fs/xfs/quota/xfs_qm.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_clnt.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_alloc.h"
  29 #include "xfs_dmapi.h"
  30 #include "xfs_quota.h"
  31 #include "xfs_mount.h"
  32 #include "xfs_bmap_btree.h"
  33 #include "xfs_alloc_btree.h"
  34 #include "xfs_ialloc_btree.h"
  35 #include "xfs_dir2_sf.h"
  36 #include "xfs_attr_sf.h"
  37 #include "xfs_dinode.h"
  38 #include "xfs_inode.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_ialloc.h"
  41 #include "xfs_itable.h"
  42 #include "xfs_rtalloc.h"
  43 #include "xfs_error.h"
  44 #include "xfs_bmap.h"
  45 #include "xfs_rw.h"
  46 #include "xfs_acl.h"
  47 #include "xfs_attr.h"
  48 #include "xfs_buf_item.h"
  49 #include "xfs_trans_space.h"
  50 #include "xfs_utils.h"
  51 #include "xfs_qm.h"
  52
  53 /*
  54  * The global quota manager. There is only one of these for the entire
  55  * system, _not_ one per file system. XQM keeps track of the overall
  56  * quota functionality, including maintaining the freelist and hash
  57  * tables of dquots.
  58  */
  59 mutex_t         xfs_Gqm_lock;
  60 struct xfs_qm   *xfs_Gqm;
  61 uint            ndquot;
  62
  63 kmem_zone_t     *qm_dqzone;
  64 kmem_zone_t     *qm_dqtrxzone;
  65
  66 static cred_t   xfs_zerocr;
  67
  68 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
  69 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
  70
  71 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
  72 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
  73 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
  74 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
  75
  76 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
  77 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
  78 STATIC int      xfs_qm_shake(int, gfp_t);
  79
  80 static struct shrinker xfs_qm_shaker = {
  81         .shrink = xfs_qm_shake,
  82         .seeks = DEFAULT_SEEKS,
  83 };
  84
  85 #ifdef DEBUG
  86 extern mutex_t  qcheck_lock;
  87 #endif
  88
  89 #ifdef QUOTADEBUG
  90 #define XQM_LIST_PRINT(l, NXT, title) \
  91 { \
  92         xfs_dquot_t     *dqp; int i = 0; \
  93         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
  94         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
  95                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
  96                                   "bcnt = %d, icnt = %d, refs = %d", \
  97                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
  98                         DQFLAGTO_TYPESTR(dqp),       \
  99                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
 100                         (int) be64_to_cpu(dqp->q_core.d_icount), \
 101                         (int) dqp->q_nrefs);  } \
 102 }
 103 #else
 104 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
 105 #endif
 106
 107 /*
 108  * Initialize the XQM structure.
 109  * Note that there is not one quota manager per file system.
 110  */
 111 STATIC struct xfs_qm *
 112 xfs_Gqm_init(void)
 113 {
 114         xfs_dqhash_t    *udqhash, *gdqhash;
 115         xfs_qm_t        *xqm;
 116         size_t          hsize;
 117         uint            i;
 118
 119         /*
 120          * Initialize the dquot hash tables.
 121          */
 122         udqhash = kmem_zalloc_greedy(&hsize,
 123                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
 124                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t),
 125                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
 126         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
 127         hsize /= sizeof(xfs_dqhash_t);
 128         ndquot = hsize << 8;
 129
 130         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
 131         xqm->qm_dqhashmask = hsize - 1;
 132         xqm->qm_usr_dqhtable = udqhash;
 133         xqm->qm_grp_dqhtable = gdqhash;
 134         ASSERT(xqm->qm_usr_dqhtable != NULL);
 135         ASSERT(xqm->qm_grp_dqhtable != NULL);
 136
 137         for (i = 0; i < hsize; i++) {
 138                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
 139                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
 140         }
 141
 142         /*
 143          * Freelist of all dquots of all file systems
 144          */
 145         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
 146
 147         /*
 148          * dquot zone. we register our own low-memory callback.
 149          */
 150         if (!qm_dqzone) {
 151                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
 152                                                 "xfs_dquots");
 153                 qm_dqzone = xqm->qm_dqzone;
 154         } else
 155                 xqm->qm_dqzone = qm_dqzone;
 156
 157         register_shrinker(&xfs_qm_shaker);
 158
 159         /*
 160          * The t_dqinfo portion of transactions.
 161          */
 162         if (!qm_dqtrxzone) {
 163                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
 164                                                    "xfs_dqtrx");
 165                 qm_dqtrxzone = xqm->qm_dqtrxzone;
 166         } else
 167                 xqm->qm_dqtrxzone = qm_dqtrxzone;
 168
 169         atomic_set(&xqm->qm_totaldquots, 0);
 170         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
 171         xqm->qm_nrefs = 0;
 172 #ifdef DEBUG
 173         mutex_init(&qcheck_lock);
 174 #endif
 175         return xqm;
 176 }
 177
 178 /*
 179  * Destroy the global quota manager when its reference count goes to zero.
 180  */
 181 STATIC void
 182 xfs_qm_destroy(
 183         struct xfs_qm   *xqm)
 184 {
 185         int             hsize, i;
 186
 187         ASSERT(xqm != NULL);
 188         ASSERT(xqm->qm_nrefs == 0);
 189         unregister_shrinker(&xfs_qm_shaker);
 190         hsize = xqm->qm_dqhashmask + 1;
 191         for (i = 0; i < hsize; i++) {
 192                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
 193                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
 194         }
 195         kmem_free(xqm->qm_usr_dqhtable);
 196         kmem_free(xqm->qm_grp_dqhtable);
 197         xqm->qm_usr_dqhtable = NULL;
 198         xqm->qm_grp_dqhtable = NULL;
 199         xqm->qm_dqhashmask = 0;
 200         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
 201 #ifdef DEBUG
 202         mutex_destroy(&qcheck_lock);
 203 #endif
 204         kmem_free(xqm);
 205 }
 206
 207 /*
 208  * Called at mount time to let XQM know that another file system is
 209  * starting quotas. This isn't crucial information as the individual mount
 210  * structures are pretty independent, but it helps the XQM keep a
 211  * global view of what's going on.
 212  */
 213 /* ARGSUSED */
 214 STATIC int
 215 xfs_qm_hold_quotafs_ref(
 216         struct xfs_mount *mp)
 217 {
 218         /*
 219          * Need to lock the xfs_Gqm structure for things like this. For example,
 220          * the structure could disappear between the entry to this routine and
 221          * a HOLD operation if not locked.
 222          */
 223         XFS_QM_LOCK(xfs_Gqm);
 224
 225         if (xfs_Gqm == NULL)
 226                 xfs_Gqm = xfs_Gqm_init();
 227         /*
 228          * We can keep a list of all filesystems with quotas mounted for
 229          * debugging and statistical purposes, but ...
 230          * Just take a reference and get out.
 231          */
 232         XFS_QM_HOLD(xfs_Gqm);
 233         XFS_QM_UNLOCK(xfs_Gqm);
 234
 235         return 0;
 236 }
 237
 238
 239 /*
 240  * Release the reference that a filesystem took at mount time,
 241  * so that we know when we need to destroy the entire quota manager.
 242  */
 243 /* ARGSUSED */
 244 STATIC void
 245 xfs_qm_rele_quotafs_ref(
 246         struct xfs_mount *mp)
 247 {
 248         xfs_dquot_t     *dqp, *nextdqp;
 249
 250         ASSERT(xfs_Gqm);
 251         ASSERT(xfs_Gqm->qm_nrefs > 0);
 252
 253         /*
 254          * Go thru the freelist and destroy all inactive dquots.
 255          */
 256         xfs_qm_freelist_lock(xfs_Gqm);
 257
 258         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
 259              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
 260                 xfs_dqlock(dqp);
 261                 nextdqp = dqp->dq_flnext;
 262                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
 263                         ASSERT(dqp->q_mount == NULL);
 264                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
 265                         ASSERT(dqp->HL_PREVP == NULL);
 266                         ASSERT(dqp->MPL_PREVP == NULL);
 267                         XQM_FREELIST_REMOVE(dqp);
 268                         xfs_dqunlock(dqp);
 269                         xfs_qm_dqdestroy(dqp);
 270                 } else {
 271                         xfs_dqunlock(dqp);
 272                 }
 273                 dqp = nextdqp;
 274         }
 275         xfs_qm_freelist_unlock(xfs_Gqm);
 276
 277         /*
 278          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
 279          * be restarted.
 280          */
 281         XFS_QM_LOCK(xfs_Gqm);
 282         XFS_QM_RELE(xfs_Gqm);
 283         if (xfs_Gqm->qm_nrefs == 0) {
 284                 xfs_qm_destroy(xfs_Gqm);
 285                 xfs_Gqm = NULL;
 286         }
 287         XFS_QM_UNLOCK(xfs_Gqm);
 288 }
 289
 290 /*
 291  * Just destroy the quotainfo structure.
 292  */
 293 void
 294 xfs_qm_unmount_quotadestroy(
 295         xfs_mount_t     *mp)
 296 {
 297         if (mp->m_quotainfo)
 298                 xfs_qm_destroy_quotainfo(mp);
 299 }
 300
 301
 302 /*
 303  * This is called from xfs_mountfs to start quotas and initialize all
 304  * necessary data structures like quotainfo.  This is also responsible for
 305  * running a quotacheck as necessary.  We are guaranteed that the superblock
 306  * is consistently read in at this point.
 307  *
 308  * If we fail here, the mount will continue with quota turned off. We don't
 309  * need to inidicate success or failure at all.
 310  */
 311 void
 312 xfs_qm_mount_quotas(
 313         xfs_mount_t     *mp)
 314 {
 315         int             error = 0;
 316         uint            sbf;
 317
 318         /*
 319          * If quotas on realtime volumes is not supported, we disable
 320          * quotas immediately.
 321          */
 322         if (mp->m_sb.sb_rextents) {
 323                 cmn_err(CE_NOTE,
 324                         "Cannot turn on quotas for realtime filesystem %s",
 325                         mp->m_fsname);
 326                 mp->m_qflags = 0;
 327                 goto write_changes;
 328         }
 329
 330         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 331
 332         /*
 333          * Allocate the quotainfo structure inside the mount struct, and
 334          * create quotainode(s), and change/rev superblock if necessary.
 335          */
 336         error = xfs_qm_init_quotainfo(mp);
 337         if (error) {
 338                 /*
 339                  * We must turn off quotas.
 340                  */
 341                 ASSERT(mp->m_quotainfo == NULL);
 342                 mp->m_qflags = 0;
 343                 goto write_changes;
 344         }
 345         /*
 346          * If any of the quotas are not consistent, do a quotacheck.
 347          */
 348         if (XFS_QM_NEED_QUOTACHECK(mp)) {
 349                 error = xfs_qm_quotacheck(mp);
 350                 if (error) {
 351                         /* Quotacheck failed and disabled quotas. */
 352                         return;
 353                 }
 354         }
 355         /*
 356          * If one type of quotas is off, then it will lose its
 357          * quotachecked status, since we won't be doing accounting for
 358          * that type anymore.
