Merge branch 'for-linus' of git://oss.sgi.com:8090/xfs/xfs-2.6
[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         if (! XFS_IS_QUOTA_ON(mp))
1012                 return 0;
1013
1014         restarts = 0;
1015         /*
1016          * We won't block unless we are asked to.
1017          */
1018         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1019
1020   again:
1021         xfs_qm_mplist_lock(mp);
1022         /*
1023          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1024          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1025          * when we have the mplist lock, we know that dquots will be consistent
1026          * as long as we have it locked.
1027          */
1028         if (! XFS_IS_QUOTA_ON(mp)) {
1029                 xfs_qm_mplist_unlock(mp);
1030                 return 0;
1031         }
1032         FOREACH_DQUOT_IN_MP(dqp, mp) {
1033                 /*
1034                  * If this is vfs_sync calling, then skip the dquots that
1035                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1036                  * This is very similar to what xfs_sync does with inodes.
1037                  */
1038                 if (flags & SYNC_BDFLUSH) {
1039                         if (! XFS_DQ_IS_DIRTY(dqp))
1040                                 continue;
1041                 }
1042
1043                 if (nowait) {
1044                         /*
1045                          * Try to acquire the dquot lock. We are NOT out of
1046                          * lock order, but we just don't want to wait for this
1047                          * lock, unless somebody wanted us to.
1048                          */
1049                         if (! xfs_qm_dqlock_nowait(dqp))
1050                                 continue;
1051                 } else {
1052                         xfs_dqlock(dqp);
1053                 }
1054
1055                 /*
1056                  * Now, find out for sure if this dquot is dirty or not.
1057                  */
1058                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1059                         xfs_dqunlock(dqp);
1060                         continue;
1061                 }
1062
1063                 /* XXX a sentinel would be better */
1064                 recl = XFS_QI_MPLRECLAIMS(mp);
1065                 if (! xfs_qm_dqflock_nowait(dqp)) {
1066                         if (nowait) {
1067                                 xfs_dqunlock(dqp);
1068                                 continue;
1069                         }
1070                         /*
1071                          * If we can't grab the flush lock then if the caller
1072                          * really wanted us to give this our best shot, so
1073                          * see if we can give a push to the buffer before we wait
1074                          * on the flush lock. At this point, we know that
1075                          * even though the dquot is being flushed,
1076                          * it has (new) dirty data.
1077                          */
1078                         xfs_qm_dqflock_pushbuf_wait(dqp);
1079                 }
1080                 /*
1081                  * Let go of the mplist lock. We don't want to hold it
1082                  * across a disk write
1083                  */
1084                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1085                 xfs_qm_mplist_unlock(mp);
1086                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1087                 error = xfs_qm_dqflush(dqp, flush_flags);
1088                 xfs_dqunlock(dqp);
1089                 if (error && XFS_FORCED_SHUTDOWN(mp))
1090                         return 0;       /* Need to prevent umount failure */
1091                 else if (error)
1092                         return error;
1093
1094                 xfs_qm_mplist_lock(mp);
1095                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1096                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1097                                 break;
1098
1099                         xfs_qm_mplist_unlock(mp);
1100                         goto again;
1101                 }
1102         }
1103
1104         xfs_qm_mplist_unlock(mp);
1105         return 0;
1106 }
1107
1108
1109 /*
1110  * This initializes all the quota information that's kept in the
1111  * mount structure
1112  */
1113 STATIC int
1114 xfs_qm_init_quotainfo(
1115         xfs_mount_t     *mp)
1116 {
1117         xfs_quotainfo_t *qinf;
1118         int             error;
1119         xfs_dquot_t     *dqp;
1120
1121         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1122
1123         /*
1124          * Tell XQM that we exist as soon as possible.
1125          */
1126         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1127                 return error;
1128         }
1129
1130         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1131
1132         /*
1133          * See if quotainodes are setup, and if not, allocate them,
1134          * and change the superblock accordingly.
1135          */
1136         if ((error = xfs_qm_init_quotainos(mp))) {
1137                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1138                 mp->m_quotainfo = NULL;
1139                 return error;
1140         }
1141
1142         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1143         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1144         qinf->qi_dqreclaims = 0;
1145
1146         /* mutex used to serialize quotaoffs */
1147         mutex_init(&qinf->qi_quotaofflock);
1148
1149         /* Precalc some constants */
1150         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1151         ASSERT(qinf->qi_dqchunklen);
1152         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1153         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1154
1155         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1156
1157         /*
1158          * We try to get the limits from the superuser's limits fields.
1159          * This is quite hacky, but it is standard quota practice.
1160          * We look at the USR dquot with id == 0 first, but if user quotas
1161          * are not enabled we goto the GRP dquot with id == 0.
1162          * We don't really care to keep separate default limits for user
1163          * and group quotas, at least not at this point.
1164          */
1165         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1166                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1167                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1168                                 XFS_DQ_PROJ),
1169                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1170                              &dqp);
1171         if (! error) {
1172                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1173
1174                 /*
1175                  * The warnings and timers set the grace period given to
1176                  * a user or group before he or she can not perform any
1177                  * more writing. If it is zero, a default is used.
1178                  */
1179                 qinf->qi_btimelimit = ddqp->d_btimer ?
1180                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1181                 qinf->qi_itimelimit = ddqp->d_itimer ?
1182                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1183                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1184                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1185                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1186                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1187                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1188                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1189                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1190                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1191                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1192                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1193                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1194                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1195                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1196                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1197  
1198                 /*
1199                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1200                  * we don't want this dquot cached. We haven't done a
1201                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1202                  */
1203                 xfs_qm_dqdestroy(dqp);
1204         } else {
1205                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1206                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1207                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1208                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1209                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1210                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1211         }
1212
1213         return 0;
1214 }
1215
1216
1217 /*
1218  * Gets called when unmounting a filesystem or when all quotas get
1219  * turned off.
1220  * This purges the quota inodes, destroys locks and frees itself.
1221  */
1222 void
1223 xfs_qm_destroy_quotainfo(
1224         xfs_mount_t     *mp)
1225 {
1226         xfs_quotainfo_t *qi;
1227
1228         qi = mp->m_quotainfo;
1229         ASSERT(qi != NULL);
1230         ASSERT(xfs_Gqm != NULL);
1231
1232         /*
1233          * Release the reference that XQM kept, so that we know
1234          * when the XQM structure should be freed. We cannot assume
1235          * that xfs_Gqm is non-null after this point.
