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, hsize * sizeof(xfs_dqhash_t));
 196         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
 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, sizeof(xfs_qm_t));
 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 int
 309 xfs_qm_mount_quotas(
 310         xfs_mount_t     *mp,
 311         int             mfsi_flags)
 312 {
 313         int             error = 0;
 314         uint            sbf;
 315
 316
 317         /*
 318          * If quotas on realtime volumes is not supported, we disable
 319          * quotas immediately.
 320          */
 321         if (mp->m_sb.sb_rextents) {
 322                 cmn_err(CE_NOTE,
 323                         "Cannot turn on quotas for realtime filesystem %s",
 324                         mp->m_fsname);
 325                 mp->m_qflags = 0;
 326                 goto write_changes;
 327         }
 328
 329         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 330
 331         /*
 332          * Allocate the quotainfo structure inside the mount struct, and
 333          * create quotainode(s), and change/rev superblock if necessary.
 334          */
 335         if ((error = xfs_qm_init_quotainfo(mp))) {
 336                 /*
 337                  * We must turn off quotas.
 338                  */
 339                 ASSERT(mp->m_quotainfo == NULL);
 340                 mp->m_qflags = 0;
 341                 goto write_changes;
 342         }
 343         /*
 344          * If any of the quotas are not consistent, do a quotacheck.
 345          */
 346         if (XFS_QM_NEED_QUOTACHECK(mp) &&
 347                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
 348                 if ((error = xfs_qm_quotacheck(mp))) {
 349                         /* Quotacheck has failed and quotas have
 350                          * been disabled.
 351                          */
 352                         return XFS_ERROR(error);
 353                 }
 354         }
 355         /*
 356          * If one type of quotas is off, then it will lose its
 357          * quotachecked status, since we won't be doing accounting for
 358          * that type anymore.
 359          */
 360         if (!XFS_IS_UQUOTA_ON(mp)) {
 361                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
 362         }
 363         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) {
 364                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
 365         }
 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 XFS_ERROR(error);
 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                  XFS_PURGE_INODE(uqp);
 449                  mp->m_quotainfo->qi_uquotaip = NULL;
 450         }
 451         if (gqp) {
 452                 XFS_PURGE_INODE(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_qm_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, flags);
 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_ISLOCKED_INODE_EXCL(ip));
 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_ISLOCKED_INODE_EXCL(ip));
 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         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
 892         if (XFS_IS_OQUOTA_ON(mp)) {
 893                 error = XFS_IS_GQUOTA_ON(mp) ?
 894                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
 895                                                 flags & XFS_QMOPT_DQALLOC,
 896                                                 flags & XFS_QMOPT_DQLOCK,
 897                                                 ip->i_udquot, &ip->i_gdquot) :
 898                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
 899                                                 flags & XFS_QMOPT_DQALLOC,
 900                                                 flags & XFS_QMOPT_DQLOCK,
 901                                                 ip->i_udquot, &ip->i_gdquot);
 902                 /*
 903                  * Don't worry about the udquot that we may have
 904                  * attached above. It'll get detached, if not already.
 905                  */
 906                 if (error)
 907                         goto done;
 908                 nquotas++;
 909         }
 910
 911         /*
 912          * Attach this group quota to the user quota as a hint.
 913          * This WON'T, in general, result in a thrash.
 914          */
 915         if (nquotas == 2) {
 916                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
 917                 ASSERT(ip->i_udquot);
 918                 ASSERT(ip->i_gdquot);
 919
 920                 /*
 921                  * We may or may not have the i_udquot locked at this point,
 922                  * but this check is OK since we don't depend on the i_gdquot to
 923                  * be accurate 100% all the time. It is just a hint, and this
 924                  * will succeed in general.
 925                  */
 926                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
 927                         goto done;
 928                 /*
 929                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
 930                  */
 931                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
 932                                          flags & XFS_QMOPT_DQLOCK);
 933         }
 934
 935       done:
 936
 937 #ifdef QUOTADEBUG
 938         if (! error) {
 939                 if (ip->i_udquot) {
 940                         if (flags & XFS_QMOPT_DQLOCK)
 941                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
 942                 }
 943                 if (ip->i_gdquot) {
 944                         if (flags & XFS_QMOPT_DQLOCK)
 945                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
 946                 }
 947                 if (XFS_IS_UQUOTA_ON(mp))
 948                         ASSERT(ip->i_udquot);
 949                 if (XFS_IS_OQUOTA_ON(mp))
 950                         ASSERT(ip->i_gdquot);
 951         }
 952 #endif
 953
 954         if (! (flags & XFS_QMOPT_ILOCKED))
 955                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 956
 957 #ifdef QUOTADEBUG
 958         else
 959                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
 960 #endif
 961         return error;
 962 }
 963
 964 /*
 965  * Release dquots (and their references) if any.
 966  * The inode should be locked EXCL except when this's called by
 967  * xfs_ireclaim.
 968  */
 969 void
 970 xfs_qm_dqdetach(
 971         xfs_inode_t     *ip)
 972 {
 973         if (!(ip->i_udquot || ip->i_gdquot))
 974                 return;
 975
 976         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
 977         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
 978         if (ip->i_udquot) {
 979                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
 980                 xfs_qm_dqrele(ip->i_udquot);
 981                 ip->i_udquot = NULL;
 982         }
 983         if (ip->i_gdquot) {
 984                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
 985                 xfs_qm_dqrele(ip->i_gdquot);
 986                 ip->i_gdquot = NULL;
 987         }
 988 }
 989
 990 /*
 991  * This is called by VFS_SYNC and flags arg determines the caller,
 992  * and its motives, as done in xfs_sync.
 993  *
 994  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
 995  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
 996  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
 997  */
 998
 999 int
1000 xfs_qm_sync(
1001         xfs_mount_t     *mp,
1002         int             flags)
1003 {
1004         int             recl, restarts;
1005         xfs_dquot_t     *dqp;
1006         uint            flush_flags;
1007         boolean_t       nowait;
1008         int             error;
1009
1010         if (! XFS_IS_QUOTA_ON(mp))
1011                 return 0;
1012
1013         restarts = 0;
1014         /*
1015          * We won't block unless we are asked to.
1016          */
1017         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1018
1019   again:
1020         xfs_qm_mplist_lock(mp);
1021         /*
1022          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1023          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1024          * when we have the mplist lock, we know that dquots will be consistent
1025          * as long as we have it locked.
1026          */
1027         if (! XFS_IS_QUOTA_ON(mp)) {
1028                 xfs_qm_mplist_unlock(mp);
1029                 return 0;
1030         }
1031         FOREACH_DQUOT_IN_MP(dqp, mp) {
1032                 /*
1033                  * If this is vfs_sync calling, then skip the dquots that
1034                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1035                  * This is very similar to what xfs_sync does with inodes.
1036                  */
1037                 if (flags & SYNC_BDFLUSH) {
1038                         if (! XFS_DQ_IS_DIRTY(dqp))
1039                                 continue;
1040                 }
1041
1042                 if (nowait) {
1043                         /*
1044                          * Try to acquire the dquot lock. We are NOT out of
1045                          * lock order, but we just don't want to wait for this
1046                          * lock, unless somebody wanted us to.
1047                          */
1048                         if (! xfs_qm_dqlock_nowait(dqp))
1049                                 continue;
1050                 } else {
1051                         xfs_dqlock(dqp);
1052                 }
1053
1054                 /*
1055                  * Now, find out for sure if this dquot is dirty or not.
