fs/xfs/quota/xfs_qm.c

   1 /*
   2  * Copyright (c) 2000-2005 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_clnt.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_alloc.h"
  29 #include "xfs_dmapi.h"
  30 #include "xfs_quota.h"
  31 #include "xfs_mount.h"
  32 #include "xfs_bmap_btree.h"
  33 #include "xfs_alloc_btree.h"
  34 #include "xfs_ialloc_btree.h"
  35 #include "xfs_dir2_sf.h"
  36 #include "xfs_attr_sf.h"
  37 #include "xfs_dinode.h"
  38 #include "xfs_inode.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_ialloc.h"
  41 #include "xfs_itable.h"
  42 #include "xfs_rtalloc.h"
  43 #include "xfs_error.h"
  44 #include "xfs_bmap.h"
  45 #include "xfs_rw.h"
  46 #include "xfs_acl.h"
  47 #include "xfs_attr.h"
  48 #include "xfs_buf_item.h"
  49 #include "xfs_trans_space.h"
  50 #include "xfs_utils.h"
  51 #include "xfs_qm.h"
  52
  53 /*
  54  * The global quota manager. There is only one of these for the entire
  55  * system, _not_ one per file system. XQM keeps track of the overall
  56  * quota functionality, including maintaining the freelist and hash
  57  * tables of dquots.
  58  */
  59 mutex_t         xfs_Gqm_lock;
  60 struct xfs_qm   *xfs_Gqm;
  61 uint            ndquot;
  62
  63 kmem_zone_t     *qm_dqzone;
  64 kmem_zone_t     *qm_dqtrxzone;
  65
  66 static cred_t   xfs_zerocr;
  67
  68 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
  69 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
  70
  71 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
  72 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
  73 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
  74 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
  75
  76 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
  77 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
  78 STATIC int      xfs_qm_shake(int, gfp_t);
  79
  80 static struct shrinker xfs_qm_shaker = {
  81         .shrink = xfs_qm_shake,
  82         .seeks = DEFAULT_SEEKS,
  83 };
  84
  85 #ifdef DEBUG
  86 extern mutex_t  qcheck_lock;
  87 #endif
  88
  89 #ifdef QUOTADEBUG
  90 #define XQM_LIST_PRINT(l, NXT, title) \
  91 { \
  92         xfs_dquot_t     *dqp; int i = 0; \
  93         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
  94         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
  95                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
  96                                   "bcnt = %d, icnt = %d, refs = %d", \
  97                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
  98                         DQFLAGTO_TYPESTR(dqp),       \
  99                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
 100                         (int) be64_to_cpu(dqp->q_core.d_icount), \
 101                         (int) dqp->q_nrefs);  } \
 102 }
 103 #else
 104 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
 105 #endif
 106
 107 /*
 108  * Initialize the XQM structure.
 109  * Note that there is not one quota manager per file system.
 110  */
 111 STATIC struct xfs_qm *
 112 xfs_Gqm_init(void)
 113 {
 114         xfs_dqhash_t    *udqhash, *gdqhash;
 115         xfs_qm_t        *xqm;
 116         size_t          hsize;
 117         uint            i;
 118
 119         /*
 120          * Initialize the dquot hash tables.
 121          */
 122         udqhash = kmem_zalloc_greedy(&hsize,
 123                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
 124                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t),
 125                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
 126         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
 127         hsize /= sizeof(xfs_dqhash_t);
 128         ndquot = hsize << 8;
 129
 130         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
 131         xqm->qm_dqhashmask = hsize - 1;
 132         xqm->qm_usr_dqhtable = udqhash;
 133         xqm->qm_grp_dqhtable = gdqhash;
 134         ASSERT(xqm->qm_usr_dqhtable != NULL);
 135         ASSERT(xqm->qm_grp_dqhtable != NULL);
 136
 137         for (i = 0; i < hsize; i++) {
 138                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
 139                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
 140         }
 141
 142         /*
 143          * Freelist of all dquots of all file systems
 144          */
 145         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
 146
 147         /*
 148          * dquot zone. we register our own low-memory callback.
 149          */
 150         if (!qm_dqzone) {
 151                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
 152                                                 "xfs_dquots");
 153                 qm_dqzone = xqm->qm_dqzone;
 154         } else
 155                 xqm->qm_dqzone = qm_dqzone;
 156
 157         register_shrinker(&xfs_qm_shaker);
 158
 159         /*
 160          * The t_dqinfo portion of transactions.
 161          */
 162         if (!qm_dqtrxzone) {
 163                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
 164                                                    "xfs_dqtrx");
 165                 qm_dqtrxzone = xqm->qm_dqtrxzone;
 166         } else
 167                 xqm->qm_dqtrxzone = qm_dqtrxzone;
 168
 169         atomic_set(&xqm->qm_totaldquots, 0);
 170         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
 171         xqm->qm_nrefs = 0;
 172 #ifdef DEBUG
 173         mutex_init(&qcheck_lock);
 174 #endif
 175         return xqm;
 176 }
 177
 178 /*
 179  * Destroy the global quota manager when its reference count goes to zero.
 180  */
 181 STATIC void
 182 xfs_qm_destroy(
 183         struct xfs_qm   *xqm)
 184 {
 185         int             hsize, i;
 186
 187         ASSERT(xqm != NULL);
 188         ASSERT(xqm->qm_nrefs == 0);
 189         unregister_shrinker(&xfs_qm_shaker);
 190         hsize = xqm->qm_dqhashmask + 1;
 191         for (i = 0; i < hsize; i++) {
 192                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
 193                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
 194         }
 195         kmem_free(xqm->qm_usr_dqhtable);
 196         kmem_free(xqm->qm_grp_dqhtable);
 197         xqm->qm_usr_dqhtable = NULL;
 198         xqm->qm_grp_dqhtable = NULL;
 199         xqm->qm_dqhashmask = 0;
 200         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
 201 #ifdef DEBUG
 202         mutex_destroy(&qcheck_lock);
 203 #endif
 204         kmem_free(xqm);
 205 }
 206
 207 /*
 208  * Called at mount time to let XQM know that another file system is
 209  * starting quotas. This isn't crucial information as the individual mount
 210  * structures are pretty independent, but it helps the XQM keep a
 211  * global view of what's going on.
 212  */
 213 /* ARGSUSED */
 214 STATIC int
 215 xfs_qm_hold_quotafs_ref(
 216         struct xfs_mount *mp)
 217 {
 218         /*
 219          * Need to lock the xfs_Gqm structure for things like this. For example,
 220          * the structure could disappear between the entry to this routine and
 221          * a HOLD operation if not locked.
 222          */
 223         XFS_QM_LOCK(xfs_Gqm);
 224
 225         if (xfs_Gqm == NULL)
 226                 xfs_Gqm = xfs_Gqm_init();
 227         /*
 228          * We can keep a list of all filesystems with quotas mounted for
 229          * debugging and statistical purposes, but ...
 230          * Just take a reference and get out.
 231          */
 232         XFS_QM_HOLD(xfs_Gqm);
 233         XFS_QM_UNLOCK(xfs_Gqm);
 234
 235         return 0;
 236 }
 237
 238
 239 /*
 240  * Release the reference that a filesystem took at mount time,
 241  * so that we know when we need to destroy the entire quota manager.
 242  */
 243 /* ARGSUSED */
 244 STATIC void
 245 xfs_qm_rele_quotafs_ref(
 246         struct xfs_mount *mp)
 247 {
 248         xfs_dquot_t     *dqp, *nextdqp;
 249
 250         ASSERT(xfs_Gqm);
 251         ASSERT(xfs_Gqm->qm_nrefs > 0);
 252
 253         /*
 254          * Go thru the freelist and destroy all inactive dquots.
 255          */
 256         xfs_qm_freelist_lock(xfs_Gqm);
 257
 258         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
 259              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
 260                 xfs_dqlock(dqp);
 261                 nextdqp = dqp->dq_flnext;
 262                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
 263                         ASSERT(dqp->q_mount == NULL);
 264                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
 265                         ASSERT(dqp->HL_PREVP == NULL);
 266                         ASSERT(dqp->MPL_PREVP == NULL);
 267                         XQM_FREELIST_REMOVE(dqp);
 268                         xfs_dqunlock(dqp);
 269                         xfs_qm_dqdestroy(dqp);
 270                 } else {
 271                         xfs_dqunlock(dqp);
 272                 }
 273                 dqp = nextdqp;
 274         }
 275         xfs_qm_freelist_unlock(xfs_Gqm);
 276
 277         /*
 278          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
 279          * be restarted.
 280          */
 281         XFS_QM_LOCK(xfs_Gqm);
 282         XFS_QM_RELE(xfs_Gqm);
 283         if (xfs_Gqm->qm_nrefs == 0) {
 284                 xfs_qm_destroy(xfs_Gqm);
 285                 xfs_Gqm = NULL;
 286         }
 287         XFS_QM_UNLOCK(xfs_Gqm);
 288 }
 289
 290 /*
 291  * Just destroy the quotainfo structure.
 292  */
 293 void
 294 xfs_qm_unmount_quotadestroy(
 295         xfs_mount_t     *mp)
 296 {
 297         if (mp->m_quotainfo)
 298                 xfs_qm_destroy_quotainfo(mp);
 299 }
 300
 301
 302 /*
 303  * This is called from xfs_mountfs to start quotas and initialize all
 304  * necessary data structures like quotainfo.  This is also responsible for
 305  * running a quotacheck as necessary.  We are guaranteed that the superblock
 306  * is consistently read in at this point.
 307  *
 308  * If we fail here, the mount will continue with quota turned off. We don't
 309  * need to inidicate success or failure at all.
 310  */
 311 void
 312 xfs_qm_mount_quotas(
 313         xfs_mount_t     *mp,
 314         int             mfsi_flags)
 315 {
 316         int             error = 0;
 317         uint            sbf;
 318
 319         /*
 320          * If quotas on realtime volumes is not supported, we disable
 321          * quotas immediately.
 322          */
 323         if (mp->m_sb.sb_rextents) {
 324                 cmn_err(CE_NOTE,
 325                         "Cannot turn on quotas for realtime filesystem %s",
 326                         mp->m_fsname);
 327                 mp->m_qflags = 0;
 328                 goto write_changes;
 329         }
 330
 331         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 332
 333         /*
 334          * Allocate the quotainfo structure inside the mount struct, and
 335          * create quotainode(s), and change/rev superblock if necessary.
 336          */
 337         error = xfs_qm_init_quotainfo(mp);
 338         if (error) {
 339                 /*
 340                  * We must turn off quotas.
 341                  */
 342                 ASSERT(mp->m_quotainfo == NULL);
 343                 mp->m_qflags = 0;
 344                 goto write_changes;
 345         }
 346         /*
 347          * If any of the quotas are not consistent, do a quotacheck.