1236          */
1237         xfs_qm_rele_quotafs_ref(mp);
1238
1239         spinlock_destroy(&qi->qi_pinlock);
1240         xfs_qm_list_destroy(&qi->qi_dqlist);
1241
1242         if (qi->qi_uquotaip) {
1243                 XFS_PURGE_INODE(qi->qi_uquotaip);
1244                 qi->qi_uquotaip = NULL; /* paranoia */
1245         }
1246         if (qi->qi_gquotaip) {
1247                 XFS_PURGE_INODE(qi->qi_gquotaip);
1248                 qi->qi_gquotaip = NULL;
1249         }
1250         mutex_destroy(&qi->qi_quotaofflock);
1251         kmem_free(qi, sizeof(xfs_quotainfo_t));
1252         mp->m_quotainfo = NULL;
1253 }
1254
1255
1256
1257 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1258
1259 /* ARGSUSED */
1260 STATIC void
1261 xfs_qm_list_init(
1262         xfs_dqlist_t    *list,
1263         char            *str,
1264         int             n)
1265 {
1266         mutex_init(&list->qh_lock);
1267         list->qh_next = NULL;
1268         list->qh_version = 0;
1269         list->qh_nelems = 0;
1270 }
1271
1272 STATIC void
1273 xfs_qm_list_destroy(
1274         xfs_dqlist_t    *list)
1275 {
1276         mutex_destroy(&(list->qh_lock));
1277 }
1278
1279
1280 /*
1281  * Stripped down version of dqattach. This doesn't attach, or even look at the
1282  * dquots attached to the inode. The rationale is that there won't be any
1283  * attached at the time this is called from quotacheck.
1284  */
1285 STATIC int
1286 xfs_qm_dqget_noattach(
1287         xfs_inode_t     *ip,
1288         xfs_dquot_t     **O_udqpp,
1289         xfs_dquot_t     **O_gdqpp)
1290 {
1291         int             error;
1292         xfs_mount_t     *mp;
1293         xfs_dquot_t     *udqp, *gdqp;
1294
1295         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1296         mp = ip->i_mount;
1297         udqp = NULL;
1298         gdqp = NULL;
1299
1300         if (XFS_IS_UQUOTA_ON(mp)) {
1301                 ASSERT(ip->i_udquot == NULL);
1302                 /*
1303                  * We want the dquot allocated if it doesn't exist.
1304                  */
1305                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1306                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1307                                          &udqp))) {
1308                         /*
1309                          * Shouldn't be able to turn off quotas here.
1310                          */
1311                         ASSERT(error != ESRCH);
1312                         ASSERT(error != ENOENT);
1313                         return error;
1314                 }
1315                 ASSERT(udqp);
1316         }
1317
1318         if (XFS_IS_OQUOTA_ON(mp)) {
1319                 ASSERT(ip->i_gdquot == NULL);
1320                 if (udqp)
1321                         xfs_dqunlock(udqp);
1322                 error = XFS_IS_GQUOTA_ON(mp) ?
1323                                 xfs_qm_dqget(mp, ip,
1324                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1325                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1326                                              &gdqp) :
1327                                 xfs_qm_dqget(mp, ip,
1328                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1329                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1330                                              &gdqp);
1331                 if (error) {
1332                         if (udqp)
1333                                 xfs_qm_dqrele(udqp);
1334                         ASSERT(error != ESRCH);
1335                         ASSERT(error != ENOENT);
1336                         return error;
1337                 }
1338                 ASSERT(gdqp);
1339
1340                 /* Reacquire the locks in the right order */
1341                 if (udqp) {
1342                         if (! xfs_qm_dqlock_nowait(udqp)) {
1343                                 xfs_dqunlock(gdqp);
1344                                 xfs_dqlock(udqp);
1345                                 xfs_dqlock(gdqp);
1346                         }
1347                 }
1348         }
1349
1350         *O_udqpp = udqp;
1351         *O_gdqpp = gdqp;
1352
1353 #ifdef QUOTADEBUG
1354         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1355         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1356 #endif
1357         return 0;
1358 }
1359
1360 /*
1361  * Create an inode and return with a reference already taken, but unlocked
1362  * This is how we create quota inodes
1363  */
1364 STATIC int
1365 xfs_qm_qino_alloc(
1366         xfs_mount_t     *mp,
1367         xfs_inode_t     **ip,
1368         __int64_t       sbfields,
1369         uint            flags)
1370 {
1371         xfs_trans_t     *tp;
1372         int             error;
1373         unsigned long   s;
1374         int             committed;
1375
1376         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1377         if ((error = xfs_trans_reserve(tp,
1378                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1379                                       XFS_CREATE_LOG_RES(mp), 0,
1380                                       XFS_TRANS_PERM_LOG_RES,
1381                                       XFS_CREATE_LOG_COUNT))) {
1382                 xfs_trans_cancel(tp, 0);
1383                 return error;
1384         }
1385
1386         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1387                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1388                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1389                                  XFS_TRANS_ABORT);
1390                 return error;
1391         }
1392
1393         /*
1394          * Keep an extra reference to this quota inode. This inode is
1395          * locked exclusively and joined to the transaction already.
1396          */
1397         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1398         VN_HOLD(XFS_ITOV((*ip)));
1399
1400         /*
1401          * Make the changes in the superblock, and log those too.
1402          * sbfields arg may contain fields other than *QUOTINO;
1403          * VERSIONNUM for example.
1404          */
1405         s = XFS_SB_LOCK(mp);
1406         if (flags & XFS_QMOPT_SBVERSION) {
1407 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1408                 unsigned oldv = mp->m_sb.sb_versionnum;
1409 #endif
1410                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1411                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1412                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1413                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1414                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1415
1416                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1417                 mp->m_sb.sb_uquotino = NULLFSINO;
1418                 mp->m_sb.sb_gquotino = NULLFSINO;
1419
1420                 /* qflags will get updated _after_ quotacheck */
1421                 mp->m_sb.sb_qflags = 0;
1422 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1423                 cmn_err(CE_NOTE,
1424                         "Old superblock version %x, converting to %x.",
1425                         oldv, mp->m_sb.sb_versionnum);
1426 #endif
1427         }
1428         if (flags & XFS_QMOPT_UQUOTA)
1429                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1430         else
1431                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1432         XFS_SB_UNLOCK(mp, s);
1433         xfs_mod_sb(tp, sbfields);
1434
1435         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1436                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1437                 return error;
1438         }
1439         return 0;
1440 }
1441
1442
1443 STATIC int
1444 xfs_qm_reset_dqcounts(
1445         xfs_mount_t     *mp,
1446         xfs_buf_t       *bp,
1447         xfs_dqid_t      id,
1448         uint            type)
1449 {
1450         xfs_disk_dquot_t        *ddq;
1451         int                     j;
1452
1453         xfs_buftrace("RESET DQUOTS", bp);
1454         /*
1455          * Reset all counters and timers. They'll be
1456          * started afresh by xfs_qm_quotacheck.