1056                  */
1057                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1058                         xfs_dqunlock(dqp);
1059                         continue;
1060                 }
1061
1062                 /* XXX a sentinel would be better */
1063                 recl = XFS_QI_MPLRECLAIMS(mp);
1064                 if (! xfs_qm_dqflock_nowait(dqp)) {
1065                         if (nowait) {
1066                                 xfs_dqunlock(dqp);
1067                                 continue;
1068                         }
1069                         /*
1070                          * If we can't grab the flush lock then if the caller
1071                          * really wanted us to give this our best shot, so
1072                          * see if we can give a push to the buffer before we wait
1073                          * on the flush lock. At this point, we know that
1074                          * even though the dquot is being flushed,
1075                          * it has (new) dirty data.
1076                          */
1077                         xfs_qm_dqflock_pushbuf_wait(dqp);
1078                 }
1079                 /*
1080                  * Let go of the mplist lock. We don't want to hold it
1081                  * across a disk write
1082                  */
1083                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1084                 xfs_qm_mplist_unlock(mp);
1085                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1086                 error = xfs_qm_dqflush(dqp, flush_flags);
1087                 xfs_dqunlock(dqp);
1088                 if (error && XFS_FORCED_SHUTDOWN(mp))
1089                         return 0;       /* Need to prevent umount failure */
1090                 else if (error)
1091                         return error;
1092
1093                 xfs_qm_mplist_lock(mp);
1094                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1095                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1096                                 break;
1097
1098                         xfs_qm_mplist_unlock(mp);
1099                         goto again;
1100                 }
1101         }
1102
1103         xfs_qm_mplist_unlock(mp);
1104         return 0;
1105 }
1106
1107
1108 /*
1109  * This initializes all the quota information that's kept in the
1110  * mount structure
1111  */
1112 STATIC int
1113 xfs_qm_init_quotainfo(
1114         xfs_mount_t     *mp)
1115 {
1116         xfs_quotainfo_t *qinf;
1117         int             error;
1118         xfs_dquot_t     *dqp;
1119
1120         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1121
1122         /*
1123          * Tell XQM that we exist as soon as possible.
1124          */
1125         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1126                 return error;
1127         }
1128
1129         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1130
1131         /*
1132          * See if quotainodes are setup, and if not, allocate them,
1133          * and change the superblock accordingly.
1134          */
1135         if ((error = xfs_qm_init_quotainos(mp))) {
1136                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1137                 mp->m_quotainfo = NULL;
1138                 return error;
1139         }
1140
1141         spin_lock_init(&qinf->qi_pinlock);
1142         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1143         qinf->qi_dqreclaims = 0;
1144
1145         /* mutex used to serialize quotaoffs */
1146         mutex_init(&qinf->qi_quotaofflock);
1147
1148         /* Precalc some constants */
1149         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1150         ASSERT(qinf->qi_dqchunklen);
1151         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1152         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1153
1154         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1155
1156         /*
1157          * We try to get the limits from the superuser's limits fields.
1158          * This is quite hacky, but it is standard quota practice.
1159          * We look at the USR dquot with id == 0 first, but if user quotas
1160          * are not enabled we goto the GRP dquot with id == 0.
1161          * We don't really care to keep separate default limits for user
1162          * and group quotas, at least not at this point.
1163          */
1164         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1165                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER :
1166                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1167                                 XFS_DQ_PROJ),
1168                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1169                              &dqp);
1170         if (! error) {
1171                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1172
1173                 /*
1174                  * The warnings and timers set the grace period given to
1175                  * a user or group before he or she can not perform any
1176                  * more writing. If it is zero, a default is used.
1177                  */
1178                 qinf->qi_btimelimit = ddqp->d_btimer ?
1179                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1180                 qinf->qi_itimelimit = ddqp->d_itimer ?
1181                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1182                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1183                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1184                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1185                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1186                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1187                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1188                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1189                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1190                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1191                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1192                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1193                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1194                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1195                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1196
1197                 /*
1198                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1199                  * we don't want this dquot cached. We haven't done a
1200                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1201                  */
1202                 xfs_qm_dqdestroy(dqp);
1203         } else {
1204                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1205                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1206                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1207                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1208                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1209                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1210         }
1211
1212         return 0;
1213 }
1214
1215
1216 /*
1217  * Gets called when unmounting a filesystem or when all quotas get
1218  * turned off.
1219  * This purges the quota inodes, destroys locks and frees itself.
1220  */
1221 void
1222 xfs_qm_destroy_quotainfo(
1223         xfs_mount_t     *mp)
1224 {
1225         xfs_quotainfo_t *qi;
1226
1227         qi = mp->m_quotainfo;
1228         ASSERT(qi != NULL);
1229         ASSERT(xfs_Gqm != NULL);
1230
1231         /*
1232          * Release the reference that XQM kept, so that we know
1233          * when the XQM structure should be freed. We cannot assume
1234          * that xfs_Gqm is non-null after this point.
1235          */
1236         xfs_qm_rele_quotafs_ref(mp);
1237
1238         spinlock_destroy(&qi->qi_pinlock);
1239         xfs_qm_list_destroy(&qi->qi_dqlist);
1240
1241         if (qi->qi_uquotaip) {
1242                 XFS_PURGE_INODE(qi->qi_uquotaip);
1243                 qi->qi_uquotaip = NULL; /* paranoia */
1244         }
1245         if (qi->qi_gquotaip) {
1246                 XFS_PURGE_INODE(qi->qi_gquotaip);
1247                 qi->qi_gquotaip = NULL;
1248         }
1249         mutex_destroy(&qi->qi_quotaofflock);
1250         kmem_free(qi, sizeof(xfs_quotainfo_t));
1251         mp->m_quotainfo = NULL;
1252 }
1253
1254
1255
1256 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1257
1258 /* ARGSUSED */
1259 STATIC void
1260 xfs_qm_list_init(
1261         xfs_dqlist_t    *list,
1262         char            *str,
1263         int             n)
1264 {
1265         mutex_init(&list->qh_lock);
1266         list->qh_next = NULL;
1267         list->qh_version = 0;
1268         list->qh_nelems = 0;
1269 }
1270
1271 STATIC void
1272 xfs_qm_list_destroy(
1273         xfs_dqlist_t    *list)
1274 {
1275         mutex_destroy(&(list->qh_lock));
1276 }
1277
1278
1279 /*
1280  * Stripped down version of dqattach. This doesn't attach, or even look at the
1281  * dquots attached to the inode. The rationale is that there won't be any
1282  * attached at the time this is called from quotacheck.
1283  */
1284 STATIC int
1285 xfs_qm_dqget_noattach(
1286         xfs_inode_t     *ip,
1287         xfs_dquot_t     **O_udqpp,
1288         xfs_dquot_t     **O_gdqpp)
1289 {
1290         int             error;
1291         xfs_mount_t     *mp;
1292         xfs_dquot_t     *udqp, *gdqp;
1293
1294         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1295         mp = ip->i_mount;
1296         udqp = NULL;
1297         gdqp = NULL;
1298
1299         if (XFS_IS_UQUOTA_ON(mp)) {
1300                 ASSERT(ip->i_udquot == NULL);
1301                 /*
1302                  * We want the dquot allocated if it doesn't exist.
1303                  */
1304                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1305                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1306                                          &udqp))) {
1307                         /*
1308                          * Shouldn't be able to turn off quotas here.
1309                          */
1310                         ASSERT(error != ESRCH);
1311                         ASSERT(error != ENOENT);
1312                         return error;
1313                 }
1314                 ASSERT(udqp);
1315         }
1316
1317         if (XFS_IS_OQUOTA_ON(mp)) {
1318                 ASSERT(ip->i_gdquot == NULL);
1319                 if (udqp)
1320                         xfs_dqunlock(udqp);
1321                 error = XFS_IS_GQUOTA_ON(mp) ?
