Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_clnt.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_alloc.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_quota.h"
31 #include "xfs_mount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_itable.h"
42 #include "xfs_rtalloc.h"
43 #include "xfs_error.h"
44 #include "xfs_bmap.h"
45 #include "xfs_rw.h"
46 #include "xfs_acl.h"
47 #include "xfs_attr.h"
48 #include "xfs_buf_item.h"
49 #include "xfs_trans_space.h"
50 #include "xfs_utils.h"
51 #include "xfs_qm.h"
52
53 /*
54  * The global quota manager. There is only one of these for the entire
55  * system, _not_ one per file system. XQM keeps track of the overall
56  * quota functionality, including maintaining the freelist and hash
57  * tables of dquots.
58  */
59 mutex_t         xfs_Gqm_lock;
60 struct xfs_qm   *xfs_Gqm;
61 uint            ndquot;
62
63 kmem_zone_t     *qm_dqzone;
64 kmem_zone_t     *qm_dqtrxzone;
65
66 static cred_t   xfs_zerocr;
67
68 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
69 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
70
71 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
72 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
73 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
74 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
75
76 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
77 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
78 STATIC int      xfs_qm_shake(int, gfp_t);
79
80 static struct shrinker xfs_qm_shaker = {
81         .shrink = xfs_qm_shake,
82         .seeks = DEFAULT_SEEKS,
83 };
84
85 #ifdef DEBUG
86 extern mutex_t  qcheck_lock;
87 #endif
88
89 #ifdef QUOTADEBUG
90 #define XQM_LIST_PRINT(l, NXT, title) \
91 { \
92         xfs_dquot_t     *dqp; int i = 0; \
93         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
94         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
95                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
96                                   "bcnt = %d, icnt = %d, refs = %d", \
97                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
98                         DQFLAGTO_TYPESTR(dqp),       \
99                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
100                         (int) be64_to_cpu(dqp->q_core.d_icount), \
101                         (int) dqp->q_nrefs);  } \
102 }
103 #else
104 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
105 #endif
106
107 /*
108  * Initialize the XQM structure.
109  * Note that there is not one quota manager per file system.
110  */
111 STATIC struct xfs_qm *
112 xfs_Gqm_init(void)
113 {
114         xfs_dqhash_t    *udqhash, *gdqhash;
115         xfs_qm_t        *xqm;
116         size_t          hsize;
117         uint            i;
118
119         /*
120          * Initialize the dquot hash tables.
121          */
122         udqhash = kmem_zalloc_greedy(&hsize,
123                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
124                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t),
125                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
126         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
127         hsize /= sizeof(xfs_dqhash_t);
128         ndquot = hsize << 8;
129
130         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
131         xqm->qm_dqhashmask = hsize - 1;
132         xqm->qm_usr_dqhtable = udqhash;
133         xqm->qm_grp_dqhtable = gdqhash;
134         ASSERT(xqm->qm_usr_dqhtable != NULL);
135         ASSERT(xqm->qm_grp_dqhtable != NULL);
136
137         for (i = 0; i < hsize; i++) {
138                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
139                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
140         }
141
142         /*
143          * Freelist of all dquots of all file systems
144          */
145         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
146
147         /*
148          * dquot zone. we register our own low-memory callback.
149          */
150         if (!qm_dqzone) {
151                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
152                                                 "xfs_dquots");
153                 qm_dqzone = xqm->qm_dqzone;
154         } else
155                 xqm->qm_dqzone = qm_dqzone;
156
157         register_shrinker(&xfs_qm_shaker);
158
159         /*
160          * The t_dqinfo portion of transactions.
161          */
162         if (!qm_dqtrxzone) {
163                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
164                                                    "xfs_dqtrx");
165                 qm_dqtrxzone = xqm->qm_dqtrxzone;
166         } else
167                 xqm->qm_dqtrxzone = qm_dqtrxzone;
168
169         atomic_set(&xqm->qm_totaldquots, 0);
170         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
171         xqm->qm_nrefs = 0;
172 #ifdef DEBUG
173         mutex_init(&qcheck_lock);
174 #endif
175         return xqm;
176 }
177
178 /*
179  * Destroy the global quota manager when its reference count goes to zero.
180  */
181 STATIC void
182 xfs_qm_destroy(
183         struct xfs_qm   *xqm)
184 {
185         int             hsize, i;
186
187         ASSERT(xqm != NULL);
188         ASSERT(xqm->qm_nrefs == 0);
189         unregister_shrinker(&xfs_qm_shaker);
190         hsize = xqm->qm_dqhashmask + 1;
191         for (i = 0; i < hsize; i++) {
192                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
193                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
194         }
195         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
196         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
197         xqm->qm_usr_dqhtable = NULL;
198         xqm->qm_grp_dqhtable = NULL;
199         xqm->qm_dqhashmask = 0;
200         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
201 #ifdef DEBUG
202         mutex_destroy(&qcheck_lock);
203 #endif
204         kmem_free(xqm, sizeof(xfs_qm_t));
205 }
206
207 /*
208  * Called at mount time to let XQM know that another file system is
209  * starting quotas. This isn't crucial information as the individual mount
210  * structures are pretty independent, but it helps the XQM keep a
211  * global view of what's going on.
212  */
213 /* ARGSUSED */
214 STATIC int
215 xfs_qm_hold_quotafs_ref(
216         struct xfs_mount *mp)
217 {
218         /*
219          * Need to lock the xfs_Gqm structure for things like this. For example,
220          * the structure could disappear between the entry to this routine and
221          * a HOLD operation if not locked.
222          */
223         XFS_QM_LOCK(xfs_Gqm);
224
225         if (xfs_Gqm == NULL)
226                 xfs_Gqm = xfs_Gqm_init();
227         /*
228          * We can keep a list of all filesystems with quotas mounted for
229          * debugging and statistical purposes, but ...
230          * Just take a reference and get out.
231          */
232         XFS_QM_HOLD(xfs_Gqm);
233         XFS_QM_UNLOCK(xfs_Gqm);
234
235         return 0;
236 }
237
238
239 /*
240  * Release the reference that a filesystem took at mount time,
241  * so that we know when we need to destroy the entire quota manager.
242  */
243 /* ARGSUSED */
244 STATIC void
245 xfs_qm_rele_quotafs_ref(
246         struct xfs_mount *mp)
247 {
248         xfs_dquot_t     *dqp, *nextdqp;
249
250         ASSERT(xfs_Gqm);
251         ASSERT(xfs_Gqm->qm_nrefs > 0);
252
253         /*
254          * Go thru the freelist and destroy all inactive dquots.
255          */
256         xfs_qm_freelist_lock(xfs_Gqm);
257
258         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
259              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
260                 xfs_dqlock(dqp);
261                 nextdqp = dqp->dq_flnext;
262                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
263                         ASSERT(dqp->q_mount == NULL);
264                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
265                         ASSERT(dqp->HL_PREVP == NULL);
266                         ASSERT(dqp->MPL_PREVP == NULL);
267                         XQM_FREELIST_REMOVE(dqp);
268                         xfs_dqunlock(dqp);
269                         xfs_qm_dqdestroy(dqp);
270                 } else {
271                         xfs_dqunlock(dqp);
272                 }
273                 dqp = nextdqp;
274         }
275         xfs_qm_freelist_unlock(xfs_Gqm);
276
277         /*
278          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
279          * be restarted.
280          */
281         XFS_QM_LOCK(xfs_Gqm);
282         XFS_QM_RELE(xfs_Gqm);
283         if (xfs_Gqm->qm_nrefs == 0) {
284                 xfs_qm_destroy(xfs_Gqm);
285                 xfs_Gqm = NULL;
286         }
287         XFS_QM_UNLOCK(xfs_Gqm);
288 }
289
290 /*
291  * Just destroy the quotainfo structure.
292  */
293 void
294 xfs_qm_unmount_quotadestroy(
295         xfs_mount_t     *mp)
296 {
297         if (mp->m_quotainfo)
298                 xfs_qm_destroy_quotainfo(mp);
299 }
300
301
302 /*
303  * This is called from xfs_mountfs to start quotas and initialize all
304  * necessary data structures like quotainfo.  This is also responsible for
305  * running a quotacheck as necessary.  We are guaranteed that the superblock
306  * is consistently read in at this point.
307  */
308 int
309 xfs_qm_mount_quotas(
310         xfs_mount_t     *mp,
311         int             mfsi_flags)
312 {
313         unsigned long   s;
314         int             error = 0;
315         uint            sbf;
316
317
318         /*
319          * If quotas on realtime volumes is not supported, we disable
320          * quotas immediately.
321          */
322         if (mp->m_sb.sb_rextents) {
323                 cmn_err(CE_NOTE,
324                         "Cannot turn on quotas for realtime filesystem %s",
325                         mp->m_fsname);
326                 mp->m_qflags = 0;
327                 goto write_changes;
328         }
329
330         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
331
332         /*
333          * Allocate the quotainfo structure inside the mount struct, and
334          * create quotainode(s), and change/rev superblock if necessary.
335          */
336         if ((error = xfs_qm_init_quotainfo(mp))) {
337                 /*
338                  * We must turn off quotas.
339                  */
340                 ASSERT(mp->m_quotainfo == NULL);
341                 mp->m_qflags = 0;
342                 goto write_changes;
343         }
344         /*
345          * If any of the quotas are not consistent, do a quotacheck.
346          */
347         if (XFS_QM_NEED_QUOTACHECK(mp) &&
348                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
349                 if ((error = xfs_qm_quotacheck(mp))) {
350                         /* Quotacheck has failed and quotas have
351                          * been disabled.
352                          */
353                         return XFS_ERROR(error);
354                 }
355         }
356         /* 
357          * If one type of quotas is off, then it will lose its
358          * quotachecked status, since we won't be doing accounting for
359          * that type anymore.
360          */
361         if (!XFS_IS_UQUOTA_ON(mp)) {
362                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
363         }
364         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) {
365                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
366         }
367
368  write_changes:
369         /*
370          * We actually don't have to acquire the SB_LOCK at all.
371          * This can only be called from mount, and that's single threaded. XXX
372          */
373         s = XFS_SB_LOCK(mp);
374         sbf = mp->m_sb.sb_qflags;
375         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
376         XFS_SB_UNLOCK(mp, s);
377
378         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
379                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
380                         /*
381                          * We could only have been turning quotas off.
382                          * We aren't in very good shape actually because
383                          * the incore structures are convinced that quotas are
384                          * off, but the on disk superblock doesn't know that !