1457          */
1458 #ifdef DEBUG
1459         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1460         do_div(j, sizeof(xfs_dqblk_t));
1461         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1462 #endif
1463         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1464         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1465                 /*
1466                  * Do a sanity check, and if needed, repair the dqblk. Don't
1467                  * output any warnings because it's perfectly possible to
1468                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1469                  */
1470                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1471                                       "xfs_quotacheck");
1472                 ddq->d_bcount = 0;
1473                 ddq->d_icount = 0;
1474                 ddq->d_rtbcount = 0;
1475                 ddq->d_btimer = 0;
1476                 ddq->d_itimer = 0;
1477                 ddq->d_rtbtimer = 0;
1478                 ddq->d_bwarns = 0;
1479                 ddq->d_iwarns = 0;
1480                 ddq->d_rtbwarns = 0;
1481                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1482         }
1483
1484         return 0;
1485 }
1486
1487 STATIC int
1488 xfs_qm_dqiter_bufs(
1489         xfs_mount_t     *mp,
1490         xfs_dqid_t      firstid,
1491         xfs_fsblock_t   bno,
1492         xfs_filblks_t   blkcnt,
1493         uint            flags)
1494 {
1495         xfs_buf_t       *bp;
1496         int             error;
1497         int             notcommitted;
1498         int             incr;
1499         int             type;
1500
1501         ASSERT(blkcnt > 0);
1502         notcommitted = 0;
1503         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1504                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1505         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1506                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1507         error = 0;
1508
1509         /*
1510          * Blkcnt arg can be a very big number, and might even be
1511          * larger than the log itself. So, we have to break it up into
1512          * manageable-sized transactions.
1513          * Note that we don't start a permanent transaction here; we might
1514          * not be able to get a log reservation for the whole thing up front,
1515          * and we don't really care to either, because we just discard
1516          * everything if we were to crash in the middle of this loop.
1517          */
1518         while (blkcnt--) {
1519                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1520                               XFS_FSB_TO_DADDR(mp, bno),
1521                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1522                 if (error)
1523                         break;
1524
1525                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1526                 xfs_bdwrite(mp, bp);
1527                 /*
1528                  * goto the next block.
1529                  */
1530                 bno++;
1531                 firstid += XFS_QM_DQPERBLK(mp);
1532         }
1533         return error;
1534 }
1535
1536 /*
1537  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1538  * caller supplied function for every chunk of dquots that we find.
1539  */
1540 STATIC int
1541 xfs_qm_dqiterate(
1542         xfs_mount_t     *mp,
1543         xfs_inode_t     *qip,
1544         uint            flags)
1545 {
1546         xfs_bmbt_irec_t         *map;
1547         int                     i, nmaps;       /* number of map entries */
1548         int                     error;          /* return value */
1549         xfs_fileoff_t           lblkno;
1550         xfs_filblks_t           maxlblkcnt;
1551         xfs_dqid_t              firstid;
1552         xfs_fsblock_t           rablkno;
1553         xfs_filblks_t           rablkcnt;
1554
1555         error = 0;
1556         /*
1557          * This looks racy, but we can't keep an inode lock across a
1558          * trans_reserve. But, this gets called during quotacheck, and that
1559          * happens only at mount time which is single threaded.
1560          */
1561         if (qip->i_d.di_nblocks == 0)
1562                 return 0;
1563
1564         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1565
1566         lblkno = 0;
1567         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1568         do {
1569                 nmaps = XFS_DQITER_MAP_SIZE;
1570                 /*
1571                  * We aren't changing the inode itself. Just changing
1572                  * some of its data. No new blocks are added here, and
1573                  * the inode is never added to the transaction.
1574                  */
1575                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1576                 error = xfs_bmapi(NULL, qip, lblkno,
1577                                   maxlblkcnt - lblkno,
1578                                   XFS_BMAPI_METADATA,
1579                                   NULL,
1580                                   0, map, &nmaps, NULL, NULL);
1581                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1582                 if (error)
1583                         break;
1584
1585                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1586                 for (i = 0; i < nmaps; i++) {
1587                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1588                         ASSERT(map[i].br_blockcount);
1589
1590
1591                         lblkno += map[i].br_blockcount;
1592
1593                         if (map[i].br_startblock == HOLESTARTBLOCK)
1594                                 continue;
1595
1596                         firstid = (xfs_dqid_t) map[i].br_startoff *
1597                                 XFS_QM_DQPERBLK(mp);
1598                         /*
1599                          * Do a read-ahead on the next extent.
1600                          */
1601                         if ((i+1 < nmaps) &&
1602                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1603                                 rablkcnt =  map[i+1].br_blockcount;
1604                                 rablkno = map[i+1].br_startblock;
1605                                 while (rablkcnt--) {
1606                                         xfs_baread(mp->m_ddev_targp,
1607                                                XFS_FSB_TO_DADDR(mp, rablkno),
1608                                                (int)XFS_QI_DQCHUNKLEN(mp));
1609                                         rablkno++;
1610                                 }
1611                         }
1612                         /*
1613                          * Iterate thru all the blks in the extent and
1614                          * reset the counters of all the dquots inside them.
1615                          */
1616                         if ((error = xfs_qm_dqiter_bufs(mp,
1617                                                        firstid,
1618                                                        map[i].br_startblock,
1619                                                        map[i].br_blockcount,
1620                                                        flags))) {
1621                                 break;
1622                         }
1623                 }
1624
1625                 if (error)
1626                         break;
1627         } while (nmaps > 0);
1628
1629         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1630
1631         return error;
1632 }
1633
1634 /*
1635  * Called by dqusage_adjust in doing a quotacheck.
1636  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1637  * this updates its incore copy as well as the buffer copy. This is
1638  * so that once the quotacheck is done, we can just log all the buffers,
1639  * as opposed to logging numerous updates to individual dquots.
1640  */
1641 STATIC void
1642 xfs_qm_quotacheck_dqadjust(
1643         xfs_dquot_t             *dqp,
1644         xfs_qcnt_t              nblks,
1645         xfs_qcnt_t              rtblks)
1646 {
1647         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1648         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1649         /*
1650          * Adjust the inode count and the block count to reflect this inode's
1651          * resource usage.
1652          */
1653         be64_add(&dqp->q_core.d_icount, 1);
1654         dqp->q_res_icount++;
1655         if (nblks) {
1656                 be64_add(&dqp->q_core.d_bcount, nblks);
1657                 dqp->q_res_bcount += nblks;
1658         }
1659         if (rtblks) {
1660                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1661                 dqp->q_res_rtbcount += rtblks;
1662         }
1663
1664         /*
1665          * Set default limits, adjust timers (since we changed usages)
1666          */
1667         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1668                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1669                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1670         }
1671
1672         dqp->dq_flags |= XFS_DQ_DIRTY;
1673 }
1674
1675 STATIC int
1676 xfs_qm_get_rtblks(
1677         xfs_inode_t     *ip,
1678         xfs_qcnt_t      *O_rtblks)
1679 {
1680         xfs_filblks_t   rtblks;                 /* total rt blks */
1681         xfs_extnum_t    idx;                    /* extent record index */
1682         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1683         xfs_extnum_t    nextents;               /* number of extent entries */
1684         int             error;
1685
1686         ASSERT(XFS_IS_REALTIME_INODE(ip));
1687         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1688         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1689                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1690                         return error;
1691         }
1692         rtblks = 0;
1693         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1694         for (idx = 0; idx < nextents; idx++)
1695                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1696         *O_rtblks = (xfs_qcnt_t)rtblks;
1697         return 0;
1698 }
1699
1700 /*
1701  * callback routine supplied to bulkstat(). Given an inumber, find its
1702  * dquots and update them to account for resources taken by that inode.