 359          */
 360         if (!XFS_IS_UQUOTA_ON(mp))
 361                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
 362         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp)))
 363                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
 364
 365  write_changes:
 366         /*
 367          * We actually don't have to acquire the m_sb_lock at all.
 368          * This can only be called from mount, and that's single threaded. XXX
 369          */
 370         spin_lock(&mp->m_sb_lock);
 371         sbf = mp->m_sb.sb_qflags;
 372         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
 373         spin_unlock(&mp->m_sb_lock);
 374
 375         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
 376                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
 377                         /*
 378                          * We could only have been turning quotas off.
 379                          * We aren't in very good shape actually because
 380                          * the incore structures are convinced that quotas are
 381                          * off, but the on disk superblock doesn't know that !
 382                          */
 383                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
 384                         xfs_fs_cmn_err(CE_ALERT, mp,
 385                                 "XFS mount_quotas: Superblock update failed!");
 386                 }
 387         }
 388
 389         if (error) {
 390                 xfs_fs_cmn_err(CE_WARN, mp,
 391                         "Failed to initialize disk quotas.");
 392         }
 393         return;
 394 }
 395
 396 /*
 397  * Called from the vfsops layer.
 398  */
 399 int
 400 xfs_qm_unmount_quotas(
 401         xfs_mount_t     *mp)
 402 {
 403         xfs_inode_t     *uqp, *gqp;
 404         int             error = 0;
 405
 406         /*
 407          * Release the dquots that root inode, et al might be holding,
 408          * before we flush quotas and blow away the quotainfo structure.
 409          */
 410         ASSERT(mp->m_rootip);
 411         xfs_qm_dqdetach(mp->m_rootip);
 412         if (mp->m_rbmip)
 413                 xfs_qm_dqdetach(mp->m_rbmip);
 414         if (mp->m_rsumip)
 415                 xfs_qm_dqdetach(mp->m_rsumip);
 416
 417         /*
 418          * Flush out the quota inodes.
 419          */
 420         uqp = gqp = NULL;
 421         if (mp->m_quotainfo) {
 422                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
 423                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
 424                         xfs_iflock(uqp);
 425                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
 426                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
 427                         if (unlikely(error == EFSCORRUPTED)) {
 428                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
 429                                                  XFS_ERRLEVEL_LOW, mp);
 430                                 goto out;
 431                         }
 432                 }
 433                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
 434                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
 435                         xfs_iflock(gqp);
 436                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
 437                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
 438                         if (unlikely(error == EFSCORRUPTED)) {
 439                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
 440                                                  XFS_ERRLEVEL_LOW, mp);
 441                                 goto out;
 442                         }
 443                 }
 444         }
 445         if (uqp) {
 446                  IRELE(uqp);
 447                  mp->m_quotainfo->qi_uquotaip = NULL;
 448         }
 449         if (gqp) {
 450                 IRELE(gqp);
 451                 mp->m_quotainfo->qi_gquotaip = NULL;
 452         }
 453 out:
 454         return XFS_ERROR(error);
 455 }
 456
 457 /*
 458  * Flush all dquots of the given file system to disk. The dquots are
 459  * _not_ purged from memory here, just their data written to disk.
 460  */
 461 STATIC int
 462 xfs_qm_dqflush_all(
 463         xfs_mount_t     *mp,
 464         int             flags)
 465 {
 466         int             recl;
 467         xfs_dquot_t     *dqp;
 468         int             niters;
 469         int             error;
 470
 471         if (mp->m_quotainfo == NULL)
 472                 return 0;
 473         niters = 0;
 474 again:
 475         xfs_qm_mplist_lock(mp);
 476         FOREACH_DQUOT_IN_MP(dqp, mp) {
 477                 xfs_dqlock(dqp);
 478                 if (! XFS_DQ_IS_DIRTY(dqp)) {
 479                         xfs_dqunlock(dqp);
 480                         continue;
 481                 }
 482                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
 483                 /* XXX a sentinel would be better */
 484                 recl = XFS_QI_MPLRECLAIMS(mp);
 485                 if (!xfs_dqflock_nowait(dqp)) {
 486                         /*
 487                          * If we can't grab the flush lock then check
 488                          * to see if the dquot has been flushed delayed
 489                          * write.  If so, grab its buffer and send it
 490                          * out immediately.  We'll be able to acquire
 491                          * the flush lock when the I/O completes.
 492                          */
 493                         xfs_qm_dqflock_pushbuf_wait(dqp);
 494                 }
 495                 /*
 496                  * Let go of the mplist lock. We don't want to hold it
 497                  * across a disk write.
 498                  */
 499                 xfs_qm_mplist_unlock(mp);
 500                 error = xfs_qm_dqflush(dqp, flags);
 501                 xfs_dqunlock(dqp);
 502                 if (error)
 503                         return error;
 504
 505                 xfs_qm_mplist_lock(mp);
 506                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
 507                         xfs_qm_mplist_unlock(mp);
 508                         /* XXX restart limit */
 509                         goto again;
 510                 }
 511         }
 512
 513         xfs_qm_mplist_unlock(mp);
 514         /* return ! busy */
 515         return 0;
 516 }
 517 /*
 518  * Release the group dquot pointers the user dquots may be
 519  * carrying around as a hint. mplist is locked on entry and exit.
 520  */
 521 STATIC void
 522 xfs_qm_detach_gdquots(
 523         xfs_mount_t     *mp)
 524 {
 525         xfs_dquot_t     *dqp, *gdqp;
 526         int             nrecl;
 527
 528  again:
 529         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
 530         dqp = XFS_QI_MPLNEXT(mp);
 531         while (dqp) {
 532                 xfs_dqlock(dqp);
 533                 if ((gdqp = dqp->q_gdquot)) {
 534                         xfs_dqlock(gdqp);
 535                         dqp->q_gdquot = NULL;
 536                 }
 537                 xfs_dqunlock(dqp);
 538
 539                 if (gdqp) {
 540                         /*
 541                          * Can't hold the mplist lock across a dqput.
 542                          * XXXmust convert to marker based iterations here.
 543                          */
 544                         nrecl = XFS_QI_MPLRECLAIMS(mp);
 545                         xfs_qm_mplist_unlock(mp);
 546                         xfs_qm_dqput(gdqp);
 547
 548                         xfs_qm_mplist_lock(mp);
 549                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
 550                                 goto again;
 551                 }
 552                 dqp = dqp->MPL_NEXT;
 553         }
 554 }
 555
 556 /*
 557  * Go through all the incore dquots of this file system and take them
 558  * off the mplist and hashlist, if the dquot type matches the dqtype
 559  * parameter. This is used when turning off quota accounting for
 560  * users and/or groups, as well as when the filesystem is unmounting.
 561  */
 562 STATIC int
 563 xfs_qm_dqpurge_int(
 564         xfs_mount_t     *mp,
 565         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
 566 {
 567         xfs_dquot_t     *dqp;
 568         uint            dqtype;
 569         int             nrecl;
 570         xfs_dquot_t     *nextdqp;
 571         int             nmisses;
 572
 573         if (mp->m_quotainfo == NULL)
 574                 return 0;
 575
 576         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
 577         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
 578         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
 579
 580         xfs_qm_mplist_lock(mp);
 581
 582         /*
 583          * In the first pass through all incore dquots of this filesystem,
 584          * we release the group dquot pointers the user dquots may be
 585          * carrying around as a hint. We need to do this irrespective of
 586          * what's being turned off.
 587          */
 588         xfs_qm_detach_gdquots(mp);
 589
 590       again:
 591         nmisses = 0;
 592         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
 593         /*
 594          * Try to get rid of all of the unwanted dquots. The idea is to
 595          * get them off mplist and hashlist, but leave them on freelist.
 596          */
 597         dqp = XFS_QI_MPLNEXT(mp);
 598         while (dqp) {
 599                 /*
 600                  * It's OK to look at the type without taking dqlock here.
 601                  * We're holding the mplist lock here, and that's needed for
 602                  * a dqreclaim.
 603                  */
 604                 if ((dqp->dq_flags & dqtype) == 0) {
 605                         dqp = dqp->MPL_NEXT;
 606                         continue;
 607                 }
 608
 609                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
 610                         nrecl = XFS_QI_MPLRECLAIMS(mp);
 611                         xfs_qm_mplist_unlock(mp);
 612                         XFS_DQ_HASH_LOCK(dqp->q_hash);
 613                         xfs_qm_mplist_lock(mp);
 614
 615                         /*
 616                          * XXXTheoretically, we can get into a very long
 617                          * ping pong game here.
 618                          * No one can be adding dquots to the mplist at
 619                          * this point, but somebody might be taking things off.
 620                          */
 621                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
 622                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
 623                                 goto again;
 624                         }
 625                 }
 626
 627                 /*
 628                  * Take the dquot off the mplist and hashlist. It may remain on
 629                  * freelist in INACTIVE state.
 630                  */
 631                 nextdqp = dqp->MPL_NEXT;
 632                 nmisses += xfs_qm_dqpurge(dqp);
 633                 dqp = nextdqp;
 634         }
 635         xfs_qm_mplist_unlock(mp);
 636         return nmisses;
 637 }
 638
 639 int
 640 xfs_qm_dqpurge_all(
 641         xfs_mount_t     *mp,
 642         uint            flags)
 643 {
 644         int             ndquots;
 645
 646         /*
 647          * Purge the dquot cache.
 648          * None of the dquots should really be busy at this point.
 649          */
 650         if (mp->m_quotainfo) {
 651                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
 652                         delay(ndquots * 10);
 653                 }
 654         }
 655         return 0;
 656 }
 657
 658 STATIC int
 659 xfs_qm_dqattach_one(
 660         xfs_inode_t     *ip,
 661         xfs_dqid_t      id,
 662         uint            type,
 663         uint            doalloc,
 664         uint            dolock,
 665         xfs_dquot_t     *udqhint, /* hint */
 666         xfs_dquot_t     **IO_idqpp)
 667 {
 668         xfs_dquot_t     *dqp;
 669         int             error;
 670
 671         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 672         error = 0;
 673         /*
 674          * See if we already have it in the inode itself. IO_idqpp is
 675          * &i_udquot or &i_gdquot. This made the code look weird, but
 676          * made the logic a lot simpler.
 677          */
 678         if ((dqp = *IO_idqpp)) {
 679                 if (dolock)
 680                         xfs_dqlock(dqp);
 681                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
 682                 goto done;
 683         }
 684
 685         /*
 686          * udqhint is the i_udquot field in inode, and is non-NULL only
 687          * when the type arg is group/project. Its purpose is to save a
 688          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
 689          * the user dquot.
 690          */
 691         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
 692         if (udqhint && !dolock)
 693                 xfs_dqlock(udqhint);
 694
 695         /*
 696          * No need to take dqlock to look at the id.