1322                                 xfs_qm_dqget(mp, ip,
1323                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1324                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1325                                              &gdqp) :
1326                                 xfs_qm_dqget(mp, ip,
1327                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1328                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1329                                              &gdqp);
1330                 if (error) {
1331                         if (udqp)
1332                                 xfs_qm_dqrele(udqp);
1333                         ASSERT(error != ESRCH);
1334                         ASSERT(error != ENOENT);
1335                         return error;
1336                 }
1337                 ASSERT(gdqp);
1338
1339                 /* Reacquire the locks in the right order */
1340                 if (udqp) {
1341                         if (! xfs_qm_dqlock_nowait(udqp)) {
1342                                 xfs_dqunlock(gdqp);
1343                                 xfs_dqlock(udqp);
1344                                 xfs_dqlock(gdqp);
1345                         }
1346                 }
1347         }
1348
1349         *O_udqpp = udqp;
1350         *O_gdqpp = gdqp;
1351
1352 #ifdef QUOTADEBUG
1353         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1354         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1355 #endif
1356         return 0;
1357 }
1358
1359 /*
1360  * Create an inode and return with a reference already taken, but unlocked
1361  * This is how we create quota inodes
1362  */
1363 STATIC int
1364 xfs_qm_qino_alloc(
1365         xfs_mount_t     *mp,
1366         xfs_inode_t     **ip,
1367         __int64_t       sbfields,
1368         uint            flags)
1369 {
1370         xfs_trans_t     *tp;
1371         int             error;
1372         int             committed;
1373
1374         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1375         if ((error = xfs_trans_reserve(tp,
1376                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1377                                       XFS_CREATE_LOG_RES(mp), 0,
1378                                       XFS_TRANS_PERM_LOG_RES,
1379                                       XFS_CREATE_LOG_COUNT))) {
1380                 xfs_trans_cancel(tp, 0);
1381                 return error;
1382         }
1383
1384         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1385                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1386                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1387                                  XFS_TRANS_ABORT);
1388                 return error;
1389         }
1390
1391         /*
1392          * Keep an extra reference to this quota inode. This inode is
1393          * locked exclusively and joined to the transaction already.
1394          */
1395         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1396         VN_HOLD(XFS_ITOV((*ip)));
1397
1398         /*
1399          * Make the changes in the superblock, and log those too.
1400          * sbfields arg may contain fields other than *QUOTINO;
1401          * VERSIONNUM for example.
1402          */
1403         spin_lock(&mp->m_sb_lock);
1404         if (flags & XFS_QMOPT_SBVERSION) {
1405 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1406                 unsigned oldv = mp->m_sb.sb_versionnum;
1407 #endif
1408                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1409                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1410                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1411                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1412                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1413
1414                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1415                 mp->m_sb.sb_uquotino = NULLFSINO;
1416                 mp->m_sb.sb_gquotino = NULLFSINO;
1417
1418                 /* qflags will get updated _after_ quotacheck */
1419                 mp->m_sb.sb_qflags = 0;
1420 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1421                 cmn_err(CE_NOTE,
1422                         "Old superblock version %x, converting to %x.",
1423                         oldv, mp->m_sb.sb_versionnum);
1424 #endif
1425         }
1426         if (flags & XFS_QMOPT_UQUOTA)
1427                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1428         else
1429                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1430         spin_unlock(&mp->m_sb_lock);
1431         xfs_mod_sb(tp, sbfields);
1432
1433         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1434                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1435                 return error;
1436         }
1437         return 0;
1438 }
1439
1440
1441 STATIC int
1442 xfs_qm_reset_dqcounts(
1443         xfs_mount_t     *mp,
1444         xfs_buf_t       *bp,
1445         xfs_dqid_t      id,
1446         uint            type)
1447 {
1448         xfs_disk_dquot_t        *ddq;
1449         int                     j;
1450
1451         xfs_buftrace("RESET DQUOTS", bp);
1452         /*
1453          * Reset all counters and timers. They'll be
1454          * started afresh by xfs_qm_quotacheck.
1455          */
1456 #ifdef DEBUG
1457         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1458         do_div(j, sizeof(xfs_dqblk_t));
1459         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1460 #endif
1461         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1462         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1463                 /*
1464                  * Do a sanity check, and if needed, repair the dqblk. Don't
1465                  * output any warnings because it's perfectly possible to
1466                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1467                  */
1468                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1469                                       "xfs_quotacheck");
1470                 ddq->d_bcount = 0;
1471                 ddq->d_icount = 0;
1472                 ddq->d_rtbcount = 0;
1473                 ddq->d_btimer = 0;
1474                 ddq->d_itimer = 0;
1475                 ddq->d_rtbtimer = 0;
1476                 ddq->d_bwarns = 0;
1477                 ddq->d_iwarns = 0;
1478                 ddq->d_rtbwarns = 0;
1479                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1480         }
1481
1482         return 0;
1483 }
1484
1485 STATIC int
1486 xfs_qm_dqiter_bufs(
1487         xfs_mount_t     *mp,
1488         xfs_dqid_t      firstid,
1489         xfs_fsblock_t   bno,
1490         xfs_filblks_t   blkcnt,
1491         uint            flags)
1492 {
1493         xfs_buf_t       *bp;
1494         int             error;
1495         int             notcommitted;
1496         int             incr;
1497         int             type;
1498
1499         ASSERT(blkcnt > 0);
1500         notcommitted = 0;
1501         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1502                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1503         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1504                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1505         error = 0;
1506
1507         /*
1508          * Blkcnt arg can be a very big number, and might even be
1509          * larger than the log itself. So, we have to break it up into
1510          * manageable-sized transactions.
1511          * Note that we don't start a permanent transaction here; we might
1512          * not be able to get a log reservation for the whole thing up front,
1513          * and we don't really care to either, because we just discard
1514          * everything if we were to crash in the middle of this loop.
1515          */
1516         while (blkcnt--) {
1517                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1518                               XFS_FSB_TO_DADDR(mp, bno),
1519                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1520                 if (error)
1521                         break;
1522
1523                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1524                 xfs_bdwrite(mp, bp);
1525                 /*
1526                  * goto the next block.
1527                  */
1528                 bno++;
1529                 firstid += XFS_QM_DQPERBLK(mp);
1530         }
1531         return error;
1532 }
1533
1534 /*
1535  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1536  * caller supplied function for every chunk of dquots that we find.
1537  */
1538 STATIC int
1539 xfs_qm_dqiterate(
1540         xfs_mount_t     *mp,
1541         xfs_inode_t     *qip,
1542         uint            flags)
1543 {
1544         xfs_bmbt_irec_t         *map;
1545         int                     i, nmaps;       /* number of map entries */
1546         int                     error;          /* return value */
1547         xfs_fileoff_t           lblkno;
1548         xfs_filblks_t           maxlblkcnt;
1549         xfs_dqid_t              firstid;
1550         xfs_fsblock_t           rablkno;
1551         xfs_filblks_t           rablkcnt;
1552
1553         error = 0;
1554         /*
1555          * This looks racy, but we can't keep an inode lock across a
1556          * trans_reserve. But, this gets called during quotacheck, and that
1557          * happens only at mount time which is single threaded.
1558          */
1559         if (qip->i_d.di_nblocks == 0)
1560                 return 0;
1561
1562         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1563
1564         lblkno = 0;
1565         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1566         do {
1567                 nmaps = XFS_DQITER_MAP_SIZE;
1568                 /*
1569                  * We aren't changing the inode itself. Just changing
1570                  * some of its data. No new blocks are added here, and
1571                  * the inode is never added to the transaction.
