fs/xfs/quota/xfs_qm.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_alloc.h"
  27 #include "xfs_quota.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_ialloc_btree.h"
  31 #include "xfs_dinode.h"
  32 #include "xfs_inode.h"
  33 #include "xfs_ialloc.h"
  34 #include "xfs_itable.h"
  35 #include "xfs_rtalloc.h"
  36 #include "xfs_error.h"
  37 #include "xfs_bmap.h"
  38 #include "xfs_attr.h"
  39 #include "xfs_buf_item.h"
  40 #include "xfs_trans_space.h"
  41 #include "xfs_utils.h"
  42 #include "xfs_qm.h"
  43 #include "xfs_trace.h"
  44
  45 /*
  46  * The global quota manager. There is only one of these for the entire
  47  * system, _not_ one per file system. XQM keeps track of the overall
  48  * quota functionality, including maintaining the freelist and hash
  49  * tables of dquots.
  50  */
  51 struct mutex    xfs_Gqm_lock;
  52 struct xfs_qm   *xfs_Gqm;
  53 uint            ndquot;
  54
  55 kmem_zone_t     *qm_dqzone;
  56 kmem_zone_t     *qm_dqtrxzone;
  57
  58 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
  59 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
  60
  61 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
  62 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
  63 STATIC int      xfs_qm_shake(struct shrinker *, struct shrink_control *);
  64
  65 static struct shrinker xfs_qm_shaker = {
  66         .shrink = xfs_qm_shake,
  67         .seeks = DEFAULT_SEEKS,
  68 };
  69
  70 /*
  71  * Initialize the XQM structure.
  72  * Note that there is not one quota manager per file system.
  73  */
  74 STATIC struct xfs_qm *
  75 xfs_Gqm_init(void)
  76 {
  77         xfs_dqhash_t    *udqhash, *gdqhash;
  78         xfs_qm_t        *xqm;
  79         size_t          hsize;
  80         uint            i;
  81
  82         /*
  83          * Initialize the dquot hash tables.
  84          */
  85         udqhash = kmem_zalloc_greedy(&hsize,
  86                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
  87                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t));
  88         if (!udqhash)
  89                 goto out;
  90
  91         gdqhash = kmem_zalloc_large(hsize);
  92         if (!gdqhash)
  93                 goto out_free_udqhash;
  94
  95         hsize /= sizeof(xfs_dqhash_t);
  96         ndquot = hsize << 8;
  97
  98         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
  99         xqm->qm_dqhashmask = hsize - 1;
 100         xqm->qm_usr_dqhtable = udqhash;
 101         xqm->qm_grp_dqhtable = gdqhash;
 102         ASSERT(xqm->qm_usr_dqhtable != NULL);
 103         ASSERT(xqm->qm_grp_dqhtable != NULL);
 104
 105         for (i = 0; i < hsize; i++) {
 106                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
 107                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
 108         }
 109
 110         /*
 111          * Freelist of all dquots of all file systems
 112          */
 113         INIT_LIST_HEAD(&xqm->qm_dqfrlist);
 114         xqm->qm_dqfrlist_cnt = 0;
 115         mutex_init(&xqm->qm_dqfrlist_lock);
 116
 117         /*
 118          * dquot zone. we register our own low-memory callback.
 119          */
 120         if (!qm_dqzone) {
 121                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
 122                                                 "xfs_dquots");
 123                 qm_dqzone = xqm->qm_dqzone;
 124         } else
 125                 xqm->qm_dqzone = qm_dqzone;
 126
 127         register_shrinker(&xfs_qm_shaker);
 128
 129         /*
 130          * The t_dqinfo portion of transactions.
 131          */
 132         if (!qm_dqtrxzone) {
 133                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
 134                                                    "xfs_dqtrx");
 135                 qm_dqtrxzone = xqm->qm_dqtrxzone;
 136         } else
 137                 xqm->qm_dqtrxzone = qm_dqtrxzone;
 138
 139         atomic_set(&xqm->qm_totaldquots, 0);
 140         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
 141         xqm->qm_nrefs = 0;
 142         return xqm;
 143
 144  out_free_udqhash:
 145         kmem_free_large(udqhash);
 146  out:
 147         return NULL;
 148 }
 149
 150 /*
 151  * Destroy the global quota manager when its reference count goes to zero.
 152  */
 153 STATIC void
 154 xfs_qm_destroy(
 155         struct xfs_qm   *xqm)
 156 {
 157         struct xfs_dquot *dqp, *n;
 158         int             hsize, i;
 159
 160         ASSERT(xqm != NULL);
 161         ASSERT(xqm->qm_nrefs == 0);
 162         unregister_shrinker(&xfs_qm_shaker);
 163         hsize = xqm->qm_dqhashmask + 1;
 164         for (i = 0; i < hsize; i++) {
 165                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
 166                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
 167         }
 168         kmem_free_large(xqm->qm_usr_dqhtable);
 169         kmem_free_large(xqm->qm_grp_dqhtable);
 170         xqm->qm_usr_dqhtable = NULL;
 171         xqm->qm_grp_dqhtable = NULL;
 172         xqm->qm_dqhashmask = 0;
 173
 174         /* frlist cleanup */
 175         mutex_lock(&xqm->qm_dqfrlist_lock);
 176         list_for_each_entry_safe(dqp, n, &xqm->qm_dqfrlist, q_freelist) {
 177                 xfs_dqlock(dqp);
 178                 list_del_init(&dqp->q_freelist);
 179                 xfs_Gqm->qm_dqfrlist_cnt--;
 180                 xfs_dqunlock(dqp);
 181                 xfs_qm_dqdestroy(dqp);
 182         }
 183         mutex_unlock(&xqm->qm_dqfrlist_lock);
 184         mutex_destroy(&xqm->qm_dqfrlist_lock);
 185         kmem_free(xqm);
 186 }
 187
 188 /*
 189  * Called at mount time to let XQM know that another file system is
 190  * starting quotas. This isn't crucial information as the individual mount
 191  * structures are pretty independent, but it helps the XQM keep a
 192  * global view of what's going on.
 193  */
 194 /* ARGSUSED */
 195 STATIC int
 196 xfs_qm_hold_quotafs_ref(
 197         struct xfs_mount *mp)
 198 {
 199         /*
 200          * Need to lock the xfs_Gqm structure for things like this. For example,
 201          * the structure could disappear between the entry to this routine and
 202          * a HOLD operation if not locked.
 203          */
 204         mutex_lock(&xfs_Gqm_lock);
 205
 206         if (!xfs_Gqm) {
 207                 xfs_Gqm = xfs_Gqm_init();
 208                 if (!xfs_Gqm) {
 209                         mutex_unlock(&xfs_Gqm_lock);
 210                         return ENOMEM;
 211                 }
 212         }
 213
 214         /*
 215          * We can keep a list of all filesystems with quotas mounted for
 216          * debugging and statistical purposes, but ...
 217          * Just take a reference and get out.
 218          */
 219         xfs_Gqm->qm_nrefs++;
 220         mutex_unlock(&xfs_Gqm_lock);
 221
 222         return 0;
 223 }
 224
 225
 226 /*
 227  * Release the reference that a filesystem took at mount time,
 228  * so that we know when we need to destroy the entire quota manager.
 229  */
 230 /* ARGSUSED */
 231 STATIC void
 232 xfs_qm_rele_quotafs_ref(
 233         struct xfs_mount *mp)
 234 {
 235         xfs_dquot_t     *dqp, *n;
 236
 237         ASSERT(xfs_Gqm);
 238         ASSERT(xfs_Gqm->qm_nrefs > 0);
 239
 240         /*
 241          * Go thru the freelist and destroy all inactive dquots.
 242          */
 243         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
 244
 245         list_for_each_entry_safe(dqp, n, &xfs_Gqm->qm_dqfrlist, q_freelist) {
 246                 xfs_dqlock(dqp);
 247                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
 248                         ASSERT(dqp->q_mount == NULL);
 249                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
 250                         ASSERT(list_empty(&dqp->q_hashlist));
 251                         ASSERT(list_empty(&dqp->q_mplist));
 252                         list_del_init(&dqp->q_freelist);
 253                         xfs_Gqm->qm_dqfrlist_cnt--;
 254                         xfs_dqunlock(dqp);
 255                         xfs_qm_dqdestroy(dqp);
 256                 } else {
 257                         xfs_dqunlock(dqp);
 258                 }
 259         }
 260         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 261
 262         /*
 263          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
 264          * be restarted.
 265          */
 266         mutex_lock(&xfs_Gqm_lock);
 267         if (--xfs_Gqm->qm_nrefs == 0) {
 268                 xfs_qm_destroy(xfs_Gqm);
 269                 xfs_Gqm = NULL;
 270         }
 271         mutex_unlock(&xfs_Gqm_lock);
 272 }
 273
 274 /*
 275  * Just destroy the quotainfo structure.
 276  */
 277 void
 278 xfs_qm_unmount(
 279         struct xfs_mount        *mp)
 280 {
 281         if (mp->m_quotainfo) {
 282                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL);
 283                 xfs_qm_destroy_quotainfo(mp);
 284         }
 285 }
 286
 287
 288 /*
 289  * This is called from xfs_mountfs to start quotas and initialize all
 290  * necessary data structures like quotainfo.  This is also responsible for
 291  * running a quotacheck as necessary.  We are guaranteed that the superblock
 292  * is consistently read in at this point.
 293  *
 294  * If we fail here, the mount will continue with quota turned off. We don't
 295  * need to inidicate success or failure at all.
 296  */
 297 void
 298 xfs_qm_mount_quotas(
 299         xfs_mount_t     *mp)
 300 {
 301         int             error = 0;
 302         uint            sbf;
 303
 304         /*
 305          * If quotas on realtime volumes is not supported, we disable
 306          * quotas immediately.
