Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[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  * This is called at mount time from xfs_mountfs to initialize the quotainfo
292  * structure and start the global quota manager (xfs_Gqm) if it hasn't done
293  * so already.  Note that the superblock has not been read in yet.
294  */
295 void
296 xfs_qm_mount_quotainit(
297         xfs_mount_t     *mp,
298         uint            flags)
299 {
300         /*
301          * User, projects or group quotas has to be on.
302          */
303         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_PQUOTA | XFSMNT_GQUOTA));
304
305         /*
306          * Initialize the flags in the mount structure. From this point
307          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
308          * Note that we enforce nothing if accounting is off.
309          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
310          * It isn't necessary to take the quotaoff lock to do this; this is
311          * called from mount.
312          */
313         if (flags & XFSMNT_UQUOTA) {
314                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
315                 if (flags & XFSMNT_UQUOTAENF)
316                         mp->m_qflags |= XFS_UQUOTA_ENFD;
317         }
318         if (flags & XFSMNT_GQUOTA) {
319                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
320                 if (flags & XFSMNT_GQUOTAENF)
321                         mp->m_qflags |= XFS_OQUOTA_ENFD;
322         } else if (flags & XFSMNT_PQUOTA) {
323                 mp->m_qflags |= (XFS_PQUOTA_ACCT | XFS_PQUOTA_ACTIVE);
324                 if (flags & XFSMNT_PQUOTAENF)
325                         mp->m_qflags |= XFS_OQUOTA_ENFD;
326         }
327 }
328
329 /*
330  * Just destroy the quotainfo structure.
331  */
332 void
333 xfs_qm_unmount_quotadestroy(
334         xfs_mount_t     *mp)
335 {
336         if (mp->m_quotainfo)
337                 xfs_qm_destroy_quotainfo(mp);
338 }
339
340
341 /*
342  * This is called from xfs_mountfs to start quotas and initialize all
343  * necessary data structures like quotainfo.  This is also responsible for
344  * running a quotacheck as necessary.  We are guaranteed that the superblock
345  * is consistently read in at this point.
346  */
347 int
348 xfs_qm_mount_quotas(
349         xfs_mount_t     *mp,
350         int             mfsi_flags)
351 {
352         unsigned long   s;
353         int             error = 0;
354         uint            sbf;
355
356
357         /*
358          * If quotas on realtime volumes is not supported, we disable
359          * quotas immediately.
360          */
361         if (mp->m_sb.sb_rextents) {
362                 cmn_err(CE_NOTE,
363                         "Cannot turn on quotas for realtime filesystem %s",
364                         mp->m_fsname);
365                 mp->m_qflags = 0;
366                 goto write_changes;
367         }
368
369         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
370
371         /*
372          * Allocate the quotainfo structure inside the mount struct, and
373          * create quotainode(s), and change/rev superblock if necessary.
374          */
375         if ((error = xfs_qm_init_quotainfo(mp))) {
376                 /*
377                  * We must turn off quotas.
378                  */
379                 ASSERT(mp->m_quotainfo == NULL);
380                 mp->m_qflags = 0;
381                 goto write_changes;
382         }
383         /*
384          * If any of the quotas are not consistent, do a quotacheck.
385          */
386         if (XFS_QM_NEED_QUOTACHECK(mp) &&
387                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
388                 if ((error = xfs_qm_quotacheck(mp))) {
389                         /* Quotacheck has failed and quotas have
390                          * been disabled.
391                          */
392                         return XFS_ERROR(error);
393                 }
394         }
395         /* 
396          * If one type of quotas is off, then it will lose its
397          * quotachecked status, since we won't be doing accounting for
398          * that type anymore.
399          */
400         if (!XFS_IS_UQUOTA_ON(mp)) {
401                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
402         }
403         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) {
404                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
405         }
406
407  write_changes:
408         /*
409          * We actually don't have to acquire the SB_LOCK at all.
410          * This can only be called from mount, and that's single threaded. XXX
411          */
412         s = XFS_SB_LOCK(mp);
413         sbf = mp->m_sb.sb_qflags;
414         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
415         XFS_SB_UNLOCK(mp, s);
416
417         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
418                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
419                         /*
420                          * We could only have been turning quotas off.
421                          * We aren't in very good shape actually because
422                          * the incore structures are convinced that quotas are
423                          * off, but the on disk superblock doesn't know that !