1572                  */
1573                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1574                 error = xfs_bmapi(NULL, qip, lblkno,
1575                                   maxlblkcnt - lblkno,
1576                                   XFS_BMAPI_METADATA,
1577                                   NULL,
1578                                   0, map, &nmaps, NULL, NULL);
1579                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1580                 if (error)
1581                         break;
1582
1583                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1584                 for (i = 0; i < nmaps; i++) {
1585                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1586                         ASSERT(map[i].br_blockcount);
1587
1588
1589                         lblkno += map[i].br_blockcount;
1590
1591                         if (map[i].br_startblock == HOLESTARTBLOCK)
1592                                 continue;
1593
1594                         firstid = (xfs_dqid_t) map[i].br_startoff *
1595                                 XFS_QM_DQPERBLK(mp);
1596                         /*
1597                          * Do a read-ahead on the next extent.
1598                          */
1599                         if ((i+1 < nmaps) &&
1600                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1601                                 rablkcnt =  map[i+1].br_blockcount;
1602                                 rablkno = map[i+1].br_startblock;
1603                                 while (rablkcnt--) {
1604                                         xfs_baread(mp->m_ddev_targp,
1605                                                XFS_FSB_TO_DADDR(mp, rablkno),
1606                                                (int)XFS_QI_DQCHUNKLEN(mp));
1607                                         rablkno++;
1608                                 }
1609                         }
1610                         /*
1611                          * Iterate thru all the blks in the extent and
1612                          * reset the counters of all the dquots inside them.
1613                          */
1614                         if ((error = xfs_qm_dqiter_bufs(mp,
1615                                                        firstid,
1616                                                        map[i].br_startblock,
1617                                                        map[i].br_blockcount,
1618                                                        flags))) {
1619                                 break;
1620                         }
1621                 }
1622
1623                 if (error)
1624                         break;
1625         } while (nmaps > 0);
1626
1627         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1628
1629         return error;
1630 }
1631
1632 /*
1633  * Called by dqusage_adjust in doing a quotacheck.
1634  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1635  * this updates its incore copy as well as the buffer copy. This is
1636  * so that once the quotacheck is done, we can just log all the buffers,
1637  * as opposed to logging numerous updates to individual dquots.
1638  */
1639 STATIC void
1640 xfs_qm_quotacheck_dqadjust(
1641         xfs_dquot_t             *dqp,
1642         xfs_qcnt_t              nblks,
1643         xfs_qcnt_t              rtblks)
1644 {
1645         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1646         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1647         /*
1648          * Adjust the inode count and the block count to reflect this inode's
1649          * resource usage.
1650          */
1651         be64_add(&dqp->q_core.d_icount, 1);
1652         dqp->q_res_icount++;
1653         if (nblks) {
1654                 be64_add(&dqp->q_core.d_bcount, nblks);
1655                 dqp->q_res_bcount += nblks;
1656         }
1657         if (rtblks) {
1658                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1659                 dqp->q_res_rtbcount += rtblks;
1660         }
1661
1662         /*
1663          * Set default limits, adjust timers (since we changed usages)
1664          */
1665         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1666                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1667                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1668         }
1669
1670         dqp->dq_flags |= XFS_DQ_DIRTY;
1671 }
1672
1673 STATIC int
1674 xfs_qm_get_rtblks(
1675         xfs_inode_t     *ip,
1676         xfs_qcnt_t      *O_rtblks)
1677 {
1678         xfs_filblks_t   rtblks;                 /* total rt blks */
1679         xfs_extnum_t    idx;                    /* extent record index */
1680         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1681         xfs_extnum_t    nextents;               /* number of extent entries */
1682         int             error;
1683
1684         ASSERT(XFS_IS_REALTIME_INODE(ip));
1685         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1686         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1687                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1688                         return error;
1689         }
1690         rtblks = 0;
1691         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1692         for (idx = 0; idx < nextents; idx++)
1693                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1694         *O_rtblks = (xfs_qcnt_t)rtblks;
1695         return 0;
1696 }
1697
1698 /*
1699  * callback routine supplied to bulkstat(). Given an inumber, find its
1700  * dquots and update them to account for resources taken by that inode.
1701  */
1702 /* ARGSUSED */
1703 STATIC int
1704 xfs_qm_dqusage_adjust(
1705         xfs_mount_t     *mp,            /* mount point for filesystem */
1706         xfs_ino_t       ino,            /* inode number to get data for */
1707         void            __user *buffer, /* not used */
1708         int             ubsize,         /* not used */
1709         void            *private_data,  /* not used */
1710         xfs_daddr_t     bno,            /* starting block of inode cluster */
1711         int             *ubused,        /* not used */
1712         void            *dip,           /* on-disk inode pointer (not used) */
1713         int             *res)           /* result code value */
1714 {
1715         xfs_inode_t     *ip;
1716         xfs_dquot_t     *udqp, *gdqp;
1717         xfs_qcnt_t      nblks, rtblks;
1718         int             error;
1719
1720         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1721
1722         /*
1723          * rootino must have its resources accounted for, not so with the quota
1724          * inodes.
1725          */
1726         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1727                 *res = BULKSTAT_RV_NOTHING;
1728                 return XFS_ERROR(EINVAL);
1729         }
1730
1731         /*
1732          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1733          * interface expects the inode to be exclusively locked because that's
1734          * the case in all other instances. It's OK that we do this because
1735          * quotacheck is done only at mount time.
1736          */
1737         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1738                 *res = BULKSTAT_RV_NOTHING;
1739                 return error;
1740         }
1741
1742         if (ip->i_d.di_mode == 0) {
1743                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1744                 *res = BULKSTAT_RV_NOTHING;
1745                 return XFS_ERROR(ENOENT);
1746         }
1747
1748         /*
1749          * Obtain the locked dquots. In case of an error (eg. allocation
1750          * fails for ENOSPC), we return the negative of the error number
1751          * to bulkstat, so that it can get propagated to quotacheck() and
1752          * making us disable quotas for the file system.
1753          */
1754         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1755                 xfs_iput(ip, XFS_ILOCK_EXCL);
1756                 *res = BULKSTAT_RV_GIVEUP;
1757                 return error;
1758         }
1759
1760         rtblks = 0;
1761         if (! XFS_IS_REALTIME_INODE(ip)) {
1762                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1763         } else {
1764                 /*
1765                  * Walk thru the extent list and count the realtime blocks.
1766                  */
1767                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1768                         xfs_iput(ip, XFS_ILOCK_EXCL);
1769                         if (udqp)
1770                                 xfs_qm_dqput(udqp);
1771                         if (gdqp)
1772                                 xfs_qm_dqput(gdqp);
1773                         *res = BULKSTAT_RV_GIVEUP;
1774                         return error;
1775                 }
1776                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1777         }
1778         ASSERT(ip->i_delayed_blks == 0);
1779
1780         /*
1781          * We can't release the inode while holding its dquot locks.
1782          * The inode can go into inactive and might try to acquire the dquotlocks.
1783          * So, just unlock here and do a vn_rele at the end.
1784          */
1785         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1786
1787         /*
1788          * Add the (disk blocks and inode) resources occupied by this
1789          * inode to its dquots. We do this adjustment in the incore dquot,
1790          * and also copy the changes to its buffer.
1791          * We don't care about putting these changes in a transaction
1792          * envelope because if we crash in the middle of a 'quotacheck'
1793          * we have to start from the beginning anyway.
1794          * Once we're done, we'll log all the dquot bufs.
1795          *
1796          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1797          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1798          */
1799         if (XFS_IS_UQUOTA_ON(mp)) {
1800                 ASSERT(udqp);
1801                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1802                 xfs_qm_dqput(udqp);
1803         }
1804         if (XFS_IS_OQUOTA_ON(mp)) {
1805                 ASSERT(gdqp);
1806                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1807                 xfs_qm_dqput(gdqp);
1808         }
1809         /*
1810          * Now release the inode. This will send it to 'inactive', and
1811          * possibly even free blocks.
1812          */
1813         VN_RELE(XFS_ITOV(ip));
1814
1815         /*
1816          * Goto next inode.