 348          */
 349         if (XFS_QM_NEED_QUOTACHECK(mp) &&
 350             !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
 351                 error = xfs_qm_quotacheck(mp);
 352                 if (error) {
 353                         /* Quotacheck failed and disabled quotas. */
 354                         return;
 355                 }
 356         }
 357         /*
 358          * If one type of quotas is off, then it will lose its
 359          * quotachecked status, since we won't be doing accounting for
 360          * that type anymore.
 361          */
 362         if (!XFS_IS_UQUOTA_ON(mp))
 363                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
 364         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp)))
 365                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
 366
 367  write_changes:
 368         /*
 369          * We actually don't have to acquire the m_sb_lock at all.
 370          * This can only be called from mount, and that's single threaded. XXX
 371          */
 372         spin_lock(&mp->m_sb_lock);
 373         sbf = mp->m_sb.sb_qflags;
 374         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
 375         spin_unlock(&mp->m_sb_lock);
 376
 377         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
 378                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
 379                         /*
 380                          * We could only have been turning quotas off.
 381                          * We aren't in very good shape actually because
 382                          * the incore structures are convinced that quotas are
 383                          * off, but the on disk superblock doesn't know that !
 384                          */
 385                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
 386                         xfs_fs_cmn_err(CE_ALERT, mp,
 387                                 "XFS mount_quotas: Superblock update failed!");
 388                 }
 389         }
 390
 391         if (error) {
 392                 xfs_fs_cmn_err(CE_WARN, mp,
 393                         "Failed to initialize disk quotas.");
 394         }
 395         return;
 396 }
 397
 398 /*
 399  * Called from the vfsops layer.
 400  */
 401 int
 402 xfs_qm_unmount_quotas(
 403         xfs_mount_t     *mp)
 404 {
 405         xfs_inode_t     *uqp, *gqp;
 406         int             error = 0;
 407
 408         /*
 409          * Release the dquots that root inode, et al might be holding,
 410          * before we flush quotas and blow away the quotainfo structure.
 411          */
 412         ASSERT(mp->m_rootip);
 413         xfs_qm_dqdetach(mp->m_rootip);
 414         if (mp->m_rbmip)
 415                 xfs_qm_dqdetach(mp->m_rbmip);
 416         if (mp->m_rsumip)
 417                 xfs_qm_dqdetach(mp->m_rsumip);
 418
 419         /*
 420          * Flush out the quota inodes.
 421          */
 422         uqp = gqp = NULL;
 423         if (mp->m_quotainfo) {
 424                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
 425                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
 426                         xfs_iflock(uqp);
 427                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
 428                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
 429                         if (unlikely(error == EFSCORRUPTED)) {
 430                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
 431                                                  XFS_ERRLEVEL_LOW, mp);
 432                                 goto out;
 433                         }
 434                 }
 435                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
 436                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
 437                         xfs_iflock(gqp);
 438                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
 439                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
 440                         if (unlikely(error == EFSCORRUPTED)) {
 441                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
 442                                                  XFS_ERRLEVEL_LOW, mp);
 443                                 goto out;
 444                         }
 445                 }
 446         }
 447         if (uqp) {
 448                  IRELE(uqp);
 449                  mp->m_quotainfo->qi_uquotaip = NULL;
 450         }
 451         if (gqp) {
 452                 IRELE(gqp);
 453                 mp->m_quotainfo->qi_gquotaip = NULL;
 454         }
 455 out:
 456         return XFS_ERROR(error);
 457 }
 458
 459 /*
 460  * Flush all dquots of the given file system to disk. The dquots are
 461  * _not_ purged from memory here, just their data written to disk.
 462  */
 463 STATIC int
 464 xfs_qm_dqflush_all(
 465         xfs_mount_t     *mp,
 466         int             flags)
 467 {
 468         int             recl;
 469         xfs_dquot_t     *dqp;
 470         int             niters;
 471         int             error;
 472
 473         if (mp->m_quotainfo == NULL)
 474                 return 0;
 475         niters = 0;
 476 again:
 477         xfs_qm_mplist_lock(mp);
 478         FOREACH_DQUOT_IN_MP(dqp, mp) {
 479                 xfs_dqlock(dqp);
 480                 if (! XFS_DQ_IS_DIRTY(dqp)) {
 481                         xfs_dqunlock(dqp);
 482                         continue;
 483                 }
 484                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
 485                 /* XXX a sentinel would be better */
 486                 recl = XFS_QI_MPLRECLAIMS(mp);
 487                 if (!xfs_dqflock_nowait(dqp)) {
 488                         /*
 489                          * If we can't grab the flush lock then check
 490                          * to see if the dquot has been flushed delayed
 491                          * write.  If so, grab its buffer and send it
 492                          * out immediately.  We'll be able to acquire
 493                          * the flush lock when the I/O completes.
 494                          */
 495                         xfs_qm_dqflock_pushbuf_wait(dqp);
 496                 }
 497                 /*
 498                  * Let go of the mplist lock. We don't want to hold it
 499                  * across a disk write.
 500                  */
 501                 xfs_qm_mplist_unlock(mp);
 502                 error = xfs_qm_dqflush(dqp, flags);
 503                 xfs_dqunlock(dqp);
 504                 if (error)
 505                         return error;
 506
 507                 xfs_qm_mplist_lock(mp);
 508                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
 509                         xfs_qm_mplist_unlock(mp);
 510                         /* XXX restart limit */
 511                         goto again;
 512                 }
 513         }
 514
 515         xfs_qm_mplist_unlock(mp);
 516         /* return ! busy */
 517         return 0;
 518 }
 519 /*
 520  * Release the group dquot pointers the user dquots may be
 521  * carrying around as a hint. mplist is locked on entry and exit.
 522  */
 523 STATIC void
 524 xfs_qm_detach_gdquots(
 525         xfs_mount_t     *mp)
 526 {
 527         xfs_dquot_t     *dqp, *gdqp;
 528         int             nrecl;
 529
 530  again:
 531         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
 532         dqp = XFS_QI_MPLNEXT(mp);
 533         while (dqp) {
 534                 xfs_dqlock(dqp);
 535                 if ((gdqp = dqp->q_gdquot)) {
 536                         xfs_dqlock(gdqp);
 537                         dqp->q_gdquot = NULL;
 538                 }
 539                 xfs_dqunlock(dqp);
 540
 541                 if (gdqp) {
 542                         /*
 543                          * Can't hold the mplist lock across a dqput.
 544                          * XXXmust convert to marker based iterations here.
 545                          */
 546                         nrecl = XFS_QI_MPLRECLAIMS(mp);
 547                         xfs_qm_mplist_unlock(mp);
 548                         xfs_qm_dqput(gdqp);
 549
 550                         xfs_qm_mplist_lock(mp);
 551                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
 552                                 goto again;
 553                 }
 554                 dqp = dqp->MPL_NEXT;
 555         }
 556 }
 557
 558 /*
 559  * Go through all the incore dquots of this file system and take them
 560  * off the mplist and hashlist, if the dquot type matches the dqtype
 561  * parameter. This is used when turning off quota accounting for
 562  * users and/or groups, as well as when the filesystem is unmounting.
 563  */
 564 STATIC int
 565 xfs_qm_dqpurge_int(
 566         xfs_mount_t     *mp,
 567         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
 568 {
 569         xfs_dquot_t     *dqp;
 570         uint            dqtype;
 571         int             nrecl;
 572         xfs_dquot_t     *nextdqp;
 573         int             nmisses;
 574
 575         if (mp->m_quotainfo == NULL)
 576                 return 0;
 577
 578         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
 579         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
 580         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
 581
 582         xfs_qm_mplist_lock(mp);
 583
 584         /*
 585          * In the first pass through all incore dquots of this filesystem,
 586          * we release the group dquot pointers the user dquots may be
 587          * carrying around as a hint. We need to do this irrespective of
 588          * what's being turned off.
 589          */
 590         xfs_qm_detach_gdquots(mp);
 591
 592       again:
 593         nmisses = 0;
 594         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
 595         /*
 596          * Try to get rid of all of the unwanted dquots. The idea is to
 597          * get them off mplist and hashlist, but leave them on freelist.
 598          */
 599         dqp = XFS_QI_MPLNEXT(mp);
 600         while (dqp) {
 601                 /*
 602                  * It's OK to look at the type without taking dqlock here.
 603                  * We're holding the mplist lock here, and that's needed for
 604                  * a dqreclaim.
 605                  */
 606                 if ((dqp->dq_flags & dqtype) == 0) {
 607                         dqp = dqp->MPL_NEXT;
 608                         continue;
 609                 }
 610
 611                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
 612                         nrecl = XFS_QI_MPLRECLAIMS(mp);
 613                         xfs_qm_mplist_unlock(mp);
 614                         XFS_DQ_HASH_LOCK(dqp->q_hash);
 615                         xfs_qm_mplist_lock(mp);
 616
 617                         /*
 618                          * XXXTheoretically, we can get into a very long
 619                          * ping pong game here.
 620                          * No one can be adding dquots to the mplist at
 621                          * this point, but somebody might be taking things off.
 622                          */
 623                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
 624                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
 625                                 goto again;
 626                         }
 627                 }
 628
 629                 /*
 630                  * Take the dquot off the mplist and hashlist. It may remain on
 631                  * freelist in INACTIVE state.
 632                  */
 633                 nextdqp = dqp->MPL_NEXT;
 634                 nmisses += xfs_qm_dqpurge(dqp);
 635                 dqp = nextdqp;
 636         }
 637         xfs_qm_mplist_unlock(mp);
 638         return nmisses;
 639 }
 640
 641 int
 642 xfs_qm_dqpurge_all(
 643         xfs_mount_t     *mp,
 644         uint            flags)
 645 {
 646         int             ndquots;
 647
 648         /*
 649          * Purge the dquot cache.
 650          * None of the dquots should really be busy at this point.
 651          */
 652         if (mp->m_quotainfo) {
 653                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
 654                         delay(ndquots * 10);
 655                 }
 656         }
 657         return 0;
 658 }
 659
 660 STATIC int
 661 xfs_qm_dqattach_one(
 662         xfs_inode_t     *ip,
 663         xfs_dqid_t      id,
 664         uint            type,
 665         uint            doalloc,
 666         uint            dolock,
 667         xfs_dquot_t     *udqhint, /* hint */
 668         xfs_dquot_t     **IO_idqpp)
 669 {
 670         xfs_dquot_t     *dqp;
 671         int             error;
 672
 673         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 674         error = 0;
 675         /*
 676          * See if we already have it in the inode itself. IO_idqpp is
 677          * &i_udquot or &i_gdquot. This made the code look weird, but
 678          * made the logic a lot simpler.
 679          */
 680         if ((dqp = *IO_idqpp)) {
 681                 if (dolock)
 682                         xfs_dqlock(dqp);
 683                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
 684                 goto done;
 685         }
 686
 687         /*
 688          * udqhint is the i_udquot field in inode, and is non-NULL only
 689          * when the type arg is group/project. Its purpose is to save a
 690          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
 691          * the user dquot.
 692          */
 693         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
 694         if (udqhint && !dolock)
 695                 xfs_dqlock(udqhint);
 696
 697         /*
 698          * No need to take dqlock to look at the id.