 697          * The ID can't change until it gets reclaimed, and it won't
 698          * be reclaimed as long as we have a ref from inode and we hold
 699          * the ilock.
 700          */
 701         if (udqhint &&
 702             (dqp = udqhint->q_gdquot) &&
 703             (be32_to_cpu(dqp->q_core.d_id) == id)) {
 704                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
 705                 xfs_dqlock(dqp);
 706                 XFS_DQHOLD(dqp);
 707                 ASSERT(*IO_idqpp == NULL);
 708                 *IO_idqpp = dqp;
 709                 if (!dolock) {
 710                         xfs_dqunlock(dqp);
 711                         xfs_dqunlock(udqhint);
 712                 }
 713                 goto done;
 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         if (udqhint)
 722                 xfs_dqunlock(udqhint);
 723         /*
 724          * Find the dquot from somewhere. This bumps the
 725          * reference count of dquot and returns it locked.
 726          * This can return ENOENT if dquot didn't exist on
 727          * disk and we didn't ask it to allocate;
 728          * ESRCH if quotas got turned off suddenly.
 729          */
 730         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
 731                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
 732                 if (udqhint && dolock)
 733                         xfs_dqlock(udqhint);
 734                 goto done;
 735         }
 736
 737         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
 738         /*
 739          * dqget may have dropped and re-acquired the ilock, but it guarantees
 740          * that the dquot returned is the one that should go in the inode.
 741          */
 742         *IO_idqpp = dqp;
 743         ASSERT(dqp);
 744         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 745         if (! dolock) {
 746                 xfs_dqunlock(dqp);
 747                 goto done;
 748         }
 749         if (! udqhint)
 750                 goto done;
 751
 752         ASSERT(udqhint);
 753         ASSERT(dolock);
 754         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 755         if (! xfs_qm_dqlock_nowait(udqhint)) {
 756                 xfs_dqunlock(dqp);
 757                 xfs_dqlock(udqhint);
 758                 xfs_dqlock(dqp);
 759         }
 760       done:
 761 #ifdef QUOTADEBUG
 762         if (udqhint) {
 763                 if (dolock)
 764                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
 765         }
 766         if (! error) {
 767                 if (dolock)
 768                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
 769         }
 770 #endif
 771         return error;
 772 }
 773
 774
 775 /*
 776  * Given a udquot and gdquot, attach a ptr to the group dquot in the
 777  * udquot as a hint for future lookups. The idea sounds simple, but the
 778  * execution isn't, because the udquot might have a group dquot attached
 779  * already and getting rid of that gets us into lock ordering constraints.
 780  * The process is complicated more by the fact that the dquots may or may not
 781  * be locked on entry.
 782  */
 783 STATIC void
 784 xfs_qm_dqattach_grouphint(
 785         xfs_dquot_t     *udq,
 786         xfs_dquot_t     *gdq,
 787         uint            locked)
 788 {
 789         xfs_dquot_t     *tmp;
 790
 791 #ifdef QUOTADEBUG
 792         if (locked) {
 793                 ASSERT(XFS_DQ_IS_LOCKED(udq));
 794                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
 795         }
 796 #endif
 797         if (! locked)
 798                 xfs_dqlock(udq);
 799
 800         if ((tmp = udq->q_gdquot)) {
 801                 if (tmp == gdq) {
 802                         if (! locked)
 803                                 xfs_dqunlock(udq);
 804                         return;
 805                 }
 806
 807                 udq->q_gdquot = NULL;
 808                 /*
 809                  * We can't keep any dqlocks when calling dqrele,
 810                  * because the freelist lock comes before dqlocks.
 811                  */
 812                 xfs_dqunlock(udq);
 813                 if (locked)
 814                         xfs_dqunlock(gdq);
 815                 /*
 816                  * we took a hard reference once upon a time in dqget,
 817                  * so give it back when the udquot no longer points at it
 818                  * dqput() does the unlocking of the dquot.
 819                  */
 820                 xfs_qm_dqrele(tmp);
 821
 822                 xfs_dqlock(udq);
 823                 xfs_dqlock(gdq);
 824
 825         } else {
 826                 ASSERT(XFS_DQ_IS_LOCKED(udq));
 827                 if (! locked) {
 828                         xfs_dqlock(gdq);
 829                 }
 830         }
 831
 832         ASSERT(XFS_DQ_IS_LOCKED(udq));
 833         ASSERT(XFS_DQ_IS_LOCKED(gdq));
 834         /*
 835          * Somebody could have attached a gdquot here,
 836          * when we dropped the uqlock. If so, just do nothing.
 837          */
 838         if (udq->q_gdquot == NULL) {
 839                 XFS_DQHOLD(gdq);
 840                 udq->q_gdquot = gdq;
 841         }
 842         if (! locked) {
 843                 xfs_dqunlock(gdq);
 844                 xfs_dqunlock(udq);
 845         }
 846 }
 847
 848
 849 /*
 850  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
 851  * into account.
 852  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
 853  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
 854  * much made this code a complete mess, but it has been pretty useful.
 855  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
 856  * Inode may get unlocked and relocked in here, and the caller must deal with
 857  * the consequences.
 858  */
 859 int
 860 xfs_qm_dqattach(
 861         xfs_inode_t     *ip,
 862         uint            flags)
 863 {
 864         xfs_mount_t     *mp = ip->i_mount;
 865         uint            nquotas = 0;
 866         int             error = 0;
 867
 868         if ((! XFS_IS_QUOTA_ON(mp)) ||
 869             (! XFS_NOT_DQATTACHED(mp, ip)) ||
 870             (ip->i_ino == mp->m_sb.sb_uquotino) ||
 871             (ip->i_ino == mp->m_sb.sb_gquotino))
 872                 return 0;
 873
 874         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
 875                xfs_isilocked(ip, XFS_ILOCK_EXCL));
 876
 877         if (! (flags & XFS_QMOPT_ILOCKED))
 878                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 879
 880         if (XFS_IS_UQUOTA_ON(mp)) {
 881                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
 882                                                 flags & XFS_QMOPT_DQALLOC,
 883                                                 flags & XFS_QMOPT_DQLOCK,
 884                                                 NULL, &ip->i_udquot);
 885                 if (error)
 886                         goto done;
 887                 nquotas++;
 888         }
 889
 890         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 891         if (XFS_IS_OQUOTA_ON(mp)) {
 892                 error = XFS_IS_GQUOTA_ON(mp) ?
 893                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
 894                                                 flags & XFS_QMOPT_DQALLOC,
 895                                                 flags & XFS_QMOPT_DQLOCK,
 896                                                 ip->i_udquot, &ip->i_gdquot) :
 897                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
 898                                                 flags & XFS_QMOPT_DQALLOC,
 899                                                 flags & XFS_QMOPT_DQLOCK,
 900                                                 ip->i_udquot, &ip->i_gdquot);
 901                 /*
 902                  * Don't worry about the udquot that we may have
 903                  * attached above. It'll get detached, if not already.
 904                  */
 905                 if (error)
 906                         goto done;
 907                 nquotas++;
 908         }
 909
 910         /*
 911          * Attach this group quota to the user quota as a hint.
 912          * This WON'T, in general, result in a thrash.
 913          */
 914         if (nquotas == 2) {
 915                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 916                 ASSERT(ip->i_udquot);
 917                 ASSERT(ip->i_gdquot);
 918
 919                 /*
 920                  * We may or may not have the i_udquot locked at this point,
 921                  * but this check is OK since we don't depend on the i_gdquot to
 922                  * be accurate 100% all the time. It is just a hint, and this
 923                  * will succeed in general.
 924                  */
 925                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
 926                         goto done;
 927                 /*
 928                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
 929                  */
 930                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
 931                                          flags & XFS_QMOPT_DQLOCK);
 932         }
 933
 934       done:
 935
 936 #ifdef QUOTADEBUG
 937         if (! error) {
 938                 if (ip->i_udquot) {
 939                         if (flags & XFS_QMOPT_DQLOCK)
 940                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
 941                 }
 942                 if (ip->i_gdquot) {
 943                         if (flags & XFS_QMOPT_DQLOCK)
 944                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
 945                 }
 946                 if (XFS_IS_UQUOTA_ON(mp))
 947                         ASSERT(ip->i_udquot);
 948                 if (XFS_IS_OQUOTA_ON(mp))
 949                         ASSERT(ip->i_gdquot);
 950         }
 951 #endif
 952
 953         if (! (flags & XFS_QMOPT_ILOCKED))
 954                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 955
 956 #ifdef QUOTADEBUG
 957         else
 958                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 959 #endif
 960         return error;
 961 }
 962
 963 /*
 964  * Release dquots (and their references) if any.
 965  * The inode should be locked EXCL except when this's called by
 966  * xfs_ireclaim.
 967  */
 968 void
 969 xfs_qm_dqdetach(
 970         xfs_inode_t     *ip)
 971 {
 972         if (!(ip->i_udquot || ip->i_gdquot))
 973                 return;
 974
 975         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
 976         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
 977         if (ip->i_udquot) {
 978                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
 979                 xfs_qm_dqrele(ip->i_udquot);
 980                 ip->i_udquot = NULL;
 981         }
 982         if (ip->i_gdquot) {
 983                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
 984                 xfs_qm_dqrele(ip->i_gdquot);
 985                 ip->i_gdquot = NULL;
 986         }
 987 }
 988
 989 /*
 990  * This is called by VFS_SYNC and flags arg determines the caller,
 991  * and its motives, as done in xfs_sync.
 992  *
 993  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
 994  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
 995  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
 996  */
 997
 998 int
 999 xfs_qm_sync(
1000         xfs_mount_t     *mp,
1001         int             flags)
1002 {
1003         int             recl, restarts;
1004         xfs_dquot_t     *dqp;
1005         uint            flush_flags;
1006         boolean_t       nowait;
1007         int             error;
1008
1009         if (! XFS_IS_QUOTA_ON(mp))
1010                 return 0;
1011
1012         restarts = 0;
1013         /*
1014          * We won't block unless we are asked to.
1015          */
1016         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1017
1018   again:
1019         xfs_qm_mplist_lock(mp);
1020         /*
1021          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1022          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1023          * when we have the mplist lock, we know that dquots will be consistent
1024          * as long as we have it locked.
1025          */
1026         if (! XFS_IS_QUOTA_ON(mp)) {
1027                 xfs_qm_mplist_unlock(mp);
1028                 return 0;
1029         }
1030         FOREACH_DQUOT_IN_MP(dqp, mp) {
1031                 /*
1032                  * If this is vfs_sync calling, then skip the dquots that
1033                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1034                  * This is very similar to what xfs_sync does with inodes.
1035                  */
1036                 if (flags & SYNC_BDFLUSH) {
1037                         if (! XFS_DQ_IS_DIRTY(dqp))
1038                                 continue;
1039                 }
1040
1041                 if (nowait) {
1042                         /*
1043                          * Try to acquire the dquot lock. We are NOT out of
1044                          * lock order, but we just don't want to wait for this
1045                          * lock, unless somebody wanted us to.