385                          */
386                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
387                         xfs_fs_cmn_err(CE_ALERT, mp,
388                                 "XFS mount_quotas: Superblock update failed!");
389                 }
390         }
391
392         if (error) {
393                 xfs_fs_cmn_err(CE_WARN, mp,
394                         "Failed to initialize disk quotas.");
395         }
396         return XFS_ERROR(error);
397 }
398
399 /*
400  * Called from the vfsops layer.
401  */
402 int
403 xfs_qm_unmount_quotas(
404         xfs_mount_t     *mp)
405 {
406         xfs_inode_t     *uqp, *gqp;
407         int             error = 0;
408
409         /*
410          * Release the dquots that root inode, et al might be holding,
411          * before we flush quotas and blow away the quotainfo structure.
412          */
413         ASSERT(mp->m_rootip);
414         xfs_qm_dqdetach(mp->m_rootip);
415         if (mp->m_rbmip)
416                 xfs_qm_dqdetach(mp->m_rbmip);
417         if (mp->m_rsumip)
418                 xfs_qm_dqdetach(mp->m_rsumip);
419
420         /*
421          * Flush out the quota inodes.
422          */
423         uqp = gqp = NULL;
424         if (mp->m_quotainfo) {
425                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
426                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
427                         xfs_iflock(uqp);
428                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
429                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
430                         if (unlikely(error == EFSCORRUPTED)) {
431                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
432                                                  XFS_ERRLEVEL_LOW, mp);
433                                 goto out;
434                         }
435                 }
436                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
437                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
438                         xfs_iflock(gqp);
439                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
440                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
441                         if (unlikely(error == EFSCORRUPTED)) {
442                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
443                                                  XFS_ERRLEVEL_LOW, mp);
444                                 goto out;
445                         }
446                 }
447         }
448         if (uqp) {
449                  XFS_PURGE_INODE(uqp);
450                  mp->m_quotainfo->qi_uquotaip = NULL;
451         }
452         if (gqp) {
453                 XFS_PURGE_INODE(gqp);
454                 mp->m_quotainfo->qi_gquotaip = NULL;
455         }
456 out:
457         return XFS_ERROR(error);
458 }
459
460 /*
461  * Flush all dquots of the given file system to disk. The dquots are
462  * _not_ purged from memory here, just their data written to disk.
463  */
464 STATIC int
465 xfs_qm_dqflush_all(
466         xfs_mount_t     *mp,
467         int             flags)
468 {
469         int             recl;
470         xfs_dquot_t     *dqp;
471         int             niters;
472         int             error;
473
474         if (mp->m_quotainfo == NULL)
475                 return 0;
476         niters = 0;
477 again:
478         xfs_qm_mplist_lock(mp);
479         FOREACH_DQUOT_IN_MP(dqp, mp) {
480                 xfs_dqlock(dqp);
481                 if (! XFS_DQ_IS_DIRTY(dqp)) {
482                         xfs_dqunlock(dqp);
483                         continue;
484                 }
485                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
486                 /* XXX a sentinel would be better */
487                 recl = XFS_QI_MPLRECLAIMS(mp);
488                 if (! xfs_qm_dqflock_nowait(dqp)) {
489                         /*
490                          * If we can't grab the flush lock then check
491                          * to see if the dquot has been flushed delayed
492                          * write.  If so, grab its buffer and send it
493                          * out immediately.  We'll be able to acquire
494                          * the flush lock when the I/O completes.
495                          */
496                         xfs_qm_dqflock_pushbuf_wait(dqp);
497                 }
498                 /*
499                  * Let go of the mplist lock. We don't want to hold it
500                  * across a disk write.
501                  */
502                 xfs_qm_mplist_unlock(mp);
503                 error = xfs_qm_dqflush(dqp, flags);
504                 xfs_dqunlock(dqp);
505                 if (error)
506                         return error;
507
508                 xfs_qm_mplist_lock(mp);
509                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
510                         xfs_qm_mplist_unlock(mp);
511                         /* XXX restart limit */
512                         goto again;
513                 }
514         }
515
516         xfs_qm_mplist_unlock(mp);
517         /* return ! busy */
518         return 0;
519 }
520 /*
521  * Release the group dquot pointers the user dquots may be
522  * carrying around as a hint. mplist is locked on entry and exit.
523  */
524 STATIC void
525 xfs_qm_detach_gdquots(
526         xfs_mount_t     *mp)
527 {
528         xfs_dquot_t     *dqp, *gdqp;
529         int             nrecl;
530
531  again:
532         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
533         dqp = XFS_QI_MPLNEXT(mp);
534         while (dqp) {
535                 xfs_dqlock(dqp);
536                 if ((gdqp = dqp->q_gdquot)) {
537                         xfs_dqlock(gdqp);
538                         dqp->q_gdquot = NULL;
539                 }
540                 xfs_dqunlock(dqp);
541
542                 if (gdqp) {
543                         /*
544                          * Can't hold the mplist lock across a dqput.
545                          * XXXmust convert to marker based iterations here.
546                          */
547                         nrecl = XFS_QI_MPLRECLAIMS(mp);
548                         xfs_qm_mplist_unlock(mp);
549                         xfs_qm_dqput(gdqp);
550
551                         xfs_qm_mplist_lock(mp);
552                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
553                                 goto again;
554                 }
555                 dqp = dqp->MPL_NEXT;
556         }
557 }
558
559 /*
560  * Go through all the incore dquots of this file system and take them
561  * off the mplist and hashlist, if the dquot type matches the dqtype
562  * parameter. This is used when turning off quota accounting for
563  * users and/or groups, as well as when the filesystem is unmounting.
564  */
565 STATIC int
566 xfs_qm_dqpurge_int(
567         xfs_mount_t     *mp,
568         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
569 {
570         xfs_dquot_t     *dqp;
571         uint            dqtype;
572         int             nrecl;
573         xfs_dquot_t     *nextdqp;
574         int             nmisses;
575
576         if (mp->m_quotainfo == NULL)
577                 return 0;
578
579         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
580         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
581         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
582
583         xfs_qm_mplist_lock(mp);
584
585         /*
586          * In the first pass through all incore dquots of this filesystem,
587          * we release the group dquot pointers the user dquots may be
588          * carrying around as a hint. We need to do this irrespective of
589          * what's being turned off.
590          */
591         xfs_qm_detach_gdquots(mp);
592
593       again:
594         nmisses = 0;
595         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
596         /*
597          * Try to get rid of all of the unwanted dquots. The idea is to
598          * get them off mplist and hashlist, but leave them on freelist.
599          */
600         dqp = XFS_QI_MPLNEXT(mp);
601         while (dqp) {
602                 /*
603                  * It's OK to look at the type without taking dqlock here.
604                  * We're holding the mplist lock here, and that's needed for
605                  * a dqreclaim.
606                  */
607                 if ((dqp->dq_flags & dqtype) == 0) {
608                         dqp = dqp->MPL_NEXT;
609                         continue;
610                 }
611
612                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
613                         nrecl = XFS_QI_MPLRECLAIMS(mp);
614                         xfs_qm_mplist_unlock(mp);
615                         XFS_DQ_HASH_LOCK(dqp->q_hash);
616                         xfs_qm_mplist_lock(mp);
617
618                         /*
619                          * XXXTheoretically, we can get into a very long
620                          * ping pong game here.
621                          * No one can be adding dquots to the mplist at
622                          * this point, but somebody might be taking things off.
623                          */
624                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
625                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
626                                 goto again;
627                         }
628                 }
629
630                 /*
631                  * Take the dquot off the mplist and hashlist. It may remain on
632                  * freelist in INACTIVE state.
633                  */
634                 nextdqp = dqp->MPL_NEXT;
635                 nmisses += xfs_qm_dqpurge(dqp, flags);
636                 dqp = nextdqp;
637         }
638         xfs_qm_mplist_unlock(mp);
639         return nmisses;
640 }
641
642 int
643 xfs_qm_dqpurge_all(
644         xfs_mount_t     *mp,
645         uint            flags)
646 {
647         int             ndquots;
648
649         /*
650          * Purge the dquot cache.
651          * None of the dquots should really be busy at this point.
652          */
653         if (mp->m_quotainfo) {
654                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
655                         delay(ndquots * 10);
656                 }
657         }
658         return 0;
659 }
660
661 STATIC int
662 xfs_qm_dqattach_one(
663         xfs_inode_t     *ip,
664         xfs_dqid_t      id,
665         uint            type,
666         uint            doalloc,
667         uint            dolock,
668         xfs_dquot_t     *udqhint, /* hint */
669         xfs_dquot_t     **IO_idqpp)
670 {
671         xfs_dquot_t     *dqp;
672         int             error;
673
674         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
675         error = 0;
676         /*
677          * See if we already have it in the inode itself. IO_idqpp is
678          * &i_udquot or &i_gdquot. This made the code look weird, but
679          * made the logic a lot simpler.
680          */
681         if ((dqp = *IO_idqpp)) {
682                 if (dolock)
683                         xfs_dqlock(dqp);
684                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
685                 goto done;
686         }
687
688         /*
689          * udqhint is the i_udquot field in inode, and is non-NULL only
690          * when the type arg is group/project. Its purpose is to save a
691          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
692          * the user dquot.
693          */
694         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
695         if (udqhint && !dolock)
696                 xfs_dqlock(udqhint);
697
698         /*
699          * No need to take dqlock to look at the id.
700          * The ID can't change until it gets reclaimed, and it won't
701          * be reclaimed as long as we have a ref from inode and we hold
702          * the ilock.
703          */
704         if (udqhint &&
705             (dqp = udqhint->q_gdquot) &&
706             (be32_to_cpu(dqp->q_core.d_id) == id)) {
707                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
708                 xfs_dqlock(dqp);
709                 XFS_DQHOLD(dqp);
710                 ASSERT(*IO_idqpp == NULL);
711                 *IO_idqpp = dqp;
712                 if (!dolock) {
713                         xfs_dqunlock(dqp);
714                         xfs_dqunlock(udqhint);
715                 }
716                 goto done;
717         }
718         /*
719          * We can't hold a dquot lock when we call the dqget code.
720          * We'll deadlock in no time, because of (not conforming to)
721          * lock ordering - the inodelock comes before any dquot lock,
722          * and we may drop and reacquire the ilock in xfs_qm_dqget().
723          */
724         if (udqhint)
725                 xfs_dqunlock(udqhint);
726         /*
727          * Find the dquot from somewhere. This bumps the
728          * reference count of dquot and returns it locked.
729          * This can return ENOENT if dquot didn't exist on
730          * disk and we didn't ask it to allocate;
731          * ESRCH if quotas got turned off suddenly.