 699          * The ID can't change until it gets reclaimed, and it won't
 700          * be reclaimed as long as we have a ref from inode and we hold
 701          * the ilock.
 702          */
 703         if (udqhint &&
 704             (dqp = udqhint->q_gdquot) &&
 705             (be32_to_cpu(dqp->q_core.d_id) == id)) {
 706                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
 707                 xfs_dqlock(dqp);
 708                 XFS_DQHOLD(dqp);
 709                 ASSERT(*IO_idqpp == NULL);
 710                 *IO_idqpp = dqp;
 711                 if (!dolock) {
 712                         xfs_dqunlock(dqp);
 713                         xfs_dqunlock(udqhint);
 714                 }
 715                 goto done;
 716         }
 717         /*
 718          * We can't hold a dquot lock when we call the dqget code.
 719          * We'll deadlock in no time, because of (not conforming to)
 720          * lock ordering - the inodelock comes before any dquot lock,
 721          * and we may drop and reacquire the ilock in xfs_qm_dqget().
 722          */
 723         if (udqhint)
 724                 xfs_dqunlock(udqhint);
 725         /*
 726          * Find the dquot from somewhere. This bumps the
 727          * reference count of dquot and returns it locked.
 728          * This can return ENOENT if dquot didn't exist on
 729          * disk and we didn't ask it to allocate;
 730          * ESRCH if quotas got turned off suddenly.
 731          */
 732         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
 733                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
 734                 if (udqhint && dolock)
 735                         xfs_dqlock(udqhint);
 736                 goto done;
 737         }
 738
 739         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
 740         /*
 741          * dqget may have dropped and re-acquired the ilock, but it guarantees
 742          * that the dquot returned is the one that should go in the inode.
 743          */
 744         *IO_idqpp = dqp;
 745         ASSERT(dqp);
 746         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 747         if (! dolock) {
 748                 xfs_dqunlock(dqp);
 749                 goto done;
 750         }
 751         if (! udqhint)
 752                 goto done;
 753
 754         ASSERT(udqhint);
 755         ASSERT(dolock);
 756         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 757         if (! xfs_qm_dqlock_nowait(udqhint)) {
 758                 xfs_dqunlock(dqp);
 759                 xfs_dqlock(udqhint);
 760                 xfs_dqlock(dqp);
 761         }
 762       done:
 763 #ifdef QUOTADEBUG
 764         if (udqhint) {
 765                 if (dolock)
 766                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
 767         }
 768         if (! error) {
 769                 if (dolock)
 770                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 771         }
 772 #endif
 773         return error;
 774 }
 775
 776
 777 /*
 778  * Given a udquot and gdquot, attach a ptr to the group dquot in the
 779  * udquot as a hint for future lookups. The idea sounds simple, but the
 780  * execution isn't, because the udquot might have a group dquot attached
 781  * already and getting rid of that gets us into lock ordering constraints.
 782  * The process is complicated more by the fact that the dquots may or may not
 783  * be locked on entry.
 784  */
 785 STATIC void
 786 xfs_qm_dqattach_grouphint(
 787         xfs_dquot_t     *udq,
 788         xfs_dquot_t     *gdq,
 789         uint            locked)
 790 {
 791         xfs_dquot_t     *tmp;
 792
 793 #ifdef QUOTADEBUG
 794         if (locked) {
 795                 ASSERT(XFS_DQ_IS_LOCKED(udq));
 796                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
 797         }
 798 #endif
 799         if (! locked)
 800                 xfs_dqlock(udq);
 801
 802         if ((tmp = udq->q_gdquot)) {
 803                 if (tmp == gdq) {
 804                         if (! locked)
 805                                 xfs_dqunlock(udq);
 806                         return;
 807                 }
 808
 809                 udq->q_gdquot = NULL;
 810                 /*
 811                  * We can't keep any dqlocks when calling dqrele,
 812                  * because the freelist lock comes before dqlocks.
 813                  */
 814                 xfs_dqunlock(udq);
 815                 if (locked)
 816                         xfs_dqunlock(gdq);
 817                 /*
 818                  * we took a hard reference once upon a time in dqget,
 819                  * so give it back when the udquot no longer points at it
 820                  * dqput() does the unlocking of the dquot.
 821                  */
 822                 xfs_qm_dqrele(tmp);
 823
 824                 xfs_dqlock(udq);
 825                 xfs_dqlock(gdq);
 826
 827         } else {
 828                 ASSERT(XFS_DQ_IS_LOCKED(udq));
 829                 if (! locked) {
 830                         xfs_dqlock(gdq);
 831                 }
 832         }
 833
 834         ASSERT(XFS_DQ_IS_LOCKED(udq));
 835         ASSERT(XFS_DQ_IS_LOCKED(gdq));
 836         /*
 837          * Somebody could have attached a gdquot here,
 838          * when we dropped the uqlock. If so, just do nothing.
 839          */
 840         if (udq->q_gdquot == NULL) {
 841                 XFS_DQHOLD(gdq);
 842                 udq->q_gdquot = gdq;
 843         }
 844         if (! locked) {
 845                 xfs_dqunlock(gdq);
 846                 xfs_dqunlock(udq);
 847         }
 848 }
 849
 850
 851 /*
 852  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
 853  * into account.
 854  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
 855  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
 856  * much made this code a complete mess, but it has been pretty useful.
 857  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
 858  * Inode may get unlocked and relocked in here, and the caller must deal with
 859  * the consequences.
 860  */
 861 int
 862 xfs_qm_dqattach(
 863         xfs_inode_t     *ip,
 864         uint            flags)
 865 {
 866         xfs_mount_t     *mp = ip->i_mount;
 867         uint            nquotas = 0;
 868         int             error = 0;
 869
 870         if ((! XFS_IS_QUOTA_ON(mp)) ||
 871             (! XFS_NOT_DQATTACHED(mp, ip)) ||
 872             (ip->i_ino == mp->m_sb.sb_uquotino) ||
 873             (ip->i_ino == mp->m_sb.sb_gquotino))
 874                 return 0;
 875
 876         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
 877                xfs_isilocked(ip, XFS_ILOCK_EXCL));
 878
 879         if (! (flags & XFS_QMOPT_ILOCKED))
 880                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 881
 882         if (XFS_IS_UQUOTA_ON(mp)) {
 883                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
 884                                                 flags & XFS_QMOPT_DQALLOC,
 885                                                 flags & XFS_QMOPT_DQLOCK,
 886                                                 NULL, &ip->i_udquot);
 887                 if (error)
 888                         goto done;
 889                 nquotas++;
 890         }
 891
 892         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 893         if (XFS_IS_OQUOTA_ON(mp)) {
 894                 error = XFS_IS_GQUOTA_ON(mp) ?
 895                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
 896                                                 flags & XFS_QMOPT_DQALLOC,
 897                                                 flags & XFS_QMOPT_DQLOCK,
 898                                                 ip->i_udquot, &ip->i_gdquot) :
 899                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
 900                                                 flags & XFS_QMOPT_DQALLOC,
 901                                                 flags & XFS_QMOPT_DQLOCK,
 902                                                 ip->i_udquot, &ip->i_gdquot);
 903                 /*
 904                  * Don't worry about the udquot that we may have
 905                  * attached above. It'll get detached, if not already.
 906                  */
 907                 if (error)
 908                         goto done;
 909                 nquotas++;
 910         }
 911
 912         /*
 913          * Attach this group quota to the user quota as a hint.
 914          * This WON'T, in general, result in a thrash.
 915          */
 916         if (nquotas == 2) {
 917                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 918                 ASSERT(ip->i_udquot);
 919                 ASSERT(ip->i_gdquot);
 920
 921                 /*
 922                  * We may or may not have the i_udquot locked at this point,
 923                  * but this check is OK since we don't depend on the i_gdquot to
 924                  * be accurate 100% all the time. It is just a hint, and this
 925                  * will succeed in general.
 926                  */
 927                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
 928                         goto done;
 929                 /*
 930                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
 931                  */
 932                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
 933                                          flags & XFS_QMOPT_DQLOCK);
 934         }
 935
 936       done:
 937
 938 #ifdef QUOTADEBUG
 939         if (! error) {
 940                 if (ip->i_udquot) {
 941                         if (flags & XFS_QMOPT_DQLOCK)
 942                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
 943                 }
 944                 if (ip->i_gdquot) {
 945                         if (flags & XFS_QMOPT_DQLOCK)
 946                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
 947                 }
 948                 if (XFS_IS_UQUOTA_ON(mp))
 949                         ASSERT(ip->i_udquot);
 950                 if (XFS_IS_OQUOTA_ON(mp))
 951                         ASSERT(ip->i_gdquot);
 952         }
 953 #endif
 954
 955         if (! (flags & XFS_QMOPT_ILOCKED))
 956                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 957
 958 #ifdef QUOTADEBUG
 959         else
 960                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 961 #endif
 962         return error;
 963 }
 964
 965 /*
 966  * Release dquots (and their references) if any.
 967  * The inode should be locked EXCL except when this's called by
 968  * xfs_ireclaim.
 969  */
 970 void
 971 xfs_qm_dqdetach(
 972         xfs_inode_t     *ip)
 973 {
 974         if (!(ip->i_udquot || ip->i_gdquot))
 975                 return;
 976
 977         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
 978         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
 979         if (ip->i_udquot) {
 980                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
 981                 xfs_qm_dqrele(ip->i_udquot);
 982                 ip->i_udquot = NULL;
 983         }
 984         if (ip->i_gdquot) {
 985                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
 986                 xfs_qm_dqrele(ip->i_gdquot);
 987                 ip->i_gdquot = NULL;
 988         }
 989 }
 990
 991 /*
 992  * This is called by VFS_SYNC and flags arg determines the caller,
 993  * and its motives, as done in xfs_sync.
 994  *
 995  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
 996  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
 997  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
 998  */
 999
1000 int
1001 xfs_qm_sync(
1002         xfs_mount_t     *mp,
1003         int             flags)
1004 {
1005         int             recl, restarts;
1006         xfs_dquot_t     *dqp;
1007         uint            flush_flags;
1008         boolean_t       nowait;
1009         int             error;
1010
1011         if (! XFS_IS_QUOTA_ON(mp))
1012                 return 0;
1013
1014         restarts = 0;
1015         /*
1016          * We won't block unless we are asked to.
1017          */
1018         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1019
1020   again:
1021         xfs_qm_mplist_lock(mp);
1022         /*
1023          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1024          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1025          * when we have the mplist lock, we know that dquots will be consistent
1026          * as long as we have it locked.
1027          */
1028         if (! XFS_IS_QUOTA_ON(mp)) {
1029                 xfs_qm_mplist_unlock(mp);
1030                 return 0;
1031         }
1032         FOREACH_DQUOT_IN_MP(dqp, mp) {
1033                 /*
1034                  * If this is vfs_sync calling, then skip the dquots that
1035                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1036                  * This is very similar to what xfs_sync does with inodes.