1046                          */
1047                         if (! xfs_qm_dqlock_nowait(dqp))
1048                                 continue;
1049                 } else {
1050                         xfs_dqlock(dqp);
1051                 }
1052
1053                 /*
1054                  * Now, find out for sure if this dquot is dirty or not.
1055                  */
1056                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1057                         xfs_dqunlock(dqp);
1058                         continue;
1059                 }
1060
1061                 /* XXX a sentinel would be better */
1062                 recl = XFS_QI_MPLRECLAIMS(mp);
1063                 if (!xfs_dqflock_nowait(dqp)) {
1064                         if (nowait) {
1065                                 xfs_dqunlock(dqp);
1066                                 continue;
1067                         }
1068                         /*
1069                          * If we can't grab the flush lock then if the caller
1070                          * really wanted us to give this our best shot, so
1071                          * see if we can give a push to the buffer before we wait
1072                          * on the flush lock. At this point, we know that
1073                          * even though the dquot is being flushed,
1074                          * it has (new) dirty data.
1075                          */
1076                         xfs_qm_dqflock_pushbuf_wait(dqp);
1077                 }
1078                 /*
1079                  * Let go of the mplist lock. We don't want to hold it
1080                  * across a disk write
1081                  */
1082                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1083                 xfs_qm_mplist_unlock(mp);
1084                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1085                 error = xfs_qm_dqflush(dqp, flush_flags);
1086                 xfs_dqunlock(dqp);
1087                 if (error && XFS_FORCED_SHUTDOWN(mp))
1088                         return 0;       /* Need to prevent umount failure */
1089                 else if (error)
1090                         return error;
1091
1092                 xfs_qm_mplist_lock(mp);
1093                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1094                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1095                                 break;
1096
1097                         xfs_qm_mplist_unlock(mp);
1098                         goto again;
1099                 }
1100         }
1101
1102         xfs_qm_mplist_unlock(mp);
1103         return 0;
1104 }
1105
1106
1107 /*
1108  * This initializes all the quota information that's kept in the
1109  * mount structure
1110  */
1111 STATIC int
1112 xfs_qm_init_quotainfo(
1113         xfs_mount_t     *mp)
1114 {
1115         xfs_quotainfo_t *qinf;
1116         int             error;
1117         xfs_dquot_t     *dqp;
1118
1119         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1120
1121         /*
1122          * Tell XQM that we exist as soon as possible.
1123          */
1124         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1125                 return error;
1126         }
1127
1128         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1129
1130         /*
1131          * See if quotainodes are setup, and if not, allocate them,
1132          * and change the superblock accordingly.
1133          */
1134         if ((error = xfs_qm_init_quotainos(mp))) {
1135                 kmem_free(qinf);
1136                 mp->m_quotainfo = NULL;
1137                 return error;
1138         }
1139
1140         spin_lock_init(&qinf->qi_pinlock);
1141         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1142         qinf->qi_dqreclaims = 0;
1143
1144         /* mutex used to serialize quotaoffs */
1145         mutex_init(&qinf->qi_quotaofflock);
1146
1147         /* Precalc some constants */
1148         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1149         ASSERT(qinf->qi_dqchunklen);
1150         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1151         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1152
1153         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1154
1155         /*
1156          * We try to get the limits from the superuser's limits fields.
1157          * This is quite hacky, but it is standard quota practice.
1158          * We look at the USR dquot with id == 0 first, but if user quotas
1159          * are not enabled we goto the GRP dquot with id == 0.
1160          * We don't really care to keep separate default limits for user
1161          * and group quotas, at least not at this point.
1162          */
1163         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1164                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER :
1165                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1166                                 XFS_DQ_PROJ),
1167                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1168                              &dqp);
1169         if (! error) {
1170                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1171
1172                 /*
1173                  * The warnings and timers set the grace period given to
1174                  * a user or group before he or she can not perform any
1175                  * more writing. If it is zero, a default is used.
1176                  */
1177                 qinf->qi_btimelimit = ddqp->d_btimer ?
1178                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1179                 qinf->qi_itimelimit = ddqp->d_itimer ?
1180                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1181                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1182                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1183                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1184                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1185                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1186                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1187                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1188                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1189                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1190                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1191                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1192                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1193                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1194                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1195
1196                 /*
1197                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1198                  * we don't want this dquot cached. We haven't done a
1199                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1200                  */
1201                 xfs_qm_dqdestroy(dqp);
1202         } else {
1203                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1204                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1205                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1206                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1207                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1208                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1209         }
1210
1211         return 0;
1212 }
1213
1214
1215 /*
1216  * Gets called when unmounting a filesystem or when all quotas get
1217  * turned off.
1218  * This purges the quota inodes, destroys locks and frees itself.
1219  */
1220 void
1221 xfs_qm_destroy_quotainfo(
1222         xfs_mount_t     *mp)
1223 {
1224         xfs_quotainfo_t *qi;
1225
1226         qi = mp->m_quotainfo;
1227         ASSERT(qi != NULL);
1228         ASSERT(xfs_Gqm != NULL);
1229
1230         /*
1231          * Release the reference that XQM kept, so that we know
1232          * when the XQM structure should be freed. We cannot assume
1233          * that xfs_Gqm is non-null after this point.
1234          */
1235         xfs_qm_rele_quotafs_ref(mp);
1236
1237         spinlock_destroy(&qi->qi_pinlock);
1238         xfs_qm_list_destroy(&qi->qi_dqlist);
1239
1240         if (qi->qi_uquotaip) {
1241                 IRELE(qi->qi_uquotaip);
1242                 qi->qi_uquotaip = NULL; /* paranoia */
1243         }
1244         if (qi->qi_gquotaip) {
1245                 IRELE(qi->qi_gquotaip);
1246                 qi->qi_gquotaip = NULL;
1247         }
1248         mutex_destroy(&qi->qi_quotaofflock);
1249         kmem_free(qi);
1250         mp->m_quotainfo = NULL;
1251 }
1252
1253
1254
1255 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1256
1257 /* ARGSUSED */
1258 STATIC void
1259 xfs_qm_list_init(
1260         xfs_dqlist_t    *list,
1261         char            *str,
1262         int             n)
1263 {
1264         mutex_init(&list->qh_lock);
1265         list->qh_next = NULL;
1266         list->qh_version = 0;
1267         list->qh_nelems = 0;
1268 }
1269
1270 STATIC void
1271 xfs_qm_list_destroy(
1272         xfs_dqlist_t    *list)
1273 {
1274         mutex_destroy(&(list->qh_lock));
1275 }
1276
1277
1278 /*
1279  * Stripped down version of dqattach. This doesn't attach, or even look at the
1280  * dquots attached to the inode. The rationale is that there won't be any
1281  * attached at the time this is called from quotacheck.
1282  */
1283 STATIC int
1284 xfs_qm_dqget_noattach(
1285         xfs_inode_t     *ip,
1286         xfs_dquot_t     **O_udqpp,
1287         xfs_dquot_t     **O_gdqpp)
1288 {
1289         int             error;
1290         xfs_mount_t     *mp;
1291         xfs_dquot_t     *udqp, *gdqp;
1292
1293         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1294         mp = ip->i_mount;
1295         udqp = NULL;
1296         gdqp = NULL;
1297
1298         if (XFS_IS_UQUOTA_ON(mp)) {
1299                 ASSERT(ip->i_udquot == NULL);
1300                 /*
1301                  * We want the dquot allocated if it doesn't exist.
1302                  */
1303                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1304                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1305                                          &udqp))) {
1306                         /*
1307                          * Shouldn't be able to turn off quotas here.
1308                          */
1309                         ASSERT(error != ESRCH);
1310                         ASSERT(error != ENOENT);
1311                         return error;
1312                 }
1313                 ASSERT(udqp);
1314         }
1315
1316         if (XFS_IS_OQUOTA_ON(mp)) {
1317                 ASSERT(ip->i_gdquot == NULL);
1318                 if (udqp)
1319                         xfs_dqunlock(udqp);
1320                 error = XFS_IS_GQUOTA_ON(mp) ?
1321                                 xfs_qm_dqget(mp, ip,
1322                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1323                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1324                                              &gdqp) :
1325                                 xfs_qm_dqget(mp, ip,
1326                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1327                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1328                                              &gdqp);
1329                 if (error) {
1330                         if (udqp)
1331                                 xfs_qm_dqrele(udqp);
1332                         ASSERT(error != ESRCH);
1333                         ASSERT(error != ENOENT);
1334                         return error;
1335                 }
1336                 ASSERT(gdqp);
1337
1338                 /* Reacquire the locks in the right order */
1339                 if (udqp) {
1340                         if (! xfs_qm_dqlock_nowait(udqp)) {
1341                                 xfs_dqunlock(gdqp);
1342                                 xfs_dqlock(udqp);
1343                                 xfs_dqlock(gdqp);
1344                         }
1345                 }
1346         }
1347
1348         *O_udqpp = udqp;
1349         *O_gdqpp = gdqp;
1350
1351 #ifdef QUOTADEBUG
1352         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1353         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1354 #endif
1355         return 0;
1356 }
1357
1358 /*
1359  * Create an inode and return with a reference already taken, but unlocked
1360  * This is how we create quota inodes
1361  */
1362 STATIC int
1363 xfs_qm_qino_alloc(
1364         xfs_mount_t     *mp,
1365         xfs_inode_t     **ip,
1366         __int64_t       sbfields,
1367         uint            flags)
1368 {
1369         xfs_trans_t     *tp;
1370         int             error;
1371         int             committed;
1372
1373         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1374         if ((error = xfs_trans_reserve(tp,
1375                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1376                                       XFS_CREATE_LOG_RES(mp), 0,
1377                                       XFS_TRANS_PERM_LOG_RES,
1378                                       XFS_CREATE_LOG_COUNT))) {
1379                 xfs_trans_cancel(tp, 0);
1380                 return error;
1381         }
1382
1383         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1384                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1385                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1386                                  XFS_TRANS_ABORT);
1387                 return error;
1388         }
1389
1390         /*
1391          * Keep an extra reference to this quota inode. This inode is
1392          * locked exclusively and joined to the transaction already.
1393          */
1394         ASSERT(xfs_isilocked(*ip, XFS_ILOCK_EXCL));
1395         IHOLD(*ip);
1396
1397         /*
1398          * Make the changes in the superblock, and log those too.
1399          * sbfields arg may contain fields other than *QUOTINO;
1400          * VERSIONNUM for example.