732          */
733         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
734                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
735                 if (udqhint && dolock)
736                         xfs_dqlock(udqhint);
737                 goto done;
738         }
739
740         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
741         /*
742          * dqget may have dropped and re-acquired the ilock, but it guarantees
743          * that the dquot returned is the one that should go in the inode.
744          */
745         *IO_idqpp = dqp;
746         ASSERT(dqp);
747         ASSERT(XFS_DQ_IS_LOCKED(dqp));
748         if (! dolock) {
749                 xfs_dqunlock(dqp);
750                 goto done;
751         }
752         if (! udqhint)
753                 goto done;
754
755         ASSERT(udqhint);
756         ASSERT(dolock);
757         ASSERT(XFS_DQ_IS_LOCKED(dqp));
758         if (! xfs_qm_dqlock_nowait(udqhint)) {
759                 xfs_dqunlock(dqp);
760                 xfs_dqlock(udqhint);
761                 xfs_dqlock(dqp);
762         }
763       done:
764 #ifdef QUOTADEBUG
765         if (udqhint) {
766                 if (dolock)
767                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
768         }
769         if (! error) {
770                 if (dolock)
771                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
772         }
773 #endif
774         return error;
775 }
776
777
778 /*
779  * Given a udquot and gdquot, attach a ptr to the group dquot in the
780  * udquot as a hint for future lookups. The idea sounds simple, but the
781  * execution isn't, because the udquot might have a group dquot attached
782  * already and getting rid of that gets us into lock ordering constraints.
783  * The process is complicated more by the fact that the dquots may or may not
784  * be locked on entry.
785  */
786 STATIC void
787 xfs_qm_dqattach_grouphint(
788         xfs_dquot_t     *udq,
789         xfs_dquot_t     *gdq,
790         uint            locked)
791 {
792         xfs_dquot_t     *tmp;
793
794 #ifdef QUOTADEBUG
795         if (locked) {
796                 ASSERT(XFS_DQ_IS_LOCKED(udq));
797                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
798         }
799 #endif
800         if (! locked)
801                 xfs_dqlock(udq);
802
803         if ((tmp = udq->q_gdquot)) {
804                 if (tmp == gdq) {
805                         if (! locked)
806                                 xfs_dqunlock(udq);
807                         return;
808                 }
809
810                 udq->q_gdquot = NULL;
811                 /*
812                  * We can't keep any dqlocks when calling dqrele,
813                  * because the freelist lock comes before dqlocks.
814                  */
815                 xfs_dqunlock(udq);
816                 if (locked)
817                         xfs_dqunlock(gdq);
818                 /*
819                  * we took a hard reference once upon a time in dqget,
820                  * so give it back when the udquot no longer points at it
821                  * dqput() does the unlocking of the dquot.
822                  */
823                 xfs_qm_dqrele(tmp);
824
825                 xfs_dqlock(udq);
826                 xfs_dqlock(gdq);
827
828         } else {
829                 ASSERT(XFS_DQ_IS_LOCKED(udq));
830                 if (! locked) {
831                         xfs_dqlock(gdq);
832                 }
833         }
834
835         ASSERT(XFS_DQ_IS_LOCKED(udq));
836         ASSERT(XFS_DQ_IS_LOCKED(gdq));
837         /*
838          * Somebody could have attached a gdquot here,
839          * when we dropped the uqlock. If so, just do nothing.
840          */
841         if (udq->q_gdquot == NULL) {
842                 XFS_DQHOLD(gdq);
843                 udq->q_gdquot = gdq;
844         }
845         if (! locked) {
846                 xfs_dqunlock(gdq);
847                 xfs_dqunlock(udq);
848         }
849 }
850
851
852 /*
853  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
854  * into account.
855  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
856  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
857  * much made this code a complete mess, but it has been pretty useful.
858  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
859  * Inode may get unlocked and relocked in here, and the caller must deal with
860  * the consequences.
861  */
862 int
863 xfs_qm_dqattach(
864         xfs_inode_t     *ip,
865         uint            flags)
866 {
867         xfs_mount_t     *mp = ip->i_mount;
868         uint            nquotas = 0;
869         int             error = 0;
870
871         if ((! XFS_IS_QUOTA_ON(mp)) ||
872             (! XFS_NOT_DQATTACHED(mp, ip)) ||
873             (ip->i_ino == mp->m_sb.sb_uquotino) ||
874             (ip->i_ino == mp->m_sb.sb_gquotino))
875                 return 0;
876
877         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
878                XFS_ISLOCKED_INODE_EXCL(ip));
879
880         if (! (flags & XFS_QMOPT_ILOCKED))
881                 xfs_ilock(ip, XFS_ILOCK_EXCL);
882
883         if (XFS_IS_UQUOTA_ON(mp)) {
884                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
885                                                 flags & XFS_QMOPT_DQALLOC,
886                                                 flags & XFS_QMOPT_DQLOCK,
887                                                 NULL, &ip->i_udquot);
888                 if (error)
889                         goto done;
890                 nquotas++;
891         }
892         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
893         if (XFS_IS_OQUOTA_ON(mp)) {
894                 error = XFS_IS_GQUOTA_ON(mp) ?
895                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
896                                                 flags & XFS_QMOPT_DQALLOC,
897                                                 flags & XFS_QMOPT_DQLOCK,
898                                                 ip->i_udquot, &ip->i_gdquot) :
899                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
900                                                 flags & XFS_QMOPT_DQALLOC,
901                                                 flags & XFS_QMOPT_DQLOCK,
902                                                 ip->i_udquot, &ip->i_gdquot);
903                 /*
904                  * Don't worry about the udquot that we may have
905                  * attached above. It'll get detached, if not already.
906                  */
907                 if (error)
908                         goto done;
909                 nquotas++;
910         }
911
912         /*
913          * Attach this group quota to the user quota as a hint.
914          * This WON'T, in general, result in a thrash.
915          */
916         if (nquotas == 2) {
917                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
918                 ASSERT(ip->i_udquot);
919                 ASSERT(ip->i_gdquot);
920
921                 /*
922                  * We may or may not have the i_udquot locked at this point,
923                  * but this check is OK since we don't depend on the i_gdquot to
924                  * be accurate 100% all the time. It is just a hint, and this
925                  * will succeed in general.
926                  */
927                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
928                         goto done;
929                 /*
930                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
931                  */
932                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
933                                          flags & XFS_QMOPT_DQLOCK);
934         }
935
936       done:
937
938 #ifdef QUOTADEBUG
939         if (! error) {
940                 if (ip->i_udquot) {
941                         if (flags & XFS_QMOPT_DQLOCK)
942                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
943                 }
944                 if (ip->i_gdquot) {
945                         if (flags & XFS_QMOPT_DQLOCK)
946                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
947                 }
948                 if (XFS_IS_UQUOTA_ON(mp))
949                         ASSERT(ip->i_udquot);
950                 if (XFS_IS_OQUOTA_ON(mp))
951                         ASSERT(ip->i_gdquot);
952         }
953 #endif
954
955         if (! (flags & XFS_QMOPT_ILOCKED))
956                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
957
958 #ifdef QUOTADEBUG
959         else
960                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
961 #endif
962         return error;
963 }
964
965 /*
966  * Release dquots (and their references) if any.
967  * The inode should be locked EXCL except when this's called by
968  * xfs_ireclaim.
969  */
970 void
971 xfs_qm_dqdetach(
972         xfs_inode_t     *ip)
973 {
974         if (!(ip->i_udquot || ip->i_gdquot))
975                 return;
976
977         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
978         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
979         if (ip->i_udquot) {
980                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
981                 xfs_qm_dqrele(ip->i_udquot);
982                 ip->i_udquot = NULL;
983         }
984         if (ip->i_gdquot) {
985                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
986                 xfs_qm_dqrele(ip->i_gdquot);
987                 ip->i_gdquot = NULL;
988         }
989 }
990
991 /*
992  * This is called by VFS_SYNC and flags arg determines the caller,
993  * and its motives, as done in xfs_sync.
994  *
995  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
996  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
997  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
998  */
999
1000 int
1001 xfs_qm_sync(
1002         xfs_mount_t     *mp,
1003         int             flags)
1004 {
1005         int             recl, restarts;
1006         xfs_dquot_t     *dqp;
1007         uint            flush_flags;
1008         boolean_t       nowait;
1009         int             error;
1010
1011         restarts = 0;
1012         /*
1013          * We won't block unless we are asked to.
1014          */
1015         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1016
1017   again:
1018         xfs_qm_mplist_lock(mp);
1019         /*
1020          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1021          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1022          * when we have the mplist lock, we know that dquots will be consistent
1023          * as long as we have it locked.
1024          */
1025         if (! XFS_IS_QUOTA_ON(mp)) {
1026                 xfs_qm_mplist_unlock(mp);
1027                 return 0;
1028         }
1029         FOREACH_DQUOT_IN_MP(dqp, mp) {
1030                 /*
1031                  * If this is vfs_sync calling, then skip the dquots that
1032                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1033                  * This is very similar to what xfs_sync does with inodes.
1034                  */
1035                 if (flags & SYNC_BDFLUSH) {
1036                         if (! XFS_DQ_IS_DIRTY(dqp))
1037                                 continue;
1038                 }
1039
1040                 if (nowait) {
1041                         /*
1042                          * Try to acquire the dquot lock. We are NOT out of
1043                          * lock order, but we just don't want to wait for this
1044                          * lock, unless somebody wanted us to.
1045                          */
1046                         if (! xfs_qm_dqlock_nowait(dqp))
1047                                 continue;
1048                 } else {
1049                         xfs_dqlock(dqp);
1050                 }
1051
1052                 /*
1053                  * Now, find out for sure if this dquot is dirty or not.
1054                  */
1055                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1056                         xfs_dqunlock(dqp);
1057                         continue;
1058                 }
1059
1060                 /* XXX a sentinel would be better */
1061                 recl = XFS_QI_MPLRECLAIMS(mp);
1062                 if (! xfs_qm_dqflock_nowait(dqp)) {
1063                         if (nowait) {
1064                                 xfs_dqunlock(dqp);
1065                                 continue;
1066                         }
1067                         /*
1068                          * If we can't grab the flush lock then if the caller
1069                          * really wanted us to give this our best shot, so
1070                          * see if we can give a push to the buffer before we wait
1071                          * on the flush lock. At this point, we know that
1072                          * even though the dquot is being flushed,
1073                          * it has (new) dirty data.