1817          */
1818         *res = BULKSTAT_RV_DIDONE;
1819         return 0;
1820 }
1821
1822 /*
1823  * Walk thru all the filesystem inodes and construct a consistent view
1824  * of the disk quota world. If the quotacheck fails, disable quotas.
1825  */
1826 int
1827 xfs_qm_quotacheck(
1828         xfs_mount_t     *mp)
1829 {
1830         int             done, count, error;
1831         xfs_ino_t       lastino;
1832         size_t          structsz;
1833         xfs_inode_t     *uip, *gip;
1834         uint            flags;
1835
1836         count = INT_MAX;
1837         structsz = 1;
1838         lastino = 0;
1839         flags = 0;
1840
1841         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1842         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1843
1844         /*
1845          * There should be no cached dquots. The (simplistic) quotacheck
1846          * algorithm doesn't like that.
1847          */
1848         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1849
1850         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1851
1852         /*
1853          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1854          * their counters to zero. We need a clean slate.
1855          * We don't log our changes till later.
1856          */
1857         if ((uip = XFS_QI_UQIP(mp))) {
1858                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1859                         goto error_return;
1860                 flags |= XFS_UQUOTA_CHKD;
1861         }
1862
1863         if ((gip = XFS_QI_GQIP(mp))) {
1864                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1865                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1866                         goto error_return;
1867                 flags |= XFS_OQUOTA_CHKD;
1868         }
1869
1870         do {
1871                 /*
1872                  * Iterate thru all the inodes in the file system,
1873                  * adjusting the corresponding dquot counters in core.
1874                  */
1875                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1876                                      xfs_qm_dqusage_adjust, NULL,
1877                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1878                         break;
1879
1880         } while (! done);
1881
1882         /*
1883          * We can get this error if we couldn't do a dquot allocation inside
1884          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1885          * dirty dquots that might be cached, we just want to get rid of them
1886          * and turn quotaoff. The dquots won't be attached to any of the inodes
1887          * at this point (because we intentionally didn't in dqget_noattach).
1888          */
1889         if (error) {
1890                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1891                 goto error_return;
1892         }
1893         /*
1894          * We've made all the changes that we need to make incore.
1895          * Now flush_them down to disk buffers.
1896          */
1897         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1898
1899         /*
1900          * We didn't log anything, because if we crashed, we'll have to
1901          * start the quotacheck from scratch anyway. However, we must make
1902          * sure that our dquot changes are secure before we put the
1903          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1904          * flush.
1905          */
1906         XFS_bflush(mp->m_ddev_targp);
1907
1908         /*
1909          * If one type of quotas is off, then it will lose its
1910          * quotachecked status, since we won't be doing accounting for
1911          * that type anymore.
1912          */
1913         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1914         mp->m_qflags |= flags;
1915
1916         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1917
1918  error_return:
1919         if (error) {
1920                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1921                         "Disabling quotas.",
1922                         mp->m_fsname, error);
1923                 /*
1924                  * We must turn off quotas.
1925                  */
1926                 ASSERT(mp->m_quotainfo != NULL);
1927                 ASSERT(xfs_Gqm != NULL);
1928                 xfs_qm_destroy_quotainfo(mp);
1929                 (void)xfs_mount_reset_sbqflags(mp);
1930         } else {
1931                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1932         }
1933         return (error);
1934 }
1935
1936 /*
1937  * This is called after the superblock has been read in and we're ready to
1938  * iget the quota inodes.
1939  */
1940 STATIC int
1941 xfs_qm_init_quotainos(
1942         xfs_mount_t     *mp)
1943 {
1944         xfs_inode_t     *uip, *gip;
1945         int             error;
1946         __int64_t       sbflags;
1947         uint            flags;
1948
1949         ASSERT(mp->m_quotainfo);
1950         uip = gip = NULL;
1951         sbflags = 0;
1952         flags = 0;
1953
1954         /*
1955          * Get the uquota and gquota inodes
1956          */
1957         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1958                 if (XFS_IS_UQUOTA_ON(mp) &&
1959                     mp->m_sb.sb_uquotino != NULLFSINO) {
1960                         ASSERT(mp->m_sb.sb_uquotino > 0);
1961                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1962                                              0, 0, &uip, 0)))
1963                                 return XFS_ERROR(error);
1964                 }
1965                 if (XFS_IS_OQUOTA_ON(mp) &&
1966                     mp->m_sb.sb_gquotino != NULLFSINO) {
1967                         ASSERT(mp->m_sb.sb_gquotino > 0);
1968                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1969                                              0, 0, &gip, 0))) {
1970                                 if (uip)
1971                                         VN_RELE(XFS_ITOV(uip));
1972                                 return XFS_ERROR(error);
1973                         }
1974                 }
1975         } else {
1976                 flags |= XFS_QMOPT_SBVERSION;
1977                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1978                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1979         }
1980
1981         /*
1982          * Create the two inodes, if they don't exist already. The changes
1983          * made above will get added to a transaction and logged in one of
1984          * the qino_alloc calls below.  If the device is readonly,
1985          * temporarily switch to read-write to do this.
1986          */
1987         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1988                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1989                                               sbflags | XFS_SB_UQUOTINO,
1990                                               flags | XFS_QMOPT_UQUOTA)))
1991                         return XFS_ERROR(error);
1992
1993                 flags &= ~XFS_QMOPT_SBVERSION;
1994         }
1995         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1996                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1997                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1998                 error = xfs_qm_qino_alloc(mp, &gip,
1999                                           sbflags | XFS_SB_GQUOTINO, flags);
2000                 if (error) {
2001                         if (uip)
2002                                 VN_RELE(XFS_ITOV(uip));
2003
2004                         return XFS_ERROR(error);
2005                 }
2006         }
2007
2008         XFS_QI_UQIP(mp) = uip;
2009         XFS_QI_GQIP(mp) = gip;
2010
2011         return 0;
2012 }
2013
2014
2015 /*
2016  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2017  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2018  * favor the lookup function ...
2019  * XXXsup merge this with qm_reclaim_one().
2020  */
2021 STATIC int
2022 xfs_qm_shake_freelist(
2023         int howmany)
2024 {
2025         int             nreclaimed;
2026         xfs_dqhash_t    *hash;
2027         xfs_dquot_t     *dqp, *nextdqp;
2028         int             restarts;
2029         int             nflushes;
2030
2031         if (howmany <= 0)
2032                 return 0;
2033
2034         nreclaimed = 0;
2035         restarts = 0;
2036         nflushes = 0;
2037
2038 #ifdef QUOTADEBUG
2039         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2040 #endif
2041         /* lock order is : hashchainlock, freelistlock, mplistlock */
2042  tryagain:
2043         xfs_qm_freelist_lock(xfs_Gqm);
2044
2045         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2046              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2047               nreclaimed < howmany); ) {
2048                 xfs_dqlock(dqp);
2049
2050                 /*
2051                  * We are racing with dqlookup here. Naturally we don't
2052                  * want to reclaim a dquot that lookup wants.
2053                  */
2054                 if (dqp->dq_flags & XFS_DQ_WANT) {
2055                         xfs_dqunlock(dqp);
2056                         xfs_qm_freelist_unlock(xfs_Gqm);
2057                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2058                                 return nreclaimed;
2059                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2060                         goto tryagain;
2061                 }
2062
2063                 /*
2064                  * If the dquot is inactive, we are assured that it is
2065                  * not on the mplist or the hashlist, and that makes our
2066                  * life easier.