 307          */
 308         if (mp->m_sb.sb_rextents) {
 309                 xfs_notice(mp, "Cannot turn on quotas for realtime filesystem");
 310                 mp->m_qflags = 0;
 311                 goto write_changes;
 312         }
 313
 314         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 315
 316         /*
 317          * Allocate the quotainfo structure inside the mount struct, and
 318          * create quotainode(s), and change/rev superblock if necessary.
 319          */
 320         error = xfs_qm_init_quotainfo(mp);
 321         if (error) {
 322                 /*
 323                  * We must turn off quotas.
 324                  */
 325                 ASSERT(mp->m_quotainfo == NULL);
 326                 mp->m_qflags = 0;
 327                 goto write_changes;
 328         }
 329         /*
 330          * If any of the quotas are not consistent, do a quotacheck.
 331          */
 332         if (XFS_QM_NEED_QUOTACHECK(mp)) {
 333                 error = xfs_qm_quotacheck(mp);
 334                 if (error) {
 335                         /* Quotacheck failed and disabled quotas. */
 336                         return;
 337                 }
 338         }
 339         /*
 340          * If one type of quotas is off, then it will lose its
 341          * quotachecked status, since we won't be doing accounting for
 342          * that type anymore.
 343          */
 344         if (!XFS_IS_UQUOTA_ON(mp))
 345                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
 346         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp)))
 347                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
 348
 349  write_changes:
 350         /*
 351          * We actually don't have to acquire the m_sb_lock at all.
 352          * This can only be called from mount, and that's single threaded. XXX
 353          */
 354         spin_lock(&mp->m_sb_lock);
 355         sbf = mp->m_sb.sb_qflags;
 356         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
 357         spin_unlock(&mp->m_sb_lock);
 358
 359         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
 360                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
 361                         /*
 362                          * We could only have been turning quotas off.
 363                          * We aren't in very good shape actually because
 364                          * the incore structures are convinced that quotas are
 365                          * off, but the on disk superblock doesn't know that !
 366                          */
 367                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
 368                         xfs_alert(mp, "%s: Superblock update failed!",
 369                                 __func__);
 370                 }
 371         }
 372
 373         if (error) {
 374                 xfs_warn(mp, "Failed to initialize disk quotas.");
 375                 return;
 376         }
 377 }
 378
 379 /*
 380  * Called from the vfsops layer.
 381  */
 382 void
 383 xfs_qm_unmount_quotas(
 384         xfs_mount_t     *mp)
 385 {
 386         /*
 387          * Release the dquots that root inode, et al might be holding,
 388          * before we flush quotas and blow away the quotainfo structure.
 389          */
 390         ASSERT(mp->m_rootip);
 391         xfs_qm_dqdetach(mp->m_rootip);
 392         if (mp->m_rbmip)
 393                 xfs_qm_dqdetach(mp->m_rbmip);
 394         if (mp->m_rsumip)
 395                 xfs_qm_dqdetach(mp->m_rsumip);
 396
 397         /*
 398          * Release the quota inodes.
 399          */
 400         if (mp->m_quotainfo) {
 401                 if (mp->m_quotainfo->qi_uquotaip) {
 402                         IRELE(mp->m_quotainfo->qi_uquotaip);
 403                         mp->m_quotainfo->qi_uquotaip = NULL;
 404                 }
 405                 if (mp->m_quotainfo->qi_gquotaip) {
 406                         IRELE(mp->m_quotainfo->qi_gquotaip);
 407                         mp->m_quotainfo->qi_gquotaip = NULL;
 408                 }
 409         }
 410 }
 411
 412 /*
 413  * Flush all dquots of the given file system to disk. The dquots are
 414  * _not_ purged from memory here, just their data written to disk.
 415  */
 416 STATIC int
 417 xfs_qm_dqflush_all(
 418         struct xfs_mount        *mp,
 419         int                     sync_mode)
 420 {
 421         struct xfs_quotainfo    *q = mp->m_quotainfo;
 422         int                     recl;
 423         struct xfs_dquot        *dqp;
 424         int                     error;
 425
 426         if (!q)
 427                 return 0;
 428 again:
 429         mutex_lock(&q->qi_dqlist_lock);
 430         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
 431                 xfs_dqlock(dqp);
 432                 if (! XFS_DQ_IS_DIRTY(dqp)) {
 433                         xfs_dqunlock(dqp);
 434                         continue;
 435                 }
 436
 437                 /* XXX a sentinel would be better */
 438                 recl = q->qi_dqreclaims;
 439                 if (!xfs_dqflock_nowait(dqp)) {
 440                         /*
 441                          * If we can't grab the flush lock then check
 442                          * to see if the dquot has been flushed delayed
 443                          * write.  If so, grab its buffer and send it
 444                          * out immediately.  We'll be able to acquire
 445                          * the flush lock when the I/O completes.
 446                          */
 447                         xfs_qm_dqflock_pushbuf_wait(dqp);
 448                 }
 449                 /*
 450                  * Let go of the mplist lock. We don't want to hold it
 451                  * across a disk write.
 452                  */
 453                 mutex_unlock(&q->qi_dqlist_lock);
 454                 error = xfs_qm_dqflush(dqp, sync_mode);
 455                 xfs_dqunlock(dqp);
 456                 if (error)
 457                         return error;
 458
 459                 mutex_lock(&q->qi_dqlist_lock);
 460                 if (recl != q->qi_dqreclaims) {
 461                         mutex_unlock(&q->qi_dqlist_lock);
 462                         /* XXX restart limit */
 463                         goto again;
 464                 }
 465         }
 466
 467         mutex_unlock(&q->qi_dqlist_lock);
 468         /* return ! busy */
 469         return 0;
 470 }
 471 /*
 472  * Release the group dquot pointers the user dquots may be
 473  * carrying around as a hint. mplist is locked on entry and exit.
 474  */
 475 STATIC void
 476 xfs_qm_detach_gdquots(
 477         struct xfs_mount        *mp)
 478 {
 479         struct xfs_quotainfo    *q = mp->m_quotainfo;
 480         struct xfs_dquot        *dqp, *gdqp;
 481         int                     nrecl;
 482
 483  again:
 484         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
 485         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
 486                 xfs_dqlock(dqp);
 487                 if ((gdqp = dqp->q_gdquot)) {
 488                         xfs_dqlock(gdqp);
 489                         dqp->q_gdquot = NULL;
 490                 }
 491                 xfs_dqunlock(dqp);
 492
 493                 if (gdqp) {
 494                         /*
 495                          * Can't hold the mplist lock across a dqput.
 496                          * XXXmust convert to marker based iterations here.
 497                          */
 498                         nrecl = q->qi_dqreclaims;
 499                         mutex_unlock(&q->qi_dqlist_lock);
 500                         xfs_qm_dqput(gdqp);
 501
 502                         mutex_lock(&q->qi_dqlist_lock);
 503                         if (nrecl != q->qi_dqreclaims)
 504                                 goto again;
 505                 }
 506         }
 507 }
 508
 509 /*
 510  * Go through all the incore dquots of this file system and take them
 511  * off the mplist and hashlist, if the dquot type matches the dqtype
 512  * parameter. This is used when turning off quota accounting for
 513  * users and/or groups, as well as when the filesystem is unmounting.
 514  */
 515 STATIC int
 516 xfs_qm_dqpurge_int(
 517         struct xfs_mount        *mp,
 518         uint                    flags)
 519 {
 520         struct xfs_quotainfo    *q = mp->m_quotainfo;
 521         struct xfs_dquot        *dqp, *n;
 522         uint                    dqtype;
 523         int                     nrecl;
 524         int                     nmisses;
 525
 526         if (!q)
 527                 return 0;
 528
 529         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
 530         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
 531         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
 532
 533         mutex_lock(&q->qi_dqlist_lock);
 534
 535         /*
 536          * In the first pass through all incore dquots of this filesystem,
 537          * we release the group dquot pointers the user dquots may be
 538          * carrying around as a hint. We need to do this irrespective of
 539          * what's being turned off.
 540          */
 541         xfs_qm_detach_gdquots(mp);
 542
 543       again:
 544         nmisses = 0;
 545         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
 546         /*
 547          * Try to get rid of all of the unwanted dquots. The idea is to
 548          * get them off mplist and hashlist, but leave them on freelist.
 549          */
 550         list_for_each_entry_safe(dqp, n, &q->qi_dqlist, q_mplist) {
 551                 /*
 552                  * It's OK to look at the type without taking dqlock here.
 553                  * We're holding the mplist lock here, and that's needed for
 554                  * a dqreclaim.
 555                  */
 556                 if ((dqp->dq_flags & dqtype) == 0)
 557                         continue;
 558
 559                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
 560                         nrecl = q->qi_dqreclaims;
 561                         mutex_unlock(&q->qi_dqlist_lock);
 562                         mutex_lock(&dqp->q_hash->qh_lock);
 563                         mutex_lock(&q->qi_dqlist_lock);
 564
 565                         /*
 566                          * XXXTheoretically, we can get into a very long
 567                          * ping pong game here.
 568                          * No one can be adding dquots to the mplist at
 569                          * this point, but somebody might be taking things off.
 570                          */
 571                         if (nrecl != q->qi_dqreclaims) {
 572                                 mutex_unlock(&dqp->q_hash->qh_lock);
 573                                 goto again;
 574                         }
 575                 }
 576
 577                 /*
 578                  * Take the dquot off the mplist and hashlist. It may remain on
 579                  * freelist in INACTIVE state.
 580                  */
 581                 nmisses += xfs_qm_dqpurge(dqp);
 582         }
 583         mutex_unlock(&q->qi_dqlist_lock);
 584         return nmisses;
 585 }
 586
 587 int
 588 xfs_qm_dqpurge_all(
 589         xfs_mount_t     *mp,
 590         uint            flags)
 591 {
 592         int             ndquots;
 593
 594         /*
 595          * Purge the dquot cache.