1074                          */
1075                         xfs_qm_dqflock_pushbuf_wait(dqp);
1076                 }
1077                 /*
1078                  * Let go of the mplist lock. We don't want to hold it
1079                  * across a disk write
1080                  */
1081                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1082                 xfs_qm_mplist_unlock(mp);
1083                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1084                 error = xfs_qm_dqflush(dqp, flush_flags);
1085                 xfs_dqunlock(dqp);
1086                 if (error && XFS_FORCED_SHUTDOWN(mp))
1087                         return 0;       /* Need to prevent umount failure */
1088                 else if (error)
1089                         return error;
1090
1091                 xfs_qm_mplist_lock(mp);
1092                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1093                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1094                                 break;
1095
1096                         xfs_qm_mplist_unlock(mp);
1097                         goto again;
1098                 }
1099         }
1100
1101         xfs_qm_mplist_unlock(mp);
1102         return 0;
1103 }
1104
1105
1106 /*
1107  * This initializes all the quota information that's kept in the
1108  * mount structure
1109  */
1110 STATIC int
1111 xfs_qm_init_quotainfo(
1112         xfs_mount_t     *mp)
1113 {
1114         xfs_quotainfo_t *qinf;
1115         int             error;
1116         xfs_dquot_t     *dqp;
1117
1118         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1119
1120         /*
1121          * Tell XQM that we exist as soon as possible.
1122          */
1123         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1124                 return error;
1125         }
1126
1127         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1128
1129         /*
1130          * See if quotainodes are setup, and if not, allocate them,
1131          * and change the superblock accordingly.
1132          */
1133         if ((error = xfs_qm_init_quotainos(mp))) {
1134                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1135                 mp->m_quotainfo = NULL;
1136                 return error;
1137         }
1138
1139         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1140         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1141         qinf->qi_dqreclaims = 0;
1142
1143         /* mutex used to serialize quotaoffs */
1144         mutex_init(&qinf->qi_quotaofflock);
1145
1146         /* Precalc some constants */
1147         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1148         ASSERT(qinf->qi_dqchunklen);
1149         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1150         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1151
1152         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1153
1154         /*
1155          * We try to get the limits from the superuser's limits fields.
1156          * This is quite hacky, but it is standard quota practice.
1157          * We look at the USR dquot with id == 0 first, but if user quotas
1158          * are not enabled we goto the GRP dquot with id == 0.
1159          * We don't really care to keep separate default limits for user
1160          * and group quotas, at least not at this point.
1161          */
1162         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1163                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1164                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1165                                 XFS_DQ_PROJ),
1166                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1167                              &dqp);
1168         if (! error) {
1169                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1170
1171                 /*
1172                  * The warnings and timers set the grace period given to
1173                  * a user or group before he or she can not perform any
1174                  * more writing. If it is zero, a default is used.
1175                  */
1176                 qinf->qi_btimelimit = ddqp->d_btimer ?
1177                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1178                 qinf->qi_itimelimit = ddqp->d_itimer ?
1179                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1180                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1181                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1182                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1183                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1184                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1185                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1186                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1187                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1188                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1189                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1190                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1191                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1192                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1193                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1194  
1195                 /*
1196                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1197                  * we don't want this dquot cached. We haven't done a
1198                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1199                  */
1200                 xfs_qm_dqdestroy(dqp);
1201         } else {
1202                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1203                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1204                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1205                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1206                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1207                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1208         }
1209
1210         return 0;
1211 }
1212
1213
1214 /*
1215  * Gets called when unmounting a filesystem or when all quotas get
1216  * turned off.
1217  * This purges the quota inodes, destroys locks and frees itself.
1218  */
1219 void
1220 xfs_qm_destroy_quotainfo(
1221         xfs_mount_t     *mp)
1222 {
1223         xfs_quotainfo_t *qi;
1224
1225         qi = mp->m_quotainfo;
1226         ASSERT(qi != NULL);
1227         ASSERT(xfs_Gqm != NULL);
1228
1229         /*
1230          * Release the reference that XQM kept, so that we know
1231          * when the XQM structure should be freed. We cannot assume
1232          * that xfs_Gqm is non-null after this point.
1233          */
1234         xfs_qm_rele_quotafs_ref(mp);
1235
1236         spinlock_destroy(&qi->qi_pinlock);
1237         xfs_qm_list_destroy(&qi->qi_dqlist);
1238
1239         if (qi->qi_uquotaip) {
1240                 XFS_PURGE_INODE(qi->qi_uquotaip);
1241                 qi->qi_uquotaip = NULL; /* paranoia */
1242         }
1243         if (qi->qi_gquotaip) {
1244                 XFS_PURGE_INODE(qi->qi_gquotaip);
1245                 qi->qi_gquotaip = NULL;
1246         }
1247         mutex_destroy(&qi->qi_quotaofflock);
1248         kmem_free(qi, sizeof(xfs_quotainfo_t));
1249         mp->m_quotainfo = NULL;
1250 }
1251
1252
1253
1254 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1255
1256 /* ARGSUSED */
1257 STATIC void
1258 xfs_qm_list_init(
1259         xfs_dqlist_t    *list,
1260         char            *str,
1261         int             n)
1262 {
1263         mutex_init(&list->qh_lock);
1264         list->qh_next = NULL;
1265         list->qh_version = 0;
1266         list->qh_nelems = 0;
1267 }
1268
1269 STATIC void
1270 xfs_qm_list_destroy(
1271         xfs_dqlist_t    *list)
1272 {
1273         mutex_destroy(&(list->qh_lock));
1274 }
1275
1276
1277 /*
1278  * Stripped down version of dqattach. This doesn't attach, or even look at the
1279  * dquots attached to the inode. The rationale is that there won't be any
1280  * attached at the time this is called from quotacheck.
1281  */
1282 STATIC int
1283 xfs_qm_dqget_noattach(
1284         xfs_inode_t     *ip,
1285         xfs_dquot_t     **O_udqpp,
1286         xfs_dquot_t     **O_gdqpp)
1287 {
1288         int             error;
1289         xfs_mount_t     *mp;
1290         xfs_dquot_t     *udqp, *gdqp;
1291
1292         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1293         mp = ip->i_mount;
1294         udqp = NULL;
1295         gdqp = NULL;
1296
1297         if (XFS_IS_UQUOTA_ON(mp)) {
1298                 ASSERT(ip->i_udquot == NULL);
1299                 /*
1300                  * We want the dquot allocated if it doesn't exist.
1301                  */
1302                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1303                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1304                                          &udqp))) {
1305                         /*
1306                          * Shouldn't be able to turn off quotas here.
1307                          */
1308                         ASSERT(error != ESRCH);
1309                         ASSERT(error != ENOENT);
1310                         return error;
1311                 }
1312                 ASSERT(udqp);
1313         }
1314
1315         if (XFS_IS_OQUOTA_ON(mp)) {
1316                 ASSERT(ip->i_gdquot == NULL);
1317                 if (udqp)
1318                         xfs_dqunlock(udqp);
1319                 error = XFS_IS_GQUOTA_ON(mp) ?
1320                                 xfs_qm_dqget(mp, ip,
1321                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1322                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1323                                              &gdqp) :
1324                                 xfs_qm_dqget(mp, ip,
1325                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1326                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1327                                              &gdqp);
1328                 if (error) {
1329                         if (udqp)
1330                                 xfs_qm_dqrele(udqp);
1331                         ASSERT(error != ESRCH);
1332                         ASSERT(error != ENOENT);
1333                         return error;
1334                 }
1335                 ASSERT(gdqp);
1336
1337                 /* Reacquire the locks in the right order */
1338                 if (udqp) {
1339                         if (! xfs_qm_dqlock_nowait(udqp)) {
1340                                 xfs_dqunlock(gdqp);
1341                                 xfs_dqlock(udqp);
1342                                 xfs_dqlock(gdqp);
1343                         }
1344                 }
1345         }
1346
1347         *O_udqpp = udqp;
1348         *O_gdqpp = gdqp;
1349
1350 #ifdef QUOTADEBUG
1351         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1352         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1353 #endif
1354         return 0;
1355 }
1356
1357 /*
1358  * Create an inode and return with a reference already taken, but unlocked
1359  * This is how we create quota inodes
1360  */
1361 STATIC int
1362 xfs_qm_qino_alloc(
1363         xfs_mount_t     *mp,
1364         xfs_inode_t     **ip,
1365         __int64_t       sbfields,
1366         uint            flags)
1367 {
1368         xfs_trans_t     *tp;
1369         int             error;
1370         unsigned long   s;
1371         int             committed;
1372
1373         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1374         if ((error = xfs_trans_reserve(tp,
1375                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1376                                       XFS_CREATE_LOG_RES(mp), 0,
1377                                       XFS_TRANS_PERM_LOG_RES,
1378                                       XFS_CREATE_LOG_COUNT))) {
1379                 xfs_trans_cancel(tp, 0);
1380                 return error;
1381         }
1382
1383         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1384                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1385                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1386                                  XFS_TRANS_ABORT);
1387                 return error;
1388         }
1389
1390         /*
1391          * Keep an extra reference to this quota inode. This inode is
1392          * locked exclusively and joined to the transaction already.
1393          */
1394         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1395         VN_HOLD(XFS_ITOV((*ip)));
1396
1397         /*
1398          * Make the changes in the superblock, and log those too.
1399          * sbfields arg may contain fields other than *QUOTINO;
1400          * VERSIONNUM for example.
1401          */
1402         s = XFS_SB_LOCK(mp);
1403         if (flags & XFS_QMOPT_SBVERSION) {
1404 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1405                 unsigned oldv = mp->m_sb.sb_versionnum;
1406 #endif
1407                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1408                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1409                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1410                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1411                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1412
1413                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1414                 mp->m_sb.sb_uquotino = NULLFSINO;
1415                 mp->m_sb.sb_gquotino = NULLFSINO;
1416
1417                 /* qflags will get updated _after_ quotacheck */
1418                 mp->m_sb.sb_qflags = 0;
1419 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1420                 cmn_err(CE_NOTE,
1421                         "Old superblock version %x, converting to %x.",
1422                         oldv, mp->m_sb.sb_versionnum);
1423 #endif
1424         }
1425         if (flags & XFS_QMOPT_UQUOTA)
1426                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1427         else
1428                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1429         XFS_SB_UNLOCK(mp, s);
1430         xfs_mod_sb(tp, sbfields);
1431
1432         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1433                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1434                 return error;
1435         }
1436         return 0;
1437 }
1438
1439
1440 STATIC int
1441 xfs_qm_reset_dqcounts(
1442         xfs_mount_t     *mp,
1443         xfs_buf_t       *bp,
1444         xfs_dqid_t      id,
1445         uint            type)
1446 {
1447         xfs_disk_dquot_t        *ddq;
1448         int                     j;
1449
1450         xfs_buftrace("RESET DQUOTS", bp);
1451         /*
1452          * Reset all counters and timers. They'll be
1453          * started afresh by xfs_qm_quotacheck.