1401          */
1402         spin_lock(&mp->m_sb_lock);
1403         if (flags & XFS_QMOPT_SBVERSION) {
1404 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1405                 unsigned oldv = mp->m_sb.sb_versionnum;
1406 #endif
1407                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1408                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1409                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1410                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1411                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1412
1413                 xfs_sb_version_addquota(&mp->m_sb);
1414                 mp->m_sb.sb_uquotino = NULLFSINO;
1415                 mp->m_sb.sb_gquotino = NULLFSINO;
1416
1417                 /* qflags will get updated _after_ quotacheck */
1418                 mp->m_sb.sb_qflags = 0;
1419 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1420                 cmn_err(CE_NOTE,
1421                         "Old superblock version %x, converting to %x.",
1422                         oldv, mp->m_sb.sb_versionnum);
1423 #endif
1424         }
1425         if (flags & XFS_QMOPT_UQUOTA)
1426                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1427         else
1428                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1429         spin_unlock(&mp->m_sb_lock);
1430         xfs_mod_sb(tp, sbfields);
1431
1432         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1433                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1434                 return error;
1435         }
1436         return 0;
1437 }
1438
1439
1440 STATIC void
1441 xfs_qm_reset_dqcounts(
1442         xfs_mount_t     *mp,
1443         xfs_buf_t       *bp,
1444         xfs_dqid_t      id,
1445         uint            type)
1446 {
1447         xfs_disk_dquot_t        *ddq;
1448         int                     j;
1449
1450         xfs_buftrace("RESET DQUOTS", bp);
1451         /*
1452          * Reset all counters and timers. They'll be
1453          * started afresh by xfs_qm_quotacheck.
1454          */
1455 #ifdef DEBUG
1456         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1457         do_div(j, sizeof(xfs_dqblk_t));
1458         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1459 #endif
1460         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1461         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1462                 /*
1463                  * Do a sanity check, and if needed, repair the dqblk. Don't
1464                  * output any warnings because it's perfectly possible to
1465                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1466                  */
1467                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1468                                       "xfs_quotacheck");
1469                 ddq->d_bcount = 0;
1470                 ddq->d_icount = 0;
1471                 ddq->d_rtbcount = 0;
1472                 ddq->d_btimer = 0;
1473                 ddq->d_itimer = 0;
1474                 ddq->d_rtbtimer = 0;
1475                 ddq->d_bwarns = 0;
1476                 ddq->d_iwarns = 0;
1477                 ddq->d_rtbwarns = 0;
1478                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1479         }
1480 }
1481
1482 STATIC int
1483 xfs_qm_dqiter_bufs(
1484         xfs_mount_t     *mp,
1485         xfs_dqid_t      firstid,
1486         xfs_fsblock_t   bno,
1487         xfs_filblks_t   blkcnt,
1488         uint            flags)
1489 {
1490         xfs_buf_t       *bp;
1491         int             error;
1492         int             notcommitted;
1493         int             incr;
1494         int             type;
1495
1496         ASSERT(blkcnt > 0);
1497         notcommitted = 0;
1498         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1499                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1500         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1501                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1502         error = 0;
1503
1504         /*
1505          * Blkcnt arg can be a very big number, and might even be
1506          * larger than the log itself. So, we have to break it up into
1507          * manageable-sized transactions.
1508          * Note that we don't start a permanent transaction here; we might
1509          * not be able to get a log reservation for the whole thing up front,
1510          * and we don't really care to either, because we just discard
1511          * everything if we were to crash in the middle of this loop.
1512          */
1513         while (blkcnt--) {
1514                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1515                               XFS_FSB_TO_DADDR(mp, bno),
1516                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1517                 if (error)
1518                         break;
1519
1520                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1521                 xfs_bdwrite(mp, bp);
1522                 /*
1523                  * goto the next block.
1524                  */
1525                 bno++;
1526                 firstid += XFS_QM_DQPERBLK(mp);
1527         }
1528         return error;
1529 }
1530
1531 /*
1532  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1533  * caller supplied function for every chunk of dquots that we find.
1534  */
1535 STATIC int
1536 xfs_qm_dqiterate(
1537         xfs_mount_t     *mp,
1538         xfs_inode_t     *qip,
1539         uint            flags)
1540 {
1541         xfs_bmbt_irec_t         *map;
1542         int                     i, nmaps;       /* number of map entries */
1543         int                     error;          /* return value */
1544         xfs_fileoff_t           lblkno;
1545         xfs_filblks_t           maxlblkcnt;
1546         xfs_dqid_t              firstid;
1547         xfs_fsblock_t           rablkno;
1548         xfs_filblks_t           rablkcnt;
1549
1550         error = 0;
1551         /*
1552          * This looks racy, but we can't keep an inode lock across a
1553          * trans_reserve. But, this gets called during quotacheck, and that
1554          * happens only at mount time which is single threaded.
1555          */
1556         if (qip->i_d.di_nblocks == 0)
1557                 return 0;
1558
1559         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1560
1561         lblkno = 0;
1562         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1563         do {
1564                 nmaps = XFS_DQITER_MAP_SIZE;
1565                 /*
1566                  * We aren't changing the inode itself. Just changing
1567                  * some of its data. No new blocks are added here, and
1568                  * the inode is never added to the transaction.
1569                  */
1570                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1571                 error = xfs_bmapi(NULL, qip, lblkno,
1572                                   maxlblkcnt - lblkno,
1573                                   XFS_BMAPI_METADATA,
1574                                   NULL,
1575                                   0, map, &nmaps, NULL, NULL);
1576                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1577                 if (error)
1578                         break;
1579
1580                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1581                 for (i = 0; i < nmaps; i++) {
1582                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1583                         ASSERT(map[i].br_blockcount);
1584
1585
1586                         lblkno += map[i].br_blockcount;
1587
1588                         if (map[i].br_startblock == HOLESTARTBLOCK)
1589                                 continue;
1590
1591                         firstid = (xfs_dqid_t) map[i].br_startoff *
1592                                 XFS_QM_DQPERBLK(mp);
1593                         /*
1594                          * Do a read-ahead on the next extent.
1595                          */
1596                         if ((i+1 < nmaps) &&
1597                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1598                                 rablkcnt =  map[i+1].br_blockcount;
1599                                 rablkno = map[i+1].br_startblock;
1600                                 while (rablkcnt--) {
1601                                         xfs_baread(mp->m_ddev_targp,
1602                                                XFS_FSB_TO_DADDR(mp, rablkno),
1603                                                (int)XFS_QI_DQCHUNKLEN(mp));
1604                                         rablkno++;
1605                                 }
1606                         }
1607                         /*
1608                          * Iterate thru all the blks in the extent and
1609                          * reset the counters of all the dquots inside them.
1610                          */
1611                         if ((error = xfs_qm_dqiter_bufs(mp,
1612                                                        firstid,
1613                                                        map[i].br_startblock,
1614                                                        map[i].br_blockcount,
1615                                                        flags))) {
1616                                 break;
1617                         }
1618                 }
1619
1620                 if (error)
1621                         break;
1622         } while (nmaps > 0);
1623
1624         kmem_free(map);
1625
1626         return error;
1627 }
1628
1629 /*
1630  * Called by dqusage_adjust in doing a quotacheck.
1631  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1632  * this updates its incore copy as well as the buffer copy. This is
1633  * so that once the quotacheck is done, we can just log all the buffers,
1634  * as opposed to logging numerous updates to individual dquots.
1635  */
1636 STATIC void
1637 xfs_qm_quotacheck_dqadjust(
1638         xfs_dquot_t             *dqp,
1639         xfs_qcnt_t              nblks,
1640         xfs_qcnt_t              rtblks)
1641 {
1642         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1643         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1644         /*
1645          * Adjust the inode count and the block count to reflect this inode's
1646          * resource usage.
1647          */
1648         be64_add_cpu(&dqp->q_core.d_icount, 1);
1649         dqp->q_res_icount++;
1650         if (nblks) {
1651                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1652                 dqp->q_res_bcount += nblks;
1653         }
1654         if (rtblks) {
1655                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1656                 dqp->q_res_rtbcount += rtblks;
1657         }
1658
1659         /*
1660          * Set default limits, adjust timers (since we changed usages)
1661          */
1662         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1663                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1664                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1665         }
1666
1667         dqp->dq_flags |= XFS_DQ_DIRTY;
1668 }
1669
1670 STATIC int
1671 xfs_qm_get_rtblks(
1672         xfs_inode_t     *ip,
1673         xfs_qcnt_t      *O_rtblks)
1674 {
1675         xfs_filblks_t   rtblks;                 /* total rt blks */
1676         xfs_extnum_t    idx;                    /* extent record index */
1677         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1678         xfs_extnum_t    nextents;               /* number of extent entries */
1679         int             error;
1680
1681         ASSERT(XFS_IS_REALTIME_INODE(ip));
1682         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1683         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1684                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1685                         return error;
1686         }
1687         rtblks = 0;
1688         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1689         for (idx = 0; idx < nextents; idx++)
1690                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1691         *O_rtblks = (xfs_qcnt_t)rtblks;
1692         return 0;
1693 }
1694
1695 /*
1696  * callback routine supplied to bulkstat(). Given an inumber, find its
1697  * dquots and update them to account for resources taken by that inode.
1698  */
1699 /* ARGSUSED */
1700 STATIC int
1701 xfs_qm_dqusage_adjust(
1702         xfs_mount_t     *mp,            /* mount point for filesystem */
1703         xfs_ino_t       ino,            /* inode number to get data for */
1704         void            __user *buffer, /* not used */
1705         int             ubsize,         /* not used */
1706         void            *private_data,  /* not used */
1707         xfs_daddr_t     bno,            /* starting block of inode cluster */
1708         int             *ubused,        /* not used */
1709         void            *dip,           /* on-disk inode pointer (not used) */
1710         int             *res)           /* result code value */
1711 {
1712         xfs_inode_t     *ip;
1713         xfs_dquot_t     *udqp, *gdqp;
1714         xfs_qcnt_t      nblks, rtblks;
1715         int             error;
1716
1717         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1718
1719         /*
1720          * rootino must have its resources accounted for, not so with the quota
1721          * inodes.
1722          */
1723         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1724                 *res = BULKSTAT_RV_NOTHING;
1725                 return XFS_ERROR(EINVAL);
1726         }
1727
1728         /*
1729          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1730          * interface expects the inode to be exclusively locked because that's
1731          * the case in all other instances. It's OK that we do this because
1732          * quotacheck is done only at mount time.
1733          */
1734         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1735                 *res = BULKSTAT_RV_NOTHING;
1736                 return error;
1737         }
1738
1739         /*
1740          * Obtain the locked dquots. In case of an error (eg. allocation
1741          * fails for ENOSPC), we return the negative of the error number
1742          * to bulkstat, so that it can get propagated to quotacheck() and
1743          * making us disable quotas for the file system.
1744          */
1745         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1746                 xfs_iput(ip, XFS_ILOCK_EXCL);
1747                 *res = BULKSTAT_RV_GIVEUP;
1748                 return error;
1749         }
1750
1751         rtblks = 0;
1752         if (! XFS_IS_REALTIME_INODE(ip)) {
1753                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1754         } else {
1755                 /*
1756                  * Walk thru the extent list and count the realtime blocks.