1703  */
1704 /* ARGSUSED */
1705 STATIC int
1706 xfs_qm_dqusage_adjust(
1707         xfs_mount_t     *mp,            /* mount point for filesystem */
1708         xfs_ino_t       ino,            /* inode number to get data for */
1709         void            __user *buffer, /* not used */
1710         int             ubsize,         /* not used */
1711         void            *private_data,  /* not used */
1712         xfs_daddr_t     bno,            /* starting block of inode cluster */
1713         int             *ubused,        /* not used */
1714         void            *dip,           /* on-disk inode pointer (not used) */
1715         int             *res)           /* result code value */
1716 {
1717         xfs_inode_t     *ip;
1718         xfs_dquot_t     *udqp, *gdqp;
1719         xfs_qcnt_t      nblks, rtblks;
1720         int             error;
1721
1722         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1723
1724         /*
1725          * rootino must have its resources accounted for, not so with the quota
1726          * inodes.
1727          */
1728         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1729                 *res = BULKSTAT_RV_NOTHING;
1730                 return XFS_ERROR(EINVAL);
1731         }
1732
1733         /*
1734          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1735          * interface expects the inode to be exclusively locked because that's
1736          * the case in all other instances. It's OK that we do this because
1737          * quotacheck is done only at mount time.
1738          */
1739         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1740                 *res = BULKSTAT_RV_NOTHING;
1741                 return error;
1742         }
1743
1744         if (ip->i_d.di_mode == 0) {
1745                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1746                 *res = BULKSTAT_RV_NOTHING;
1747                 return XFS_ERROR(ENOENT);
1748         }
1749
1750         /*
1751          * Obtain the locked dquots. In case of an error (eg. allocation
1752          * fails for ENOSPC), we return the negative of the error number
1753          * to bulkstat, so that it can get propagated to quotacheck() and
1754          * making us disable quotas for the file system.
1755          */
1756         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1757                 xfs_iput(ip, XFS_ILOCK_EXCL);
1758                 *res = BULKSTAT_RV_GIVEUP;
1759                 return error;
1760         }
1761
1762         rtblks = 0;
1763         if (! XFS_IS_REALTIME_INODE(ip)) {
1764                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1765         } else {
1766                 /*
1767                  * Walk thru the extent list and count the realtime blocks.
1768                  */
1769                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1770                         xfs_iput(ip, XFS_ILOCK_EXCL);
1771                         if (udqp)
1772                                 xfs_qm_dqput(udqp);
1773                         if (gdqp)
1774                                 xfs_qm_dqput(gdqp);
1775                         *res = BULKSTAT_RV_GIVEUP;
1776                         return error;
1777                 }
1778                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1779         }
1780         ASSERT(ip->i_delayed_blks == 0);
1781
1782         /*
1783          * We can't release the inode while holding its dquot locks.
1784          * The inode can go into inactive and might try to acquire the dquotlocks.
1785          * So, just unlock here and do a vn_rele at the end.
1786          */
1787         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1788
1789         /*
1790          * Add the (disk blocks and inode) resources occupied by this
1791          * inode to its dquots. We do this adjustment in the incore dquot,
1792          * and also copy the changes to its buffer.
1793          * We don't care about putting these changes in a transaction
1794          * envelope because if we crash in the middle of a 'quotacheck'
1795          * we have to start from the beginning anyway.
1796          * Once we're done, we'll log all the dquot bufs.
1797          *
1798          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1799          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1800          */
1801         if (XFS_IS_UQUOTA_ON(mp)) {
1802                 ASSERT(udqp);
1803                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1804                 xfs_qm_dqput(udqp);
1805         }
1806         if (XFS_IS_OQUOTA_ON(mp)) {
1807                 ASSERT(gdqp);
1808                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1809                 xfs_qm_dqput(gdqp);
1810         }
1811         /*
1812          * Now release the inode. This will send it to 'inactive', and
1813          * possibly even free blocks.
1814          */
1815         VN_RELE(XFS_ITOV(ip));
1816
1817         /*
1818          * Goto next inode.
1819          */
1820         *res = BULKSTAT_RV_DIDONE;
1821         return 0;
1822 }
1823
1824 /*
1825  * Walk thru all the filesystem inodes and construct a consistent view
1826  * of the disk quota world. If the quotacheck fails, disable quotas.
1827  */
1828 int
1829 xfs_qm_quotacheck(
1830         xfs_mount_t     *mp)
1831 {
1832         int             done, count, error;
1833         xfs_ino_t       lastino;
1834         size_t          structsz;
1835         xfs_inode_t     *uip, *gip;
1836         uint            flags;
1837
1838         count = INT_MAX;
1839         structsz = 1;
1840         lastino = 0;
1841         flags = 0;
1842
1843         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1844         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1845
1846         /*
1847          * There should be no cached dquots. The (simplistic) quotacheck
1848          * algorithm doesn't like that.
1849          */
1850         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1851
1852         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1853
1854         /*
1855          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1856          * their counters to zero. We need a clean slate.
1857          * We don't log our changes till later.
1858          */
1859         if ((uip = XFS_QI_UQIP(mp))) {
1860                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1861                         goto error_return;
1862                 flags |= XFS_UQUOTA_CHKD;
1863         }
1864
1865         if ((gip = XFS_QI_GQIP(mp))) {
1866                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1867                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1868                         goto error_return;
1869                 flags |= XFS_OQUOTA_CHKD;
1870         }
1871
1872         do {
1873                 /*
1874                  * Iterate thru all the inodes in the file system,
1875                  * adjusting the corresponding dquot counters in core.
1876                  */
1877                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1878                                      xfs_qm_dqusage_adjust, NULL,
1879                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1880                         break;
1881
1882         } while (! done);
1883
1884         /*
1885          * We can get this error if we couldn't do a dquot allocation inside
1886          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1887          * dirty dquots that might be cached, we just want to get rid of them
1888          * and turn quotaoff. The dquots won't be attached to any of the inodes
1889          * at this point (because we intentionally didn't in dqget_noattach).
1890          */
1891         if (error) {
1892                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1893                 goto error_return;
1894         }
1895         /*
1896          * We've made all the changes that we need to make incore.
1897          * Now flush_them down to disk buffers.
1898          */
1899         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1900
1901         /*
1902          * We didn't log anything, because if we crashed, we'll have to
1903          * start the quotacheck from scratch anyway. However, we must make
1904          * sure that our dquot changes are secure before we put the
1905          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1906          * flush.
1907          */
1908         XFS_bflush(mp->m_ddev_targp);
1909
1910         /*
1911          * If one type of quotas is off, then it will lose its
1912          * quotachecked status, since we won't be doing accounting for
1913          * that type anymore.
1914          */
1915         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1916         mp->m_qflags |= flags;
1917
1918         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1919
1920  error_return:
1921         if (error) {
1922                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1923                         "Disabling quotas.",
1924                         mp->m_fsname, error);
1925                 /*
1926                  * We must turn off quotas.