1454          */
1455 #ifdef DEBUG
1456         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1457         do_div(j, sizeof(xfs_dqblk_t));
1458         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1459 #endif
1460         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1461         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1462                 /*
1463                  * Do a sanity check, and if needed, repair the dqblk. Don't
1464                  * output any warnings because it's perfectly possible to
1465                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1466                  */
1467                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1468                                       "xfs_quotacheck");
1469                 ddq->d_bcount = 0;
1470                 ddq->d_icount = 0;
1471                 ddq->d_rtbcount = 0;
1472                 ddq->d_btimer = 0;
1473                 ddq->d_itimer = 0;
1474                 ddq->d_rtbtimer = 0;
1475                 ddq->d_bwarns = 0;
1476                 ddq->d_iwarns = 0;
1477                 ddq->d_rtbwarns = 0;
1478                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1479         }
1480
1481         return 0;
1482 }
1483
1484 STATIC int
1485 xfs_qm_dqiter_bufs(
1486         xfs_mount_t     *mp,
1487         xfs_dqid_t      firstid,
1488         xfs_fsblock_t   bno,
1489         xfs_filblks_t   blkcnt,
1490         uint            flags)
1491 {
1492         xfs_buf_t       *bp;
1493         int             error;
1494         int             notcommitted;
1495         int             incr;
1496         int             type;
1497
1498         ASSERT(blkcnt > 0);
1499         notcommitted = 0;
1500         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1501                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1502         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1503                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1504         error = 0;
1505
1506         /*
1507          * Blkcnt arg can be a very big number, and might even be
1508          * larger than the log itself. So, we have to break it up into
1509          * manageable-sized transactions.
1510          * Note that we don't start a permanent transaction here; we might
1511          * not be able to get a log reservation for the whole thing up front,
1512          * and we don't really care to either, because we just discard
1513          * everything if we were to crash in the middle of this loop.
1514          */
1515         while (blkcnt--) {
1516                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1517                               XFS_FSB_TO_DADDR(mp, bno),
1518                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1519                 if (error)
1520                         break;
1521
1522                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1523                 xfs_bdwrite(mp, bp);
1524                 /*
1525                  * goto the next block.
1526                  */
1527                 bno++;
1528                 firstid += XFS_QM_DQPERBLK(mp);
1529         }
1530         return error;
1531 }
1532
1533 /*
1534  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1535  * caller supplied function for every chunk of dquots that we find.
1536  */
1537 STATIC int
1538 xfs_qm_dqiterate(
1539         xfs_mount_t     *mp,
1540         xfs_inode_t     *qip,
1541         uint            flags)
1542 {
1543         xfs_bmbt_irec_t         *map;
1544         int                     i, nmaps;       /* number of map entries */
1545         int                     error;          /* return value */
1546         xfs_fileoff_t           lblkno;
1547         xfs_filblks_t           maxlblkcnt;
1548         xfs_dqid_t              firstid;
1549         xfs_fsblock_t           rablkno;
1550         xfs_filblks_t           rablkcnt;
1551
1552         error = 0;
1553         /*
1554          * This looks racy, but we can't keep an inode lock across a
1555          * trans_reserve. But, this gets called during quotacheck, and that
1556          * happens only at mount time which is single threaded.
1557          */
1558         if (qip->i_d.di_nblocks == 0)
1559                 return 0;
1560
1561         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1562
1563         lblkno = 0;
1564         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1565         do {
1566                 nmaps = XFS_DQITER_MAP_SIZE;
1567                 /*
1568                  * We aren't changing the inode itself. Just changing
1569                  * some of its data. No new blocks are added here, and
1570                  * the inode is never added to the transaction.
1571                  */
1572                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1573                 error = xfs_bmapi(NULL, qip, lblkno,
1574                                   maxlblkcnt - lblkno,
1575                                   XFS_BMAPI_METADATA,
1576                                   NULL,
1577                                   0, map, &nmaps, NULL, NULL);
1578                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1579                 if (error)
1580                         break;
1581
1582                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1583                 for (i = 0; i < nmaps; i++) {
1584                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1585                         ASSERT(map[i].br_blockcount);
1586
1587
1588                         lblkno += map[i].br_blockcount;
1589
1590                         if (map[i].br_startblock == HOLESTARTBLOCK)
1591                                 continue;
1592
1593                         firstid = (xfs_dqid_t) map[i].br_startoff *
1594                                 XFS_QM_DQPERBLK(mp);
1595                         /*
1596                          * Do a read-ahead on the next extent.
1597                          */
1598                         if ((i+1 < nmaps) &&
1599                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1600                                 rablkcnt =  map[i+1].br_blockcount;
1601                                 rablkno = map[i+1].br_startblock;
1602                                 while (rablkcnt--) {
1603                                         xfs_baread(mp->m_ddev_targp,
1604                                                XFS_FSB_TO_DADDR(mp, rablkno),
1605                                                (int)XFS_QI_DQCHUNKLEN(mp));
1606                                         rablkno++;
1607                                 }
1608                         }
1609                         /*
1610                          * Iterate thru all the blks in the extent and
1611                          * reset the counters of all the dquots inside them.
1612                          */
1613                         if ((error = xfs_qm_dqiter_bufs(mp,
1614                                                        firstid,
1615                                                        map[i].br_startblock,
1616                                                        map[i].br_blockcount,
1617                                                        flags))) {
1618                                 break;
1619                         }
1620                 }
1621
1622                 if (error)
1623                         break;
1624         } while (nmaps > 0);
1625
1626         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1627
1628         return error;
1629 }
1630
1631 /*
1632  * Called by dqusage_adjust in doing a quotacheck.
1633  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1634  * this updates its incore copy as well as the buffer copy. This is
1635  * so that once the quotacheck is done, we can just log all the buffers,
1636  * as opposed to logging numerous updates to individual dquots.
1637  */
1638 STATIC void
1639 xfs_qm_quotacheck_dqadjust(
1640         xfs_dquot_t             *dqp,
1641         xfs_qcnt_t              nblks,
1642         xfs_qcnt_t              rtblks)
1643 {
1644         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1645         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1646         /*
1647          * Adjust the inode count and the block count to reflect this inode's
1648          * resource usage.
1649          */
1650         be64_add(&dqp->q_core.d_icount, 1);
1651         dqp->q_res_icount++;
1652         if (nblks) {
1653                 be64_add(&dqp->q_core.d_bcount, nblks);
1654                 dqp->q_res_bcount += nblks;
1655         }
1656         if (rtblks) {
1657                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1658                 dqp->q_res_rtbcount += rtblks;
1659         }
1660
1661         /*
1662          * Set default limits, adjust timers (since we changed usages)
1663          */
1664         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1665                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1666                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1667         }
1668
1669         dqp->dq_flags |= XFS_DQ_DIRTY;
1670 }
1671
1672 STATIC int
1673 xfs_qm_get_rtblks(
1674         xfs_inode_t     *ip,
1675         xfs_qcnt_t      *O_rtblks)
1676 {
1677         xfs_filblks_t   rtblks;                 /* total rt blks */
1678         xfs_extnum_t    idx;                    /* extent record index */
1679         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1680         xfs_extnum_t    nextents;               /* number of extent entries */
1681         int             error;
1682
1683         ASSERT(XFS_IS_REALTIME_INODE(ip));
1684         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1685         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1686                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1687                         return error;
1688         }
1689         rtblks = 0;
1690         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1691         for (idx = 0; idx < nextents; idx++)
1692                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1693         *O_rtblks = (xfs_qcnt_t)rtblks;
1694         return 0;
1695 }
1696
1697 /*
1698  * callback routine supplied to bulkstat(). Given an inumber, find its
1699  * dquots and update them to account for resources taken by that inode.
1700  */
1701 /* ARGSUSED */
1702 STATIC int
1703 xfs_qm_dqusage_adjust(
1704         xfs_mount_t     *mp,            /* mount point for filesystem */
1705         xfs_ino_t       ino,            /* inode number to get data for */
1706         void            __user *buffer, /* not used */
1707         int             ubsize,         /* not used */
1708         void            *private_data,  /* not used */
1709         xfs_daddr_t     bno,            /* starting block of inode cluster */
1710         int             *ubused,        /* not used */
1711         void            *dip,           /* on-disk inode pointer (not used) */
1712         int             *res)           /* result code value */
1713 {
1714         xfs_inode_t     *ip;
1715         xfs_dquot_t     *udqp, *gdqp;
1716         xfs_qcnt_t      nblks, rtblks;
1717         int             error;
1718
1719         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1720
1721         /*
1722          * rootino must have its resources accounted for, not so with the quota
1723          * inodes.
1724          */
1725         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1726                 *res = BULKSTAT_RV_NOTHING;
1727                 return XFS_ERROR(EINVAL);
1728         }
1729
1730         /*
1731          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1732          * interface expects the inode to be exclusively locked because that's
1733          * the case in all other instances. It's OK that we do this because
1734          * quotacheck is done only at mount time.
1735          */
1736         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1737                 *res = BULKSTAT_RV_NOTHING;
1738                 return error;
1739         }
1740
1741         if (ip->i_d.di_mode == 0) {
1742                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1743                 *res = BULKSTAT_RV_NOTHING;
1744                 return XFS_ERROR(ENOENT);
1745         }
1746
1747         /*
1748          * Obtain the locked dquots. In case of an error (eg. allocation
1749          * fails for ENOSPC), we return the negative of the error number
1750          * to bulkstat, so that it can get propagated to quotacheck() and
1751          * making us disable quotas for the file system.
1752          */
1753         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1754                 xfs_iput(ip, XFS_ILOCK_EXCL);
1755                 *res = BULKSTAT_RV_GIVEUP;
1756                 return error;
1757         }
1758
1759         rtblks = 0;
1760         if (! XFS_IS_REALTIME_INODE(ip)) {
1761                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1762         } else {
1763                 /*
1764                  * Walk thru the extent list and count the realtime blocks.
1765                  */
1766                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1767                         xfs_iput(ip, XFS_ILOCK_EXCL);
1768                         if (udqp)
1769                                 xfs_qm_dqput(udqp);
1770                         if (gdqp)
1771                                 xfs_qm_dqput(gdqp);
1772                         *res = BULKSTAT_RV_GIVEUP;
1773                         return error;
1774                 }
1775                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1776         }
1777         ASSERT(ip->i_delayed_blks == 0);
1778
1779         /*
1780          * We can't release the inode while holding its dquot locks.
1781          * The inode can go into inactive and might try to acquire the dquotlocks.
1782          * So, just unlock here and do a vn_rele at the end.