1757                  */
1758                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1759                         xfs_iput(ip, XFS_ILOCK_EXCL);
1760                         if (udqp)
1761                                 xfs_qm_dqput(udqp);
1762                         if (gdqp)
1763                                 xfs_qm_dqput(gdqp);
1764                         *res = BULKSTAT_RV_GIVEUP;
1765                         return error;
1766                 }
1767                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1768         }
1769         ASSERT(ip->i_delayed_blks == 0);
1770
1771         /*
1772          * We can't release the inode while holding its dquot locks.
1773          * The inode can go into inactive and might try to acquire the dquotlocks.
1774          * So, just unlock here and do a vn_rele at the end.
1775          */
1776         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1777
1778         /*
1779          * Add the (disk blocks and inode) resources occupied by this
1780          * inode to its dquots. We do this adjustment in the incore dquot,
1781          * and also copy the changes to its buffer.
1782          * We don't care about putting these changes in a transaction
1783          * envelope because if we crash in the middle of a 'quotacheck'
1784          * we have to start from the beginning anyway.
1785          * Once we're done, we'll log all the dquot bufs.
1786          *
1787          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1788          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1789          */
1790         if (XFS_IS_UQUOTA_ON(mp)) {
1791                 ASSERT(udqp);
1792                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1793                 xfs_qm_dqput(udqp);
1794         }
1795         if (XFS_IS_OQUOTA_ON(mp)) {
1796                 ASSERT(gdqp);
1797                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1798                 xfs_qm_dqput(gdqp);
1799         }
1800         /*
1801          * Now release the inode. This will send it to 'inactive', and
1802          * possibly even free blocks.
1803          */
1804         IRELE(ip);
1805
1806         /*
1807          * Goto next inode.
1808          */
1809         *res = BULKSTAT_RV_DIDONE;
1810         return 0;
1811 }
1812
1813 /*
1814  * Walk thru all the filesystem inodes and construct a consistent view
1815  * of the disk quota world. If the quotacheck fails, disable quotas.
1816  */
1817 int
1818 xfs_qm_quotacheck(
1819         xfs_mount_t     *mp)
1820 {
1821         int             done, count, error;
1822         xfs_ino_t       lastino;
1823         size_t          structsz;
1824         xfs_inode_t     *uip, *gip;
1825         uint            flags;
1826
1827         count = INT_MAX;
1828         structsz = 1;
1829         lastino = 0;
1830         flags = 0;
1831
1832         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1833         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1834
1835         /*
1836          * There should be no cached dquots. The (simplistic) quotacheck
1837          * algorithm doesn't like that.
1838          */
1839         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1840
1841         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1842
1843         /*
1844          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1845          * their counters to zero. We need a clean slate.
1846          * We don't log our changes till later.
1847          */
1848         if ((uip = XFS_QI_UQIP(mp))) {
1849                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1850                         goto error_return;
1851                 flags |= XFS_UQUOTA_CHKD;
1852         }
1853
1854         if ((gip = XFS_QI_GQIP(mp))) {
1855                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1856                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1857                         goto error_return;
1858                 flags |= XFS_OQUOTA_CHKD;
1859         }
1860
1861         do {
1862                 /*
1863                  * Iterate thru all the inodes in the file system,
1864                  * adjusting the corresponding dquot counters in core.
1865                  */
1866                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1867                                      xfs_qm_dqusage_adjust, NULL,
1868                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1869                         break;
1870
1871         } while (! done);
1872
1873         /*
1874          * We've made all the changes that we need to make incore.
1875          * Flush them down to disk buffers if everything was updated
1876          * successfully.
1877          */
1878         if (!error)
1879                 error = xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1880
1881         /*
1882          * We can get this error if we couldn't do a dquot allocation inside
1883          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1884          * dirty dquots that might be cached, we just want to get rid of them
1885          * and turn quotaoff. The dquots won't be attached to any of the inodes
1886          * at this point (because we intentionally didn't in dqget_noattach).
1887          */
1888         if (error) {
1889                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1890                 goto error_return;
1891         }
1892
1893         /*
1894          * We didn't log anything, because if we crashed, we'll have to
1895          * start the quotacheck from scratch anyway. However, we must make
1896          * sure that our dquot changes are secure before we put the
1897          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1898          * flush.
1899          */
1900         XFS_bflush(mp->m_ddev_targp);
1901
1902         /*
1903          * If one type of quotas is off, then it will lose its
1904          * quotachecked status, since we won't be doing accounting for
1905          * that type anymore.
1906          */
1907         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1908         mp->m_qflags |= flags;
1909
1910         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1911
1912  error_return:
1913         if (error) {
1914                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1915                         "Disabling quotas.",
1916                         mp->m_fsname, error);
1917                 /*
1918                  * We must turn off quotas.
1919                  */
1920                 ASSERT(mp->m_quotainfo != NULL);
1921                 ASSERT(xfs_Gqm != NULL);
1922                 xfs_qm_destroy_quotainfo(mp);
1923                 if (xfs_mount_reset_sbqflags(mp)) {
1924                         cmn_err(CE_WARN, "XFS quotacheck %s: "
1925                                 "Failed to reset quota flags.", mp->m_fsname);
1926                 }
1927         } else {
1928                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1929         }
1930         return (error);
1931 }
1932
1933 /*
1934  * This is called after the superblock has been read in and we're ready to
1935  * iget the quota inodes.
1936  */
1937 STATIC int
1938 xfs_qm_init_quotainos(
1939         xfs_mount_t     *mp)
1940 {
1941         xfs_inode_t     *uip, *gip;
1942         int             error;
1943         __int64_t       sbflags;
1944         uint            flags;
1945
1946         ASSERT(mp->m_quotainfo);
1947         uip = gip = NULL;
1948         sbflags = 0;
1949         flags = 0;
1950
1951         /*
1952          * Get the uquota and gquota inodes
1953          */
1954         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1955                 if (XFS_IS_UQUOTA_ON(mp) &&
1956                     mp->m_sb.sb_uquotino != NULLFSINO) {
1957                         ASSERT(mp->m_sb.sb_uquotino > 0);
1958                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1959                                              0, 0, &uip, 0)))
1960                                 return XFS_ERROR(error);
1961                 }
1962                 if (XFS_IS_OQUOTA_ON(mp) &&
1963                     mp->m_sb.sb_gquotino != NULLFSINO) {
1964                         ASSERT(mp->m_sb.sb_gquotino > 0);
1965                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1966                                              0, 0, &gip, 0))) {
1967                                 if (uip)
1968                                         IRELE(uip);
1969                                 return XFS_ERROR(error);
1970                         }
1971                 }
1972         } else {
1973                 flags |= XFS_QMOPT_SBVERSION;
1974                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1975                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1976         }
1977
1978         /*
1979          * Create the two inodes, if they don't exist already. The changes
1980          * made above will get added to a transaction and logged in one of
1981          * the qino_alloc calls below.  If the device is readonly,
1982          * temporarily switch to read-write to do this.
1983          */
1984         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1985                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1986                                               sbflags | XFS_SB_UQUOTINO,
1987                                               flags | XFS_QMOPT_UQUOTA)))
1988                         return XFS_ERROR(error);
1989
1990                 flags &= ~XFS_QMOPT_SBVERSION;
1991         }
1992         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1993                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1994                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1995                 error = xfs_qm_qino_alloc(mp, &gip,
1996                                           sbflags | XFS_SB_GQUOTINO, flags);
1997                 if (error) {
1998                         if (uip)
1999                                 IRELE(uip);
2000
2001                         return XFS_ERROR(error);
2002                 }
2003         }
2004
2005         XFS_QI_UQIP(mp) = uip;
2006         XFS_QI_GQIP(mp) = gip;
2007
2008         return 0;
2009 }
2010
2011
2012 /*
2013  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2014  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2015  * favor the lookup function ...
2016  * XXXsup merge this with qm_reclaim_one().
2017  */
2018 STATIC int
2019 xfs_qm_shake_freelist(
2020         int howmany)
2021 {
2022         int             nreclaimed;
2023         xfs_dqhash_t    *hash;
2024         xfs_dquot_t     *dqp, *nextdqp;
2025         int             restarts;
2026         int             nflushes;
2027
2028         if (howmany <= 0)
2029                 return 0;
2030
2031         nreclaimed = 0;
2032         restarts = 0;
2033         nflushes = 0;
2034
2035 #ifdef QUOTADEBUG
2036         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2037 #endif
2038         /* lock order is : hashchainlock, freelistlock, mplistlock */
2039  tryagain:
2040         xfs_qm_freelist_lock(xfs_Gqm);
2041
2042         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2043              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2044               nreclaimed < howmany); ) {
2045                 xfs_dqlock(dqp);
2046
2047                 /*
2048                  * We are racing with dqlookup here. Naturally we don't
2049                  * want to reclaim a dquot that lookup wants.
2050                  */
2051                 if (dqp->dq_flags & XFS_DQ_WANT) {
2052                         xfs_dqunlock(dqp);
2053                         xfs_qm_freelist_unlock(xfs_Gqm);
2054                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2055                                 return nreclaimed;
2056                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2057                         goto tryagain;
2058                 }
2059
2060                 /*
2061                  * If the dquot is inactive, we are assured that it is
2062                  * not on the mplist or the hashlist, and that makes our
2063                  * life easier.
2064                  */
2065                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2066                         ASSERT(dqp->q_mount == NULL);
2067                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2068                         ASSERT(dqp->HL_PREVP == NULL);
2069                         ASSERT(dqp->MPL_PREVP == NULL);
2070                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2071                         nextdqp = dqp->dq_flnext;
2072                         goto off_freelist;
2073                 }
2074
2075                 ASSERT(dqp->MPL_PREVP);
2076                 /*
2077                  * Try to grab the flush lock. If this dquot is in the process of
2078                  * getting flushed to disk, we don't want to reclaim it.
2079                  */
2080                 if (!xfs_dqflock_nowait(dqp)) {
2081                         xfs_dqunlock(dqp);
2082                         dqp = dqp->dq_flnext;
2083                         continue;
2084                 }
2085
2086                 /*
2087                  * We have the flush lock so we know that this is not in the
2088                  * process of being flushed. So, if this is dirty, flush it
2089                  * DELWRI so that we don't get a freelist infested with
2090                  * dirty dquots.
2091                  */
2092                 if (XFS_DQ_IS_DIRTY(dqp)) {
2093                         int     error;
2094                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2095                         /*
2096                          * We flush it delayed write, so don't bother
2097                          * releasing the mplock.
2098                          */
2099                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2100                         if (error) {
2101                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2102                         "xfs_qm_dqflush_all: dquot %p flush failed", dqp);
2103                         }
2104                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2105                         dqp = dqp->dq_flnext;
2106                         continue;
2107                 }
2108                 /*
2109                  * We're trying to get the hashlock out of order. This races
2110                  * with dqlookup; so, we giveup and goto the next dquot if
2111                  * we couldn't get the hashlock. This way, we won't starve
2112                  * a dqlookup process that holds the hashlock that is
2113                  * waiting for the freelist lock.