1037                  */
1038                 if (flags & SYNC_BDFLUSH) {
1039                         if (! XFS_DQ_IS_DIRTY(dqp))
1040                                 continue;
1041                 }
1042
1043                 if (nowait) {
1044                         /*
1045                          * Try to acquire the dquot lock. We are NOT out of
1046                          * lock order, but we just don't want to wait for this
1047                          * lock, unless somebody wanted us to.
1048                          */
1049                         if (! xfs_qm_dqlock_nowait(dqp))
1050                                 continue;
1051                 } else {
1052                         xfs_dqlock(dqp);
1053                 }
1054
1055                 /*
1056                  * Now, find out for sure if this dquot is dirty or not.
1057                  */
1058                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1059                         xfs_dqunlock(dqp);
1060                         continue;
1061                 }
1062
1063                 /* XXX a sentinel would be better */
1064                 recl = XFS_QI_MPLRECLAIMS(mp);
1065                 if (!xfs_dqflock_nowait(dqp)) {
1066                         if (nowait) {
1067                                 xfs_dqunlock(dqp);
1068                                 continue;
1069                         }
1070                         /*
1071                          * If we can't grab the flush lock then if the caller
1072                          * really wanted us to give this our best shot, so
1073                          * see if we can give a push to the buffer before we wait
1074                          * on the flush lock. At this point, we know that
1075                          * even though the dquot is being flushed,
1076                          * it has (new) dirty data.
1077                          */
1078                         xfs_qm_dqflock_pushbuf_wait(dqp);
1079                 }
1080                 /*
1081                  * Let go of the mplist lock. We don't want to hold it
1082                  * across a disk write
1083                  */
1084                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1085                 xfs_qm_mplist_unlock(mp);
1086                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1087                 error = xfs_qm_dqflush(dqp, flush_flags);
1088                 xfs_dqunlock(dqp);
1089                 if (error && XFS_FORCED_SHUTDOWN(mp))
1090                         return 0;       /* Need to prevent umount failure */
1091                 else if (error)
1092                         return error;
1093
1094                 xfs_qm_mplist_lock(mp);
1095                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1096                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1097                                 break;
1098
1099                         xfs_qm_mplist_unlock(mp);
1100                         goto again;
1101                 }
1102         }
1103
1104         xfs_qm_mplist_unlock(mp);
1105         return 0;
1106 }
1107
1108
1109 /*
1110  * This initializes all the quota information that's kept in the
1111  * mount structure
1112  */
1113 STATIC int
1114 xfs_qm_init_quotainfo(
1115         xfs_mount_t     *mp)
1116 {
1117         xfs_quotainfo_t *qinf;
1118         int             error;
1119         xfs_dquot_t     *dqp;
1120
1121         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1122
1123         /*
1124          * Tell XQM that we exist as soon as possible.
1125          */
1126         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1127                 return error;
1128         }
1129
1130         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1131
1132         /*
1133          * See if quotainodes are setup, and if not, allocate them,
1134          * and change the superblock accordingly.
1135          */
1136         if ((error = xfs_qm_init_quotainos(mp))) {
1137                 kmem_free(qinf);
1138                 mp->m_quotainfo = NULL;
1139                 return error;
1140         }
1141
1142         spin_lock_init(&qinf->qi_pinlock);
1143         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1144         qinf->qi_dqreclaims = 0;
1145
1146         /* mutex used to serialize quotaoffs */
1147         mutex_init(&qinf->qi_quotaofflock);
1148
1149         /* Precalc some constants */
1150         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1151         ASSERT(qinf->qi_dqchunklen);
1152         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1153         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1154
1155         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1156
1157         /*
1158          * We try to get the limits from the superuser's limits fields.
1159          * This is quite hacky, but it is standard quota practice.
1160          * We look at the USR dquot with id == 0 first, but if user quotas
1161          * are not enabled we goto the GRP dquot with id == 0.
1162          * We don't really care to keep separate default limits for user
1163          * and group quotas, at least not at this point.
1164          */
1165         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1166                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER :
1167                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1168                                 XFS_DQ_PROJ),
1169                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1170                              &dqp);
1171         if (! error) {
1172                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1173
1174                 /*
1175                  * The warnings and timers set the grace period given to
1176                  * a user or group before he or she can not perform any
1177                  * more writing. If it is zero, a default is used.
1178                  */
1179                 qinf->qi_btimelimit = ddqp->d_btimer ?
1180                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1181                 qinf->qi_itimelimit = ddqp->d_itimer ?
1182                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1183                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1184                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1185                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1186                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1187                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1188                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1189                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1190                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1191                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1192                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1193                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1194                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1195                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1196                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1197
1198                 /*
1199                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1200                  * we don't want this dquot cached. We haven't done a
1201                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1202                  */
1203                 xfs_qm_dqdestroy(dqp);
1204         } else {
1205                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1206                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1207                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1208                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1209                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1210                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1211         }
1212
1213         return 0;
1214 }
1215
1216
1217 /*
1218  * Gets called when unmounting a filesystem or when all quotas get
1219  * turned off.
1220  * This purges the quota inodes, destroys locks and frees itself.
1221  */
1222 void
1223 xfs_qm_destroy_quotainfo(
1224         xfs_mount_t     *mp)
1225 {
1226         xfs_quotainfo_t *qi;
1227
1228         qi = mp->m_quotainfo;
1229         ASSERT(qi != NULL);
1230         ASSERT(xfs_Gqm != NULL);
1231
1232         /*
1233          * Release the reference that XQM kept, so that we know
1234          * when the XQM structure should be freed. We cannot assume
1235          * that xfs_Gqm is non-null after this point.
1236          */
1237         xfs_qm_rele_quotafs_ref(mp);
1238
1239         spinlock_destroy(&qi->qi_pinlock);
1240         xfs_qm_list_destroy(&qi->qi_dqlist);
1241
1242         if (qi->qi_uquotaip) {
1243                 IRELE(qi->qi_uquotaip);
1244                 qi->qi_uquotaip = NULL; /* paranoia */
1245         }
1246         if (qi->qi_gquotaip) {
1247                 IRELE(qi->qi_gquotaip);
1248                 qi->qi_gquotaip = NULL;
1249         }
1250         mutex_destroy(&qi->qi_quotaofflock);
1251         kmem_free(qi);
1252         mp->m_quotainfo = NULL;
1253 }
1254
1255
1256
1257 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1258
1259 /* ARGSUSED */
1260 STATIC void
1261 xfs_qm_list_init(
1262         xfs_dqlist_t    *list,
1263         char            *str,
1264         int             n)
1265 {
1266         mutex_init(&list->qh_lock);
1267         list->qh_next = NULL;
1268         list->qh_version = 0;
1269         list->qh_nelems = 0;
1270 }
1271
1272 STATIC void
1273 xfs_qm_list_destroy(
1274         xfs_dqlist_t    *list)
1275 {
1276         mutex_destroy(&(list->qh_lock));
1277 }
1278
1279
1280 /*
1281  * Stripped down version of dqattach. This doesn't attach, or even look at the
1282  * dquots attached to the inode. The rationale is that there won't be any
1283  * attached at the time this is called from quotacheck.
1284  */
1285 STATIC int
1286 xfs_qm_dqget_noattach(
1287         xfs_inode_t     *ip,
1288         xfs_dquot_t     **O_udqpp,
1289         xfs_dquot_t     **O_gdqpp)
1290 {
1291         int             error;
1292         xfs_mount_t     *mp;
1293         xfs_dquot_t     *udqp, *gdqp;
1294
1295         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1296         mp = ip->i_mount;
1297         udqp = NULL;
1298         gdqp = NULL;
1299
1300         if (XFS_IS_UQUOTA_ON(mp)) {
1301                 ASSERT(ip->i_udquot == NULL);
1302                 /*
1303                  * We want the dquot allocated if it doesn't exist.
1304                  */
1305                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1306                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1307                                          &udqp))) {
1308                         /*
1309                          * Shouldn't be able to turn off quotas here.
1310                          */
1311                         ASSERT(error != ESRCH);
1312                         ASSERT(error != ENOENT);
1313                         return error;
1314                 }
1315                 ASSERT(udqp);
1316         }
1317
1318         if (XFS_IS_OQUOTA_ON(mp)) {
1319                 ASSERT(ip->i_gdquot == NULL);
1320                 if (udqp)
1321                         xfs_dqunlock(udqp);
1322                 error = XFS_IS_GQUOTA_ON(mp) ?
1323                                 xfs_qm_dqget(mp, ip,
1324                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1325                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1326                                              &gdqp) :
1327                                 xfs_qm_dqget(mp, ip,
1328                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1329                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1330                                              &gdqp);
1331                 if (error) {
1332                         if (udqp)
1333                                 xfs_qm_dqrele(udqp);
1334                         ASSERT(error != ESRCH);
1335                         ASSERT(error != ENOENT);
1336                         return error;
1337                 }
1338                 ASSERT(gdqp);
1339
1340                 /* Reacquire the locks in the right order */
1341                 if (udqp) {
1342                         if (! xfs_qm_dqlock_nowait(udqp)) {
1343                                 xfs_dqunlock(gdqp);
1344                                 xfs_dqlock(udqp);
1345                                 xfs_dqlock(gdqp);
1346                         }
1347                 }
1348         }
1349
1350         *O_udqpp = udqp;
1351         *O_gdqpp = gdqp;
1352
1353 #ifdef QUOTADEBUG
1354         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1355         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1356 #endif
1357         return 0;
1358 }
1359
1360 /*
1361  * Create an inode and return with a reference already taken, but unlocked
1362  * This is how we create quota inodes
1363  */
1364 STATIC int
1365 xfs_qm_qino_alloc(
1366         xfs_mount_t     *mp,
1367         xfs_inode_t     **ip,
1368         __int64_t       sbfields,
1369         uint            flags)
1370 {
1371         xfs_trans_t     *tp;
1372         int             error;
1373         int             committed;
1374
1375         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1376         if ((error = xfs_trans_reserve(tp,
1377                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1378                                       XFS_CREATE_LOG_RES(mp), 0,
1379                                       XFS_TRANS_PERM_LOG_RES,
1380                                       XFS_CREATE_LOG_COUNT))) {
1381                 xfs_trans_cancel(tp, 0);
1382                 return error;
1383         }
1384
1385         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1386                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1387                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1388                                  XFS_TRANS_ABORT);
1389                 return error;
1390         }
1391
1392         /*
1393          * Keep an extra reference to this quota inode. This inode is
1394          * locked exclusively and joined to the transaction already.
1395          */
1396         ASSERT(xfs_isilocked(*ip, XFS_ILOCK_EXCL));
1397         IHOLD(*ip);
1398
1399         /*
1400          * Make the changes in the superblock, and log those too.
1401          * sbfields arg may contain fields other than *QUOTINO;
1402          * VERSIONNUM for example.