 596          * None of the dquots should really be busy at this point.
 597          */
 598         if (mp->m_quotainfo) {
 599                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
 600                         delay(ndquots * 10);
 601                 }
 602         }
 603         return 0;
 604 }
 605
 606 STATIC int
 607 xfs_qm_dqattach_one(
 608         xfs_inode_t     *ip,
 609         xfs_dqid_t      id,
 610         uint            type,
 611         uint            doalloc,
 612         xfs_dquot_t     *udqhint, /* hint */
 613         xfs_dquot_t     **IO_idqpp)
 614 {
 615         xfs_dquot_t     *dqp;
 616         int             error;
 617
 618         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 619         error = 0;
 620
 621         /*
 622          * See if we already have it in the inode itself. IO_idqpp is
 623          * &i_udquot or &i_gdquot. This made the code look weird, but
 624          * made the logic a lot simpler.
 625          */
 626         dqp = *IO_idqpp;
 627         if (dqp) {
 628                 trace_xfs_dqattach_found(dqp);
 629                 return 0;
 630         }
 631
 632         /*
 633          * udqhint is the i_udquot field in inode, and is non-NULL only
 634          * when the type arg is group/project. Its purpose is to save a
 635          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
 636          * the user dquot.
 637          */
 638         if (udqhint) {
 639                 ASSERT(type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
 640                 xfs_dqlock(udqhint);
 641
 642                 /*
 643                  * No need to take dqlock to look at the id.
 644                  *
 645                  * The ID can't change until it gets reclaimed, and it won't
 646                  * be reclaimed as long as we have a ref from inode and we
 647                  * hold the ilock.
 648                  */
 649                 dqp = udqhint->q_gdquot;
 650                 if (dqp && be32_to_cpu(dqp->q_core.d_id) == id) {
 651                         xfs_dqlock(dqp);
 652                         XFS_DQHOLD(dqp);
 653                         ASSERT(*IO_idqpp == NULL);
 654                         *IO_idqpp = dqp;
 655
 656                         xfs_dqunlock(dqp);
 657                         xfs_dqunlock(udqhint);
 658                         return 0;
 659                 }
 660
 661                 /*
 662                  * We can't hold a dquot lock when we call the dqget code.
 663                  * We'll deadlock in no time, because of (not conforming to)
 664                  * lock ordering - the inodelock comes before any dquot lock,
 665                  * and we may drop and reacquire the ilock in xfs_qm_dqget().
 666                  */
 667                 xfs_dqunlock(udqhint);
 668         }
 669
 670         /*
 671          * Find the dquot from somewhere. This bumps the
 672          * reference count of dquot and returns it locked.
 673          * This can return ENOENT if dquot didn't exist on
 674          * disk and we didn't ask it to allocate;
 675          * ESRCH if quotas got turned off suddenly.
 676          */
 677         error = xfs_qm_dqget(ip->i_mount, ip, id, type, XFS_QMOPT_DOWARN, &dqp);
 678         if (error)
 679                 return error;
 680
 681         trace_xfs_dqattach_get(dqp);
 682
 683         /*
 684          * dqget may have dropped and re-acquired the ilock, but it guarantees
 685          * that the dquot returned is the one that should go in the inode.
 686          */
 687         *IO_idqpp = dqp;
 688         xfs_dqunlock(dqp);
 689         return 0;
 690 }
 691
 692
 693 /*
 694  * Given a udquot and gdquot, attach a ptr to the group dquot in the
 695  * udquot as a hint for future lookups. The idea sounds simple, but the
 696  * execution isn't, because the udquot might have a group dquot attached
 697  * already and getting rid of that gets us into lock ordering constraints.
 698  * The process is complicated more by the fact that the dquots may or may not
 699  * be locked on entry.
 700  */
 701 STATIC void
 702 xfs_qm_dqattach_grouphint(
 703         xfs_dquot_t     *udq,
 704         xfs_dquot_t     *gdq)
 705 {
 706         xfs_dquot_t     *tmp;
 707
 708         xfs_dqlock(udq);
 709
 710         if ((tmp = udq->q_gdquot)) {
 711                 if (tmp == gdq) {
 712                         xfs_dqunlock(udq);
 713                         return;
 714                 }
 715
 716                 udq->q_gdquot = NULL;
 717                 /*
 718                  * We can't keep any dqlocks when calling dqrele,
 719                  * because the freelist lock comes before dqlocks.
 720                  */
 721                 xfs_dqunlock(udq);
 722                 /*
 723                  * we took a hard reference once upon a time in dqget,
 724                  * so give it back when the udquot no longer points at it
 725                  * dqput() does the unlocking of the dquot.
 726                  */
 727                 xfs_qm_dqrele(tmp);
 728
 729                 xfs_dqlock(udq);
 730                 xfs_dqlock(gdq);
 731
 732         } else {
 733                 ASSERT(XFS_DQ_IS_LOCKED(udq));
 734                 xfs_dqlock(gdq);
 735         }
 736
 737         ASSERT(XFS_DQ_IS_LOCKED(udq));
 738         ASSERT(XFS_DQ_IS_LOCKED(gdq));
 739         /*
 740          * Somebody could have attached a gdquot here,
 741          * when we dropped the uqlock. If so, just do nothing.
 742          */
 743         if (udq->q_gdquot == NULL) {
 744                 XFS_DQHOLD(gdq);
 745                 udq->q_gdquot = gdq;
 746         }
 747
 748         xfs_dqunlock(gdq);
 749         xfs_dqunlock(udq);
 750 }
 751
 752
 753 /*
 754  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
 755  * into account.
 756  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
 757  * Inode may get unlocked and relocked in here, and the caller must deal with
 758  * the consequences.
 759  */
 760 int
 761 xfs_qm_dqattach_locked(
 762         xfs_inode_t     *ip,
 763         uint            flags)
 764 {
 765         xfs_mount_t     *mp = ip->i_mount;
 766         uint            nquotas = 0;
 767         int             error = 0;
 768
 769         if (!XFS_IS_QUOTA_RUNNING(mp) ||
 770             !XFS_IS_QUOTA_ON(mp) ||
 771             !XFS_NOT_DQATTACHED(mp, ip) ||
 772             ip->i_ino == mp->m_sb.sb_uquotino ||
 773             ip->i_ino == mp->m_sb.sb_gquotino)
 774                 return 0;
 775
 776         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 777
 778         if (XFS_IS_UQUOTA_ON(mp)) {
 779                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
 780                                                 flags & XFS_QMOPT_DQALLOC,
 781                                                 NULL, &ip->i_udquot);
 782                 if (error)
 783                         goto done;
 784                 nquotas++;
 785         }
 786
 787         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 788         if (XFS_IS_OQUOTA_ON(mp)) {
 789                 error = XFS_IS_GQUOTA_ON(mp) ?
 790                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
 791                                                 flags & XFS_QMOPT_DQALLOC,
 792                                                 ip->i_udquot, &ip->i_gdquot) :
 793                         xfs_qm_dqattach_one(ip, xfs_get_projid(ip), XFS_DQ_PROJ,
 794                                                 flags & XFS_QMOPT_DQALLOC,
 795                                                 ip->i_udquot, &ip->i_gdquot);
 796                 /*
 797                  * Don't worry about the udquot that we may have
 798                  * attached above. It'll get detached, if not already.
 799                  */
 800                 if (error)
 801                         goto done;
 802                 nquotas++;
 803         }
 804
 805         /*
 806          * Attach this group quota to the user quota as a hint.
 807          * This WON'T, in general, result in a thrash.
 808          */
 809         if (nquotas == 2) {
 810                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 811                 ASSERT(ip->i_udquot);
 812                 ASSERT(ip->i_gdquot);
 813
 814                 /*
 815                  * We may or may not have the i_udquot locked at this point,
 816                  * but this check is OK since we don't depend on the i_gdquot to
 817                  * be accurate 100% all the time. It is just a hint, and this
 818                  * will succeed in general.
 819                  */
 820                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
 821                         goto done;
 822                 /*
 823                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
 824                  */
 825                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot);
 826         }
 827
 828  done:
 829 #ifdef DEBUG
 830         if (!error) {
 831                 if (XFS_IS_UQUOTA_ON(mp))
 832                         ASSERT(ip->i_udquot);
 833                 if (XFS_IS_OQUOTA_ON(mp))
 834                         ASSERT(ip->i_gdquot);
 835         }
 836         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 837 #endif
 838         return error;
 839 }
 840
 841 int
 842 xfs_qm_dqattach(
 843         struct xfs_inode        *ip,
 844         uint                    flags)
 845 {
 846         int                     error;
 847
 848         xfs_ilock(ip, XFS_ILOCK_EXCL);
 849         error = xfs_qm_dqattach_locked(ip, flags);
 850         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 851
 852         return error;
 853 }
 854
 855 /*
 856  * Release dquots (and their references) if any.
 857  * The inode should be locked EXCL except when this's called by
 858  * xfs_ireclaim.
 859  */
 860 void
 861 xfs_qm_dqdetach(
 862         xfs_inode_t     *ip)
 863 {
 864         if (!(ip->i_udquot || ip->i_gdquot))
 865                 return;
 866
 867         trace_xfs_dquot_dqdetach(ip);
 868
 869         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
 870         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
 871         if (ip->i_udquot) {
 872                 xfs_qm_dqrele(ip->i_udquot);
 873                 ip->i_udquot = NULL;
 874         }
 875         if (ip->i_gdquot) {
 876                 xfs_qm_dqrele(ip->i_gdquot);
 877                 ip->i_gdquot = NULL;
 878         }
 879 }
 880
 881 int
 882 xfs_qm_sync(
 883         struct xfs_mount        *mp,
 884         int                     flags)
 885 {
 886         struct xfs_quotainfo    *q = mp->m_quotainfo;
 887         int                     recl, restarts;
 888         struct xfs_dquot        *dqp;
 889         int                     error;
 890
 891         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
 892                 return 0;
 893
 894         restarts = 0;
 895
 896   again:
 897         mutex_lock(&q->qi_dqlist_lock);
 898         /*
 899          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
 900          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
 901          * when we have the mplist lock, we know that dquots will be consistent
 902          * as long as we have it locked.