2114                  */
2115                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2116                         xfs_dqfunlock(dqp);
2117                         xfs_dqunlock(dqp);
2118                         dqp = dqp->dq_flnext;
2119                         continue;
2120                 }
2121                 /*
2122                  * This races with dquot allocation code as well as dqflush_all
2123                  * and reclaim code. So, if we failed to grab the mplist lock,
2124                  * giveup everything and start over.
2125                  */
2126                 hash = dqp->q_hash;
2127                 ASSERT(hash);
2128                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2129                         /* XXX put a sentinel so that we can come back here */
2130                         xfs_dqfunlock(dqp);
2131                         xfs_dqunlock(dqp);
2132                         XFS_DQ_HASH_UNLOCK(hash);
2133                         xfs_qm_freelist_unlock(xfs_Gqm);
2134                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2135                                 return nreclaimed;
2136                         goto tryagain;
2137                 }
2138                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2139 #ifdef QUOTADEBUG
2140                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2141                         dqp, be32_to_cpu(dqp->q_core.d_id));
2142 #endif
2143                 ASSERT(dqp->q_nrefs == 0);
2144                 nextdqp = dqp->dq_flnext;
2145                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2146                 XQM_HASHLIST_REMOVE(hash, dqp);
2147                 xfs_dqfunlock(dqp);
2148                 xfs_qm_mplist_unlock(dqp->q_mount);
2149                 XFS_DQ_HASH_UNLOCK(hash);
2150
2151  off_freelist:
2152                 XQM_FREELIST_REMOVE(dqp);
2153                 xfs_dqunlock(dqp);
2154                 nreclaimed++;
2155                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2156                 xfs_qm_dqdestroy(dqp);
2157                 dqp = nextdqp;
2158         }
2159         xfs_qm_freelist_unlock(xfs_Gqm);
2160         return nreclaimed;
2161 }
2162
2163
2164 /*
2165  * The kmem_shake interface is invoked when memory is running low.
2166  */
2167 /* ARGSUSED */
2168 STATIC int
2169 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2170 {
2171         int     ndqused, nfree, n;
2172
2173         if (!kmem_shake_allow(gfp_mask))
2174                 return 0;
2175         if (!xfs_Gqm)
2176                 return 0;
2177
2178         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2179         /* incore dquots in all f/s's */
2180         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2181
2182         ASSERT(ndqused >= 0);
2183
2184         if (nfree <= ndqused && nfree < ndquot)
2185                 return 0;
2186
2187         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2188         n = nfree - ndqused - ndquot;           /* # over target */
2189
2190         return xfs_qm_shake_freelist(MAX(nfree, n));
2191 }
2192
2193
2194 /*
2195  * Just pop the least recently used dquot off the freelist and
2196  * recycle it. The returned dquot is locked.
2197  */
2198 STATIC xfs_dquot_t *
2199 xfs_qm_dqreclaim_one(void)
2200 {
2201         xfs_dquot_t     *dqpout;
2202         xfs_dquot_t     *dqp;
2203         int             restarts;
2204         int             nflushes;
2205
2206         restarts = 0;
2207         dqpout = NULL;
2208         nflushes = 0;
2209
2210         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2211  startagain:
2212         xfs_qm_freelist_lock(xfs_Gqm);
2213
2214         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2215                 xfs_dqlock(dqp);
2216
2217                 /*
2218                  * We are racing with dqlookup here. Naturally we don't
2219                  * want to reclaim a dquot that lookup wants. We release the
2220                  * freelist lock and start over, so that lookup will grab
2221                  * both the dquot and the freelistlock.
2222                  */
2223                 if (dqp->dq_flags & XFS_DQ_WANT) {
2224                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2225                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2226                         xfs_dqunlock(dqp);
2227                         xfs_qm_freelist_unlock(xfs_Gqm);
2228                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2229                                 return NULL;
2230                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2231                         goto startagain;
2232                 }
2233
2234                 /*
2235                  * If the dquot is inactive, we are assured that it is
2236                  * not on the mplist or the hashlist, and that makes our
2237                  * life easier.
2238                  */
2239                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2240                         ASSERT(dqp->q_mount == NULL);
2241                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2242                         ASSERT(dqp->HL_PREVP == NULL);
2243                         ASSERT(dqp->MPL_PREVP == NULL);
2244                         XQM_FREELIST_REMOVE(dqp);
2245                         xfs_dqunlock(dqp);
2246                         dqpout = dqp;
2247                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2248                         break;
2249                 }
2250
2251                 ASSERT(dqp->q_hash);
2252                 ASSERT(dqp->MPL_PREVP);
2253
2254                 /*
2255                  * Try to grab the flush lock. If this dquot is in the process of
2256                  * getting flushed to disk, we don't want to reclaim it.
2257                  */
2258                 if (!xfs_dqflock_nowait(dqp)) {
2259                         xfs_dqunlock(dqp);
2260                         continue;
2261                 }
2262
2263                 /*
2264                  * We have the flush lock so we know that this is not in the
2265                  * process of being flushed. So, if this is dirty, flush it
2266                  * DELWRI so that we don't get a freelist infested with
2267                  * dirty dquots.
2268                  */
2269                 if (XFS_DQ_IS_DIRTY(dqp)) {
2270                         int     error;
2271                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2272                         /*
2273                          * We flush it delayed write, so don't bother
2274                          * releasing the freelist lock.
2275                          */
2276                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2277                         if (error) {
2278                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2279                         "xfs_qm_dqreclaim: dquot %p flush failed", dqp);
2280                         }
2281                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2282                         continue;
2283                 }
2284
2285                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2286                         xfs_dqfunlock(dqp);
2287                         xfs_dqunlock(dqp);
2288                         continue;
2289                 }
2290
2291                 if (! xfs_qm_dqhashlock_nowait(dqp))
2292                         goto mplistunlock;
2293
2294                 ASSERT(dqp->q_nrefs == 0);
2295                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2296                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2297                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2298                 XQM_FREELIST_REMOVE(dqp);
2299                 dqpout = dqp;
2300                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2301  mplistunlock:
2302                 xfs_qm_mplist_unlock(dqp->q_mount);
2303                 xfs_dqfunlock(dqp);
2304                 xfs_dqunlock(dqp);
2305                 if (dqpout)
2306                         break;
2307         }
2308
2309         xfs_qm_freelist_unlock(xfs_Gqm);
2310         return dqpout;
2311 }
2312
2313
2314 /*------------------------------------------------------------------*/
2315
2316 /*
2317  * Return a new incore dquot. Depending on the number of
2318  * dquots in the system, we either allocate a new one on the kernel heap,
2319  * or reclaim a free one.
2320  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2321  * to reclaim an existing one from the freelist.
2322  */
2323 boolean_t
2324 xfs_qm_dqalloc_incore(
2325         xfs_dquot_t **O_dqpp)
2326 {
2327         xfs_dquot_t     *dqp;
2328
2329         /*
2330          * Check against high water mark to see if we want to pop
2331          * a nincompoop dquot off the freelist.
2332          */
2333         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2334                 /*
2335                  * Try to recycle a dquot from the freelist.
2336                  */
2337                 if ((dqp = xfs_qm_dqreclaim_one())) {
2338                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2339                         /*
2340                          * Just zero the core here. The rest will get
2341                          * reinitialized by caller. XXX we shouldn't even
2342                          * do this zero ...
2343                          */
2344                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2345                         *O_dqpp = dqp;
2346                         return B_FALSE;
2347                 }
2348                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2349         }
2350
2351         /*
2352          * Allocate a brand new dquot on the kernel heap and return it
2353          * to the caller to initialize.
2354          */
2355         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2356         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2357         atomic_inc(&xfs_Gqm->qm_totaldquots);
2358
2359         return B_TRUE;
2360 }
2361
2362
2363 /*
2364  * Start a transaction and write the incore superblock changes to
2365  * disk. flags parameter indicates which fields have changed.
2366  */
2367 int
2368 xfs_qm_write_sb_changes(
2369         xfs_mount_t     *mp,
2370         __int64_t       flags)
2371 {
2372         xfs_trans_t     *tp;
2373         int             error;
2374
2375 #ifdef QUOTADEBUG
2376         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2377 #endif
2378         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2379         if ((error = xfs_trans_reserve(tp, 0,
2380                                       mp->m_sb.sb_sectsize + 128, 0,
2381                                       0,
2382                                       XFS_DEFAULT_LOG_COUNT))) {
2383                 xfs_trans_cancel(tp, 0);
2384                 return error;
2385         }
2386
2387         xfs_mod_sb(tp, flags);
2388         error = xfs_trans_commit(tp, 0);
2389
2390         return error;
2391 }
2392
2393
2394 /* --------------- utility functions for vnodeops ---------------- */
2395
2396
2397 /*
2398  * Given an inode, a uid and gid (from cred_t) make sure that we have
2399  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2400  * quotas by creating this file.
2401  * This also attaches dquot(s) to the given inode after locking it,
2402  * and returns the dquots corresponding to the uid and/or gid.
2403  *
2404  * in   : inode (unlocked)
2405  * out  : udquot, gdquot with references taken and unlocked
2406  */
2407 int
2408 xfs_qm_vop_dqalloc(
2409         xfs_mount_t     *mp,
2410         xfs_inode_t     *ip,
2411         uid_t           uid,
2412         gid_t           gid,
2413         prid_t          prid,
2414         uint            flags,
2415         xfs_dquot_t     **O_udqpp,
2416         xfs_dquot_t     **O_gdqpp)
2417 {
2418         int             error;
2419         xfs_dquot_t     *uq, *gq;
2420         uint            lockflags;
2421
2422         if (!XFS_IS_QUOTA_ON(mp))
2423                 return 0;
2424
2425         lockflags = XFS_ILOCK_EXCL;
2426         xfs_ilock(ip, lockflags);
2427
2428         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2429                 gid = ip->i_d.di_gid;
2430
2431         /*
2432          * Attach the dquot(s) to this inode, doing a dquot allocation
2433          * if necessary. The dquot(s) will not be locked.
2434          */
2435         if (XFS_NOT_DQATTACHED(mp, ip)) {
2436                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2437                                             XFS_QMOPT_ILOCKED))) {
2438                         xfs_iunlock(ip, lockflags);
2439                         return error;
2440                 }
2441         }
2442
2443         uq = gq = NULL;
2444         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2445                 if (ip->i_d.di_uid != uid) {
2446                         /*
2447                          * What we need is the dquot that has this uid, and
2448                          * if we send the inode to dqget, the uid of the inode
2449                          * takes priority over what's sent in the uid argument.
2450                          * We must unlock inode here before calling dqget if
2451                          * we're not sending the inode, because otherwise
2452                          * we'll deadlock by doing trans_reserve while
2453                          * holding ilock.