424                          */
425                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
426                         xfs_fs_cmn_err(CE_ALERT, mp,
427                                 "XFS mount_quotas: Superblock update failed!");
428                 }
429         }
430
431         if (error) {
432                 xfs_fs_cmn_err(CE_WARN, mp,
433                         "Failed to initialize disk quotas.");
434         }
435         return XFS_ERROR(error);
436 }
437
438 /*
439  * Called from the vfsops layer.
440  */
441 int
442 xfs_qm_unmount_quotas(
443         xfs_mount_t     *mp)
444 {
445         xfs_inode_t     *uqp, *gqp;
446         int             error = 0;
447
448         /*
449          * Release the dquots that root inode, et al might be holding,
450          * before we flush quotas and blow away the quotainfo structure.
451          */
452         ASSERT(mp->m_rootip);
453         xfs_qm_dqdetach(mp->m_rootip);
454         if (mp->m_rbmip)
455                 xfs_qm_dqdetach(mp->m_rbmip);
456         if (mp->m_rsumip)
457                 xfs_qm_dqdetach(mp->m_rsumip);
458
459         /*
460          * Flush out the quota inodes.
461          */
462         uqp = gqp = NULL;
463         if (mp->m_quotainfo) {
464                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
465                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
466                         xfs_iflock(uqp);
467                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
468                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
469                         if (unlikely(error == EFSCORRUPTED)) {
470                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
471                                                  XFS_ERRLEVEL_LOW, mp);
472                                 goto out;
473                         }
474                 }
475                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
476                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
477                         xfs_iflock(gqp);
478                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
479                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
480                         if (unlikely(error == EFSCORRUPTED)) {
481                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
482                                                  XFS_ERRLEVEL_LOW, mp);
483                                 goto out;
484                         }
485                 }
486         }
487         if (uqp) {
488                  XFS_PURGE_INODE(uqp);
489                  mp->m_quotainfo->qi_uquotaip = NULL;
490         }
491         if (gqp) {
492                 XFS_PURGE_INODE(gqp);
493                 mp->m_quotainfo->qi_gquotaip = NULL;
494         }
495 out:
496         return XFS_ERROR(error);
497 }
498
499 /*
500  * Flush all dquots of the given file system to disk. The dquots are
501  * _not_ purged from memory here, just their data written to disk.
502  */
503 STATIC int
504 xfs_qm_dqflush_all(
505         xfs_mount_t     *mp,
506         int             flags)
507 {
508         int             recl;
509         xfs_dquot_t     *dqp;
510         int             niters;
511         int             error;
512
513         if (mp->m_quotainfo == NULL)
514                 return 0;
515         niters = 0;
516 again:
517         xfs_qm_mplist_lock(mp);
518         FOREACH_DQUOT_IN_MP(dqp, mp) {
519                 xfs_dqlock(dqp);
520                 if (! XFS_DQ_IS_DIRTY(dqp)) {
521                         xfs_dqunlock(dqp);
522                         continue;
523                 }
524                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
525                 /* XXX a sentinel would be better */
526                 recl = XFS_QI_MPLRECLAIMS(mp);
527                 if (! xfs_qm_dqflock_nowait(dqp)) {
528                         /*
529                          * If we can't grab the flush lock then check
530                          * to see if the dquot has been flushed delayed
531                          * write.  If so, grab its buffer and send it
532                          * out immediately.  We'll be able to acquire
533                          * the flush lock when the I/O completes.
534                          */
535                         xfs_qm_dqflock_pushbuf_wait(dqp);
536                 }
537                 /*
538                  * Let go of the mplist lock. We don't want to hold it
539                  * across a disk write.
540                  */
541                 xfs_qm_mplist_unlock(mp);
542                 error = xfs_qm_dqflush(dqp, flags);
543                 xfs_dqunlock(dqp);
544                 if (error)
545                         return error;
546
547                 xfs_qm_mplist_lock(mp);
548                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
549                         xfs_qm_mplist_unlock(mp);
550                         /* XXX restart limit */
551                         goto again;
552                 }
553         }
554
555         xfs_qm_mplist_unlock(mp);
556         /* return ! busy */
557         return 0;
558 }
559 /*
560  * Release the group dquot pointers the user dquots may be
561  * carrying around as a hint. mplist is locked on entry and exit.