1783          */
1784         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1785
1786         /*
1787          * Add the (disk blocks and inode) resources occupied by this
1788          * inode to its dquots. We do this adjustment in the incore dquot,
1789          * and also copy the changes to its buffer.
1790          * We don't care about putting these changes in a transaction
1791          * envelope because if we crash in the middle of a 'quotacheck'
1792          * we have to start from the beginning anyway.
1793          * Once we're done, we'll log all the dquot bufs.
1794          *
1795          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1796          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1797          */
1798         if (XFS_IS_UQUOTA_ON(mp)) {
1799                 ASSERT(udqp);
1800                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1801                 xfs_qm_dqput(udqp);
1802         }
1803         if (XFS_IS_OQUOTA_ON(mp)) {
1804                 ASSERT(gdqp);
1805                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1806                 xfs_qm_dqput(gdqp);
1807         }
1808         /*
1809          * Now release the inode. This will send it to 'inactive', and
1810          * possibly even free blocks.
1811          */
1812         VN_RELE(XFS_ITOV(ip));
1813
1814         /*
1815          * Goto next inode.
1816          */
1817         *res = BULKSTAT_RV_DIDONE;
1818         return 0;
1819 }
1820
1821 /*
1822  * Walk thru all the filesystem inodes and construct a consistent view
1823  * of the disk quota world. If the quotacheck fails, disable quotas.
1824  */
1825 int
1826 xfs_qm_quotacheck(
1827         xfs_mount_t     *mp)
1828 {
1829         int             done, count, error;
1830         xfs_ino_t       lastino;
1831         size_t          structsz;
1832         xfs_inode_t     *uip, *gip;
1833         uint            flags;
1834
1835         count = INT_MAX;
1836         structsz = 1;
1837         lastino = 0;
1838         flags = 0;
1839
1840         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1841         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1842
1843         /*
1844          * There should be no cached dquots. The (simplistic) quotacheck
1845          * algorithm doesn't like that.
1846          */
1847         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1848
1849         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1850
1851         /*
1852          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1853          * their counters to zero. We need a clean slate.
1854          * We don't log our changes till later.
1855          */
1856         if ((uip = XFS_QI_UQIP(mp))) {
1857                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1858                         goto error_return;
1859                 flags |= XFS_UQUOTA_CHKD;
1860         }
1861
1862         if ((gip = XFS_QI_GQIP(mp))) {
1863                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1864                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1865                         goto error_return;
1866                 flags |= XFS_OQUOTA_CHKD;
1867         }
1868
1869         do {
1870                 /*
1871                  * Iterate thru all the inodes in the file system,
1872                  * adjusting the corresponding dquot counters in core.
1873                  */
1874                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1875                                      xfs_qm_dqusage_adjust, NULL,
1876                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1877                         break;
1878
1879         } while (! done);
1880
1881         /*
1882          * We can get this error if we couldn't do a dquot allocation inside
1883          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1884          * dirty dquots that might be cached, we just want to get rid of them
1885          * and turn quotaoff. The dquots won't be attached to any of the inodes
1886          * at this point (because we intentionally didn't in dqget_noattach).
1887          */
1888         if (error) {
1889                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1890                 goto error_return;
1891         }
1892         /*
1893          * We've made all the changes that we need to make incore.
1894          * Now flush_them down to disk buffers.
1895          */
1896         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1897
1898         /*
1899          * We didn't log anything, because if we crashed, we'll have to
1900          * start the quotacheck from scratch anyway. However, we must make
1901          * sure that our dquot changes are secure before we put the
1902          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1903          * flush.
1904          */
1905         XFS_bflush(mp->m_ddev_targp);
1906
1907         /*
1908          * If one type of quotas is off, then it will lose its
1909          * quotachecked status, since we won't be doing accounting for
1910          * that type anymore.
1911          */
1912         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1913         mp->m_qflags |= flags;
1914
1915         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1916
1917  error_return:
1918         if (error) {
1919                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1920                         "Disabling quotas.",
1921                         mp->m_fsname, error);
1922                 /*
1923                  * We must turn off quotas.
1924                  */
1925                 ASSERT(mp->m_quotainfo != NULL);
1926                 ASSERT(xfs_Gqm != NULL);
1927                 xfs_qm_destroy_quotainfo(mp);
1928                 (void)xfs_mount_reset_sbqflags(mp);
1929         } else {
1930                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1931         }
1932         return (error);
1933 }
1934
1935 /*
1936  * This is called after the superblock has been read in and we're ready to
1937  * iget the quota inodes.
1938  */
1939 STATIC int
1940 xfs_qm_init_quotainos(
1941         xfs_mount_t     *mp)
1942 {
1943         xfs_inode_t     *uip, *gip;
1944         int             error;
1945         __int64_t       sbflags;
1946         uint            flags;
1947
1948         ASSERT(mp->m_quotainfo);
1949         uip = gip = NULL;
1950         sbflags = 0;
1951         flags = 0;
1952
1953         /*
1954          * Get the uquota and gquota inodes
1955          */
1956         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1957                 if (XFS_IS_UQUOTA_ON(mp) &&
1958                     mp->m_sb.sb_uquotino != NULLFSINO) {
1959                         ASSERT(mp->m_sb.sb_uquotino > 0);
1960                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1961                                              0, 0, &uip, 0)))
1962                                 return XFS_ERROR(error);
1963                 }
1964                 if (XFS_IS_OQUOTA_ON(mp) &&
1965                     mp->m_sb.sb_gquotino != NULLFSINO) {
1966                         ASSERT(mp->m_sb.sb_gquotino > 0);
1967                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1968                                              0, 0, &gip, 0))) {
1969                                 if (uip)
1970                                         VN_RELE(XFS_ITOV(uip));
1971                                 return XFS_ERROR(error);
1972                         }
1973                 }
1974         } else {
1975                 flags |= XFS_QMOPT_SBVERSION;
1976                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1977                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1978         }
1979
1980         /*
1981          * Create the two inodes, if they don't exist already. The changes
1982          * made above will get added to a transaction and logged in one of
1983          * the qino_alloc calls below.  If the device is readonly,
1984          * temporarily switch to read-write to do this.
1985          */
1986         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1987                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1988                                               sbflags | XFS_SB_UQUOTINO,
1989                                               flags | XFS_QMOPT_UQUOTA)))
1990                         return XFS_ERROR(error);
1991
1992                 flags &= ~XFS_QMOPT_SBVERSION;
1993         }
1994         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1995                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1996                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1997                 error = xfs_qm_qino_alloc(mp, &gip,
1998                                           sbflags | XFS_SB_GQUOTINO, flags);
1999                 if (error) {
2000                         if (uip)
2001                                 VN_RELE(XFS_ITOV(uip));
2002
2003                         return XFS_ERROR(error);
2004                 }
2005         }
2006
2007         XFS_QI_UQIP(mp) = uip;
2008         XFS_QI_GQIP(mp) = gip;
2009
2010         return 0;
2011 }
2012
2013
2014 /*
2015  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2016  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2017  * favor the lookup function ...
2018  * XXXsup merge this with qm_reclaim_one().
2019  */
2020 STATIC int
2021 xfs_qm_shake_freelist(
2022         int howmany)
2023 {
2024         int             nreclaimed;
2025         xfs_dqhash_t    *hash;
2026         xfs_dquot_t     *dqp, *nextdqp;
2027         int             restarts;
2028         int             nflushes;
2029
2030         if (howmany <= 0)
2031                 return 0;
2032
2033         nreclaimed = 0;
2034         restarts = 0;
2035         nflushes = 0;
2036
2037 #ifdef QUOTADEBUG
2038         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2039 #endif
2040         /* lock order is : hashchainlock, freelistlock, mplistlock */
2041  tryagain:
2042         xfs_qm_freelist_lock(xfs_Gqm);
2043
2044         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2045              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2046               nreclaimed < howmany); ) {
2047                 xfs_dqlock(dqp);
2048
2049                 /*
2050                  * We are racing with dqlookup here. Naturally we don't
2051                  * want to reclaim a dquot that lookup wants.
2052                  */
2053                 if (dqp->dq_flags & XFS_DQ_WANT) {
2054                         xfs_dqunlock(dqp);
2055                         xfs_qm_freelist_unlock(xfs_Gqm);
2056                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2057                                 return nreclaimed;
2058                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2059                         goto tryagain;
2060                 }
2061
2062                 /*
2063                  * If the dquot is inactive, we are assured that it is
2064                  * not on the mplist or the hashlist, and that makes our
2065                  * life easier.
2066                  */
2067                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2068                         ASSERT(dqp->q_mount == NULL);
2069                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2070                         ASSERT(dqp->HL_PREVP == NULL);
2071                         ASSERT(dqp->MPL_PREVP == NULL);
2072                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2073                         nextdqp = dqp->dq_flnext;
2074                         goto off_freelist;
2075                 }
2076
2077                 ASSERT(dqp->MPL_PREVP);
2078                 /*
2079                  * Try to grab the flush lock. If this dquot is in the process of
2080                  * getting flushed to disk, we don't want to reclaim it.
2081                  */
2082                 if (! xfs_qm_dqflock_nowait(dqp)) {
2083                         xfs_dqunlock(dqp);
2084                         dqp = dqp->dq_flnext;
2085                         continue;
2086                 }
2087
2088                 /*
2089                  * We have the flush lock so we know that this is not in the
2090                  * process of being flushed. So, if this is dirty, flush it
2091                  * DELWRI so that we don't get a freelist infested with
2092                  * dirty dquots.
2093                  */
2094                 if (XFS_DQ_IS_DIRTY(dqp)) {
2095                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2096                         /*
2097                          * We flush it delayed write, so don't bother
2098                          * releasing the mplock.
2099                          */
2100                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2101                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2102                         dqp = dqp->dq_flnext;
2103                         continue;
2104                 }
2105                 /*
2106                  * We're trying to get the hashlock out of order. This races
2107                  * with dqlookup; so, we giveup and goto the next dquot if
2108                  * we couldn't get the hashlock. This way, we won't starve
2109                  * a dqlookup process that holds the hashlock that is
2110                  * waiting for the freelist lock.
2111                  */
2112                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2113                         xfs_dqfunlock(dqp);
2114                         xfs_dqunlock(dqp);
2115                         dqp = dqp->dq_flnext;
2116                         continue;
2117                 }
2118                 /*
2119                  * This races with dquot allocation code as well as dqflush_all
2120                  * and reclaim code. So, if we failed to grab the mplist lock,
2121                  * giveup everything and start over.