1403          */
1404         spin_lock(&mp->m_sb_lock);
1405         if (flags & XFS_QMOPT_SBVERSION) {
1406 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1407                 unsigned oldv = mp->m_sb.sb_versionnum;
1408 #endif
1409                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1410                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1411                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1412                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1413                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1414
1415                 xfs_sb_version_addquota(&mp->m_sb);
1416                 mp->m_sb.sb_uquotino = NULLFSINO;
1417                 mp->m_sb.sb_gquotino = NULLFSINO;
1418
1419                 /* qflags will get updated _after_ quotacheck */
1420                 mp->m_sb.sb_qflags = 0;
1421 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1422                 cmn_err(CE_NOTE,
1423                         "Old superblock version %x, converting to %x.",
1424                         oldv, mp->m_sb.sb_versionnum);
1425 #endif
1426         }
1427         if (flags & XFS_QMOPT_UQUOTA)
1428                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1429         else
1430                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1431         spin_unlock(&mp->m_sb_lock);
1432         xfs_mod_sb(tp, sbfields);
1433
1434         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1435                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1436                 return error;
1437         }
1438         return 0;
1439 }
1440
1441
1442 STATIC void
1443 xfs_qm_reset_dqcounts(
1444         xfs_mount_t     *mp,
1445         xfs_buf_t       *bp,
1446         xfs_dqid_t      id,
1447         uint            type)
1448 {
1449         xfs_disk_dquot_t        *ddq;
1450         int                     j;
1451
1452         xfs_buftrace("RESET DQUOTS", bp);
1453         /*
1454          * Reset all counters and timers. They'll be
1455          * started afresh by xfs_qm_quotacheck.
1456          */
1457 #ifdef DEBUG
1458         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1459         do_div(j, sizeof(xfs_dqblk_t));
1460         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1461 #endif
1462         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1463         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1464                 /*
1465                  * Do a sanity check, and if needed, repair the dqblk. Don't
1466                  * output any warnings because it's perfectly possible to
1467                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1468                  */
1469                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1470                                       "xfs_quotacheck");
1471                 ddq->d_bcount = 0;
1472                 ddq->d_icount = 0;
1473                 ddq->d_rtbcount = 0;
1474                 ddq->d_btimer = 0;
1475                 ddq->d_itimer = 0;
1476                 ddq->d_rtbtimer = 0;
1477                 ddq->d_bwarns = 0;
1478                 ddq->d_iwarns = 0;
1479                 ddq->d_rtbwarns = 0;
1480                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1481         }
1482 }
1483
1484 STATIC int
1485 xfs_qm_dqiter_bufs(
1486         xfs_mount_t     *mp,
1487         xfs_dqid_t      firstid,
1488         xfs_fsblock_t   bno,
1489         xfs_filblks_t   blkcnt,
1490         uint            flags)
1491 {
1492         xfs_buf_t       *bp;
1493         int             error;
1494         int             notcommitted;
1495         int             incr;
1496         int             type;
1497
1498         ASSERT(blkcnt > 0);
1499         notcommitted = 0;
1500         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1501                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1502         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1503                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1504         error = 0;
1505
1506         /*
1507          * Blkcnt arg can be a very big number, and might even be
1508          * larger than the log itself. So, we have to break it up into
1509          * manageable-sized transactions.
1510          * Note that we don't start a permanent transaction here; we might
1511          * not be able to get a log reservation for the whole thing up front,
1512          * and we don't really care to either, because we just discard
1513          * everything if we were to crash in the middle of this loop.
1514          */
1515         while (blkcnt--) {
1516                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1517                               XFS_FSB_TO_DADDR(mp, bno),
1518                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1519                 if (error)
1520                         break;
1521
1522                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1523                 xfs_bdwrite(mp, bp);
1524                 /*
1525                  * goto the next block.
1526                  */
1527                 bno++;
1528                 firstid += XFS_QM_DQPERBLK(mp);
1529         }
1530         return error;
1531 }
1532
1533 /*
1534  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1535  * caller supplied function for every chunk of dquots that we find.
1536  */
1537 STATIC int
1538 xfs_qm_dqiterate(
1539         xfs_mount_t     *mp,
1540         xfs_inode_t     *qip,
1541         uint            flags)
1542 {
1543         xfs_bmbt_irec_t         *map;
1544         int                     i, nmaps;       /* number of map entries */
1545         int                     error;          /* return value */
1546         xfs_fileoff_t           lblkno;
1547         xfs_filblks_t           maxlblkcnt;
1548         xfs_dqid_t              firstid;
1549         xfs_fsblock_t           rablkno;
1550         xfs_filblks_t           rablkcnt;
1551
1552         error = 0;
1553         /*
1554          * This looks racy, but we can't keep an inode lock across a
1555          * trans_reserve. But, this gets called during quotacheck, and that
1556          * happens only at mount time which is single threaded.
1557          */
1558         if (qip->i_d.di_nblocks == 0)
1559                 return 0;
1560
1561         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1562
1563         lblkno = 0;
1564         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1565         do {
1566                 nmaps = XFS_DQITER_MAP_SIZE;
1567                 /*
1568                  * We aren't changing the inode itself. Just changing
1569                  * some of its data. No new blocks are added here, and
1570                  * the inode is never added to the transaction.
1571                  */
1572                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1573                 error = xfs_bmapi(NULL, qip, lblkno,
1574                                   maxlblkcnt - lblkno,
1575                                   XFS_BMAPI_METADATA,
1576                                   NULL,
1577                                   0, map, &nmaps, NULL, NULL);
1578                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1579                 if (error)
1580                         break;
1581
1582                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1583                 for (i = 0; i < nmaps; i++) {
1584                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1585                         ASSERT(map[i].br_blockcount);
1586
1587
1588                         lblkno += map[i].br_blockcount;
1589
1590                         if (map[i].br_startblock == HOLESTARTBLOCK)
1591                                 continue;
1592
1593                         firstid = (xfs_dqid_t) map[i].br_startoff *
1594                                 XFS_QM_DQPERBLK(mp);
1595                         /*
1596                          * Do a read-ahead on the next extent.
1597                          */
1598                         if ((i+1 < nmaps) &&
1599                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1600                                 rablkcnt =  map[i+1].br_blockcount;
1601                                 rablkno = map[i+1].br_startblock;
1602                                 while (rablkcnt--) {
1603                                         xfs_baread(mp->m_ddev_targp,
1604                                                XFS_FSB_TO_DADDR(mp, rablkno),
1605                                                (int)XFS_QI_DQCHUNKLEN(mp));
1606                                         rablkno++;
1607                                 }
1608                         }
1609                         /*
1610                          * Iterate thru all the blks in the extent and
1611                          * reset the counters of all the dquots inside them.
1612                          */
1613                         if ((error = xfs_qm_dqiter_bufs(mp,
1614                                                        firstid,
1615                                                        map[i].br_startblock,
1616                                                        map[i].br_blockcount,
1617                                                        flags))) {
1618                                 break;
1619                         }
1620                 }
1621
1622                 if (error)
1623                         break;
1624         } while (nmaps > 0);
1625
1626         kmem_free(map);
1627
1628         return error;
1629 }
1630
1631 /*
1632  * Called by dqusage_adjust in doing a quotacheck.
1633  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1634  * this updates its incore copy as well as the buffer copy. This is
1635  * so that once the quotacheck is done, we can just log all the buffers,
1636  * as opposed to logging numerous updates to individual dquots.
1637  */
1638 STATIC void
1639 xfs_qm_quotacheck_dqadjust(
1640         xfs_dquot_t             *dqp,
1641         xfs_qcnt_t              nblks,
1642         xfs_qcnt_t              rtblks)
1643 {
1644         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1645         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1646         /*
1647          * Adjust the inode count and the block count to reflect this inode's
1648          * resource usage.
1649          */
1650         be64_add_cpu(&dqp->q_core.d_icount, 1);
1651         dqp->q_res_icount++;
1652         if (nblks) {
1653                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1654                 dqp->q_res_bcount += nblks;
1655         }
1656         if (rtblks) {
1657                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1658                 dqp->q_res_rtbcount += rtblks;
1659         }
1660
1661         /*
1662          * Set default limits, adjust timers (since we changed usages)
1663          */
1664         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1665                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1666                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1667         }
1668
1669         dqp->dq_flags |= XFS_DQ_DIRTY;
1670 }
1671
1672 STATIC int
1673 xfs_qm_get_rtblks(
1674         xfs_inode_t     *ip,
1675         xfs_qcnt_t      *O_rtblks)
1676 {
1677         xfs_filblks_t   rtblks;                 /* total rt blks */
1678         xfs_extnum_t    idx;                    /* extent record index */
1679         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1680         xfs_extnum_t    nextents;               /* number of extent entries */
1681         int             error;
1682
1683         ASSERT(XFS_IS_REALTIME_INODE(ip));
1684         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1685         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1686                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1687                         return error;
1688         }
1689         rtblks = 0;
1690         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1691         for (idx = 0; idx < nextents; idx++)
1692                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1693         *O_rtblks = (xfs_qcnt_t)rtblks;
1694         return 0;
1695 }
1696
1697 /*
1698  * callback routine supplied to bulkstat(). Given an inumber, find its
1699  * dquots and update them to account for resources taken by that inode.
1700  */
1701 /* ARGSUSED */
1702 STATIC int
1703 xfs_qm_dqusage_adjust(
1704         xfs_mount_t     *mp,            /* mount point for filesystem */
1705         xfs_ino_t       ino,            /* inode number to get data for */
1706         void            __user *buffer, /* not used */
1707         int             ubsize,         /* not used */
1708         void            *private_data,  /* not used */
1709         xfs_daddr_t     bno,            /* starting block of inode cluster */
1710         int             *ubused,        /* not used */
1711         void            *dip,           /* on-disk inode pointer (not used) */
1712         int             *res)           /* result code value */
1713 {
1714         xfs_inode_t     *ip;
1715         xfs_dquot_t     *udqp, *gdqp;
1716         xfs_qcnt_t      nblks, rtblks;
1717         int             error;
1718
1719         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1720
1721         /*
1722          * rootino must have its resources accounted for, not so with the quota
1723          * inodes.
1724          */
1725         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1726                 *res = BULKSTAT_RV_NOTHING;
1727                 return XFS_ERROR(EINVAL);
1728         }
1729
1730         /*
1731          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1732          * interface expects the inode to be exclusively locked because that's
1733          * the case in all other instances. It's OK that we do this because
1734          * quotacheck is done only at mount time.
1735          */
1736         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1737                 *res = BULKSTAT_RV_NOTHING;
1738                 return error;
1739         }
1740
1741         /*
1742          * Obtain the locked dquots. In case of an error (eg. allocation
1743          * fails for ENOSPC), we return the negative of the error number
1744          * to bulkstat, so that it can get propagated to quotacheck() and
1745          * making us disable quotas for the file system.
1746          */
1747         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1748                 xfs_iput(ip, XFS_ILOCK_EXCL);
1749                 *res = BULKSTAT_RV_GIVEUP;
1750                 return error;
1751         }
1752
1753         rtblks = 0;
1754         if (! XFS_IS_REALTIME_INODE(ip)) {
1755                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1756         } else {
1757                 /*
1758                  * Walk thru the extent list and count the realtime blocks.