562  */
563 STATIC void
564 xfs_qm_detach_gdquots(
565         xfs_mount_t     *mp)
566 {
567         xfs_dquot_t     *dqp, *gdqp;
568         int             nrecl;
569
570  again:
571         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
572         dqp = XFS_QI_MPLNEXT(mp);
573         while (dqp) {
574                 xfs_dqlock(dqp);
575                 if ((gdqp = dqp->q_gdquot)) {
576                         xfs_dqlock(gdqp);
577                         dqp->q_gdquot = NULL;
578                 }
579                 xfs_dqunlock(dqp);
580
581                 if (gdqp) {
582                         /*
583                          * Can't hold the mplist lock across a dqput.
584                          * XXXmust convert to marker based iterations here.
585                          */
586                         nrecl = XFS_QI_MPLRECLAIMS(mp);
587                         xfs_qm_mplist_unlock(mp);
588                         xfs_qm_dqput(gdqp);
589
590                         xfs_qm_mplist_lock(mp);
591                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
592                                 goto again;
593                 }
594                 dqp = dqp->MPL_NEXT;
595         }
596 }
597
598 /*
599  * Go through all the incore dquots of this file system and take them
600  * off the mplist and hashlist, if the dquot type matches the dqtype
601  * parameter. This is used when turning off quota accounting for
602  * users and/or groups, as well as when the filesystem is unmounting.
603  */
604 STATIC int
605 xfs_qm_dqpurge_int(
606         xfs_mount_t     *mp,
607         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
608 {
609         xfs_dquot_t     *dqp;
610         uint            dqtype;
611         int             nrecl;
612         xfs_dquot_t     *nextdqp;
613         int             nmisses;
614
615         if (mp->m_quotainfo == NULL)
616                 return 0;
617
618         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
619         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
620         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
621
622         xfs_qm_mplist_lock(mp);
623
624         /*
625          * In the first pass through all incore dquots of this filesystem,
626          * we release the group dquot pointers the user dquots may be
627          * carrying around as a hint. We need to do this irrespective of
628          * what's being turned off.
629          */
630         xfs_qm_detach_gdquots(mp);
631
632       again:
633         nmisses = 0;
634         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
635         /*
636          * Try to get rid of all of the unwanted dquots. The idea is to
637          * get them off mplist and hashlist, but leave them on freelist.
638          */
639         dqp = XFS_QI_MPLNEXT(mp);
640         while (dqp) {
641                 /*
642                  * It's OK to look at the type without taking dqlock here.
643                  * We're holding the mplist lock here, and that's needed for
644                  * a dqreclaim.
645                  */
646                 if ((dqp->dq_flags & dqtype) == 0) {
647                         dqp = dqp->MPL_NEXT;
648                         continue;
649                 }
650
651                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
652                         nrecl = XFS_QI_MPLRECLAIMS(mp);
653                         xfs_qm_mplist_unlock(mp);
654                         XFS_DQ_HASH_LOCK(dqp->q_hash);
655                         xfs_qm_mplist_lock(mp);
656
657                         /*
658                          * XXXTheoretically, we can get into a very long
659                          * ping pong game here.
660                          * No one can be adding dquots to the mplist at
661                          * this point, but somebody might be taking things off.
662                          */
663                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
664                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
665                                 goto again;
666                         }
667                 }
668
669                 /*
670                  * Take the dquot off the mplist and hashlist. It may remain on
671                  * freelist in INACTIVE state.
672                  */
673                 nextdqp = dqp->MPL_NEXT;
674                 nmisses += xfs_qm_dqpurge(dqp, flags);
675                 dqp = nextdqp;
676         }
677         xfs_qm_mplist_unlock(mp);
678         return nmisses;
679 }
680
681 int
682 xfs_qm_dqpurge_all(
683         xfs_mount_t     *mp,
684         uint            flags)
685 {
686         int             ndquots;
687
688         /*
689          * Purge the dquot cache.
690          * None of the dquots should really be busy at this point.
691          */
692         if (mp->m_quotainfo) {
693                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
694                         delay(ndquots * 10);
695                 }
696         }
697         return 0;
698 }
699
700 STATIC int
701 xfs_qm_dqattach_one(
702         xfs_inode_t     *ip,
703         xfs_dqid_t      id,
704         uint            type,
705         uint            doalloc,
706         uint            dolock,
707         xfs_dquot_t     *udqhint, /* hint */
708         xfs_dquot_t     **IO_idqpp)
709 {
710         xfs_dquot_t     *dqp;
711         int             error;
712
713         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
714         error = 0;
715         /*
716          * See if we already have it in the inode itself. IO_idqpp is
717          * &i_udquot or &i_gdquot. This made the code look weird, but
718          * made the logic a lot simpler.