 903          */
 904         if (!XFS_IS_QUOTA_ON(mp)) {
 905                 mutex_unlock(&q->qi_dqlist_lock);
 906                 return 0;
 907         }
 908         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
 909         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
 910                 /*
 911                  * If this is vfs_sync calling, then skip the dquots that
 912                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
 913                  * This is very similar to what xfs_sync does with inodes.
 914                  */
 915                 if (flags & SYNC_TRYLOCK) {
 916                         if (!XFS_DQ_IS_DIRTY(dqp))
 917                                 continue;
 918                         if (!xfs_qm_dqlock_nowait(dqp))
 919                                 continue;
 920                 } else {
 921                         xfs_dqlock(dqp);
 922                 }
 923
 924                 /*
 925                  * Now, find out for sure if this dquot is dirty or not.
 926                  */
 927                 if (! XFS_DQ_IS_DIRTY(dqp)) {
 928                         xfs_dqunlock(dqp);
 929                         continue;
 930                 }
 931
 932                 /* XXX a sentinel would be better */
 933                 recl = q->qi_dqreclaims;
 934                 if (!xfs_dqflock_nowait(dqp)) {
 935                         if (flags & SYNC_TRYLOCK) {
 936                                 xfs_dqunlock(dqp);
 937                                 continue;
 938                         }
 939                         /*
 940                          * If we can't grab the flush lock then if the caller
 941                          * really wanted us to give this our best shot, so
 942                          * see if we can give a push to the buffer before we wait
 943                          * on the flush lock. At this point, we know that
 944                          * even though the dquot is being flushed,
 945                          * it has (new) dirty data.
 946                          */
 947                         xfs_qm_dqflock_pushbuf_wait(dqp);
 948                 }
 949                 /*
 950                  * Let go of the mplist lock. We don't want to hold it
 951                  * across a disk write
 952                  */
 953                 mutex_unlock(&q->qi_dqlist_lock);
 954                 error = xfs_qm_dqflush(dqp, flags);
 955                 xfs_dqunlock(dqp);
 956                 if (error && XFS_FORCED_SHUTDOWN(mp))
 957                         return 0;       /* Need to prevent umount failure */
 958                 else if (error)
 959                         return error;
 960
 961                 mutex_lock(&q->qi_dqlist_lock);
 962                 if (recl != q->qi_dqreclaims) {
 963                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
 964                                 break;
 965
 966                         mutex_unlock(&q->qi_dqlist_lock);
 967                         goto again;
 968                 }
 969         }
 970
 971         mutex_unlock(&q->qi_dqlist_lock);
 972         return 0;
 973 }
 974
 975 /*
 976  * The hash chains and the mplist use the same xfs_dqhash structure as
 977  * their list head, but we can take the mplist qh_lock and one of the
 978  * hash qh_locks at the same time without any problem as they aren't
 979  * related.
 980  */
 981 static struct lock_class_key xfs_quota_mplist_class;
 982
 983 /*
 984  * This initializes all the quota information that's kept in the
 985  * mount structure
 986  */
 987 STATIC int
 988 xfs_qm_init_quotainfo(
 989         xfs_mount_t     *mp)
 990 {
 991         xfs_quotainfo_t *qinf;
 992         int             error;
 993         xfs_dquot_t     *dqp;
 994
 995         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
 996
 997         /*
 998          * Tell XQM that we exist as soon as possible.
 999          */
1000         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1001                 return error;
1002         }
1003
1004         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1005
1006         /*
1007          * See if quotainodes are setup, and if not, allocate them,
1008          * and change the superblock accordingly.
1009          */
1010         if ((error = xfs_qm_init_quotainos(mp))) {
1011                 kmem_free(qinf);
1012                 mp->m_quotainfo = NULL;
1013                 return error;
1014         }
1015
1016         INIT_LIST_HEAD(&qinf->qi_dqlist);
1017         mutex_init(&qinf->qi_dqlist_lock);
1018         lockdep_set_class(&qinf->qi_dqlist_lock, &xfs_quota_mplist_class);
1019
1020         qinf->qi_dqreclaims = 0;
1021
1022         /* mutex used to serialize quotaoffs */
1023         mutex_init(&qinf->qi_quotaofflock);
1024
1025         /* Precalc some constants */
1026         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1027         ASSERT(qinf->qi_dqchunklen);
1028         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1029         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1030
1031         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1032
1033         /*
1034          * We try to get the limits from the superuser's limits fields.
1035          * This is quite hacky, but it is standard quota practice.
1036          * We look at the USR dquot with id == 0 first, but if user quotas
1037          * are not enabled we goto the GRP dquot with id == 0.
1038          * We don't really care to keep separate default limits for user
1039          * and group quotas, at least not at this point.
1040          */
1041         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1042                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER :
1043                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1044                                 XFS_DQ_PROJ),
1045                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1046                              &dqp);
1047         if (! error) {
1048                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1049
1050                 /*
1051                  * The warnings and timers set the grace period given to
1052                  * a user or group before he or she can not perform any
1053                  * more writing. If it is zero, a default is used.
1054                  */
1055                 qinf->qi_btimelimit = ddqp->d_btimer ?
1056                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1057                 qinf->qi_itimelimit = ddqp->d_itimer ?
1058                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1059                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1060                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1061                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1062                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1063                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1064                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1065                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1066                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1067                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1068                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1069                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1070                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1071                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1072                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1073
1074                 /*
1075                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1076                  * we don't want this dquot cached. We haven't done a
1077                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1078                  */
1079                 xfs_qm_dqdestroy(dqp);
1080         } else {
1081                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1082                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1083                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1084                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1085                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1086                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1087         }
1088
1089         return 0;
1090 }
1091
1092
1093 /*
1094  * Gets called when unmounting a filesystem or when all quotas get
1095  * turned off.
1096  * This purges the quota inodes, destroys locks and frees itself.
1097  */
1098 void
1099 xfs_qm_destroy_quotainfo(
1100         xfs_mount_t     *mp)
1101 {
1102         xfs_quotainfo_t *qi;
1103
1104         qi = mp->m_quotainfo;
1105         ASSERT(qi != NULL);
1106         ASSERT(xfs_Gqm != NULL);
1107
1108         /*
1109          * Release the reference that XQM kept, so that we know
1110          * when the XQM structure should be freed. We cannot assume
1111          * that xfs_Gqm is non-null after this point.
1112          */
1113         xfs_qm_rele_quotafs_ref(mp);
1114
1115         ASSERT(list_empty(&qi->qi_dqlist));
1116         mutex_destroy(&qi->qi_dqlist_lock);
1117
1118         if (qi->qi_uquotaip) {
1119                 IRELE(qi->qi_uquotaip);
1120                 qi->qi_uquotaip = NULL; /* paranoia */
1121         }
1122         if (qi->qi_gquotaip) {
1123                 IRELE(qi->qi_gquotaip);
1124                 qi->qi_gquotaip = NULL;
1125         }
1126         mutex_destroy(&qi->qi_quotaofflock);
1127         kmem_free(qi);
1128         mp->m_quotainfo = NULL;
1129 }
1130
1131
1132
1133 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1134
1135 /* ARGSUSED */
1136 STATIC void
1137 xfs_qm_list_init(
1138         xfs_dqlist_t    *list,
1139         char            *str,
1140         int             n)
1141 {
1142         mutex_init(&list->qh_lock);
1143         INIT_LIST_HEAD(&list->qh_list);
1144         list->qh_version = 0;
1145         list->qh_nelems = 0;
1146 }
1147
1148 STATIC void
1149 xfs_qm_list_destroy(
1150         xfs_dqlist_t    *list)
1151 {
1152         mutex_destroy(&(list->qh_lock));
1153 }
1154
1155 /*
1156  * Create an inode and return with a reference already taken, but unlocked
1157  * This is how we create quota inodes
1158  */
1159 STATIC int
1160 xfs_qm_qino_alloc(
1161         xfs_mount_t     *mp,
1162         xfs_inode_t     **ip,
1163         __int64_t       sbfields,
1164         uint            flags)
1165 {
1166         xfs_trans_t     *tp;
1167         int             error;
1168         int             committed;
1169
1170         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1171         if ((error = xfs_trans_reserve(tp,
1172                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1173                                       XFS_CREATE_LOG_RES(mp), 0,
1174                                       XFS_TRANS_PERM_LOG_RES,
1175                                       XFS_CREATE_LOG_COUNT))) {
1176                 xfs_trans_cancel(tp, 0);
1177                 return error;
1178         }
1179
1180         error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0, 0, 1, ip, &committed);
1181         if (error) {
1182                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1183                                  XFS_TRANS_ABORT);
1184                 return error;
1185         }
1186
1187         /*
1188          * Make the changes in the superblock, and log those too.
1189          * sbfields arg may contain fields other than *QUOTINO;
1190          * VERSIONNUM for example.