2067                  */
2068                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2069                         ASSERT(dqp->q_mount == NULL);
2070                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2071                         ASSERT(dqp->HL_PREVP == NULL);
2072                         ASSERT(dqp->MPL_PREVP == NULL);
2073                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2074                         nextdqp = dqp->dq_flnext;
2075                         goto off_freelist;
2076                 }
2077
2078                 ASSERT(dqp->MPL_PREVP);
2079                 /*
2080                  * Try to grab the flush lock. If this dquot is in the process of
2081                  * getting flushed to disk, we don't want to reclaim it.
2082                  */
2083                 if (! xfs_qm_dqflock_nowait(dqp)) {
2084                         xfs_dqunlock(dqp);
2085                         dqp = dqp->dq_flnext;
2086                         continue;
2087                 }
2088
2089                 /*
2090                  * We have the flush lock so we know that this is not in the
2091                  * process of being flushed. So, if this is dirty, flush it
2092                  * DELWRI so that we don't get a freelist infested with
2093                  * dirty dquots.
2094                  */
2095                 if (XFS_DQ_IS_DIRTY(dqp)) {
2096                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2097                         /*
2098                          * We flush it delayed write, so don't bother
2099                          * releasing the mplock.
2100                          */
2101                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2102                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2103                         dqp = dqp->dq_flnext;
2104                         continue;
2105                 }
2106                 /*
2107                  * We're trying to get the hashlock out of order. This races
2108                  * with dqlookup; so, we giveup and goto the next dquot if
2109                  * we couldn't get the hashlock. This way, we won't starve
2110                  * a dqlookup process that holds the hashlock that is
2111                  * waiting for the freelist lock.
2112                  */
2113                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2114                         xfs_dqfunlock(dqp);
2115                         xfs_dqunlock(dqp);
2116                         dqp = dqp->dq_flnext;
2117                         continue;
2118                 }
2119                 /*
2120                  * This races with dquot allocation code as well as dqflush_all
2121                  * and reclaim code. So, if we failed to grab the mplist lock,
2122                  * giveup everything and start over.
2123                  */
2124                 hash = dqp->q_hash;
2125                 ASSERT(hash);
2126                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2127                         /* XXX put a sentinel so that we can come back here */
2128                         xfs_dqfunlock(dqp);
2129                         xfs_dqunlock(dqp);
2130                         XFS_DQ_HASH_UNLOCK(hash);
2131                         xfs_qm_freelist_unlock(xfs_Gqm);
2132                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2133                                 return nreclaimed;
2134                         goto tryagain;
2135                 }
2136                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2137 #ifdef QUOTADEBUG
2138                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2139                         dqp, be32_to_cpu(dqp->q_core.d_id));
2140 #endif
2141                 ASSERT(dqp->q_nrefs == 0);
2142                 nextdqp = dqp->dq_flnext;
2143                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2144                 XQM_HASHLIST_REMOVE(hash, dqp);
2145                 xfs_dqfunlock(dqp);
2146                 xfs_qm_mplist_unlock(dqp->q_mount);
2147                 XFS_DQ_HASH_UNLOCK(hash);
2148
2149  off_freelist:
2150                 XQM_FREELIST_REMOVE(dqp);
2151                 xfs_dqunlock(dqp);
2152                 nreclaimed++;
2153                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2154                 xfs_qm_dqdestroy(dqp);
2155                 dqp = nextdqp;
2156         }
2157         xfs_qm_freelist_unlock(xfs_Gqm);
2158         return nreclaimed;
2159 }
2160
2161
2162 /*
2163  * The kmem_shake interface is invoked when memory is running low.
2164  */
2165 /* ARGSUSED */
2166 STATIC int
2167 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2168 {
2169         int     ndqused, nfree, n;
2170
2171         if (!kmem_shake_allow(gfp_mask))
2172                 return 0;
2173         if (!xfs_Gqm)
2174                 return 0;
2175
2176         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2177         /* incore dquots in all f/s's */
2178         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2179
2180         ASSERT(ndqused >= 0);
2181
2182         if (nfree <= ndqused && nfree < ndquot)
2183                 return 0;
2184
2185         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2186         n = nfree - ndqused - ndquot;           /* # over target */
2187
2188         return xfs_qm_shake_freelist(MAX(nfree, n));
2189 }
2190
2191
2192 /*
2193  * Just pop the least recently used dquot off the freelist and
2194  * recycle it. The returned dquot is locked.
2195  */
2196 STATIC xfs_dquot_t *
2197 xfs_qm_dqreclaim_one(void)
2198 {
2199         xfs_dquot_t     *dqpout;
2200         xfs_dquot_t     *dqp;
2201         int             restarts;
2202         int             nflushes;
2203
2204         restarts = 0;
2205         dqpout = NULL;
2206         nflushes = 0;
2207
2208         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2209  startagain:
2210         xfs_qm_freelist_lock(xfs_Gqm);
2211
2212         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2213                 xfs_dqlock(dqp);
2214
2215                 /*
2216                  * We are racing with dqlookup here. Naturally we don't
2217                  * want to reclaim a dquot that lookup wants. We release the
2218                  * freelist lock and start over, so that lookup will grab
2219                  * both the dquot and the freelistlock.
2220                  */
2221                 if (dqp->dq_flags & XFS_DQ_WANT) {
2222                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2223                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2224                         xfs_dqunlock(dqp);
2225                         xfs_qm_freelist_unlock(xfs_Gqm);
2226                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2227                                 return NULL;
2228                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2229                         goto startagain;
2230                 }
2231
2232                 /*
2233                  * If the dquot is inactive, we are assured that it is
2234                  * not on the mplist or the hashlist, and that makes our
2235                  * life easier.
2236                  */
2237                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2238                         ASSERT(dqp->q_mount == NULL);
2239                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2240                         ASSERT(dqp->HL_PREVP == NULL);
2241                         ASSERT(dqp->MPL_PREVP == NULL);
2242                         XQM_FREELIST_REMOVE(dqp);
2243                         xfs_dqunlock(dqp);
2244                         dqpout = dqp;
2245                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2246                         break;
2247                 }
2248
2249                 ASSERT(dqp->q_hash);
2250                 ASSERT(dqp->MPL_PREVP);
2251
2252                 /*
2253                  * Try to grab the flush lock. If this dquot is in the process of
2254                  * getting flushed to disk, we don't want to reclaim it.
2255                  */
2256                 if (! xfs_qm_dqflock_nowait(dqp)) {
2257                         xfs_dqunlock(dqp);
2258                         continue;
2259                 }
2260
2261                 /*
2262                  * We have the flush lock so we know that this is not in the
2263                  * process of being flushed. So, if this is dirty, flush it
2264                  * DELWRI so that we don't get a freelist infested with
2265                  * dirty dquots.
2266                  */
2267                 if (XFS_DQ_IS_DIRTY(dqp)) {
2268                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2269                         /*
2270                          * We flush it delayed write, so don't bother
2271                          * releasing the freelist lock.
2272                          */
2273                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2274                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2275                         continue;
2276                 }
2277
2278                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2279                         xfs_dqfunlock(dqp);
2280                         xfs_dqunlock(dqp);
2281                         continue;
2282                 }
2283
2284                 if (! xfs_qm_dqhashlock_nowait(dqp))
2285                         goto mplistunlock;
2286
2287                 ASSERT(dqp->q_nrefs == 0);
2288                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2289                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2290                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2291                 XQM_FREELIST_REMOVE(dqp);
2292                 dqpout = dqp;
2293                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2294  mplistunlock:
2295                 xfs_qm_mplist_unlock(dqp->q_mount);
2296                 xfs_dqfunlock(dqp);
2297                 xfs_dqunlock(dqp);
2298                 if (dqpout)
2299                         break;
2300         }
2301
2302         xfs_qm_freelist_unlock(xfs_Gqm);
2303         return dqpout;
2304 }
2305
2306
2307 /*------------------------------------------------------------------*/
2308
2309 /*
2310  * Return a new incore dquot. Depending on the number of
2311  * dquots in the system, we either allocate a new one on the kernel heap,
2312  * or reclaim a free one.