2122                  */
2123                 hash = dqp->q_hash;
2124                 ASSERT(hash);
2125                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2126                         /* XXX put a sentinel so that we can come back here */
2127                         xfs_dqfunlock(dqp);
2128                         xfs_dqunlock(dqp);
2129                         XFS_DQ_HASH_UNLOCK(hash);
2130                         xfs_qm_freelist_unlock(xfs_Gqm);
2131                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2132                                 return nreclaimed;
2133                         goto tryagain;
2134                 }
2135                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2136 #ifdef QUOTADEBUG
2137                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2138                         dqp, be32_to_cpu(dqp->q_core.d_id));
2139 #endif
2140                 ASSERT(dqp->q_nrefs == 0);
2141                 nextdqp = dqp->dq_flnext;
2142                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2143                 XQM_HASHLIST_REMOVE(hash, dqp);
2144                 xfs_dqfunlock(dqp);
2145                 xfs_qm_mplist_unlock(dqp->q_mount);
2146                 XFS_DQ_HASH_UNLOCK(hash);
2147
2148  off_freelist:
2149                 XQM_FREELIST_REMOVE(dqp);
2150                 xfs_dqunlock(dqp);
2151                 nreclaimed++;
2152                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2153                 xfs_qm_dqdestroy(dqp);
2154                 dqp = nextdqp;
2155         }
2156         xfs_qm_freelist_unlock(xfs_Gqm);
2157         return nreclaimed;
2158 }
2159
2160
2161 /*
2162  * The kmem_shake interface is invoked when memory is running low.
2163  */
2164 /* ARGSUSED */
2165 STATIC int
2166 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2167 {
2168         int     ndqused, nfree, n;
2169
2170         if (!kmem_shake_allow(gfp_mask))
2171                 return 0;
2172         if (!xfs_Gqm)
2173                 return 0;
2174
2175         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2176         /* incore dquots in all f/s's */
2177         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2178
2179         ASSERT(ndqused >= 0);
2180
2181         if (nfree <= ndqused && nfree < ndquot)
2182                 return 0;
2183
2184         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2185         n = nfree - ndqused - ndquot;           /* # over target */
2186
2187         return xfs_qm_shake_freelist(MAX(nfree, n));
2188 }
2189
2190
2191 /*
2192  * Just pop the least recently used dquot off the freelist and
2193  * recycle it. The returned dquot is locked.
2194  */
2195 STATIC xfs_dquot_t *
2196 xfs_qm_dqreclaim_one(void)
2197 {
2198         xfs_dquot_t     *dqpout;
2199         xfs_dquot_t     *dqp;
2200         int             restarts;
2201         int             nflushes;
2202
2203         restarts = 0;
2204         dqpout = NULL;
2205         nflushes = 0;
2206
2207         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2208  startagain:
2209         xfs_qm_freelist_lock(xfs_Gqm);
2210
2211         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2212                 xfs_dqlock(dqp);
2213
2214                 /*
2215                  * We are racing with dqlookup here. Naturally we don't
2216                  * want to reclaim a dquot that lookup wants. We release the
2217                  * freelist lock and start over, so that lookup will grab
2218                  * both the dquot and the freelistlock.
2219                  */
2220                 if (dqp->dq_flags & XFS_DQ_WANT) {
2221                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2222                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2223                         xfs_dqunlock(dqp);
2224                         xfs_qm_freelist_unlock(xfs_Gqm);
2225                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2226                                 return NULL;
2227                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2228                         goto startagain;
2229                 }
2230
2231                 /*
2232                  * If the dquot is inactive, we are assured that it is
2233                  * not on the mplist or the hashlist, and that makes our
2234                  * life easier.
2235                  */
2236                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2237                         ASSERT(dqp->q_mount == NULL);
2238                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2239                         ASSERT(dqp->HL_PREVP == NULL);
2240                         ASSERT(dqp->MPL_PREVP == NULL);
2241                         XQM_FREELIST_REMOVE(dqp);
2242                         xfs_dqunlock(dqp);
2243                         dqpout = dqp;
2244                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2245                         break;
2246                 }
2247
2248                 ASSERT(dqp->q_hash);
2249                 ASSERT(dqp->MPL_PREVP);
2250
2251                 /*
2252                  * Try to grab the flush lock. If this dquot is in the process of
2253                  * getting flushed to disk, we don't want to reclaim it.
2254                  */
2255                 if (! xfs_qm_dqflock_nowait(dqp)) {
2256                         xfs_dqunlock(dqp);
2257                         continue;
2258                 }
2259
2260                 /*
2261                  * We have the flush lock so we know that this is not in the
2262                  * process of being flushed. So, if this is dirty, flush it
2263                  * DELWRI so that we don't get a freelist infested with
2264                  * dirty dquots.
2265                  */
2266                 if (XFS_DQ_IS_DIRTY(dqp)) {
2267                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2268                         /*
2269                          * We flush it delayed write, so don't bother
2270                          * releasing the freelist lock.
2271                          */
2272                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2273                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2274                         continue;
2275                 }
2276
2277                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2278                         xfs_dqfunlock(dqp);
2279                         xfs_dqunlock(dqp);
2280                         continue;
2281                 }
2282
2283                 if (! xfs_qm_dqhashlock_nowait(dqp))
2284                         goto mplistunlock;
2285
2286                 ASSERT(dqp->q_nrefs == 0);
2287                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2288                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2289                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2290                 XQM_FREELIST_REMOVE(dqp);
2291                 dqpout = dqp;
2292                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2293  mplistunlock:
2294                 xfs_qm_mplist_unlock(dqp->q_mount);
2295                 xfs_dqfunlock(dqp);
2296                 xfs_dqunlock(dqp);
2297                 if (dqpout)
2298                         break;
2299         }
2300
2301         xfs_qm_freelist_unlock(xfs_Gqm);
2302         return dqpout;
2303 }
2304
2305
2306 /*------------------------------------------------------------------*/
2307
2308 /*
2309  * Return a new incore dquot. Depending on the number of
2310  * dquots in the system, we either allocate a new one on the kernel heap,
2311  * or reclaim a free one.
2312  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2313  * to reclaim an existing one from the freelist.
2314  */
2315 boolean_t
2316 xfs_qm_dqalloc_incore(
2317         xfs_dquot_t **O_dqpp)
2318 {
2319         xfs_dquot_t     *dqp;
2320
2321         /*
2322          * Check against high water mark to see if we want to pop
2323          * a nincompoop dquot off the freelist.
2324          */
2325         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2326                 /*
2327                  * Try to recycle a dquot from the freelist.
2328                  */
2329                 if ((dqp = xfs_qm_dqreclaim_one())) {
2330                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2331                         /*
2332                          * Just zero the core here. The rest will get
2333                          * reinitialized by caller. XXX we shouldn't even
2334                          * do this zero ...
2335                          */
2336                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2337                         *O_dqpp = dqp;
2338                         return B_FALSE;
2339                 }
2340                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2341         }
2342
2343         /*
2344          * Allocate a brand new dquot on the kernel heap and return it
2345          * to the caller to initialize.
2346          */
2347         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2348         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2349         atomic_inc(&xfs_Gqm->qm_totaldquots);
2350
2351         return B_TRUE;
2352 }
2353
2354
2355 /*
2356  * Start a transaction and write the incore superblock changes to
2357  * disk. flags parameter indicates which fields have changed.
2358  */
2359 int
2360 xfs_qm_write_sb_changes(
2361         xfs_mount_t     *mp,
2362         __int64_t       flags)
2363 {
2364         xfs_trans_t     *tp;
2365         int             error;
2366
2367 #ifdef QUOTADEBUG
2368         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2369 #endif
2370         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2371         if ((error = xfs_trans_reserve(tp, 0,
2372                                       mp->m_sb.sb_sectsize + 128, 0,
2373                                       0,
2374                                       XFS_DEFAULT_LOG_COUNT))) {
2375                 xfs_trans_cancel(tp, 0);
2376                 return error;
2377         }
2378
2379         xfs_mod_sb(tp, flags);
2380         (void) xfs_trans_commit(tp, 0);
2381
2382         return 0;
2383 }
2384
2385
2386 /* --------------- utility functions for vnodeops ---------------- */
2387
2388
2389 /*
2390  * Given an inode, a uid and gid (from cred_t) make sure that we have
2391  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2392  * quotas by creating this file.
2393  * This also attaches dquot(s) to the given inode after locking it,
2394  * and returns the dquots corresponding to the uid and/or gid.
2395  *
2396  * in   : inode (unlocked)
2397  * out  : udquot, gdquot with references taken and unlocked
2398  */
2399 int
2400 xfs_qm_vop_dqalloc(
2401         xfs_mount_t     *mp,
2402         xfs_inode_t     *ip,
2403         uid_t           uid,
2404         gid_t           gid,
2405         prid_t          prid,
2406         uint            flags,
2407         xfs_dquot_t     **O_udqpp,
2408         xfs_dquot_t     **O_gdqpp)
2409 {
2410         int             error;
2411         xfs_dquot_t     *uq, *gq;
2412         uint            lockflags;
2413
2414         if (!XFS_IS_QUOTA_ON(mp))
2415                 return 0;
2416
2417         lockflags = XFS_ILOCK_EXCL;
2418         xfs_ilock(ip, lockflags);
2419
2420         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2421                 gid = ip->i_d.di_gid;
2422
2423         /*
2424          * Attach the dquot(s) to this inode, doing a dquot allocation
2425          * if necessary. The dquot(s) will not be locked.
2426          */
2427         if (XFS_NOT_DQATTACHED(mp, ip)) {
2428                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2429                                             XFS_QMOPT_ILOCKED))) {
2430                         xfs_iunlock(ip, lockflags);
2431                         return error;
2432                 }
2433         }
2434
2435         uq = gq = NULL;
2436         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2437                 if (ip->i_d.di_uid != uid) {
2438                         /*
2439                          * What we need is the dquot that has this uid, and
2440                          * if we send the inode to dqget, the uid of the inode
2441                          * takes priority over what's sent in the uid argument.
2442                          * We must unlock inode here before calling dqget if
2443                          * we're not sending the inode, because otherwise
2444                          * we'll deadlock by doing trans_reserve while
2445                          * holding ilock.
2446                          */
2447                         xfs_iunlock(ip, lockflags);
2448                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2449                                                  XFS_DQ_USER,
2450                                                  XFS_QMOPT_DQALLOC |
2451                                                  XFS_QMOPT_DOWARN,
2452                                                  &uq))) {
2453                                 ASSERT(error != ENOENT);
2454                                 return error;
2455                         }
2456                         /*
2457                          * Get the ilock in the right order.