1191          */
1192         spin_lock(&mp->m_sb_lock);
1193         if (flags & XFS_QMOPT_SBVERSION) {
1194                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1195                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1196                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1197                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1198                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1199
1200                 xfs_sb_version_addquota(&mp->m_sb);
1201                 mp->m_sb.sb_uquotino = NULLFSINO;
1202                 mp->m_sb.sb_gquotino = NULLFSINO;
1203
1204                 /* qflags will get updated _after_ quotacheck */
1205                 mp->m_sb.sb_qflags = 0;
1206         }
1207         if (flags & XFS_QMOPT_UQUOTA)
1208                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1209         else
1210                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1211         spin_unlock(&mp->m_sb_lock);
1212         xfs_mod_sb(tp, sbfields);
1213
1214         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1215                 xfs_alert(mp, "%s failed (error %d)!", __func__, error);
1216                 return error;
1217         }
1218         return 0;
1219 }
1220
1221
1222 STATIC void
1223 xfs_qm_reset_dqcounts(
1224         xfs_mount_t     *mp,
1225         xfs_buf_t       *bp,
1226         xfs_dqid_t      id,
1227         uint            type)
1228 {
1229         xfs_disk_dquot_t        *ddq;
1230         int                     j;
1231
1232         trace_xfs_reset_dqcounts(bp, _RET_IP_);
1233
1234         /*
1235          * Reset all counters and timers. They'll be
1236          * started afresh by xfs_qm_quotacheck.
1237          */
1238 #ifdef DEBUG
1239         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1240         do_div(j, sizeof(xfs_dqblk_t));
1241         ASSERT(mp->m_quotainfo->qi_dqperchunk == j);
1242 #endif
1243         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1244         for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) {
1245                 /*
1246                  * Do a sanity check, and if needed, repair the dqblk. Don't
1247                  * output any warnings because it's perfectly possible to
1248                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1249                  */
1250                 (void) xfs_qm_dqcheck(mp, ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1251                                       "xfs_quotacheck");
1252                 ddq->d_bcount = 0;
1253                 ddq->d_icount = 0;
1254                 ddq->d_rtbcount = 0;
1255                 ddq->d_btimer = 0;
1256                 ddq->d_itimer = 0;
1257                 ddq->d_rtbtimer = 0;
1258                 ddq->d_bwarns = 0;
1259                 ddq->d_iwarns = 0;
1260                 ddq->d_rtbwarns = 0;
1261                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1262         }
1263 }
1264
1265 STATIC int
1266 xfs_qm_dqiter_bufs(
1267         xfs_mount_t     *mp,
1268         xfs_dqid_t      firstid,
1269         xfs_fsblock_t   bno,
1270         xfs_filblks_t   blkcnt,
1271         uint            flags)
1272 {
1273         xfs_buf_t       *bp;
1274         int             error;
1275         int             type;
1276
1277         ASSERT(blkcnt > 0);
1278         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1279                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1280         error = 0;
1281
1282         /*
1283          * Blkcnt arg can be a very big number, and might even be
1284          * larger than the log itself. So, we have to break it up into
1285          * manageable-sized transactions.
1286          * Note that we don't start a permanent transaction here; we might
1287          * not be able to get a log reservation for the whole thing up front,
1288          * and we don't really care to either, because we just discard
1289          * everything if we were to crash in the middle of this loop.
1290          */
1291         while (blkcnt--) {
1292                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1293                               XFS_FSB_TO_DADDR(mp, bno),
1294                               mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1295                 if (error)
1296                         break;
1297
1298                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1299                 xfs_bdwrite(mp, bp);
1300                 /*
1301                  * goto the next block.
1302                  */
1303                 bno++;
1304                 firstid += mp->m_quotainfo->qi_dqperchunk;
1305         }
1306         return error;
1307 }
1308
1309 /*
1310  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1311  * caller supplied function for every chunk of dquots that we find.
1312  */
1313 STATIC int
1314 xfs_qm_dqiterate(
1315         xfs_mount_t     *mp,
1316         xfs_inode_t     *qip,
1317         uint            flags)
1318 {
1319         xfs_bmbt_irec_t         *map;
1320         int                     i, nmaps;       /* number of map entries */
1321         int                     error;          /* return value */
1322         xfs_fileoff_t           lblkno;
1323         xfs_filblks_t           maxlblkcnt;
1324         xfs_dqid_t              firstid;
1325         xfs_fsblock_t           rablkno;
1326         xfs_filblks_t           rablkcnt;
1327
1328         error = 0;
1329         /*
1330          * This looks racy, but we can't keep an inode lock across a
1331          * trans_reserve. But, this gets called during quotacheck, and that
1332          * happens only at mount time which is single threaded.
1333          */
1334         if (qip->i_d.di_nblocks == 0)
1335                 return 0;
1336
1337         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1338
1339         lblkno = 0;
1340         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1341         do {
1342                 nmaps = XFS_DQITER_MAP_SIZE;
1343                 /*
1344                  * We aren't changing the inode itself. Just changing
1345                  * some of its data. No new blocks are added here, and
1346                  * the inode is never added to the transaction.
1347                  */
1348                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1349                 error = xfs_bmapi(NULL, qip, lblkno,
1350                                   maxlblkcnt - lblkno,
1351                                   XFS_BMAPI_METADATA,
1352                                   NULL,
1353                                   0, map, &nmaps, NULL);
1354                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1355                 if (error)
1356                         break;
1357
1358                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1359                 for (i = 0; i < nmaps; i++) {
1360                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1361                         ASSERT(map[i].br_blockcount);
1362
1363
1364                         lblkno += map[i].br_blockcount;
1365
1366                         if (map[i].br_startblock == HOLESTARTBLOCK)
1367                                 continue;
1368
1369                         firstid = (xfs_dqid_t) map[i].br_startoff *
1370                                 mp->m_quotainfo->qi_dqperchunk;
1371                         /*
1372                          * Do a read-ahead on the next extent.
1373                          */
1374                         if ((i+1 < nmaps) &&
1375                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1376                                 rablkcnt =  map[i+1].br_blockcount;
1377                                 rablkno = map[i+1].br_startblock;
1378                                 while (rablkcnt--) {
1379                                         xfs_buf_readahead(mp->m_ddev_targp,
1380                                                XFS_FSB_TO_DADDR(mp, rablkno),
1381                                                mp->m_quotainfo->qi_dqchunklen);
1382                                         rablkno++;
1383                                 }
1384                         }
1385                         /*
1386                          * Iterate thru all the blks in the extent and
1387                          * reset the counters of all the dquots inside them.
1388                          */
1389                         if ((error = xfs_qm_dqiter_bufs(mp,
1390                                                        firstid,
1391                                                        map[i].br_startblock,
1392                                                        map[i].br_blockcount,
1393                                                        flags))) {
1394                                 break;
1395                         }
1396                 }
1397
1398                 if (error)
1399                         break;
1400         } while (nmaps > 0);
1401
1402         kmem_free(map);
1403
1404         return error;
1405 }
1406
1407 /*
1408  * Called by dqusage_adjust in doing a quotacheck.
1409  *
1410  * Given the inode, and a dquot id this updates both the incore dqout as well
1411  * as the buffer copy. This is so that once the quotacheck is done, we can
1412  * just log all the buffers, as opposed to logging numerous updates to
1413  * individual dquots.
1414  */
1415 STATIC int
1416 xfs_qm_quotacheck_dqadjust(
1417         struct xfs_inode        *ip,
1418         xfs_dqid_t              id,
1419         uint                    type,
1420         xfs_qcnt_t              nblks,
1421         xfs_qcnt_t              rtblks)
1422 {
1423         struct xfs_mount        *mp = ip->i_mount;
1424         struct xfs_dquot        *dqp;
1425         int                     error;
1426
1427         error = xfs_qm_dqget(mp, ip, id, type,
1428                              XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN, &dqp);
1429         if (error) {
1430                 /*
1431                  * Shouldn't be able to turn off quotas here.
1432                  */
1433                 ASSERT(error != ESRCH);
1434                 ASSERT(error != ENOENT);
1435                 return error;
1436         }
1437
1438         trace_xfs_dqadjust(dqp);
1439
1440         /*
1441          * Adjust the inode count and the block count to reflect this inode's
1442          * resource usage.
1443          */
1444         be64_add_cpu(&dqp->q_core.d_icount, 1);
1445         dqp->q_res_icount++;
1446         if (nblks) {
1447                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1448                 dqp->q_res_bcount += nblks;
1449         }
1450         if (rtblks) {
1451                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1452                 dqp->q_res_rtbcount += rtblks;
1453         }
1454
1455         /*
1456          * Set default limits, adjust timers (since we changed usages)
1457          *
1458          * There are no timers for the default values set in the root dquot.
1459          */
1460         if (dqp->q_core.d_id) {
1461                 xfs_qm_adjust_dqlimits(mp, &dqp->q_core);
1462                 xfs_qm_adjust_dqtimers(mp, &dqp->q_core);
1463         }
1464
1465         dqp->dq_flags |= XFS_DQ_DIRTY;
1466         xfs_qm_dqput(dqp);
1467         return 0;
1468 }
1469
1470 STATIC int
1471 xfs_qm_get_rtblks(
1472         xfs_inode_t     *ip,
1473         xfs_qcnt_t      *O_rtblks)
1474 {
1475         xfs_filblks_t   rtblks;                 /* total rt blks */
1476         xfs_extnum_t    idx;                    /* extent record index */
1477         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1478         xfs_extnum_t    nextents;               /* number of extent entries */
1479         int             error;
1480
1481         ASSERT(XFS_IS_REALTIME_INODE(ip));
1482         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1483         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1484                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1485                         return error;
1486         }
1487         rtblks = 0;
1488         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1489         for (idx = 0; idx < nextents; idx++)
1490                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1491         *O_rtblks = (xfs_qcnt_t)rtblks;
1492         return 0;
1493 }
1494
1495 /*
1496  * callback routine supplied to bulkstat(). Given an inumber, find its
1497  * dquots and update them to account for resources taken by that inode.