1927                  */
1928                 ASSERT(mp->m_quotainfo != NULL);
1929                 ASSERT(xfs_Gqm != NULL);
1930                 xfs_qm_destroy_quotainfo(mp);
1931                 (void)xfs_mount_reset_sbqflags(mp);
1932         } else {
1933                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1934         }
1935         return (error);
1936 }
1937
1938 /*
1939  * This is called after the superblock has been read in and we're ready to
1940  * iget the quota inodes.
1941  */
1942 STATIC int
1943 xfs_qm_init_quotainos(
1944         xfs_mount_t     *mp)
1945 {
1946         xfs_inode_t     *uip, *gip;
1947         int             error;
1948         __int64_t       sbflags;
1949         uint            flags;
1950
1951         ASSERT(mp->m_quotainfo);
1952         uip = gip = NULL;
1953         sbflags = 0;
1954         flags = 0;
1955
1956         /*
1957          * Get the uquota and gquota inodes
1958          */
1959         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1960                 if (XFS_IS_UQUOTA_ON(mp) &&
1961                     mp->m_sb.sb_uquotino != NULLFSINO) {
1962                         ASSERT(mp->m_sb.sb_uquotino > 0);
1963                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1964                                              0, 0, &uip, 0)))
1965                                 return XFS_ERROR(error);
1966                 }
1967                 if (XFS_IS_OQUOTA_ON(mp) &&
1968                     mp->m_sb.sb_gquotino != NULLFSINO) {
1969                         ASSERT(mp->m_sb.sb_gquotino > 0);
1970                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1971                                              0, 0, &gip, 0))) {
1972                                 if (uip)
1973                                         VN_RELE(XFS_ITOV(uip));
1974                                 return XFS_ERROR(error);
1975                         }
1976                 }
1977         } else {
1978                 flags |= XFS_QMOPT_SBVERSION;
1979                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1980                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1981         }
1982
1983         /*
1984          * Create the two inodes, if they don't exist already. The changes
1985          * made above will get added to a transaction and logged in one of
1986          * the qino_alloc calls below.  If the device is readonly,
1987          * temporarily switch to read-write to do this.
1988          */
1989         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1990                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1991                                               sbflags | XFS_SB_UQUOTINO,
1992                                               flags | XFS_QMOPT_UQUOTA)))
1993                         return XFS_ERROR(error);
1994
1995                 flags &= ~XFS_QMOPT_SBVERSION;
1996         }
1997         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1998                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1999                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2000                 error = xfs_qm_qino_alloc(mp, &gip,
2001                                           sbflags | XFS_SB_GQUOTINO, flags);
2002                 if (error) {
2003                         if (uip)
2004                                 VN_RELE(XFS_ITOV(uip));
2005
2006                         return XFS_ERROR(error);
2007                 }
2008         }
2009
2010         XFS_QI_UQIP(mp) = uip;
2011         XFS_QI_GQIP(mp) = gip;
2012
2013         return 0;
2014 }
2015
2016
2017 /*
2018  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2019  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2020  * favor the lookup function ...
2021  * XXXsup merge this with qm_reclaim_one().
2022  */
2023 STATIC int
2024 xfs_qm_shake_freelist(
2025         int howmany)
2026 {
2027         int             nreclaimed;
2028         xfs_dqhash_t    *hash;
2029         xfs_dquot_t     *dqp, *nextdqp;
2030         int             restarts;
2031         int             nflushes;
2032
2033         if (howmany <= 0)
2034                 return 0;
2035
2036         nreclaimed = 0;
2037         restarts = 0;
2038         nflushes = 0;
2039
2040 #ifdef QUOTADEBUG
2041         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2042 #endif
2043         /* lock order is : hashchainlock, freelistlock, mplistlock */
2044  tryagain:
2045         xfs_qm_freelist_lock(xfs_Gqm);
2046
2047         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2048              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2049               nreclaimed < howmany); ) {
2050                 xfs_dqlock(dqp);
2051
2052                 /*
2053                  * We are racing with dqlookup here. Naturally we don't
2054                  * want to reclaim a dquot that lookup wants.
2055                  */
2056                 if (dqp->dq_flags & XFS_DQ_WANT) {
2057                         xfs_dqunlock(dqp);
2058                         xfs_qm_freelist_unlock(xfs_Gqm);
2059                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2060                                 return nreclaimed;
2061                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2062                         goto tryagain;
2063                 }
2064
2065                 /*
2066                  * If the dquot is inactive, we are assured that it is
2067                  * not on the mplist or the hashlist, and that makes our
2068                  * life easier.
2069                  */
2070                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2071                         ASSERT(dqp->q_mount == NULL);
2072                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2073                         ASSERT(dqp->HL_PREVP == NULL);
2074                         ASSERT(dqp->MPL_PREVP == NULL);
2075                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2076                         nextdqp = dqp->dq_flnext;
2077                         goto off_freelist;
2078                 }
2079
2080                 ASSERT(dqp->MPL_PREVP);
2081                 /*
2082                  * Try to grab the flush lock. If this dquot is in the process of
2083                  * getting flushed to disk, we don't want to reclaim it.
2084                  */
2085                 if (! xfs_qm_dqflock_nowait(dqp)) {
2086                         xfs_dqunlock(dqp);
2087                         dqp = dqp->dq_flnext;
2088                         continue;
2089                 }
2090
2091                 /*
2092                  * We have the flush lock so we know that this is not in the
2093                  * process of being flushed. So, if this is dirty, flush it
2094                  * DELWRI so that we don't get a freelist infested with
2095                  * dirty dquots.
2096                  */
2097                 if (XFS_DQ_IS_DIRTY(dqp)) {
2098                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2099                         /*
2100                          * We flush it delayed write, so don't bother
2101                          * releasing the mplock.
2102                          */
2103                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2104                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2105                         dqp = dqp->dq_flnext;
2106                         continue;
2107                 }
2108                 /*
2109                  * We're trying to get the hashlock out of order. This races
2110                  * with dqlookup; so, we giveup and goto the next dquot if
2111                  * we couldn't get the hashlock. This way, we won't starve
2112                  * a dqlookup process that holds the hashlock that is
2113                  * waiting for the freelist lock.
2114                  */
2115                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2116                         xfs_dqfunlock(dqp);
2117                         xfs_dqunlock(dqp);
2118                         dqp = dqp->dq_flnext;
2119                         continue;
2120                 }
2121                 /*
2122                  * This races with dquot allocation code as well as dqflush_all
2123                  * and reclaim code. So, if we failed to grab the mplist lock,
2124                  * giveup everything and start over.