2313  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2314  * to reclaim an existing one from the freelist.
2315  */
2316 boolean_t
2317 xfs_qm_dqalloc_incore(
2318         xfs_dquot_t **O_dqpp)
2319 {
2320         xfs_dquot_t     *dqp;
2321
2322         /*
2323          * Check against high water mark to see if we want to pop
2324          * a nincompoop dquot off the freelist.
2325          */
2326         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2327                 /*
2328                  * Try to recycle a dquot from the freelist.
2329                  */
2330                 if ((dqp = xfs_qm_dqreclaim_one())) {
2331                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2332                         /*
2333                          * Just zero the core here. The rest will get
2334                          * reinitialized by caller. XXX we shouldn't even
2335                          * do this zero ...
2336                          */
2337                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2338                         *O_dqpp = dqp;
2339                         return B_FALSE;
2340                 }
2341                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2342         }
2343
2344         /*
2345          * Allocate a brand new dquot on the kernel heap and return it
2346          * to the caller to initialize.
2347          */
2348         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2349         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2350         atomic_inc(&xfs_Gqm->qm_totaldquots);
2351
2352         return B_TRUE;
2353 }
2354
2355
2356 /*
2357  * Start a transaction and write the incore superblock changes to
2358  * disk. flags parameter indicates which fields have changed.
2359  */
2360 int
2361 xfs_qm_write_sb_changes(
2362         xfs_mount_t     *mp,
2363         __int64_t       flags)
2364 {
2365         xfs_trans_t     *tp;
2366         int             error;
2367
2368 #ifdef QUOTADEBUG
2369         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2370 #endif
2371         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2372         if ((error = xfs_trans_reserve(tp, 0,
2373                                       mp->m_sb.sb_sectsize + 128, 0,
2374                                       0,
2375                                       XFS_DEFAULT_LOG_COUNT))) {
2376                 xfs_trans_cancel(tp, 0);
2377                 return error;
2378         }
2379
2380         xfs_mod_sb(tp, flags);
2381         (void) xfs_trans_commit(tp, 0);
2382
2383         return 0;
2384 }
2385
2386
2387 /* --------------- utility functions for vnodeops ---------------- */
2388
2389
2390 /*
2391  * Given an inode, a uid and gid (from cred_t) make sure that we have
2392  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2393  * quotas by creating this file.
2394  * This also attaches dquot(s) to the given inode after locking it,
2395  * and returns the dquots corresponding to the uid and/or gid.
2396  *
2397  * in   : inode (unlocked)
2398  * out  : udquot, gdquot with references taken and unlocked
2399  */
2400 int
2401 xfs_qm_vop_dqalloc(
2402         xfs_mount_t     *mp,
2403         xfs_inode_t     *ip,
2404         uid_t           uid,
2405         gid_t           gid,
2406         prid_t          prid,
2407         uint            flags,
2408         xfs_dquot_t     **O_udqpp,
2409         xfs_dquot_t     **O_gdqpp)
2410 {
2411         int             error;
2412         xfs_dquot_t     *uq, *gq;
2413         uint            lockflags;
2414
2415         if (!XFS_IS_QUOTA_ON(mp))
2416                 return 0;
2417
2418         lockflags = XFS_ILOCK_EXCL;
2419         xfs_ilock(ip, lockflags);
2420
2421         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2422                 gid = ip->i_d.di_gid;
2423
2424         /*
2425          * Attach the dquot(s) to this inode, doing a dquot allocation
2426          * if necessary. The dquot(s) will not be locked.
2427          */
2428         if (XFS_NOT_DQATTACHED(mp, ip)) {
2429                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2430                                             XFS_QMOPT_ILOCKED))) {
2431                         xfs_iunlock(ip, lockflags);
2432                         return error;
2433                 }
2434         }
2435
2436         uq = gq = NULL;
2437         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2438                 if (ip->i_d.di_uid != uid) {
2439                         /*
2440                          * What we need is the dquot that has this uid, and
2441                          * if we send the inode to dqget, the uid of the inode
2442                          * takes priority over what's sent in the uid argument.
2443                          * We must unlock inode here before calling dqget if
2444                          * we're not sending the inode, because otherwise
2445                          * we'll deadlock by doing trans_reserve while
2446                          * holding ilock.
2447                          */
2448                         xfs_iunlock(ip, lockflags);
2449                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2450                                                  XFS_DQ_USER,
2451                                                  XFS_QMOPT_DQALLOC |
2452                                                  XFS_QMOPT_DOWARN,
2453                                                  &uq))) {
2454                                 ASSERT(error != ENOENT);
2455                                 return error;
2456                         }
2457                         /*
2458                          * Get the ilock in the right order.
2459                          */
2460                         xfs_dqunlock(uq);
2461                         lockflags = XFS_ILOCK_SHARED;
2462                         xfs_ilock(ip, lockflags);
2463                 } else {
2464                         /*
2465                          * Take an extra reference, because we'll return
2466                          * this to caller
2467                          */
2468                         ASSERT(ip->i_udquot);
2469                         uq = ip->i_udquot;
2470                         xfs_dqlock(uq);
2471                         XFS_DQHOLD(uq);
2472                         xfs_dqunlock(uq);
2473                 }
2474         }
2475         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2476                 if (ip->i_d.di_gid != gid) {
2477                         xfs_iunlock(ip, lockflags);
2478                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2479                                                  XFS_DQ_GROUP,
2480                                                  XFS_QMOPT_DQALLOC |
2481                                                  XFS_QMOPT_DOWARN,
2482                                                  &gq))) {
2483                                 if (uq)
2484                                         xfs_qm_dqrele(uq);
2485                                 ASSERT(error != ENOENT);
2486                                 return error;
2487                         }
2488                         xfs_dqunlock(gq);
2489                         lockflags = XFS_ILOCK_SHARED;
2490                         xfs_ilock(ip, lockflags);
2491                 } else {
2492                         ASSERT(ip->i_gdquot);
2493                         gq = ip->i_gdquot;
2494                         xfs_dqlock(gq);
2495                         XFS_DQHOLD(gq);
2496                         xfs_dqunlock(gq);
2497                 }
2498         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2499                 if (ip->i_d.di_projid != prid) {
2500                         xfs_iunlock(ip, lockflags);
2501                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2502                                                  XFS_DQ_PROJ,
2503                                                  XFS_QMOPT_DQALLOC |
2504                                                  XFS_QMOPT_DOWARN,
2505                                                  &gq))) {
2506                                 if (uq)
2507                                         xfs_qm_dqrele(uq);
2508                                 ASSERT(error != ENOENT);
2509                                 return (error);
2510                         }
2511                         xfs_dqunlock(gq);
2512                         lockflags = XFS_ILOCK_SHARED;
2513                         xfs_ilock(ip, lockflags);
2514                 } else {
2515                         ASSERT(ip->i_gdquot);
2516                         gq = ip->i_gdquot;
2517                         xfs_dqlock(gq);
2518                         XFS_DQHOLD(gq);
2519                         xfs_dqunlock(gq);
2520                 }
2521         }
2522         if (uq)
2523                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2524
2525         xfs_iunlock(ip, lockflags);
2526         if (O_udqpp)
2527                 *O_udqpp = uq;
2528         else if (uq)
2529                 xfs_qm_dqrele(uq);
2530         if (O_gdqpp)
2531                 *O_gdqpp = gq;
2532         else if (gq)
2533                 xfs_qm_dqrele(gq);
2534         return 0;
2535 }
2536
2537 /*
2538  * Actually transfer ownership, and do dquot modifications.
2539  * These were already reserved.