1759                  */
1760                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1761                         xfs_iput(ip, XFS_ILOCK_EXCL);
1762                         if (udqp)
1763                                 xfs_qm_dqput(udqp);
1764                         if (gdqp)
1765                                 xfs_qm_dqput(gdqp);
1766                         *res = BULKSTAT_RV_GIVEUP;
1767                         return error;
1768                 }
1769                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1770         }
1771         ASSERT(ip->i_delayed_blks == 0);
1772
1773         /*
1774          * We can't release the inode while holding its dquot locks.
1775          * The inode can go into inactive and might try to acquire the dquotlocks.
1776          * So, just unlock here and do a vn_rele at the end.
1777          */
1778         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1779
1780         /*
1781          * Add the (disk blocks and inode) resources occupied by this
1782          * inode to its dquots. We do this adjustment in the incore dquot,
1783          * and also copy the changes to its buffer.
1784          * We don't care about putting these changes in a transaction
1785          * envelope because if we crash in the middle of a 'quotacheck'
1786          * we have to start from the beginning anyway.
1787          * Once we're done, we'll log all the dquot bufs.
1788          *
1789          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1790          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1791          */
1792         if (XFS_IS_UQUOTA_ON(mp)) {
1793                 ASSERT(udqp);
1794                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1795                 xfs_qm_dqput(udqp);
1796         }
1797         if (XFS_IS_OQUOTA_ON(mp)) {
1798                 ASSERT(gdqp);
1799                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1800                 xfs_qm_dqput(gdqp);
1801         }
1802         /*
1803          * Now release the inode. This will send it to 'inactive', and
1804          * possibly even free blocks.
1805          */
1806         IRELE(ip);
1807
1808         /*
1809          * Goto next inode.
1810          */
1811         *res = BULKSTAT_RV_DIDONE;
1812         return 0;
1813 }
1814
1815 /*
1816  * Walk thru all the filesystem inodes and construct a consistent view
1817  * of the disk quota world. If the quotacheck fails, disable quotas.
1818  */
1819 int
1820 xfs_qm_quotacheck(
1821         xfs_mount_t     *mp)
1822 {
1823         int             done, count, error;
1824         xfs_ino_t       lastino;
1825         size_t          structsz;
1826         xfs_inode_t     *uip, *gip;
1827         uint            flags;
1828
1829         count = INT_MAX;
1830         structsz = 1;
1831         lastino = 0;
1832         flags = 0;
1833
1834         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1835         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1836
1837         /*
1838          * There should be no cached dquots. The (simplistic) quotacheck
1839          * algorithm doesn't like that.
1840          */
1841         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1842
1843         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1844
1845         /*
1846          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1847          * their counters to zero. We need a clean slate.
1848          * We don't log our changes till later.
1849          */
1850         if ((uip = XFS_QI_UQIP(mp))) {
1851                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1852                         goto error_return;
1853                 flags |= XFS_UQUOTA_CHKD;
1854         }
1855
1856         if ((gip = XFS_QI_GQIP(mp))) {
1857                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1858                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1859                         goto error_return;
1860                 flags |= XFS_OQUOTA_CHKD;
1861         }
1862
1863         do {
1864                 /*
1865                  * Iterate thru all the inodes in the file system,
1866                  * adjusting the corresponding dquot counters in core.
1867                  */
1868                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1869                                      xfs_qm_dqusage_adjust, NULL,
1870                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1871                         break;
1872
1873         } while (! done);
1874
1875         /*
1876          * We've made all the changes that we need to make incore.
1877          * Flush them down to disk buffers if everything was updated
1878          * successfully.
1879          */
1880         if (!error)
1881                 error = xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1882
1883         /*
1884          * We can get this error if we couldn't do a dquot allocation inside
1885          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1886          * dirty dquots that might be cached, we just want to get rid of them
1887          * and turn quotaoff. The dquots won't be attached to any of the inodes
1888          * at this point (because we intentionally didn't in dqget_noattach).
1889          */
1890         if (error) {
1891                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1892                 goto error_return;
1893         }
1894
1895         /*
1896          * We didn't log anything, because if we crashed, we'll have to
1897          * start the quotacheck from scratch anyway. However, we must make
1898          * sure that our dquot changes are secure before we put the
1899          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1900          * flush.
1901          */
1902         XFS_bflush(mp->m_ddev_targp);
1903
1904         /*
1905          * If one type of quotas is off, then it will lose its
1906          * quotachecked status, since we won't be doing accounting for
1907          * that type anymore.
1908          */
1909         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1910         mp->m_qflags |= flags;
1911
1912         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1913
1914  error_return:
1915         if (error) {
1916                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1917                         "Disabling quotas.",
1918                         mp->m_fsname, error);
1919                 /*
1920                  * We must turn off quotas.
1921                  */
1922                 ASSERT(mp->m_quotainfo != NULL);
1923                 ASSERT(xfs_Gqm != NULL);
1924                 xfs_qm_destroy_quotainfo(mp);
1925                 if (xfs_mount_reset_sbqflags(mp)) {
1926                         cmn_err(CE_WARN, "XFS quotacheck %s: "
1927                                 "Failed to reset quota flags.", mp->m_fsname);
1928                 }
1929         } else {
1930                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1931         }
1932         return (error);
1933 }
1934
1935 /*
1936  * This is called after the superblock has been read in and we're ready to
1937  * iget the quota inodes.
1938  */
1939 STATIC int
1940 xfs_qm_init_quotainos(
1941         xfs_mount_t     *mp)
1942 {
1943         xfs_inode_t     *uip, *gip;
1944         int             error;
1945         __int64_t       sbflags;
1946         uint            flags;
1947
1948         ASSERT(mp->m_quotainfo);
1949         uip = gip = NULL;
1950         sbflags = 0;
1951         flags = 0;
1952
1953         /*
1954          * Get the uquota and gquota inodes
1955          */
1956         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1957                 if (XFS_IS_UQUOTA_ON(mp) &&
1958                     mp->m_sb.sb_uquotino != NULLFSINO) {
1959                         ASSERT(mp->m_sb.sb_uquotino > 0);
1960                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1961                                              0, 0, &uip, 0)))
1962                                 return XFS_ERROR(error);
1963                 }
1964                 if (XFS_IS_OQUOTA_ON(mp) &&
1965                     mp->m_sb.sb_gquotino != NULLFSINO) {
1966                         ASSERT(mp->m_sb.sb_gquotino > 0);
1967                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1968                                              0, 0, &gip, 0))) {
1969                                 if (uip)
1970                                         IRELE(uip);
1971                                 return XFS_ERROR(error);
1972                         }
1973                 }
1974         } else {
1975                 flags |= XFS_QMOPT_SBVERSION;
1976                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1977                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1978         }
1979
1980         /*
1981          * Create the two inodes, if they don't exist already. The changes
1982          * made above will get added to a transaction and logged in one of
1983          * the qino_alloc calls below.  If the device is readonly,
1984          * temporarily switch to read-write to do this.
1985          */
1986         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1987                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1988                                               sbflags | XFS_SB_UQUOTINO,
1989                                               flags | XFS_QMOPT_UQUOTA)))
1990                         return XFS_ERROR(error);
1991
1992                 flags &= ~XFS_QMOPT_SBVERSION;
1993         }
1994         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1995                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1996                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1997                 error = xfs_qm_qino_alloc(mp, &gip,
1998                                           sbflags | XFS_SB_GQUOTINO, flags);
1999                 if (error) {
2000                         if (uip)
2001                                 IRELE(uip);
2002
2003                         return XFS_ERROR(error);
2004                 }
2005         }
2006
2007         XFS_QI_UQIP(mp) = uip;
2008         XFS_QI_GQIP(mp) = gip;
2009
2010         return 0;
2011 }
2012
2013
2014 /*
2015  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2016  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2017  * favor the lookup function ...
2018  * XXXsup merge this with qm_reclaim_one().
2019  */
2020 STATIC int
2021 xfs_qm_shake_freelist(
2022         int howmany)
2023 {
2024         int             nreclaimed;
2025         xfs_dqhash_t    *hash;
2026         xfs_dquot_t     *dqp, *nextdqp;
2027         int             restarts;
2028         int             nflushes;
2029
2030         if (howmany <= 0)
2031                 return 0;
2032
2033         nreclaimed = 0;
2034         restarts = 0;
2035         nflushes = 0;
2036
2037 #ifdef QUOTADEBUG
2038         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2039 #endif
2040         /* lock order is : hashchainlock, freelistlock, mplistlock */
2041  tryagain:
2042         xfs_qm_freelist_lock(xfs_Gqm);
2043
2044         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2045              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2046               nreclaimed < howmany); ) {
2047                 xfs_dqlock(dqp);
2048
2049                 /*
2050                  * We are racing with dqlookup here. Naturally we don't
2051                  * want to reclaim a dquot that lookup wants.
2052                  */
2053                 if (dqp->dq_flags & XFS_DQ_WANT) {
2054                         xfs_dqunlock(dqp);
2055                         xfs_qm_freelist_unlock(xfs_Gqm);
2056                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2057                                 return nreclaimed;
2058                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2059                         goto tryagain;
2060                 }
2061
2062                 /*
2063                  * If the dquot is inactive, we are assured that it is
2064                  * not on the mplist or the hashlist, and that makes our
2065                  * life easier.
2066                  */
2067                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2068                         ASSERT(dqp->q_mount == NULL);
2069                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2070                         ASSERT(dqp->HL_PREVP == NULL);
2071                         ASSERT(dqp->MPL_PREVP == NULL);
2072                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2073                         nextdqp = dqp->dq_flnext;
2074                         goto off_freelist;
2075                 }
2076
2077                 ASSERT(dqp->MPL_PREVP);
2078                 /*
2079                  * Try to grab the flush lock. If this dquot is in the process of
2080                  * getting flushed to disk, we don't want to reclaim it.
2081                  */
2082                 if (!xfs_dqflock_nowait(dqp)) {
2083                         xfs_dqunlock(dqp);
2084                         dqp = dqp->dq_flnext;
2085                         continue;
2086                 }
2087
2088                 /*
2089                  * We have the flush lock so we know that this is not in the
2090                  * process of being flushed. So, if this is dirty, flush it
2091                  * DELWRI so that we don't get a freelist infested with
2092                  * dirty dquots.
2093                  */
2094                 if (XFS_DQ_IS_DIRTY(dqp)) {
2095                         int     error;
2096                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2097                         /*
2098                          * We flush it delayed write, so don't bother
2099                          * releasing the mplock.
2100                          */
2101                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2102                         if (error) {
2103                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2104                         "xfs_qm_dqflush_all: dquot %p flush failed", dqp);
2105                         }
2106                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2107                         dqp = dqp->dq_flnext;
2108                         continue;
2109                 }
2110                 /*
2111                  * We're trying to get the hashlock out of order. This races
2112                  * with dqlookup; so, we giveup and goto the next dquot if
2113                  * we couldn't get the hashlock. This way, we won't starve
2114                  * a dqlookup process that holds the hashlock that is
2115                  * waiting for the freelist lock.