1498  */
1499 /* ARGSUSED */
1500 STATIC int
1501 xfs_qm_dqusage_adjust(
1502         xfs_mount_t     *mp,            /* mount point for filesystem */
1503         xfs_ino_t       ino,            /* inode number to get data for */
1504         void            __user *buffer, /* not used */
1505         int             ubsize,         /* not used */
1506         int             *ubused,        /* not used */
1507         int             *res)           /* result code value */
1508 {
1509         xfs_inode_t     *ip;
1510         xfs_qcnt_t      nblks, rtblks = 0;
1511         int             error;
1512
1513         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1514
1515         /*
1516          * rootino must have its resources accounted for, not so with the quota
1517          * inodes.
1518          */
1519         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1520                 *res = BULKSTAT_RV_NOTHING;
1521                 return XFS_ERROR(EINVAL);
1522         }
1523
1524         /*
1525          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1526          * interface expects the inode to be exclusively locked because that's
1527          * the case in all other instances. It's OK that we do this because
1528          * quotacheck is done only at mount time.
1529          */
1530         error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip);
1531         if (error) {
1532                 *res = BULKSTAT_RV_NOTHING;
1533                 return error;
1534         }
1535
1536         ASSERT(ip->i_delayed_blks == 0);
1537
1538         if (XFS_IS_REALTIME_INODE(ip)) {
1539                 /*
1540                  * Walk thru the extent list and count the realtime blocks.
1541                  */
1542                 error = xfs_qm_get_rtblks(ip, &rtblks);
1543                 if (error)
1544                         goto error0;
1545         }
1546
1547         nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1548
1549         /*
1550          * Add the (disk blocks and inode) resources occupied by this
1551          * inode to its dquots. We do this adjustment in the incore dquot,
1552          * and also copy the changes to its buffer.
1553          * We don't care about putting these changes in a transaction
1554          * envelope because if we crash in the middle of a 'quotacheck'
1555          * we have to start from the beginning anyway.
1556          * Once we're done, we'll log all the dquot bufs.
1557          *
1558          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1559          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1560          */
1561         if (XFS_IS_UQUOTA_ON(mp)) {
1562                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_uid,
1563                                                    XFS_DQ_USER, nblks, rtblks);
1564                 if (error)
1565                         goto error0;
1566         }
1567
1568         if (XFS_IS_GQUOTA_ON(mp)) {
1569                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_gid,
1570                                                    XFS_DQ_GROUP, nblks, rtblks);
1571                 if (error)
1572                         goto error0;
1573         }
1574
1575         if (XFS_IS_PQUOTA_ON(mp)) {
1576                 error = xfs_qm_quotacheck_dqadjust(ip, xfs_get_projid(ip),
1577                                                    XFS_DQ_PROJ, nblks, rtblks);
1578                 if (error)
1579                         goto error0;
1580         }
1581
1582         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1583         IRELE(ip);
1584         *res = BULKSTAT_RV_DIDONE;
1585         return 0;
1586
1587 error0:
1588         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1589         IRELE(ip);
1590         *res = BULKSTAT_RV_GIVEUP;
1591         return error;
1592 }
1593
1594 /*
1595  * Walk thru all the filesystem inodes and construct a consistent view
1596  * of the disk quota world. If the quotacheck fails, disable quotas.
1597  */
1598 int
1599 xfs_qm_quotacheck(
1600         xfs_mount_t     *mp)
1601 {
1602         int             done, count, error;
1603         xfs_ino_t       lastino;
1604         size_t          structsz;
1605         xfs_inode_t     *uip, *gip;
1606         uint            flags;
1607
1608         count = INT_MAX;
1609         structsz = 1;
1610         lastino = 0;
1611         flags = 0;
1612
1613         ASSERT(mp->m_quotainfo->qi_uquotaip || mp->m_quotainfo->qi_gquotaip);
1614         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1615
1616         /*
1617          * There should be no cached dquots. The (simplistic) quotacheck
1618          * algorithm doesn't like that.
1619          */
1620         ASSERT(list_empty(&mp->m_quotainfo->qi_dqlist));
1621
1622         xfs_notice(mp, "Quotacheck needed: Please wait.");
1623
1624         /*
1625          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1626          * their counters to zero. We need a clean slate.
1627          * We don't log our changes till later.
1628          */
1629         uip = mp->m_quotainfo->qi_uquotaip;
1630         if (uip) {
1631                 error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA);
1632                 if (error)
1633                         goto error_return;
1634                 flags |= XFS_UQUOTA_CHKD;
1635         }
1636
1637         gip = mp->m_quotainfo->qi_gquotaip;
1638         if (gip) {
1639                 error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1640                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1641                 if (error)
1642                         goto error_return;
1643                 flags |= XFS_OQUOTA_CHKD;
1644         }
1645
1646         do {
1647                 /*
1648                  * Iterate thru all the inodes in the file system,
1649                  * adjusting the corresponding dquot counters in core.
1650                  */
1651                 error = xfs_bulkstat(mp, &lastino, &count,
1652                                      xfs_qm_dqusage_adjust,
1653                                      structsz, NULL, &done);
1654                 if (error)
1655                         break;
1656
1657         } while (!done);
1658
1659         /*
1660          * We've made all the changes that we need to make incore.
1661          * Flush them down to disk buffers if everything was updated
1662          * successfully.
1663          */
1664         if (!error)
1665                 error = xfs_qm_dqflush_all(mp, 0);
1666
1667         /*
1668          * We can get this error if we couldn't do a dquot allocation inside
1669          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1670          * dirty dquots that might be cached, we just want to get rid of them
1671          * and turn quotaoff. The dquots won't be attached to any of the inodes
1672          * at this point (because we intentionally didn't in dqget_noattach).
1673          */
1674         if (error) {
1675                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL);
1676                 goto error_return;
1677         }
1678
1679         /*
1680          * We didn't log anything, because if we crashed, we'll have to
1681          * start the quotacheck from scratch anyway. However, we must make
1682          * sure that our dquot changes are secure before we put the
1683          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1684          * flush.
1685          */
1686         XFS_bflush(mp->m_ddev_targp);
1687
1688         /*
1689          * If one type of quotas is off, then it will lose its
1690          * quotachecked status, since we won't be doing accounting for
1691          * that type anymore.
1692          */
1693         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1694         mp->m_qflags |= flags;
1695
1696  error_return:
1697         if (error) {
1698                 xfs_warn(mp,
1699         "Quotacheck: Unsuccessful (Error %d): Disabling quotas.",
1700                         error);
1701                 /*
1702                  * We must turn off quotas.
1703                  */
1704                 ASSERT(mp->m_quotainfo != NULL);
1705                 ASSERT(xfs_Gqm != NULL);
1706                 xfs_qm_destroy_quotainfo(mp);
1707                 if (xfs_mount_reset_sbqflags(mp)) {
1708                         xfs_warn(mp,
1709                                 "Quotacheck: Failed to reset quota flags.");
1710                 }
1711         } else
1712                 xfs_notice(mp, "Quotacheck: Done.");
1713         return (error);
1714 }
1715
1716 /*
1717  * This is called after the superblock has been read in and we're ready to
1718  * iget the quota inodes.
1719  */
1720 STATIC int
1721 xfs_qm_init_quotainos(
1722         xfs_mount_t     *mp)
1723 {
1724         xfs_inode_t     *uip, *gip;
1725         int             error;
1726         __int64_t       sbflags;
1727         uint            flags;
1728
1729         ASSERT(mp->m_quotainfo);
1730         uip = gip = NULL;
1731         sbflags = 0;
1732         flags = 0;
1733
1734         /*
1735          * Get the uquota and gquota inodes
1736          */
1737         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1738                 if (XFS_IS_UQUOTA_ON(mp) &&
1739                     mp->m_sb.sb_uquotino != NULLFSINO) {
1740                         ASSERT(mp->m_sb.sb_uquotino > 0);
1741                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1742                                              0, 0, &uip)))
1743                                 return XFS_ERROR(error);
1744                 }
1745                 if (XFS_IS_OQUOTA_ON(mp) &&
1746                     mp->m_sb.sb_gquotino != NULLFSINO) {
1747                         ASSERT(mp->m_sb.sb_gquotino > 0);
1748                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1749                                              0, 0, &gip))) {
1750                                 if (uip)
1751                                         IRELE(uip);
1752                                 return XFS_ERROR(error);
1753                         }
1754                 }
1755         } else {
1756                 flags |= XFS_QMOPT_SBVERSION;
1757                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1758                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1759         }
1760
1761         /*
1762          * Create the two inodes, if they don't exist already. The changes
1763          * made above will get added to a transaction and logged in one of
1764          * the qino_alloc calls below.  If the device is readonly,
1765          * temporarily switch to read-write to do this.
1766          */
1767         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1768                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1769                                               sbflags | XFS_SB_UQUOTINO,
1770                                               flags | XFS_QMOPT_UQUOTA)))
1771                         return XFS_ERROR(error);
1772
1773                 flags &= ~XFS_QMOPT_SBVERSION;
1774         }
1775         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1776                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1777                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1778                 error = xfs_qm_qino_alloc(mp, &gip,
1779                                           sbflags | XFS_SB_GQUOTINO, flags);
1780                 if (error) {
1781                         if (uip)
1782                                 IRELE(uip);
1783
1784                         return XFS_ERROR(error);
1785                 }
1786         }
1787
1788         mp->m_quotainfo->qi_uquotaip = uip;
1789         mp->m_quotainfo->qi_gquotaip = gip;
1790
1791         return 0;
1792 }
1793
1794
1795
1796 /*
1797  * Just pop the least recently used dquot off the freelist and
1798  * recycle it. The returned dquot is locked.