719          */
720         if ((dqp = *IO_idqpp)) {
721                 if (dolock)
722                         xfs_dqlock(dqp);
723                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
724                 goto done;
725         }
726
727         /*
728          * udqhint is the i_udquot field in inode, and is non-NULL only
729          * when the type arg is group/project. Its purpose is to save a
730          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
731          * the user dquot.
732          */
733         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
734         if (udqhint && !dolock)
735                 xfs_dqlock(udqhint);
736
737         /*
738          * No need to take dqlock to look at the id.
739          * The ID can't change until it gets reclaimed, and it won't
740          * be reclaimed as long as we have a ref from inode and we hold
741          * the ilock.
742          */
743         if (udqhint &&
744             (dqp = udqhint->q_gdquot) &&
745             (be32_to_cpu(dqp->q_core.d_id) == id)) {
746                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
747                 xfs_dqlock(dqp);
748                 XFS_DQHOLD(dqp);
749                 ASSERT(*IO_idqpp == NULL);
750                 *IO_idqpp = dqp;
751                 if (!dolock) {
752                         xfs_dqunlock(dqp);
753                         xfs_dqunlock(udqhint);
754                 }
755                 goto done;
756         }
757         /*
758          * We can't hold a dquot lock when we call the dqget code.
759          * We'll deadlock in no time, because of (not conforming to)
760          * lock ordering - the inodelock comes before any dquot lock,
761          * and we may drop and reacquire the ilock in xfs_qm_dqget().
762          */
763         if (udqhint)
764                 xfs_dqunlock(udqhint);
765         /*
766          * Find the dquot from somewhere. This bumps the
767          * reference count of dquot and returns it locked.
768          * This can return ENOENT if dquot didn't exist on
769          * disk and we didn't ask it to allocate;
770          * ESRCH if quotas got turned off suddenly.
771          */
772         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
773                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
774                 if (udqhint && dolock)
775                         xfs_dqlock(udqhint);
776                 goto done;
777         }
778
779         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
780         /*
781          * dqget may have dropped and re-acquired the ilock, but it guarantees
782          * that the dquot returned is the one that should go in the inode.
783          */
784         *IO_idqpp = dqp;
785         ASSERT(dqp);
786         ASSERT(XFS_DQ_IS_LOCKED(dqp));
787         if (! dolock) {
788                 xfs_dqunlock(dqp);
789                 goto done;
790         }
791         if (! udqhint)
792                 goto done;
793
794         ASSERT(udqhint);
795         ASSERT(dolock);
796         ASSERT(XFS_DQ_IS_LOCKED(dqp));
797         if (! xfs_qm_dqlock_nowait(udqhint)) {
798                 xfs_dqunlock(dqp);
799                 xfs_dqlock(udqhint);
800                 xfs_dqlock(dqp);
801         }
802       done:
803 #ifdef QUOTADEBUG
804         if (udqhint) {
805                 if (dolock)
806                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
807         }
808         if (! error) {
809                 if (dolock)
810                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
811         }
812 #endif
813         return error;
814 }
815
816
817 /*
818  * Given a udquot and gdquot, attach a ptr to the group dquot in the
819  * udquot as a hint for future lookups. The idea sounds simple, but the
820  * execution isn't, because the udquot might have a group dquot attached
821  * already and getting rid of that gets us into lock ordering constraints.
822  * The process is complicated more by the fact that the dquots may or may not
823  * be locked on entry.
824  */
825 STATIC void
826 xfs_qm_dqattach_grouphint(
827         xfs_dquot_t     *udq,
828         xfs_dquot_t     *gdq,
829         uint            locked)
830 {
831         xfs_dquot_t     *tmp;
832
833 #ifdef QUOTADEBUG
834         if (locked) {
835                 ASSERT(XFS_DQ_IS_LOCKED(udq));
836                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
837         }
838 #endif
839         if (! locked)
840                 xfs_dqlock(udq);
841
842         if ((tmp = udq->q_gdquot)) {
843                 if (tmp == gdq) {
844                         if (! locked)
845                                 xfs_dqunlock(udq);
846                         return;
847                 }
848
849                 udq->q_gdquot = NULL;
850                 /*
851                  * We can't keep any dqlocks when calling dqrele,
852                  * because the freelist lock comes before dqlocks.