2125                  */
2126                 hash = dqp->q_hash;
2127                 ASSERT(hash);
2128                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2129                         /* XXX put a sentinel so that we can come back here */
2130                         xfs_dqfunlock(dqp);
2131                         xfs_dqunlock(dqp);
2132                         XFS_DQ_HASH_UNLOCK(hash);
2133                         xfs_qm_freelist_unlock(xfs_Gqm);
2134                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2135                                 return nreclaimed;
2136                         goto tryagain;
2137                 }
2138                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2139 #ifdef QUOTADEBUG
2140                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2141                         dqp, be32_to_cpu(dqp->q_core.d_id));
2142 #endif
2143                 ASSERT(dqp->q_nrefs == 0);
2144                 nextdqp = dqp->dq_flnext;
2145                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2146                 XQM_HASHLIST_REMOVE(hash, dqp);
2147                 xfs_dqfunlock(dqp);
2148                 xfs_qm_mplist_unlock(dqp->q_mount);
2149                 XFS_DQ_HASH_UNLOCK(hash);
2150
2151  off_freelist:
2152                 XQM_FREELIST_REMOVE(dqp);
2153                 xfs_dqunlock(dqp);
2154                 nreclaimed++;
2155                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2156                 xfs_qm_dqdestroy(dqp);
2157                 dqp = nextdqp;
2158         }
2159         xfs_qm_freelist_unlock(xfs_Gqm);
2160         return nreclaimed;
2161 }
2162
2163
2164 /*
2165  * The kmem_shake interface is invoked when memory is running low.
2166  */
2167 /* ARGSUSED */
2168 STATIC int
2169 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2170 {
2171         int     ndqused, nfree, n;
2172
2173         if (!kmem_shake_allow(gfp_mask))
2174                 return 0;
2175         if (!xfs_Gqm)
2176                 return 0;
2177
2178         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2179         /* incore dquots in all f/s's */
2180         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2181
2182         ASSERT(ndqused >= 0);
2183
2184         if (nfree <= ndqused && nfree < ndquot)
2185                 return 0;
2186
2187         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2188         n = nfree - ndqused - ndquot;           /* # over target */
2189
2190         return xfs_qm_shake_freelist(MAX(nfree, n));
2191 }
2192
2193
2194 /*
2195  * Just pop the least recently used dquot off the freelist and
2196  * recycle it. The returned dquot is locked.
2197  */
2198 STATIC xfs_dquot_t *
2199 xfs_qm_dqreclaim_one(void)
2200 {
2201         xfs_dquot_t     *dqpout;
2202         xfs_dquot_t     *dqp;
2203         int             restarts;
2204         int             nflushes;
2205
2206         restarts = 0;
2207         dqpout = NULL;
2208         nflushes = 0;
2209
2210         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2211  startagain:
2212         xfs_qm_freelist_lock(xfs_Gqm);
2213
2214         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2215                 xfs_dqlock(dqp);
2216
2217                 /*
2218                  * We are racing with dqlookup here. Naturally we don't
2219                  * want to reclaim a dquot that lookup wants. We release the
2220                  * freelist lock and start over, so that lookup will grab
2221                  * both the dquot and the freelistlock.
2222                  */
2223                 if (dqp->dq_flags & XFS_DQ_WANT) {
2224                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2225                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2226                         xfs_dqunlock(dqp);
2227                         xfs_qm_freelist_unlock(xfs_Gqm);
2228                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2229                                 return NULL;
2230                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2231                         goto startagain;
2232                 }
2233
2234                 /*
2235                  * If the dquot is inactive, we are assured that it is
2236                  * not on the mplist or the hashlist, and that makes our
2237                  * life easier.
2238                  */
2239                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2240                         ASSERT(dqp->q_mount == NULL);
2241                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2242                         ASSERT(dqp->HL_PREVP == NULL);
2243                         ASSERT(dqp->MPL_PREVP == NULL);
2244                         XQM_FREELIST_REMOVE(dqp);
2245                         xfs_dqunlock(dqp);
2246                         dqpout = dqp;
2247                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2248                         break;
2249                 }
2250
2251                 ASSERT(dqp->q_hash);
2252                 ASSERT(dqp->MPL_PREVP);
2253
2254                 /*
2255                  * Try to grab the flush lock. If this dquot is in the process of
2256                  * getting flushed to disk, we don't want to reclaim it.
2257                  */
2258                 if (! xfs_qm_dqflock_nowait(dqp)) {
2259                         xfs_dqunlock(dqp);
2260                         continue;
2261                 }
2262
2263                 /*
2264                  * We have the flush lock so we know that this is not in the
2265                  * process of being flushed. So, if this is dirty, flush it
2266                  * DELWRI so that we don't get a freelist infested with
2267                  * dirty dquots.
2268                  */
2269                 if (XFS_DQ_IS_DIRTY(dqp)) {
2270                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2271                         /*
2272                          * We flush it delayed write, so don't bother
2273                          * releasing the freelist lock.
2274                          */
2275                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2276                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2277                         continue;
2278                 }
2279
2280                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2281                         xfs_dqfunlock(dqp);
2282                         xfs_dqunlock(dqp);
2283                         continue;
2284                 }
2285
2286                 if (! xfs_qm_dqhashlock_nowait(dqp))
2287                         goto mplistunlock;
2288
2289                 ASSERT(dqp->q_nrefs == 0);
2290                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2291                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2292                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2293                 XQM_FREELIST_REMOVE(dqp);
2294                 dqpout = dqp;
2295                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2296  mplistunlock:
2297                 xfs_qm_mplist_unlock(dqp->q_mount);
2298                 xfs_dqfunlock(dqp);
2299                 xfs_dqunlock(dqp);
2300                 if (dqpout)
2301                         break;
2302         }
2303
2304         xfs_qm_freelist_unlock(xfs_Gqm);
2305         return dqpout;
2306 }
2307
2308
2309 /*------------------------------------------------------------------*/
2310
2311 /*
2312  * Return a new incore dquot. Depending on the number of
2313  * dquots in the system, we either allocate a new one on the kernel heap,
2314  * or reclaim a free one.
2315  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2316  * to reclaim an existing one from the freelist.
2317  */
2318 boolean_t
2319 xfs_qm_dqalloc_incore(
2320         xfs_dquot_t **O_dqpp)
2321 {
2322         xfs_dquot_t     *dqp;
2323
2324         /*
2325          * Check against high water mark to see if we want to pop
2326          * a nincompoop dquot off the freelist.
2327          */
2328         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2329                 /*
2330                  * Try to recycle a dquot from the freelist.
2331                  */
2332                 if ((dqp = xfs_qm_dqreclaim_one())) {
2333                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2334                         /*
2335                          * Just zero the core here. The rest will get
2336                          * reinitialized by caller. XXX we shouldn't even
2337                          * do this zero ...
2338                          */
2339                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2340                         *O_dqpp = dqp;
2341                         return B_FALSE;
2342                 }
2343                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2344         }
2345
2346         /*
2347          * Allocate a brand new dquot on the kernel heap and return it
2348          * to the caller to initialize.
2349          */
2350         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2351         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2352         atomic_inc(&xfs_Gqm->qm_totaldquots);
2353
2354         return B_TRUE;
2355 }
2356
2357
2358 /*
2359  * Start a transaction and write the incore superblock changes to
2360  * disk. flags parameter indicates which fields have changed.