1799  */
1800 STATIC xfs_dquot_t *
1801 xfs_qm_dqreclaim_one(void)
1802 {
1803         xfs_dquot_t     *dqpout;
1804         xfs_dquot_t     *dqp;
1805         int             restarts;
1806         int             startagain;
1807
1808         restarts = 0;
1809         dqpout = NULL;
1810
1811         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
1812 again:
1813         startagain = 0;
1814         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1815
1816         list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) {
1817                 struct xfs_mount *mp = dqp->q_mount;
1818                 xfs_dqlock(dqp);
1819
1820                 /*
1821                  * We are racing with dqlookup here. Naturally we don't
1822                  * want to reclaim a dquot that lookup wants. We release the
1823                  * freelist lock and start over, so that lookup will grab
1824                  * both the dquot and the freelistlock.
1825                  */
1826                 if (dqp->dq_flags & XFS_DQ_WANT) {
1827                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
1828
1829                         trace_xfs_dqreclaim_want(dqp);
1830                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
1831                         restarts++;
1832                         startagain = 1;
1833                         goto dqunlock;
1834                 }
1835
1836                 /*
1837                  * If the dquot is inactive, we are assured that it is
1838                  * not on the mplist or the hashlist, and that makes our
1839                  * life easier.
1840                  */
1841                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
1842                         ASSERT(mp == NULL);
1843                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
1844                         ASSERT(list_empty(&dqp->q_hashlist));
1845                         ASSERT(list_empty(&dqp->q_mplist));
1846                         list_del_init(&dqp->q_freelist);
1847                         xfs_Gqm->qm_dqfrlist_cnt--;
1848                         dqpout = dqp;
1849                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
1850                         goto dqunlock;
1851                 }
1852
1853                 ASSERT(dqp->q_hash);
1854                 ASSERT(!list_empty(&dqp->q_mplist));
1855
1856                 /*
1857                  * Try to grab the flush lock. If this dquot is in the process
1858                  * of getting flushed to disk, we don't want to reclaim it.
1859                  */
1860                 if (!xfs_dqflock_nowait(dqp))
1861                         goto dqunlock;
1862
1863                 /*
1864                  * We have the flush lock so we know that this is not in the
1865                  * process of being flushed. So, if this is dirty, flush it
1866                  * DELWRI so that we don't get a freelist infested with
1867                  * dirty dquots.
1868                  */
1869                 if (XFS_DQ_IS_DIRTY(dqp)) {
1870                         int     error;
1871
1872                         trace_xfs_dqreclaim_dirty(dqp);
1873
1874                         /*
1875                          * We flush it delayed write, so don't bother
1876                          * releasing the freelist lock.
1877                          */
1878                         error = xfs_qm_dqflush(dqp, 0);
1879                         if (error) {
1880                                 xfs_warn(mp, "%s: dquot %p flush failed",
1881                                         __func__, dqp);
1882                         }
1883                         goto dqunlock;
1884                 }
1885
1886                 /*
1887                  * We're trying to get the hashlock out of order. This races
1888                  * with dqlookup; so, we giveup and goto the next dquot if
1889                  * we couldn't get the hashlock. This way, we won't starve
1890                  * a dqlookup process that holds the hashlock that is
1891                  * waiting for the freelist lock.
1892                  */
1893                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
1894                         restarts++;
1895                         goto dqfunlock;
1896                 }
1897
1898                 /*
1899                  * This races with dquot allocation code as well as dqflush_all
1900                  * and reclaim code. So, if we failed to grab the mplist lock,
1901                  * giveup everything and start over.
1902                  */
1903                 if (!mutex_trylock(&mp->m_quotainfo->qi_dqlist_lock)) {
1904                         restarts++;
1905                         startagain = 1;
1906                         goto qhunlock;
1907                 }
1908
1909                 ASSERT(dqp->q_nrefs == 0);
1910                 list_del_init(&dqp->q_mplist);
1911                 mp->m_quotainfo->qi_dquots--;
1912                 mp->m_quotainfo->qi_dqreclaims++;
1913                 list_del_init(&dqp->q_hashlist);
1914                 dqp->q_hash->qh_version++;
1915                 list_del_init(&dqp->q_freelist);
1916                 xfs_Gqm->qm_dqfrlist_cnt--;
1917                 dqpout = dqp;
1918                 mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1919 qhunlock:
1920                 mutex_unlock(&dqp->q_hash->qh_lock);
1921 dqfunlock:
1922                 xfs_dqfunlock(dqp);
1923 dqunlock:
1924                 xfs_dqunlock(dqp);
1925                 if (dqpout)
1926                         break;
1927                 if (restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
1928                         break;
1929                 if (startagain) {
1930                         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1931                         goto again;
1932                 }
1933         }
1934         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1935         return dqpout;
1936 }
1937
1938 /*
1939  * Traverse the freelist of dquots and attempt to reclaim a maximum of
1940  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
1941  * favor the lookup function ...
1942  */
1943 STATIC int
1944 xfs_qm_shake_freelist(
1945         int     howmany)
1946 {
1947         int             nreclaimed = 0;
1948         xfs_dquot_t     *dqp;
1949
1950         if (howmany <= 0)
1951                 return 0;
1952
1953         while (nreclaimed < howmany) {
1954                 dqp = xfs_qm_dqreclaim_one();
1955                 if (!dqp)
1956                         return nreclaimed;
1957                 xfs_qm_dqdestroy(dqp);
1958                 nreclaimed++;
1959         }
1960         return nreclaimed;
1961 }
1962
1963 /*
1964  * The kmem_shake interface is invoked when memory is running low.
1965  */
1966 /* ARGSUSED */
1967 STATIC int
1968 xfs_qm_shake(
1969         struct shrinker *shrink,
1970         struct shrink_control *sc)
1971 {
1972         int     ndqused, nfree, n;
1973         gfp_t gfp_mask = sc->gfp_mask;
1974
1975         if (!kmem_shake_allow(gfp_mask))
1976                 return 0;
1977         if (!xfs_Gqm)
1978                 return 0;
1979
1980         nfree = xfs_Gqm->qm_dqfrlist_cnt; /* free dquots */
1981         /* incore dquots in all f/s's */
1982         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
1983
1984         ASSERT(ndqused >= 0);
1985
1986         if (nfree <= ndqused && nfree < ndquot)
1987                 return 0;
1988
1989         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
1990         n = nfree - ndqused - ndquot;           /* # over target */
1991
1992         return xfs_qm_shake_freelist(MAX(nfree, n));
1993 }
1994
1995
1996 /*------------------------------------------------------------------*/
1997
1998 /*
1999  * Return a new incore dquot. Depending on the number of
2000  * dquots in the system, we either allocate a new one on the kernel heap,
2001  * or reclaim a free one.
2002  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2003  * to reclaim an existing one from the freelist.
2004  */
2005 boolean_t
2006 xfs_qm_dqalloc_incore(
2007         xfs_dquot_t **O_dqpp)
2008 {
2009         xfs_dquot_t     *dqp;
2010
2011         /*
2012          * Check against high water mark to see if we want to pop
2013          * a nincompoop dquot off the freelist.
2014          */
2015         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2016                 /*
2017                  * Try to recycle a dquot from the freelist.
2018                  */
2019                 if ((dqp = xfs_qm_dqreclaim_one())) {
2020                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2021                         /*
2022                          * Just zero the core here. The rest will get
2023                          * reinitialized by caller. XXX we shouldn't even
2024                          * do this zero ...
2025                          */
2026                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2027                         *O_dqpp = dqp;
2028                         return B_FALSE;
2029                 }
2030                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2031         }
2032
2033         /*
2034          * Allocate a brand new dquot on the kernel heap and return it
2035          * to the caller to initialize.
2036          */
2037         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2038         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2039         atomic_inc(&xfs_Gqm->qm_totaldquots);
2040
2041         return B_TRUE;
2042 }
2043
2044
2045 /*
2046  * Start a transaction and write the incore superblock changes to
2047  * disk. flags parameter indicates which fields have changed.
2048  */
2049 int
2050 xfs_qm_write_sb_changes(
2051         xfs_mount_t     *mp,
2052         __int64_t       flags)
2053 {
2054         xfs_trans_t     *tp;
2055         int             error;
2056
2057         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2058         if ((error = xfs_trans_reserve(tp, 0,
2059                                       mp->m_sb.sb_sectsize + 128, 0,
2060                                       0,
2061                                       XFS_DEFAULT_LOG_COUNT))) {
2062                 xfs_trans_cancel(tp, 0);
2063                 return error;
2064         }
2065
2066         xfs_mod_sb(tp, flags);
2067         error = xfs_trans_commit(tp, 0);
2068
2069         return error;
2070 }
2071
2072
2073 /* --------------- utility functions for vnodeops ---------------- */
2074
2075
2076 /*
2077  * Given an inode, a uid, gid and prid make sure that we have
2078  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2079  * quotas by creating this file.
2080  * This also attaches dquot(s) to the given inode after locking it,
2081  * and returns the dquots corresponding to the uid and/or gid.
2082  *
2083  * in   : inode (unlocked)
2084  * out  : udquot, gdquot with references taken and unlocked
2085  */
2086 int
2087 xfs_qm_vop_dqalloc(
2088         struct xfs_inode        *ip,
2089         uid_t                   uid,
2090         gid_t                   gid,
2091         prid_t                  prid,
2092         uint                    flags,
2093         struct xfs_dquot        **O_udqpp,
2094         struct xfs_dquot        **O_gdqpp)
2095 {
2096         struct xfs_mount        *mp = ip->i_mount;
2097         struct xfs_dquot        *uq, *gq;
2098         int                     error;
2099         uint                    lockflags;
2100
2101         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2102                 return 0;
2103
2104         lockflags = XFS_ILOCK_EXCL;
2105         xfs_ilock(ip, lockflags);
2106
2107         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2108                 gid = ip->i_d.di_gid;
2109
2110         /*
2111          * Attach the dquot(s) to this inode, doing a dquot allocation
2112          * if necessary. The dquot(s) will not be locked.