2116                  */
2117                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2118                         xfs_dqfunlock(dqp);
2119                         xfs_dqunlock(dqp);
2120                         dqp = dqp->dq_flnext;
2121                         continue;
2122                 }
2123                 /*
2124                  * This races with dquot allocation code as well as dqflush_all
2125                  * and reclaim code. So, if we failed to grab the mplist lock,
2126                  * giveup everything and start over.
2127                  */
2128                 hash = dqp->q_hash;
2129                 ASSERT(hash);
2130                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2131                         /* XXX put a sentinel so that we can come back here */
2132                         xfs_dqfunlock(dqp);
2133                         xfs_dqunlock(dqp);
2134                         XFS_DQ_HASH_UNLOCK(hash);
2135                         xfs_qm_freelist_unlock(xfs_Gqm);
2136                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2137                                 return nreclaimed;
2138                         goto tryagain;
2139                 }
2140                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2141 #ifdef QUOTADEBUG
2142                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2143                         dqp, be32_to_cpu(dqp->q_core.d_id));
2144 #endif
2145                 ASSERT(dqp->q_nrefs == 0);
2146                 nextdqp = dqp->dq_flnext;
2147                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2148                 XQM_HASHLIST_REMOVE(hash, dqp);
2149                 xfs_dqfunlock(dqp);
2150                 xfs_qm_mplist_unlock(dqp->q_mount);
2151                 XFS_DQ_HASH_UNLOCK(hash);
2152
2153  off_freelist:
2154                 XQM_FREELIST_REMOVE(dqp);
2155                 xfs_dqunlock(dqp);
2156                 nreclaimed++;
2157                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2158                 xfs_qm_dqdestroy(dqp);
2159                 dqp = nextdqp;
2160         }
2161         xfs_qm_freelist_unlock(xfs_Gqm);
2162         return nreclaimed;
2163 }
2164
2165
2166 /*
2167  * The kmem_shake interface is invoked when memory is running low.
2168  */
2169 /* ARGSUSED */
2170 STATIC int
2171 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2172 {
2173         int     ndqused, nfree, n;
2174
2175         if (!kmem_shake_allow(gfp_mask))
2176                 return 0;
2177         if (!xfs_Gqm)
2178                 return 0;
2179
2180         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2181         /* incore dquots in all f/s's */
2182         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2183
2184         ASSERT(ndqused >= 0);
2185
2186         if (nfree <= ndqused && nfree < ndquot)
2187                 return 0;
2188
2189         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2190         n = nfree - ndqused - ndquot;           /* # over target */
2191
2192         return xfs_qm_shake_freelist(MAX(nfree, n));
2193 }
2194
2195
2196 /*
2197  * Just pop the least recently used dquot off the freelist and
2198  * recycle it. The returned dquot is locked.
2199  */
2200 STATIC xfs_dquot_t *
2201 xfs_qm_dqreclaim_one(void)
2202 {
2203         xfs_dquot_t     *dqpout;
2204         xfs_dquot_t     *dqp;
2205         int             restarts;
2206         int             nflushes;
2207
2208         restarts = 0;
2209         dqpout = NULL;
2210         nflushes = 0;
2211
2212         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2213  startagain:
2214         xfs_qm_freelist_lock(xfs_Gqm);
2215
2216         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2217                 xfs_dqlock(dqp);
2218
2219                 /*
2220                  * We are racing with dqlookup here. Naturally we don't
2221                  * want to reclaim a dquot that lookup wants. We release the
2222                  * freelist lock and start over, so that lookup will grab
2223                  * both the dquot and the freelistlock.
2224                  */
2225                 if (dqp->dq_flags & XFS_DQ_WANT) {
2226                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2227                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2228                         xfs_dqunlock(dqp);
2229                         xfs_qm_freelist_unlock(xfs_Gqm);
2230                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2231                                 return NULL;
2232                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2233                         goto startagain;
2234                 }
2235
2236                 /*
2237                  * If the dquot is inactive, we are assured that it is
2238                  * not on the mplist or the hashlist, and that makes our
2239                  * life easier.
2240                  */
2241                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2242                         ASSERT(dqp->q_mount == NULL);
2243                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2244                         ASSERT(dqp->HL_PREVP == NULL);
2245                         ASSERT(dqp->MPL_PREVP == NULL);
2246                         XQM_FREELIST_REMOVE(dqp);
2247                         xfs_dqunlock(dqp);
2248                         dqpout = dqp;
2249                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2250                         break;
2251                 }
2252
2253                 ASSERT(dqp->q_hash);
2254                 ASSERT(dqp->MPL_PREVP);
2255
2256                 /*
2257                  * Try to grab the flush lock. If this dquot is in the process of
2258                  * getting flushed to disk, we don't want to reclaim it.
2259                  */
2260                 if (!xfs_dqflock_nowait(dqp)) {
2261                         xfs_dqunlock(dqp);
2262                         continue;
2263                 }
2264
2265                 /*
2266                  * We have the flush lock so we know that this is not in the
2267                  * process of being flushed. So, if this is dirty, flush it
2268                  * DELWRI so that we don't get a freelist infested with
2269                  * dirty dquots.
2270                  */
2271                 if (XFS_DQ_IS_DIRTY(dqp)) {
2272                         int     error;
2273                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2274                         /*
2275                          * We flush it delayed write, so don't bother
2276                          * releasing the freelist lock.
2277                          */
2278                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2279                         if (error) {
2280                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2281                         "xfs_qm_dqreclaim: dquot %p flush failed", dqp);
2282                         }
2283                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2284                         continue;
2285                 }
2286
2287                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2288                         xfs_dqfunlock(dqp);
2289                         xfs_dqunlock(dqp);
2290                         continue;
2291                 }
2292
2293                 if (! xfs_qm_dqhashlock_nowait(dqp))
2294                         goto mplistunlock;
2295
2296                 ASSERT(dqp->q_nrefs == 0);
2297                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2298                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2299                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2300                 XQM_FREELIST_REMOVE(dqp);
2301                 dqpout = dqp;
2302                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2303  mplistunlock:
2304                 xfs_qm_mplist_unlock(dqp->q_mount);
2305                 xfs_dqfunlock(dqp);
2306                 xfs_dqunlock(dqp);
2307                 if (dqpout)
2308                         break;
2309         }
2310
2311         xfs_qm_freelist_unlock(xfs_Gqm);
2312         return dqpout;
2313 }
2314
2315
2316 /*------------------------------------------------------------------*/
2317
2318 /*
2319  * Return a new incore dquot. Depending on the number of
2320  * dquots in the system, we either allocate a new one on the kernel heap,
2321  * or reclaim a free one.
2322  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2323  * to reclaim an existing one from the freelist.
2324  */
2325 boolean_t
2326 xfs_qm_dqalloc_incore(
2327         xfs_dquot_t **O_dqpp)
2328 {
2329         xfs_dquot_t     *dqp;
2330
2331         /*
2332          * Check against high water mark to see if we want to pop
2333          * a nincompoop dquot off the freelist.
2334          */
2335         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2336                 /*
2337                  * Try to recycle a dquot from the freelist.
2338                  */
2339                 if ((dqp = xfs_qm_dqreclaim_one())) {
2340                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2341                         /*
2342                          * Just zero the core here. The rest will get
2343                          * reinitialized by caller. XXX we shouldn't even
2344                          * do this zero ...
2345                          */
2346                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2347                         *O_dqpp = dqp;
2348                         return B_FALSE;
2349                 }
2350                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2351         }
2352
2353         /*
2354          * Allocate a brand new dquot on the kernel heap and return it
2355          * to the caller to initialize.
2356          */
2357         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2358         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2359         atomic_inc(&xfs_Gqm->qm_totaldquots);
2360
2361         return B_TRUE;
2362 }
2363
2364
2365 /*
2366  * Start a transaction and write the incore superblock changes to
2367  * disk. flags parameter indicates which fields have changed.
2368  */
2369 int
2370 xfs_qm_write_sb_changes(
2371         xfs_mount_t     *mp,
2372         __int64_t       flags)
2373 {
2374         xfs_trans_t     *tp;
2375         int             error;
2376
2377 #ifdef QUOTADEBUG
2378         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2379 #endif
2380         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2381         if ((error = xfs_trans_reserve(tp, 0,
2382                                       mp->m_sb.sb_sectsize + 128, 0,
2383                                       0,
2384                                       XFS_DEFAULT_LOG_COUNT))) {
2385                 xfs_trans_cancel(tp, 0);
2386                 return error;
2387         }
2388
2389         xfs_mod_sb(tp, flags);
2390         error = xfs_trans_commit(tp, 0);
2391
2392         return error;
2393 }
2394
2395
2396 /* --------------- utility functions for vnodeops ---------------- */
2397
2398
2399 /*
2400  * Given an inode, a uid and gid (from cred_t) make sure that we have
2401  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2402  * quotas by creating this file.
2403  * This also attaches dquot(s) to the given inode after locking it,
2404  * and returns the dquots corresponding to the uid and/or gid.
2405  *
2406  * in   : inode (unlocked)
2407  * out  : udquot, gdquot with references taken and unlocked
2408  */
2409 int
2410 xfs_qm_vop_dqalloc(
2411         xfs_mount_t     *mp,
2412         xfs_inode_t     *ip,
2413         uid_t           uid,
2414         gid_t           gid,
2415         prid_t          prid,
2416         uint            flags,
2417         xfs_dquot_t     **O_udqpp,
2418         xfs_dquot_t     **O_gdqpp)
2419 {
2420         int             error;
2421         xfs_dquot_t     *uq, *gq;
2422         uint            lockflags;
2423
2424         if (!XFS_IS_QUOTA_ON(mp))
2425                 return 0;
2426
2427         lockflags = XFS_ILOCK_EXCL;
2428         xfs_ilock(ip, lockflags);
2429
2430         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2431                 gid = ip->i_d.di_gid;
2432
2433         /*
2434          * Attach the dquot(s) to this inode, doing a dquot allocation
2435          * if necessary. The dquot(s) will not be locked.
2436          */
2437         if (XFS_NOT_DQATTACHED(mp, ip)) {
2438                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2439                                             XFS_QMOPT_ILOCKED))) {
2440                         xfs_iunlock(ip, lockflags);
2441                         return error;
2442                 }
2443         }
2444
2445         uq = gq = NULL;
2446         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2447                 if (ip->i_d.di_uid != uid) {
2448                         /*
2449                          * What we need is the dquot that has this uid, and
2450                          * if we send the inode to dqget, the uid of the inode
2451                          * takes priority over what's sent in the uid argument.
2452                          * We must unlock inode here before calling dqget if
2453                          * we're not sending the inode, because otherwise
2454                          * we'll deadlock by doing trans_reserve while
2455                          * holding ilock.