2361  */
2362 int
2363 xfs_qm_write_sb_changes(
2364         xfs_mount_t     *mp,
2365         __int64_t       flags)
2366 {
2367         xfs_trans_t     *tp;
2368         int             error;
2369
2370 #ifdef QUOTADEBUG
2371         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2372 #endif
2373         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2374         if ((error = xfs_trans_reserve(tp, 0,
2375                                       mp->m_sb.sb_sectsize + 128, 0,
2376                                       0,
2377                                       XFS_DEFAULT_LOG_COUNT))) {
2378                 xfs_trans_cancel(tp, 0);
2379                 return error;
2380         }
2381
2382         xfs_mod_sb(tp, flags);
2383         (void) xfs_trans_commit(tp, 0);
2384
2385         return 0;
2386 }
2387
2388
2389 /* --------------- utility functions for vnodeops ---------------- */
2390
2391
2392 /*
2393  * Given an inode, a uid and gid (from cred_t) make sure that we have
2394  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2395  * quotas by creating this file.
2396  * This also attaches dquot(s) to the given inode after locking it,
2397  * and returns the dquots corresponding to the uid and/or gid.
2398  *
2399  * in   : inode (unlocked)
2400  * out  : udquot, gdquot with references taken and unlocked
2401  */
2402 int
2403 xfs_qm_vop_dqalloc(
2404         xfs_mount_t     *mp,
2405         xfs_inode_t     *ip,
2406         uid_t           uid,
2407         gid_t           gid,
2408         prid_t          prid,
2409         uint            flags,
2410         xfs_dquot_t     **O_udqpp,
2411         xfs_dquot_t     **O_gdqpp)
2412 {
2413         int             error;
2414         xfs_dquot_t     *uq, *gq;
2415         uint            lockflags;
2416
2417         if (!XFS_IS_QUOTA_ON(mp))
2418                 return 0;
2419
2420         lockflags = XFS_ILOCK_EXCL;
2421         xfs_ilock(ip, lockflags);
2422
2423         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2424                 gid = ip->i_d.di_gid;
2425
2426         /*
2427          * Attach the dquot(s) to this inode, doing a dquot allocation
2428          * if necessary. The dquot(s) will not be locked.
2429          */
2430         if (XFS_NOT_DQATTACHED(mp, ip)) {
2431                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2432                                             XFS_QMOPT_ILOCKED))) {
2433                         xfs_iunlock(ip, lockflags);
2434                         return error;
2435                 }
2436         }
2437
2438         uq = gq = NULL;
2439         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2440                 if (ip->i_d.di_uid != uid) {
2441                         /*
2442                          * What we need is the dquot that has this uid, and
2443                          * if we send the inode to dqget, the uid of the inode
2444                          * takes priority over what's sent in the uid argument.
2445                          * We must unlock inode here before calling dqget if
2446                          * we're not sending the inode, because otherwise
2447                          * we'll deadlock by doing trans_reserve while
2448                          * holding ilock.
2449                          */
2450                         xfs_iunlock(ip, lockflags);
2451                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2452                                                  XFS_DQ_USER,
2453                                                  XFS_QMOPT_DQALLOC |
2454                                                  XFS_QMOPT_DOWARN,
2455                                                  &uq))) {
2456                                 ASSERT(error != ENOENT);
2457                                 return error;
2458                         }
2459                         /*
2460                          * Get the ilock in the right order.
2461                          */
2462                         xfs_dqunlock(uq);
2463                         lockflags = XFS_ILOCK_SHARED;
2464                         xfs_ilock(ip, lockflags);
2465                 } else {
2466                         /*
2467                          * Take an extra reference, because we'll return
2468                          * this to caller
2469                          */
2470                         ASSERT(ip->i_udquot);
2471                         uq = ip->i_udquot;
2472                         xfs_dqlock(uq);
2473                         XFS_DQHOLD(uq);
2474                         xfs_dqunlock(uq);
2475                 }
2476         }
2477         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2478                 if (ip->i_d.di_gid != gid) {
2479                         xfs_iunlock(ip, lockflags);
2480                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2481                                                  XFS_DQ_GROUP,
2482                                                  XFS_QMOPT_DQALLOC |
2483                                                  XFS_QMOPT_DOWARN,
2484                                                  &gq))) {
2485                                 if (uq)
2486                                         xfs_qm_dqrele(uq);
2487                                 ASSERT(error != ENOENT);
2488                                 return error;
2489                         }
2490                         xfs_dqunlock(gq);
2491                         lockflags = XFS_ILOCK_SHARED;
2492                         xfs_ilock(ip, lockflags);
2493                 } else {
2494                         ASSERT(ip->i_gdquot);
2495                         gq = ip->i_gdquot;
2496                         xfs_dqlock(gq);
2497                         XFS_DQHOLD(gq);
2498                         xfs_dqunlock(gq);
2499                 }
2500         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2501                 if (ip->i_d.di_projid != prid) {
2502                         xfs_iunlock(ip, lockflags);
2503                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2504                                                  XFS_DQ_PROJ,
2505                                                  XFS_QMOPT_DQALLOC |
2506                                                  XFS_QMOPT_DOWARN,
2507                                                  &gq))) {
2508                                 if (uq)
2509                                         xfs_qm_dqrele(uq);
2510                                 ASSERT(error != ENOENT);
2511                                 return (error);
2512                         }
2513                         xfs_dqunlock(gq);
2514                         lockflags = XFS_ILOCK_SHARED;
2515                         xfs_ilock(ip, lockflags);
2516                 } else {
2517                         ASSERT(ip->i_gdquot);
2518                         gq = ip->i_gdquot;
2519                         xfs_dqlock(gq);
2520                         XFS_DQHOLD(gq);
2521                         xfs_dqunlock(gq);
2522                 }
2523         }
2524         if (uq)
2525                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2526
2527         xfs_iunlock(ip, lockflags);
2528         if (O_udqpp)
2529                 *O_udqpp = uq;
2530         else if (uq)
2531                 xfs_qm_dqrele(uq);
2532         if (O_gdqpp)
2533                 *O_gdqpp = gq;
2534         else if (gq)
2535                 xfs_qm_dqrele(gq);
2536         return 0;
2537 }
2538
2539 /*
2540  * Actually transfer ownership, and do dquot modifications.
2541  * These were already reserved.
2542  */
2543 xfs_dquot_t *
2544 xfs_qm_vop_chown(
2545         xfs_trans_t     *tp,
2546         xfs_inode_t     *ip,
2547         xfs_dquot_t     **IO_olddq,
2548         xfs_dquot_t     *newdq)
2549 {
2550         xfs_dquot_t     *prevdq;
2551         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2552                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2553
2554         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2555         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2556
2557         /* old dquot */
2558         prevdq = *IO_olddq;
2559         ASSERT(prevdq);
2560         ASSERT(prevdq != newdq);
2561
2562         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2563         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2564
2565         /* the sparkling new dquot */
2566         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2567         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2568
2569         /*
2570          * Take an extra reference, because the inode
2571          * is going to keep this dquot pointer even
2572          * after the trans_commit.