2454                          */
2455                         xfs_iunlock(ip, lockflags);
2456                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2457                                                  XFS_DQ_USER,
2458                                                  XFS_QMOPT_DQALLOC |
2459                                                  XFS_QMOPT_DOWARN,
2460                                                  &uq))) {
2461                                 ASSERT(error != ENOENT);
2462                                 return error;
2463                         }
2464                         /*
2465                          * Get the ilock in the right order.
2466                          */
2467                         xfs_dqunlock(uq);
2468                         lockflags = XFS_ILOCK_SHARED;
2469                         xfs_ilock(ip, lockflags);
2470                 } else {
2471                         /*
2472                          * Take an extra reference, because we'll return
2473                          * this to caller
2474                          */
2475                         ASSERT(ip->i_udquot);
2476                         uq = ip->i_udquot;
2477                         xfs_dqlock(uq);
2478                         XFS_DQHOLD(uq);
2479                         xfs_dqunlock(uq);
2480                 }
2481         }
2482         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2483                 if (ip->i_d.di_gid != gid) {
2484                         xfs_iunlock(ip, lockflags);
2485                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2486                                                  XFS_DQ_GROUP,
2487                                                  XFS_QMOPT_DQALLOC |
2488                                                  XFS_QMOPT_DOWARN,
2489                                                  &gq))) {
2490                                 if (uq)
2491                                         xfs_qm_dqrele(uq);
2492                                 ASSERT(error != ENOENT);
2493                                 return error;
2494                         }
2495                         xfs_dqunlock(gq);
2496                         lockflags = XFS_ILOCK_SHARED;
2497                         xfs_ilock(ip, lockflags);
2498                 } else {
2499                         ASSERT(ip->i_gdquot);
2500                         gq = ip->i_gdquot;
2501                         xfs_dqlock(gq);
2502                         XFS_DQHOLD(gq);
2503                         xfs_dqunlock(gq);
2504                 }
2505         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2506                 if (ip->i_d.di_projid != prid) {
2507                         xfs_iunlock(ip, lockflags);
2508                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2509                                                  XFS_DQ_PROJ,
2510                                                  XFS_QMOPT_DQALLOC |
2511                                                  XFS_QMOPT_DOWARN,
2512                                                  &gq))) {
2513                                 if (uq)
2514                                         xfs_qm_dqrele(uq);
2515                                 ASSERT(error != ENOENT);
2516                                 return (error);
2517                         }
2518                         xfs_dqunlock(gq);
2519                         lockflags = XFS_ILOCK_SHARED;
2520                         xfs_ilock(ip, lockflags);
2521                 } else {
2522                         ASSERT(ip->i_gdquot);
2523                         gq = ip->i_gdquot;
2524                         xfs_dqlock(gq);
2525                         XFS_DQHOLD(gq);
2526                         xfs_dqunlock(gq);
2527                 }
2528         }
2529         if (uq)
2530                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2531
2532         xfs_iunlock(ip, lockflags);
2533         if (O_udqpp)
2534                 *O_udqpp = uq;
2535         else if (uq)
2536                 xfs_qm_dqrele(uq);
2537         if (O_gdqpp)
2538                 *O_gdqpp = gq;
2539         else if (gq)
2540                 xfs_qm_dqrele(gq);
2541         return 0;
2542 }
2543
2544 /*
2545  * Actually transfer ownership, and do dquot modifications.
2546  * These were already reserved.
2547  */
2548 xfs_dquot_t *
2549 xfs_qm_vop_chown(
2550         xfs_trans_t     *tp,
2551         xfs_inode_t     *ip,
2552         xfs_dquot_t     **IO_olddq,
2553         xfs_dquot_t     *newdq)
2554 {
2555         xfs_dquot_t     *prevdq;
2556         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2557                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2558
2559         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2560         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2561
2562         /* old dquot */
2563         prevdq = *IO_olddq;
2564         ASSERT(prevdq);
2565         ASSERT(prevdq != newdq);
2566
2567         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2568         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2569
2570         /* the sparkling new dquot */
2571         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2572         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2573
2574         /*
2575          * Take an extra reference, because the inode
2576          * is going to keep this dquot pointer even
2577          * after the trans_commit.
2578          */
2579         xfs_dqlock(newdq);
2580         XFS_DQHOLD(newdq);
2581         xfs_dqunlock(newdq);
2582         *IO_olddq = newdq;
2583
2584         return prevdq;
2585 }
2586
2587 /*
2588  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2589  */
2590 int
2591 xfs_qm_vop_chown_reserve(
2592         xfs_trans_t     *tp,
2593         xfs_inode_t     *ip,
2594         xfs_dquot_t     *udqp,
2595         xfs_dquot_t     *gdqp,
2596         uint            flags)
2597 {
2598         int             error;
2599         xfs_mount_t     *mp;
2600         uint            delblks, blkflags, prjflags = 0;
2601         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2602
2603         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2604         mp = ip->i_mount;
2605         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2606
2607         delblks = ip->i_delayed_blks;
2608         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2609         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2610                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2611
2612         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2613             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2614                 delblksudq = udqp;
2615                 /*
2616                  * If there are delayed allocation blocks, then we have to
2617                  * unreserve those from the old dquot, and add them to the
2618                  * new dquot.
2619                  */
2620                 if (delblks) {
2621                         ASSERT(ip->i_udquot);
2622                         unresudq = ip->i_udquot;
2623                 }
2624         }
2625         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2626                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2627                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2628                         prjflags = XFS_QMOPT_ENOSPC;
2629
2630                 if (prjflags ||
2631                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2632                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2633                         delblksgdq = gdqp;
2634                         if (delblks) {
2635                                 ASSERT(ip->i_gdquot);
2636                                 unresgdq = ip->i_gdquot;
2637                         }
2638                 }
2639         }
2640
2641         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2642                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2643                                 flags | blkflags | prjflags)))
2644                 return (error);
2645
2646         /*
2647          * Do the delayed blks reservations/unreservations now. Since, these
2648          * are done without the help of a transaction, if a reservation fails
2649          * its previous reservations won't be automatically undone by trans
2650          * code. So, we have to do it manually here.
2651          */
2652         if (delblks) {
2653                 /*
2654                  * Do the reservations first. Unreservation can't fail.
2655                  */
2656                 ASSERT(delblksudq || delblksgdq);
2657                 ASSERT(unresudq || unresgdq);
2658                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2659                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2660                                 flags | blkflags | prjflags)))
2661                         return (error);
2662                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2663                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2664                                 blkflags);
2665         }
2666
2667         return (0);
2668 }
2669
2670 int
2671 xfs_qm_vop_rename_dqattach(
2672         xfs_inode_t     **i_tab)
2673 {
2674         xfs_inode_t     *ip;
2675         int             i;
2676         int             error;
2677
2678         ip = i_tab[0];
2679
2680         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2681                 return 0;
2682
2683         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2684                 error = xfs_qm_dqattach(ip, 0);
2685                 if (error)
2686                         return error;
2687         }
2688         for (i = 1; (i < 4 && i_tab[i]); i++) {
2689                 /*
2690                  * Watch out for duplicate entries in the table.
2691                  */
2692                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2693                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2694                                 error = xfs_qm_dqattach(ip, 0);
2695                                 if (error)
2696                                         return error;
2697                         }
2698                 }
2699         }
2700         return 0;
2701 }
2702
2703 void
2704 xfs_qm_vop_dqattach_and_dqmod_newinode(
2705         xfs_trans_t     *tp,
2706         xfs_inode_t     *ip,
2707         xfs_dquot_t     *udqp,
2708         xfs_dquot_t     *gdqp)
2709 {
2710         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2711                 return;
2712
2713         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2714         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2715
2716         if (udqp) {
2717                 xfs_dqlock(udqp);
2718                 XFS_DQHOLD(udqp);
2719                 xfs_dqunlock(udqp);
2720                 ASSERT(ip->i_udquot == NULL);
2721                 ip->i_udquot = udqp;
2722                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2723                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2724                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2725         }
2726         if (gdqp) {
2727                 xfs_dqlock(gdqp);
2728                 XFS_DQHOLD(gdqp);
2729                 xfs_dqunlock(gdqp);
2730                 ASSERT(ip->i_gdquot == NULL);
2731                 ip->i_gdquot = gdqp;
2732                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2733                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2734                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2735                                 be32_to_cpu(gdqp->q_core.d_id));
2736                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2737         }
2738 }
2739
2740 /* ------------- list stuff -----------------*/
2741 STATIC void
2742 xfs_qm_freelist_init(xfs_frlist_t *ql)
2743 {
2744         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2745         mutex_init(&ql->qh_lock);
2746         ql->qh_version = 0;
2747         ql->qh_nelems = 0;
2748 }
2749
2750 STATIC void
2751 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2752 {
2753         xfs_dquot_t     *dqp, *nextdqp;
2754
2755         mutex_lock(&ql->qh_lock);
2756         for (dqp = ql->qh_next;
2757              dqp != (xfs_dquot_t *)ql; ) {
2758                 xfs_dqlock(dqp);
2759                 nextdqp = dqp->dq_flnext;
2760 #ifdef QUOTADEBUG
2761                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2762 #endif
2763                 XQM_FREELIST_REMOVE(dqp);
2764                 xfs_dqunlock(dqp);
2765                 xfs_qm_dqdestroy(dqp);
2766                 dqp = nextdqp;
2767         }
2768         mutex_unlock(&ql->qh_lock);
2769         mutex_destroy(&ql->qh_lock);
2770
2771         ASSERT(ql->qh_nelems == 0);
2772 }
2773
2774 STATIC void
2775 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2776 {
2777         dq->dq_flnext = ql->qh_next;
2778         dq->dq_flprev = (xfs_dquot_t *)ql;
2779         ql->qh_next = dq;
2780         dq->dq_flnext->dq_flprev = dq;
2781         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2782         xfs_Gqm->qm_dqfreelist.qh_version++;
2783 }
2784
2785 void
2786 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2787 {
2788         xfs_dquot_t *next = dq->dq_flnext;
2789         xfs_dquot_t *prev = dq->dq_flprev;
2790
2791         next->dq_flprev = prev;
2792         prev->dq_flnext = next;
2793         dq->dq_flnext = dq->dq_flprev = dq;
2794         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2795         xfs_Gqm->qm_dqfreelist.qh_version++;
2796 }
2797
2798 void
2799 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2800 {
2801         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2802 }
2803
2804 STATIC int
2805 xfs_qm_dqhashlock_nowait(
2806         xfs_dquot_t *dqp)
2807 {
2808         int locked;
2809
2810         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2811         return locked;
2812 }
2813
2814 int
2815 xfs_qm_freelist_lock_nowait(
2816         xfs_qm_t *xqm)
2817 {
2818         int locked;
2819
2820         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2821         return locked;
2822 }
2823
2824 STATIC int
2825 xfs_qm_mplist_nowait(
2826         xfs_mount_t     *mp)
2827 {
2828         int locked;
2829
2830         ASSERT(mp->m_quotainfo);
2831         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2832         return locked;
2833 }