2540  */
2541 xfs_dquot_t *
2542 xfs_qm_vop_chown(
2543         xfs_trans_t     *tp,
2544         xfs_inode_t     *ip,
2545         xfs_dquot_t     **IO_olddq,
2546         xfs_dquot_t     *newdq)
2547 {
2548         xfs_dquot_t     *prevdq;
2549         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2550                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2551
2552         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2553         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2554
2555         /* old dquot */
2556         prevdq = *IO_olddq;
2557         ASSERT(prevdq);
2558         ASSERT(prevdq != newdq);
2559
2560         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2561         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2562
2563         /* the sparkling new dquot */
2564         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2565         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2566
2567         /*
2568          * Take an extra reference, because the inode
2569          * is going to keep this dquot pointer even
2570          * after the trans_commit.
2571          */
2572         xfs_dqlock(newdq);
2573         XFS_DQHOLD(newdq);
2574         xfs_dqunlock(newdq);
2575         *IO_olddq = newdq;
2576
2577         return prevdq;
2578 }
2579
2580 /*
2581  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2582  */
2583 int
2584 xfs_qm_vop_chown_reserve(
2585         xfs_trans_t     *tp,
2586         xfs_inode_t     *ip,
2587         xfs_dquot_t     *udqp,
2588         xfs_dquot_t     *gdqp,
2589         uint            flags)
2590 {
2591         int             error;
2592         xfs_mount_t     *mp;
2593         uint            delblks, blkflags, prjflags = 0;
2594         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2595
2596         ASSERT(XFS_ISLOCKED_INODE(ip));
2597         mp = ip->i_mount;
2598         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2599
2600         delblks = ip->i_delayed_blks;
2601         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2602         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2603                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2604
2605         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2606             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2607                 delblksudq = udqp;
2608                 /*
2609                  * If there are delayed allocation blocks, then we have to
2610                  * unreserve those from the old dquot, and add them to the
2611                  * new dquot.
2612                  */
2613                 if (delblks) {
2614                         ASSERT(ip->i_udquot);
2615                         unresudq = ip->i_udquot;
2616                 }
2617         }
2618         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2619                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2620                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2621                         prjflags = XFS_QMOPT_ENOSPC;
2622
2623                 if (prjflags ||
2624                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2625                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2626                         delblksgdq = gdqp;
2627                         if (delblks) {
2628                                 ASSERT(ip->i_gdquot);
2629                                 unresgdq = ip->i_gdquot;
2630                         }
2631                 }
2632         }
2633
2634         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2635                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2636                                 flags | blkflags | prjflags)))
2637                 return (error);
2638
2639         /*
2640          * Do the delayed blks reservations/unreservations now. Since, these
2641          * are done without the help of a transaction, if a reservation fails
2642          * its previous reservations won't be automatically undone by trans
2643          * code. So, we have to do it manually here.
2644          */
2645         if (delblks) {
2646                 /*
2647                  * Do the reservations first. Unreservation can't fail.
2648                  */
2649                 ASSERT(delblksudq || delblksgdq);
2650                 ASSERT(unresudq || unresgdq);
2651                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2652                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2653                                 flags | blkflags | prjflags)))
2654                         return (error);
2655                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2656                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2657                                 blkflags);
2658         }
2659
2660         return (0);
2661 }
2662
2663 int
2664 xfs_qm_vop_rename_dqattach(
2665         xfs_inode_t     **i_tab)
2666 {
2667         xfs_inode_t     *ip;
2668         int             i;
2669         int             error;
2670
2671         ip = i_tab[0];
2672
2673         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2674                 return 0;
2675
2676         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2677                 error = xfs_qm_dqattach(ip, 0);
2678                 if (error)
2679                         return error;
2680         }
2681         for (i = 1; (i < 4 && i_tab[i]); i++) {
2682                 /*
2683                  * Watch out for duplicate entries in the table.
2684                  */
2685                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2686                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2687                                 error = xfs_qm_dqattach(ip, 0);
2688                                 if (error)
2689                                         return error;
2690                         }
2691                 }
2692         }
2693         return 0;
2694 }
2695
2696 void
2697 xfs_qm_vop_dqattach_and_dqmod_newinode(
2698         xfs_trans_t     *tp,
2699         xfs_inode_t     *ip,
2700         xfs_dquot_t     *udqp,
2701         xfs_dquot_t     *gdqp)
2702 {
2703         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2704                 return;
2705
2706         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2707         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2708
2709         if (udqp) {
2710                 xfs_dqlock(udqp);
2711                 XFS_DQHOLD(udqp);
2712                 xfs_dqunlock(udqp);
2713                 ASSERT(ip->i_udquot == NULL);
2714                 ip->i_udquot = udqp;
2715                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2716                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2717                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2718         }
2719         if (gdqp) {
2720                 xfs_dqlock(gdqp);
2721                 XFS_DQHOLD(gdqp);
2722                 xfs_dqunlock(gdqp);
2723                 ASSERT(ip->i_gdquot == NULL);
2724                 ip->i_gdquot = gdqp;
2725                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2726                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2727                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2728                                 be32_to_cpu(gdqp->q_core.d_id));
2729                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2730         }
2731 }
2732
2733 /* ------------- list stuff -----------------*/
2734 STATIC void
2735 xfs_qm_freelist_init(xfs_frlist_t *ql)
2736 {
2737         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2738         mutex_init(&ql->qh_lock);
2739         ql->qh_version = 0;
2740         ql->qh_nelems = 0;
2741 }
2742
2743 STATIC void
2744 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2745 {
2746         xfs_dquot_t     *dqp, *nextdqp;
2747
2748         mutex_lock(&ql->qh_lock);
2749         for (dqp = ql->qh_next;
2750              dqp != (xfs_dquot_t *)ql; ) {
2751                 xfs_dqlock(dqp);
2752                 nextdqp = dqp->dq_flnext;
2753 #ifdef QUOTADEBUG
2754                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2755 #endif
2756                 XQM_FREELIST_REMOVE(dqp);
2757                 xfs_dqunlock(dqp);
2758                 xfs_qm_dqdestroy(dqp);
2759                 dqp = nextdqp;
2760         }
2761         mutex_unlock(&ql->qh_lock);
2762         mutex_destroy(&ql->qh_lock);
2763
2764         ASSERT(ql->qh_nelems == 0);
2765 }
2766
2767 STATIC void
2768 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2769 {
2770         dq->dq_flnext = ql->qh_next;
2771         dq->dq_flprev = (xfs_dquot_t *)ql;
2772         ql->qh_next = dq;
2773         dq->dq_flnext->dq_flprev = dq;
2774         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2775         xfs_Gqm->qm_dqfreelist.qh_version++;
2776 }
2777
2778 void
2779 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2780 {
2781         xfs_dquot_t *next = dq->dq_flnext;
2782         xfs_dquot_t *prev = dq->dq_flprev;
2783
2784         next->dq_flprev = prev;
2785         prev->dq_flnext = next;
2786         dq->dq_flnext = dq->dq_flprev = dq;
2787         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2788         xfs_Gqm->qm_dqfreelist.qh_version++;
2789 }
2790
2791 void
2792 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2793 {
2794         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2795 }
2796
2797 STATIC int
2798 xfs_qm_dqhashlock_nowait(
2799         xfs_dquot_t *dqp)
2800 {
2801         int locked;
2802
2803         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2804         return locked;
2805 }
2806
2807 int
2808 xfs_qm_freelist_lock_nowait(
2809         xfs_qm_t *xqm)
2810 {
2811         int locked;
2812
2813         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2814         return locked;
2815 }
2816
2817 STATIC int
2818 xfs_qm_mplist_nowait(
2819         xfs_mount_t     *mp)
2820 {
2821         int locked;
2822
2823         ASSERT(mp->m_quotainfo);
2824         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2825         return locked;
2826 }