2573          */
2574         xfs_dqlock(newdq);
2575         XFS_DQHOLD(newdq);
2576         xfs_dqunlock(newdq);
2577         *IO_olddq = newdq;
2578
2579         return prevdq;
2580 }
2581
2582 /*
2583  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2584  */
2585 int
2586 xfs_qm_vop_chown_reserve(
2587         xfs_trans_t     *tp,
2588         xfs_inode_t     *ip,
2589         xfs_dquot_t     *udqp,
2590         xfs_dquot_t     *gdqp,
2591         uint            flags)
2592 {
2593         int             error;
2594         xfs_mount_t     *mp;
2595         uint            delblks, blkflags, prjflags = 0;
2596         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2597
2598         ASSERT(XFS_ISLOCKED_INODE(ip));
2599         mp = ip->i_mount;
2600         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2601
2602         delblks = ip->i_delayed_blks;
2603         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2604         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2605                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2606
2607         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2608             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2609                 delblksudq = udqp;
2610                 /*
2611                  * If there are delayed allocation blocks, then we have to
2612                  * unreserve those from the old dquot, and add them to the
2613                  * new dquot.
2614                  */
2615                 if (delblks) {
2616                         ASSERT(ip->i_udquot);
2617                         unresudq = ip->i_udquot;
2618                 }
2619         }
2620         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2621                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2622                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2623                         prjflags = XFS_QMOPT_ENOSPC;
2624
2625                 if (prjflags ||
2626                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2627                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2628                         delblksgdq = gdqp;
2629                         if (delblks) {
2630                                 ASSERT(ip->i_gdquot);
2631                                 unresgdq = ip->i_gdquot;
2632                         }
2633                 }
2634         }
2635
2636         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2637                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2638                                 flags | blkflags | prjflags)))
2639                 return (error);
2640
2641         /*
2642          * Do the delayed blks reservations/unreservations now. Since, these
2643          * are done without the help of a transaction, if a reservation fails
2644          * its previous reservations won't be automatically undone by trans
2645          * code. So, we have to do it manually here.
2646          */
2647         if (delblks) {
2648                 /*
2649                  * Do the reservations first. Unreservation can't fail.
2650                  */
2651                 ASSERT(delblksudq || delblksgdq);
2652                 ASSERT(unresudq || unresgdq);
2653                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2654                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2655                                 flags | blkflags | prjflags)))
2656                         return (error);
2657                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2658                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2659                                 blkflags);
2660         }
2661
2662         return (0);
2663 }
2664
2665 int
2666 xfs_qm_vop_rename_dqattach(
2667         xfs_inode_t     **i_tab)
2668 {
2669         xfs_inode_t     *ip;
2670         int             i;
2671         int             error;
2672
2673         ip = i_tab[0];
2674
2675         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2676                 return 0;
2677
2678         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2679                 error = xfs_qm_dqattach(ip, 0);
2680                 if (error)
2681                         return error;
2682         }
2683         for (i = 1; (i < 4 && i_tab[i]); i++) {
2684                 /*
2685                  * Watch out for duplicate entries in the table.
2686                  */
2687                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2688                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2689                                 error = xfs_qm_dqattach(ip, 0);
2690                                 if (error)
2691                                         return error;
2692                         }
2693                 }
2694         }
2695         return 0;
2696 }
2697
2698 void
2699 xfs_qm_vop_dqattach_and_dqmod_newinode(
2700         xfs_trans_t     *tp,
2701         xfs_inode_t     *ip,
2702         xfs_dquot_t     *udqp,
2703         xfs_dquot_t     *gdqp)
2704 {
2705         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2706                 return;
2707
2708         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2709         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2710
2711         if (udqp) {
2712                 xfs_dqlock(udqp);
2713                 XFS_DQHOLD(udqp);
2714                 xfs_dqunlock(udqp);
2715                 ASSERT(ip->i_udquot == NULL);
2716                 ip->i_udquot = udqp;
2717                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2718                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2719                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2720         }
2721         if (gdqp) {
2722                 xfs_dqlock(gdqp);
2723                 XFS_DQHOLD(gdqp);
2724                 xfs_dqunlock(gdqp);
2725                 ASSERT(ip->i_gdquot == NULL);
2726                 ip->i_gdquot = gdqp;
2727                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2728                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2729                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2730                                 be32_to_cpu(gdqp->q_core.d_id));
2731                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2732         }
2733 }
2734
2735 /* ------------- list stuff -----------------*/
2736 STATIC void
2737 xfs_qm_freelist_init(xfs_frlist_t *ql)
2738 {
2739         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2740         mutex_init(&ql->qh_lock);
2741         ql->qh_version = 0;
2742         ql->qh_nelems = 0;
2743 }
2744
2745 STATIC void
2746 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2747 {
2748         xfs_dquot_t     *dqp, *nextdqp;
2749
2750         mutex_lock(&ql->qh_lock);
2751         for (dqp = ql->qh_next;
2752              dqp != (xfs_dquot_t *)ql; ) {
2753                 xfs_dqlock(dqp);
2754                 nextdqp = dqp->dq_flnext;
2755 #ifdef QUOTADEBUG
2756                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2757 #endif
2758                 XQM_FREELIST_REMOVE(dqp);
2759                 xfs_dqunlock(dqp);
2760                 xfs_qm_dqdestroy(dqp);
2761                 dqp = nextdqp;
2762         }
2763         mutex_unlock(&ql->qh_lock);
2764         mutex_destroy(&ql->qh_lock);
2765
2766         ASSERT(ql->qh_nelems == 0);
2767 }
2768
2769 STATIC void
2770 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2771 {
2772         dq->dq_flnext = ql->qh_next;
2773         dq->dq_flprev = (xfs_dquot_t *)ql;
2774         ql->qh_next = dq;
2775         dq->dq_flnext->dq_flprev = dq;
2776         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2777         xfs_Gqm->qm_dqfreelist.qh_version++;
2778 }
2779
2780 void
2781 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2782 {
2783         xfs_dquot_t *next = dq->dq_flnext;
2784         xfs_dquot_t *prev = dq->dq_flprev;
2785
2786         next->dq_flprev = prev;
2787         prev->dq_flnext = next;
2788         dq->dq_flnext = dq->dq_flprev = dq;
2789         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2790         xfs_Gqm->qm_dqfreelist.qh_version++;
2791 }
2792
2793 void
2794 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2795 {
2796         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2797 }
2798
2799 STATIC int
2800 xfs_qm_dqhashlock_nowait(
2801         xfs_dquot_t *dqp)
2802 {
2803         int locked;
2804
2805         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2806         return locked;
2807 }
2808
2809 int
2810 xfs_qm_freelist_lock_nowait(
2811         xfs_qm_t *xqm)
2812 {
2813         int locked;
2814
2815         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2816         return locked;
2817 }
2818
2819 STATIC int
2820 xfs_qm_mplist_nowait(
2821         xfs_mount_t     *mp)
2822 {
2823         int locked;
2824
2825         ASSERT(mp->m_quotainfo);
2826         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2827         return locked;
2828 }