                 mutex_unlock(&dqp->q_hash->qh_lock);
1459                 return (1);
1460         }
1461
1462         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1463
1464         /*
1465          * If we're turning off quotas, we have to make sure that, for
1466          * example, we don't delete quota disk blocks while dquots are
1467          * in the process of getting written to those disk blocks.
1468          * This dquot might well be on AIL, and we can't leave it there
1469          * if we're turning off quotas. Basically, we need this flush
1470          * lock, and are willing to block on it.
1471          */
1472         if (!xfs_dqflock_nowait(dqp)) {
1473                 /*
1474                  * Block on the flush lock after nudging dquot buffer,
1475                  * if it is incore.
1476                  */
1477                 xfs_qm_dqflock_pushbuf_wait(dqp);
1478         }
1479
1480         /*
1481          * XXXIf we're turning this type of quotas off, we don't care
1482          * about the dirty metadata sitting in this dquot. OTOH, if
1483          * we're unmounting, we do care, so we flush it and wait.
1484          */
1485         if (XFS_DQ_IS_DIRTY(dqp)) {
1486                 int     error;
1487                 xfs_dqtrace_entry(dqp, "DQPURGE ->DQFLUSH: DQDIRTY");
1488                 /* dqflush unlocks dqflock */
1489                 /*
1490                  * Given that dqpurge is a very rare occurrence, it is OK
1491                  * that we're holding the hashlist and mplist locks
1492                  * across the disk write. But, ... XXXsup
1493                  *
1494                  * We don't care about getting disk errors here. We need
1495                  * to purge this dquot anyway, so we go ahead regardless.
1496                  */
1497                 error = xfs_qm_dqflush(dqp, XFS_QMOPT_SYNC);
1498                 if (error)
1499                         xfs_fs_cmn_err(CE_WARN, mp,
1500                                 "xfs_qm_dqpurge: dquot %p flush failed", dqp);
1501                 xfs_dqflock(dqp);
1502         }
1503         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1504         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1505                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1506
1507         thishash = dqp->q_hash;
1508         XQM_HASHLIST_REMOVE(thishash, dqp);
1509         XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(mp)), dqp);
1510         /*
1511          * XXX Move this to the front of the freelist, if we can get the
1512          * freelist lock.
1513          */
1514         ASSERT(XFS_DQ_IS_ON_FREELIST(dqp));
1515
1516         dqp->q_mount = NULL;
1517         dqp->q_hash = NULL;
1518         dqp->dq_flags = XFS_DQ_INACTIVE;
1519         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
1520         xfs_dqfunlock(dqp);
1521         xfs_dqunlock(dqp);
1522         mutex_unlock(&thishash->qh_lock);
1523         return (0);
1524 }
1525
1526
1527 #ifdef QUOTADEBUG
1528 void
1529 xfs_qm_dqprint(xfs_dquot_t *dqp)
1530 {
1531         cmn_err(CE_DEBUG, "-----------KERNEL DQUOT----------------");
1532         cmn_err(CE_DEBUG, "---- dquotID =  %d",
1533                 (int)be32_to_cpu(dqp->q_core.d_id));
1534         cmn_err(CE_DEBUG, "---- type    =  %s", DQFLAGTO_TYPESTR(dqp));
1535         cmn_err(CE_DEBUG, "---- fs      =  0x%p", dqp->q_mount);
1536         cmn_err(CE_DEBUG, "---- blkno   =  0x%x", (int) dqp->q_blkno);
1537         cmn_err(CE_DEBUG, "---- boffset =  0x%x", (int) dqp->q_bufoffset);
1538         cmn_err(CE_DEBUG, "---- blkhlimit =  %Lu (0x%x)",
1539                 be64_to_cpu(dqp->q_core.d_blk_hardlimit),
1540                 (int)be64_to_cpu(dqp->q_core.d_blk_hardlimit));
1541         cmn_err(CE_DEBUG, "---- blkslimit =  %Lu (0x%x)",
1542                 be64_to_cpu(dqp->q_core.d_blk_softlimit),
1543                 (int)be64_to_cpu(dqp->q_core.d_blk_softlimit));
1544         cmn_err(CE_DEBUG, "---- inohlimit =  %Lu (0x%x)",
1545                 be64_to_cpu(dqp->q_core.d_ino_hardlimit),
1546                 (int)be64_to_cpu(dqp->q_core.d_ino_hardlimit));
1547         cmn_err(CE_DEBUG, "---- inoslimit =  %Lu (0x%x)",
1548                 be64_to_cpu(dqp->q_core.d_ino_softlimit),
1549                 (int)be64_to_cpu(dqp->q_core.d_ino_softlimit));
1550         cmn_err(CE_DEBUG, "---- bcount  =  %Lu (0x%x)",
1551                 be64_to_cpu(dqp->q_core.d_bcount),
1552                 (int)be64_to_cpu(dqp->q_core.d_bcount));
1553         cmn_err(CE_DEBUG, "---- icount  =  %Lu (0x%x)",
1554                 be64_to_cpu(dqp->q_core.d_icount),
1555                 (int)be64_to_cpu(dqp->q_core.d_icount));
1556         cmn_err(CE_DEBUG, "---- btimer  =  %d",
1557                 (int)be32_to_cpu(dqp->q_core.d_btimer));
1558         cmn_err(CE_DEBUG, "---- itimer  =  %d",
1559                 (int)be32_to_cpu(dqp->q_core.d_itimer));
1560         cmn_err(CE_DEBUG, "---------------------------");
1561 }
1562 #endif
1563
1564 /*
1565  * Give the buffer a little push if it is incore and
1566  * wait on the flush lock.
1567  */
1568 void
1569 xfs_qm_dqflock_pushbuf_wait(
1570         xfs_dquot_t     *dqp)
1571 {
1572         xfs_buf_t       *bp;
1573
1574         /*
1575          * Check to see if the dquot has been flushed delayed
1576          * write.  If so, grab its buffer and send it
1577          * out immediately.  We'll be able to acquire
1578          * the flush lock when the I/O completes.
1579          */
1580         bp = xfs_incore(dqp->q_mount->m_ddev_targp, dqp->q_blkno,
1581                     XFS_QI_DQCHUNKLEN(dqp->q_mount),
1582                     XFS_INCORE_TRYLOCK);
1583         if (bp != NULL) {
1584                 if (XFS_BUF_ISDELAYWRITE(bp)) {
1585                         int     error;
1586                         if (XFS_BUF_ISPINNED(bp)) {
1587                                 xfs_log_force(dqp->q_mount,
1588                                               (xfs_lsn_t)0,
1589                                               XFS_LOG_FORCE);
1590                         }
1591                         error = xfs_bawrite(dqp->q_mount, bp);
1592                         if (error)
1593                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
1594                                         "xfs_qm_dqflock_pushbuf_wait: "
1595                                         "pushbuf error %d on dqp %p, bp %p",
1596                                         error, dqp, bp);
1597                 } else {
1598                         xfs_buf_relse(bp);
1599                 }
1600         }
1601         xfs_dqflock(dqp);
1602 }