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