2456                          */
2457                         xfs_iunlock(ip, lockflags);
2458                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2459                                                  XFS_DQ_USER,
2460                                                  XFS_QMOPT_DQALLOC |
2461                                                  XFS_QMOPT_DOWARN,
2462                                                  &uq))) {
2463                                 ASSERT(error != ENOENT);
2464                                 return error;
2465                         }
2466                         /*
2467                          * Get the ilock in the right order.
2468                          */
2469                         xfs_dqunlock(uq);
2470                         lockflags = XFS_ILOCK_SHARED;
2471                         xfs_ilock(ip, lockflags);
2472                 } else {
2473                         /*
2474                          * Take an extra reference, because we'll return
2475                          * this to caller
2476                          */
2477                         ASSERT(ip->i_udquot);
2478                         uq = ip->i_udquot;
2479                         xfs_dqlock(uq);
2480                         XFS_DQHOLD(uq);
2481                         xfs_dqunlock(uq);
2482                 }
2483         }
2484         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2485                 if (ip->i_d.di_gid != gid) {
2486                         xfs_iunlock(ip, lockflags);
2487                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2488                                                  XFS_DQ_GROUP,
2489                                                  XFS_QMOPT_DQALLOC |
2490                                                  XFS_QMOPT_DOWARN,
2491                                                  &gq))) {
2492                                 if (uq)
2493                                         xfs_qm_dqrele(uq);
2494                                 ASSERT(error != ENOENT);
2495                                 return error;
2496                         }
2497                         xfs_dqunlock(gq);
2498                         lockflags = XFS_ILOCK_SHARED;
2499                         xfs_ilock(ip, lockflags);
2500                 } else {
2501                         ASSERT(ip->i_gdquot);
2502                         gq = ip->i_gdquot;
2503                         xfs_dqlock(gq);
2504                         XFS_DQHOLD(gq);
2505                         xfs_dqunlock(gq);
2506                 }
2507         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2508                 if (ip->i_d.di_projid != prid) {
2509                         xfs_iunlock(ip, lockflags);
2510                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2511                                                  XFS_DQ_PROJ,
2512                                                  XFS_QMOPT_DQALLOC |
2513                                                  XFS_QMOPT_DOWARN,
2514                                                  &gq))) {
2515                                 if (uq)
2516                                         xfs_qm_dqrele(uq);
2517                                 ASSERT(error != ENOENT);
2518                                 return (error);
2519                         }
2520                         xfs_dqunlock(gq);
2521                         lockflags = XFS_ILOCK_SHARED;
2522                         xfs_ilock(ip, lockflags);
2523                 } else {
2524                         ASSERT(ip->i_gdquot);
2525                         gq = ip->i_gdquot;
2526                         xfs_dqlock(gq);
2527                         XFS_DQHOLD(gq);
2528                         xfs_dqunlock(gq);
2529                 }
2530         }
2531         if (uq)
2532                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2533
2534         xfs_iunlock(ip, lockflags);
2535         if (O_udqpp)
2536                 *O_udqpp = uq;
2537         else if (uq)
2538                 xfs_qm_dqrele(uq);
2539         if (O_gdqpp)
2540                 *O_gdqpp = gq;
2541         else if (gq)
2542                 xfs_qm_dqrele(gq);
2543         return 0;
2544 }
2545
2546 /*
2547  * Actually transfer ownership, and do dquot modifications.
2548  * These were already reserved.
2549  */
2550 xfs_dquot_t *
2551 xfs_qm_vop_chown(
2552         xfs_trans_t     *tp,
2553         xfs_inode_t     *ip,
2554         xfs_dquot_t     **IO_olddq,
2555         xfs_dquot_t     *newdq)
2556 {
2557         xfs_dquot_t     *prevdq;
2558         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2559                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2560
2561         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2562         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2563
2564         /* old dquot */
2565         prevdq = *IO_olddq;
2566         ASSERT(prevdq);
2567         ASSERT(prevdq != newdq);
2568
2569         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2570         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2571
2572         /* the sparkling new dquot */
2573         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2574         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2575
2576         /*
2577          * Take an extra reference, because the inode
2578          * is going to keep this dquot pointer even
2579          * after the trans_commit.
2580          */
2581         xfs_dqlock(newdq);
2582         XFS_DQHOLD(newdq);
2583         xfs_dqunlock(newdq);
2584         *IO_olddq = newdq;
2585
2586         return prevdq;
2587 }
2588
2589 /*
2590  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2591  */
2592 int
2593 xfs_qm_vop_chown_reserve(
2594         xfs_trans_t     *tp,
2595         xfs_inode_t     *ip,
2596         xfs_dquot_t     *udqp,
2597         xfs_dquot_t     *gdqp,
2598         uint            flags)
2599 {
2600         int             error;
2601         xfs_mount_t     *mp;
2602         uint            delblks, blkflags, prjflags = 0;
2603         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2604
2605         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2606         mp = ip->i_mount;
2607         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2608
2609         delblks = ip->i_delayed_blks;
2610         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2611         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2612                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2613
2614         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2615             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2616                 delblksudq = udqp;
2617                 /*
2618                  * If there are delayed allocation blocks, then we have to
2619                  * unreserve those from the old dquot, and add them to the
2620                  * new dquot.
2621                  */
2622                 if (delblks) {
2623                         ASSERT(ip->i_udquot);
2624                         unresudq = ip->i_udquot;
2625                 }
2626         }
2627         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2628                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2629                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2630                         prjflags = XFS_QMOPT_ENOSPC;
2631
2632                 if (prjflags ||
2633                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2634                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2635                         delblksgdq = gdqp;
2636                         if (delblks) {
2637                                 ASSERT(ip->i_gdquot);
2638                                 unresgdq = ip->i_gdquot;
2639                         }
2640                 }
2641         }
2642
2643         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2644                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2645                                 flags | blkflags | prjflags)))
2646                 return (error);
2647
2648         /*
2649          * Do the delayed blks reservations/unreservations now. Since, these
2650          * are done without the help of a transaction, if a reservation fails
2651          * its previous reservations won't be automatically undone by trans
2652          * code. So, we have to do it manually here.
2653          */
2654         if (delblks) {
2655                 /*
2656                  * Do the reservations first. Unreservation can't fail.
2657                  */
2658                 ASSERT(delblksudq || delblksgdq);
2659                 ASSERT(unresudq || unresgdq);
2660                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2661                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2662                                 flags | blkflags | prjflags)))
2663                         return (error);
2664                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2665                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2666                                 blkflags);
2667         }
2668
2669         return (0);
2670 }
2671
2672 int
2673 xfs_qm_vop_rename_dqattach(
2674         xfs_inode_t     **i_tab)
2675 {
2676         xfs_inode_t     *ip;
2677         int             i;
2678         int             error;
2679
2680         ip = i_tab[0];
2681
2682         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2683                 return 0;
2684
2685         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2686                 error = xfs_qm_dqattach(ip, 0);
2687                 if (error)
2688                         return error;
2689         }
2690         for (i = 1; (i < 4 && i_tab[i]); i++) {
2691                 /*
2692                  * Watch out for duplicate entries in the table.
2693                  */
2694                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2695                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2696                                 error = xfs_qm_dqattach(ip, 0);
2697                                 if (error)
2698                                         return error;
2699                         }
2700                 }
2701         }
2702         return 0;
2703 }
2704
2705 void
2706 xfs_qm_vop_dqattach_and_dqmod_newinode(
2707         xfs_trans_t     *tp,
2708         xfs_inode_t     *ip,
2709         xfs_dquot_t     *udqp,
2710         xfs_dquot_t     *gdqp)
2711 {
2712         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2713                 return;
2714
2715         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2716         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2717
2718         if (udqp) {
2719                 xfs_dqlock(udqp);
2720                 XFS_DQHOLD(udqp);
2721                 xfs_dqunlock(udqp);
2722                 ASSERT(ip->i_udquot == NULL);
2723                 ip->i_udquot = udqp;
2724                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2725                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2726                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2727         }
2728         if (gdqp) {
2729                 xfs_dqlock(gdqp);
2730                 XFS_DQHOLD(gdqp);
2731                 xfs_dqunlock(gdqp);
2732                 ASSERT(ip->i_gdquot == NULL);
2733                 ip->i_gdquot = gdqp;
2734                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2735                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2736                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2737                                 be32_to_cpu(gdqp->q_core.d_id));
2738                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2739         }
2740 }
2741
2742 /* ------------- list stuff -----------------*/
2743 STATIC void
2744 xfs_qm_freelist_init(xfs_frlist_t *ql)
2745 {
2746         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2747         mutex_init(&ql->qh_lock);
2748         ql->qh_version = 0;
2749         ql->qh_nelems = 0;
2750 }
2751
2752 STATIC void
2753 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2754 {
2755         xfs_dquot_t     *dqp, *nextdqp;
2756
2757         mutex_lock(&ql->qh_lock);
2758         for (dqp = ql->qh_next;
2759              dqp != (xfs_dquot_t *)ql; ) {
2760                 xfs_dqlock(dqp);
2761                 nextdqp = dqp->dq_flnext;
2762 #ifdef QUOTADEBUG
2763                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2764 #endif
2765                 XQM_FREELIST_REMOVE(dqp);
2766                 xfs_dqunlock(dqp);
2767                 xfs_qm_dqdestroy(dqp);
2768                 dqp = nextdqp;
2769         }
2770         mutex_unlock(&ql->qh_lock);
2771         mutex_destroy(&ql->qh_lock);
2772
2773         ASSERT(ql->qh_nelems == 0);
2774 }
2775
2776 STATIC void
2777 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2778 {
2779         dq->dq_flnext = ql->qh_next;
2780         dq->dq_flprev = (xfs_dquot_t *)ql;
2781         ql->qh_next = dq;
2782         dq->dq_flnext->dq_flprev = dq;
2783         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2784         xfs_Gqm->qm_dqfreelist.qh_version++;
2785 }
2786
2787 void
2788 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2789 {
2790         xfs_dquot_t *next = dq->dq_flnext;
2791         xfs_dquot_t *prev = dq->dq_flprev;
2792
2793         next->dq_flprev = prev;
2794         prev->dq_flnext = next;
2795         dq->dq_flnext = dq->dq_flprev = dq;
2796         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2797         xfs_Gqm->qm_dqfreelist.qh_version++;
2798 }
2799
2800 void
2801 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2802 {
2803         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2804 }
2805
2806 STATIC int
2807 xfs_qm_dqhashlock_nowait(
2808         xfs_dquot_t *dqp)
2809 {
2810         int locked;
2811
2812         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2813         return locked;
2814 }
2815
2816 int
2817 xfs_qm_freelist_lock_nowait(
2818         xfs_qm_t *xqm)
2819 {
2820         int locked;
2821
2822         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2823         return locked;
2824 }
2825
2826 STATIC int
2827 xfs_qm_mplist_nowait(
2828         xfs_mount_t     *mp)
2829 {
2830         int locked;
2831
2832         ASSERT(mp->m_quotainfo);
2833         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2834         return locked;
2835 }