2458                          */
2459                         xfs_dqunlock(uq);
2460                         lockflags = XFS_ILOCK_SHARED;
2461                         xfs_ilock(ip, lockflags);
2462                 } else {
2463                         /*
2464                          * Take an extra reference, because we'll return
2465                          * this to caller
2466                          */
2467                         ASSERT(ip->i_udquot);
2468                         uq = ip->i_udquot;
2469                         xfs_dqlock(uq);
2470                         XFS_DQHOLD(uq);
2471                         xfs_dqunlock(uq);
2472                 }
2473         }
2474         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2475                 if (ip->i_d.di_gid != gid) {
2476                         xfs_iunlock(ip, lockflags);
2477                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2478                                                  XFS_DQ_GROUP,
2479                                                  XFS_QMOPT_DQALLOC |
2480                                                  XFS_QMOPT_DOWARN,
2481                                                  &gq))) {
2482                                 if (uq)
2483                                         xfs_qm_dqrele(uq);
2484                                 ASSERT(error != ENOENT);
2485                                 return error;
2486                         }
2487                         xfs_dqunlock(gq);
2488                         lockflags = XFS_ILOCK_SHARED;
2489                         xfs_ilock(ip, lockflags);
2490                 } else {
2491                         ASSERT(ip->i_gdquot);
2492                         gq = ip->i_gdquot;
2493                         xfs_dqlock(gq);
2494                         XFS_DQHOLD(gq);
2495                         xfs_dqunlock(gq);
2496                 }
2497         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2498                 if (ip->i_d.di_projid != prid) {
2499                         xfs_iunlock(ip, lockflags);
2500                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2501                                                  XFS_DQ_PROJ,
2502                                                  XFS_QMOPT_DQALLOC |
2503                                                  XFS_QMOPT_DOWARN,
2504                                                  &gq))) {
2505                                 if (uq)
2506                                         xfs_qm_dqrele(uq);
2507                                 ASSERT(error != ENOENT);
2508                                 return (error);
2509                         }
2510                         xfs_dqunlock(gq);
2511                         lockflags = XFS_ILOCK_SHARED;
2512                         xfs_ilock(ip, lockflags);
2513                 } else {
2514                         ASSERT(ip->i_gdquot);
2515                         gq = ip->i_gdquot;
2516                         xfs_dqlock(gq);
2517                         XFS_DQHOLD(gq);
2518                         xfs_dqunlock(gq);
2519                 }
2520         }
2521         if (uq)
2522                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2523
2524         xfs_iunlock(ip, lockflags);
2525         if (O_udqpp)
2526                 *O_udqpp = uq;
2527         else if (uq)
2528                 xfs_qm_dqrele(uq);
2529         if (O_gdqpp)
2530                 *O_gdqpp = gq;
2531         else if (gq)
2532                 xfs_qm_dqrele(gq);
2533         return 0;
2534 }
2535
2536 /*
2537  * Actually transfer ownership, and do dquot modifications.
2538  * These were already reserved.
2539  */
2540 xfs_dquot_t *
2541 xfs_qm_vop_chown(
2542         xfs_trans_t     *tp,
2543         xfs_inode_t     *ip,
2544         xfs_dquot_t     **IO_olddq,
2545         xfs_dquot_t     *newdq)
2546 {
2547         xfs_dquot_t     *prevdq;
2548         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2549                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2550
2551         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2552         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2553
2554         /* old dquot */
2555         prevdq = *IO_olddq;
2556         ASSERT(prevdq);
2557         ASSERT(prevdq != newdq);
2558
2559         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2560         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2561
2562         /* the sparkling new dquot */
2563         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2564         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2565
2566         /*
2567          * Take an extra reference, because the inode
2568          * is going to keep this dquot pointer even
2569          * after the trans_commit.
2570          */
2571         xfs_dqlock(newdq);
2572         XFS_DQHOLD(newdq);
2573         xfs_dqunlock(newdq);
2574         *IO_olddq = newdq;
2575
2576         return prevdq;
2577 }
2578
2579 /*
2580  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2581  */
2582 int
2583 xfs_qm_vop_chown_reserve(
2584         xfs_trans_t     *tp,
2585         xfs_inode_t     *ip,
2586         xfs_dquot_t     *udqp,
2587         xfs_dquot_t     *gdqp,
2588         uint            flags)
2589 {
2590         int             error;
2591         xfs_mount_t     *mp;
2592         uint            delblks, blkflags, prjflags = 0;
2593         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2594
2595         ASSERT(XFS_ISLOCKED_INODE(ip));
2596         mp = ip->i_mount;
2597         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2598
2599         delblks = ip->i_delayed_blks;
2600         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2601         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2602                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2603
2604         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2605             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2606                 delblksudq = udqp;
2607                 /*
2608                  * If there are delayed allocation blocks, then we have to
2609                  * unreserve those from the old dquot, and add them to the
2610                  * new dquot.
2611                  */
2612                 if (delblks) {
2613                         ASSERT(ip->i_udquot);
2614                         unresudq = ip->i_udquot;
2615                 }
2616         }
2617         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2618                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2619                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2620                         prjflags = XFS_QMOPT_ENOSPC;
2621
2622                 if (prjflags ||
2623                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2624                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2625                         delblksgdq = gdqp;
2626                         if (delblks) {
2627                                 ASSERT(ip->i_gdquot);
2628                                 unresgdq = ip->i_gdquot;
2629                         }
2630                 }
2631         }
2632
2633         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2634                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2635                                 flags | blkflags | prjflags)))
2636                 return (error);
2637
2638         /*
2639          * Do the delayed blks reservations/unreservations now. Since, these
2640          * are done without the help of a transaction, if a reservation fails
2641          * its previous reservations won't be automatically undone by trans
2642          * code. So, we have to do it manually here.
2643          */
2644         if (delblks) {
2645                 /*
2646                  * Do the reservations first. Unreservation can't fail.
2647                  */
2648                 ASSERT(delblksudq || delblksgdq);
2649                 ASSERT(unresudq || unresgdq);
2650                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2651                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2652                                 flags | blkflags | prjflags)))
2653                         return (error);
2654                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2655                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2656                                 blkflags);
2657         }
2658
2659         return (0);
2660 }
2661
2662 int
2663 xfs_qm_vop_rename_dqattach(
2664         xfs_inode_t     **i_tab)
2665 {
2666         xfs_inode_t     *ip;
2667         int             i;
2668         int             error;
2669
2670         ip = i_tab[0];
2671
2672         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2673                 return 0;
2674
2675         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2676                 error = xfs_qm_dqattach(ip, 0);
2677                 if (error)
2678                         return error;
2679         }
2680         for (i = 1; (i < 4 && i_tab[i]); i++) {
2681                 /*
2682                  * Watch out for duplicate entries in the table.
2683                  */
2684                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2685                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2686                                 error = xfs_qm_dqattach(ip, 0);
2687                                 if (error)
2688                                         return error;
2689                         }
2690                 }
2691         }
2692         return 0;
2693 }
2694
2695 void
2696 xfs_qm_vop_dqattach_and_dqmod_newinode(
2697         xfs_trans_t     *tp,
2698         xfs_inode_t     *ip,
2699         xfs_dquot_t     *udqp,
2700         xfs_dquot_t     *gdqp)
2701 {
2702         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2703                 return;
2704
2705         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2706         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2707
2708         if (udqp) {
2709                 xfs_dqlock(udqp);
2710                 XFS_DQHOLD(udqp);
2711                 xfs_dqunlock(udqp);
2712                 ASSERT(ip->i_udquot == NULL);
2713                 ip->i_udquot = udqp;
2714                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2715                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2716                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2717         }
2718         if (gdqp) {
2719                 xfs_dqlock(gdqp);
2720                 XFS_DQHOLD(gdqp);
2721                 xfs_dqunlock(gdqp);
2722                 ASSERT(ip->i_gdquot == NULL);
2723                 ip->i_gdquot = gdqp;
2724                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2725                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2726                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2727                                 be32_to_cpu(gdqp->q_core.d_id));
2728                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2729         }
2730 }
2731
2732 /* ------------- list stuff -----------------*/
2733 STATIC void
2734 xfs_qm_freelist_init(xfs_frlist_t *ql)
2735 {
2736         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2737         mutex_init(&ql->qh_lock);
2738         ql->qh_version = 0;
2739         ql->qh_nelems = 0;
2740 }
2741
2742 STATIC void
2743 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2744 {
2745         xfs_dquot_t     *dqp, *nextdqp;
2746
2747         mutex_lock(&ql->qh_lock);
2748         for (dqp = ql->qh_next;
2749              dqp != (xfs_dquot_t *)ql; ) {
2750                 xfs_dqlock(dqp);
2751                 nextdqp = dqp->dq_flnext;
2752 #ifdef QUOTADEBUG
2753                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2754 #endif
2755                 XQM_FREELIST_REMOVE(dqp);
2756                 xfs_dqunlock(dqp);
2757                 xfs_qm_dqdestroy(dqp);
2758                 dqp = nextdqp;
2759         }
2760         mutex_unlock(&ql->qh_lock);
2761         mutex_destroy(&ql->qh_lock);
2762
2763         ASSERT(ql->qh_nelems == 0);
2764 }
2765
2766 STATIC void
2767 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2768 {
2769         dq->dq_flnext = ql->qh_next;
2770         dq->dq_flprev = (xfs_dquot_t *)ql;
2771         ql->qh_next = dq;
2772         dq->dq_flnext->dq_flprev = dq;
2773         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2774         xfs_Gqm->qm_dqfreelist.qh_version++;
2775 }
2776
2777 void
2778 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2779 {
2780         xfs_dquot_t *next = dq->dq_flnext;
2781         xfs_dquot_t *prev = dq->dq_flprev;
2782
2783         next->dq_flprev = prev;
2784         prev->dq_flnext = next;
2785         dq->dq_flnext = dq->dq_flprev = dq;
2786         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2787         xfs_Gqm->qm_dqfreelist.qh_version++;
2788 }
2789
2790 void
2791 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2792 {
2793         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2794 }
2795
2796 STATIC int
2797 xfs_qm_dqhashlock_nowait(
2798         xfs_dquot_t *dqp)
2799 {
2800         int locked;
2801
2802         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2803         return locked;
2804 }
2805
2806 int
2807 xfs_qm_freelist_lock_nowait(
2808         xfs_qm_t *xqm)
2809 {
2810         int locked;
2811
2812         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2813         return locked;
2814 }
2815
2816 STATIC int
2817 xfs_qm_mplist_nowait(
2818         xfs_mount_t     *mp)
2819 {
2820         int locked;
2821
2822         ASSERT(mp->m_quotainfo);
2823         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2824         return locked;
2825 }