2113          */
2114         if (XFS_NOT_DQATTACHED(mp, ip)) {
2115                 error = xfs_qm_dqattach_locked(ip, XFS_QMOPT_DQALLOC);
2116                 if (error) {
2117                         xfs_iunlock(ip, lockflags);
2118                         return error;
2119                 }
2120         }
2121
2122         uq = gq = NULL;
2123         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2124                 if (ip->i_d.di_uid != uid) {
2125                         /*
2126                          * What we need is the dquot that has this uid, and
2127                          * if we send the inode to dqget, the uid of the inode
2128                          * takes priority over what's sent in the uid argument.
2129                          * We must unlock inode here before calling dqget if
2130                          * we're not sending the inode, because otherwise
2131                          * we'll deadlock by doing trans_reserve while
2132                          * holding ilock.
2133                          */
2134                         xfs_iunlock(ip, lockflags);
2135                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2136                                                  XFS_DQ_USER,
2137                                                  XFS_QMOPT_DQALLOC |
2138                                                  XFS_QMOPT_DOWARN,
2139                                                  &uq))) {
2140                                 ASSERT(error != ENOENT);
2141                                 return error;
2142                         }
2143                         /*
2144                          * Get the ilock in the right order.
2145                          */
2146                         xfs_dqunlock(uq);
2147                         lockflags = XFS_ILOCK_SHARED;
2148                         xfs_ilock(ip, lockflags);
2149                 } else {
2150                         /*
2151                          * Take an extra reference, because we'll return
2152                          * this to caller
2153                          */
2154                         ASSERT(ip->i_udquot);
2155                         uq = ip->i_udquot;
2156                         xfs_dqlock(uq);
2157                         XFS_DQHOLD(uq);
2158                         xfs_dqunlock(uq);
2159                 }
2160         }
2161         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2162                 if (ip->i_d.di_gid != gid) {
2163                         xfs_iunlock(ip, lockflags);
2164                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2165                                                  XFS_DQ_GROUP,
2166                                                  XFS_QMOPT_DQALLOC |
2167                                                  XFS_QMOPT_DOWARN,
2168                                                  &gq))) {
2169                                 if (uq)
2170                                         xfs_qm_dqrele(uq);
2171                                 ASSERT(error != ENOENT);
2172                                 return error;
2173                         }
2174                         xfs_dqunlock(gq);
2175                         lockflags = XFS_ILOCK_SHARED;
2176                         xfs_ilock(ip, lockflags);
2177                 } else {
2178                         ASSERT(ip->i_gdquot);
2179                         gq = ip->i_gdquot;
2180                         xfs_dqlock(gq);
2181                         XFS_DQHOLD(gq);
2182                         xfs_dqunlock(gq);
2183                 }
2184         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2185                 if (xfs_get_projid(ip) != prid) {
2186                         xfs_iunlock(ip, lockflags);
2187                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2188                                                  XFS_DQ_PROJ,
2189                                                  XFS_QMOPT_DQALLOC |
2190                                                  XFS_QMOPT_DOWARN,
2191                                                  &gq))) {
2192                                 if (uq)
2193                                         xfs_qm_dqrele(uq);
2194                                 ASSERT(error != ENOENT);
2195                                 return (error);
2196                         }
2197                         xfs_dqunlock(gq);
2198                         lockflags = XFS_ILOCK_SHARED;
2199                         xfs_ilock(ip, lockflags);
2200                 } else {
2201                         ASSERT(ip->i_gdquot);
2202                         gq = ip->i_gdquot;
2203                         xfs_dqlock(gq);
2204                         XFS_DQHOLD(gq);
2205                         xfs_dqunlock(gq);
2206                 }
2207         }
2208         if (uq)
2209                 trace_xfs_dquot_dqalloc(ip);
2210
2211         xfs_iunlock(ip, lockflags);
2212         if (O_udqpp)
2213                 *O_udqpp = uq;
2214         else if (uq)
2215                 xfs_qm_dqrele(uq);
2216         if (O_gdqpp)
2217                 *O_gdqpp = gq;
2218         else if (gq)
2219                 xfs_qm_dqrele(gq);
2220         return 0;
2221 }
2222
2223 /*
2224  * Actually transfer ownership, and do dquot modifications.
2225  * These were already reserved.
2226  */
2227 xfs_dquot_t *
2228 xfs_qm_vop_chown(
2229         xfs_trans_t     *tp,
2230         xfs_inode_t     *ip,
2231         xfs_dquot_t     **IO_olddq,
2232         xfs_dquot_t     *newdq)
2233 {
2234         xfs_dquot_t     *prevdq;
2235         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2236                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2237
2238
2239         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2240         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2241
2242         /* old dquot */
2243         prevdq = *IO_olddq;
2244         ASSERT(prevdq);
2245         ASSERT(prevdq != newdq);
2246
2247         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2248         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2249
2250         /* the sparkling new dquot */
2251         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2252         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2253
2254         /*
2255          * Take an extra reference, because the inode
2256          * is going to keep this dquot pointer even
2257          * after the trans_commit.
2258          */
2259         xfs_dqlock(newdq);
2260         XFS_DQHOLD(newdq);
2261         xfs_dqunlock(newdq);
2262         *IO_olddq = newdq;
2263
2264         return prevdq;
2265 }
2266
2267 /*
2268  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2269  */
2270 int
2271 xfs_qm_vop_chown_reserve(
2272         xfs_trans_t     *tp,
2273         xfs_inode_t     *ip,
2274         xfs_dquot_t     *udqp,
2275         xfs_dquot_t     *gdqp,
2276         uint            flags)
2277 {
2278         xfs_mount_t     *mp = ip->i_mount;
2279         uint            delblks, blkflags, prjflags = 0;
2280         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2281         int             error;
2282
2283
2284         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2285         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2286
2287         delblks = ip->i_delayed_blks;
2288         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2289         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2290                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2291
2292         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2293             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2294                 delblksudq = udqp;
2295                 /*
2296                  * If there are delayed allocation blocks, then we have to
2297                  * unreserve those from the old dquot, and add them to the
2298                  * new dquot.
2299                  */
2300                 if (delblks) {
2301                         ASSERT(ip->i_udquot);
2302                         unresudq = ip->i_udquot;
2303                 }
2304         }
2305         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2306                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2307                      xfs_get_projid(ip) != be32_to_cpu(gdqp->q_core.d_id))
2308                         prjflags = XFS_QMOPT_ENOSPC;
2309
2310                 if (prjflags ||
2311                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2312                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2313                         delblksgdq = gdqp;
2314                         if (delblks) {
2315                                 ASSERT(ip->i_gdquot);
2316                                 unresgdq = ip->i_gdquot;
2317                         }
2318                 }
2319         }
2320
2321         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2322                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2323                                 flags | blkflags | prjflags)))
2324                 return (error);
2325
2326         /*
2327          * Do the delayed blks reservations/unreservations now. Since, these
2328          * are done without the help of a transaction, if a reservation fails
2329          * its previous reservations won't be automatically undone by trans
2330          * code. So, we have to do it manually here.
2331          */
2332         if (delblks) {
2333                 /*
2334                  * Do the reservations first. Unreservation can't fail.
2335                  */
2336                 ASSERT(delblksudq || delblksgdq);
2337                 ASSERT(unresudq || unresgdq);
2338                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2339                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2340                                 flags | blkflags | prjflags)))
2341                         return (error);
2342                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2343                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2344                                 blkflags);
2345         }
2346
2347         return (0);
2348 }
2349
2350 int
2351 xfs_qm_vop_rename_dqattach(
2352         struct xfs_inode        **i_tab)
2353 {
2354         struct xfs_mount        *mp = i_tab[0]->i_mount;
2355         int                     i;
2356
2357         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2358                 return 0;
2359
2360         for (i = 0; (i < 4 && i_tab[i]); i++) {
2361                 struct xfs_inode        *ip = i_tab[i];
2362                 int                     error;
2363
2364                 /*
2365                  * Watch out for duplicate entries in the table.
2366                  */
2367                 if (i == 0 || ip != i_tab[i-1]) {
2368                         if (XFS_NOT_DQATTACHED(mp, ip)) {
2369                                 error = xfs_qm_dqattach(ip, 0);
2370                                 if (error)
2371                                         return error;
2372                         }
2373                 }
2374         }
2375         return 0;
2376 }
2377
2378 void
2379 xfs_qm_vop_create_dqattach(
2380         struct xfs_trans        *tp,
2381         struct xfs_inode        *ip,
2382         struct xfs_dquot        *udqp,
2383         struct xfs_dquot        *gdqp)
2384 {
2385         struct xfs_mount        *mp = tp->t_mountp;
2386
2387         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2388                 return;
2389
2390         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2391         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2392
2393         if (udqp) {
2394                 xfs_dqlock(udqp);
2395                 XFS_DQHOLD(udqp);
2396                 xfs_dqunlock(udqp);
2397                 ASSERT(ip->i_udquot == NULL);
2398                 ip->i_udquot = udqp;
2399                 ASSERT(XFS_IS_UQUOTA_ON(mp));
2400                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2401                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2402         }
2403         if (gdqp) {
2404                 xfs_dqlock(gdqp);
2405                 XFS_DQHOLD(gdqp);
2406                 xfs_dqunlock(gdqp);
2407                 ASSERT(ip->i_gdquot == NULL);
2408                 ip->i_gdquot = gdqp;
2409                 ASSERT(XFS_IS_OQUOTA_ON(mp));
2410                 ASSERT((XFS_IS_GQUOTA_ON(mp) ?
2411                         ip->i_d.di_gid : xfs_get_projid(ip)) ==
2412                                 be32_to_cpu(gdqp->q_core.d_id));
2413                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2414         }
2415 }
2416