853                  */
854                 xfs_dqunlock(udq);
855                 if (locked)
856                         xfs_dqunlock(gdq);
857                 /*
858                  * we took a hard reference once upon a time in dqget,
859                  * so give it back when the udquot no longer points at it
860                  * dqput() does the unlocking of the dquot.
861                  */
862                 xfs_qm_dqrele(tmp);
863
864                 xfs_dqlock(udq);
865                 xfs_dqlock(gdq);
866
867         } else {
868                 ASSERT(XFS_DQ_IS_LOCKED(udq));
869                 if (! locked) {
870                         xfs_dqlock(gdq);
871                 }
872         }
873
874         ASSERT(XFS_DQ_IS_LOCKED(udq));
875         ASSERT(XFS_DQ_IS_LOCKED(gdq));
876         /*
877          * Somebody could have attached a gdquot here,
878          * when we dropped the uqlock. If so, just do nothing.
879          */
880         if (udq->q_gdquot == NULL) {
881                 XFS_DQHOLD(gdq);
882                 udq->q_gdquot = gdq;
883         }
884         if (! locked) {
885                 xfs_dqunlock(gdq);
886                 xfs_dqunlock(udq);
887         }
888 }
889
890
891 /*
892  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
893  * into account.
894  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
895  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
896  * much made this code a complete mess, but it has been pretty useful.
897  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
898  * Inode may get unlocked and relocked in here, and the caller must deal with
899  * the consequences.
900  */
901 int
902 xfs_qm_dqattach(
903         xfs_inode_t     *ip,
904         uint            flags)
905 {
906         xfs_mount_t     *mp = ip->i_mount;
907         uint            nquotas = 0;
908         int             error = 0;
909
910         if ((! XFS_IS_QUOTA_ON(mp)) ||
911             (! XFS_NOT_DQATTACHED(mp, ip)) ||
912             (ip->i_ino == mp->m_sb.sb_uquotino) ||
913             (ip->i_ino == mp->m_sb.sb_gquotino))
914                 return 0;
915
916         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
917                XFS_ISLOCKED_INODE_EXCL(ip));
918
919         if (! (flags & XFS_QMOPT_ILOCKED))
920                 xfs_ilock(ip, XFS_ILOCK_EXCL);
921
922         if (XFS_IS_UQUOTA_ON(mp)) {
923                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
924                                                 flags & XFS_QMOPT_DQALLOC,
925                                                 flags & XFS_QMOPT_DQLOCK,
926                                                 NULL, &ip->i_udquot);
927                 if (error)
928                         goto done;
929                 nquotas++;
930         }
931         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
932         if (XFS_IS_OQUOTA_ON(mp)) {
933                 error = XFS_IS_GQUOTA_ON(mp) ?
934                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
935                                                 flags & XFS_QMOPT_DQALLOC,
936                                                 flags & XFS_QMOPT_DQLOCK,
937                                                 ip->i_udquot, &ip->i_gdquot) :
938                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
939                                                 flags & XFS_QMOPT_DQALLOC,
940                                                 flags & XFS_QMOPT_DQLOCK,
941                                                 ip->i_udquot, &ip->i_gdquot);
942                 /*
943                  * Don't worry about the udquot that we may have
944                  * attached above. It'll get detached, if not already.
945                  */
946                 if (error)
947                         goto done;
948                 nquotas++;
949         }
950
951         /*
952          * Attach this group quota to the user quota as a hint.
953          * This WON'T, in general, result in a thrash.
954          */
955         if (nquotas == 2) {
956                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
957                 ASSERT(ip->i_udquot);
958                 ASSERT(ip->i_gdquot);
959
960                 /*
961                  * We may or may not have the i_udquot locked at this point,
962                  * but this check is OK since we don't depend on the i_gdquot to
963                  * be accurate 100% all the time. It is just a hint, and this
964                  * will succeed in general.
965                  */
966                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
967                         goto done;
968                 /*
969                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
970                  */
971                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
972                                          flags & XFS_QMOPT_DQLOCK);
973         }
974
975       done:
976
977 #ifdef QUOTADEBUG
978         if (! error) {
979                 if (ip->i_udquot) {
980                         if (flags & XFS_QMOPT_DQLOCK)
981                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
982                 }
983                 if (ip->i_gdquot) {
984                         if (flags & XFS_QMOPT_DQLOCK)
985                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
986                 }
987                 if (XFS_IS_UQUOTA_ON(mp))
988                         ASSERT(ip->i_udquot);
989                 if (XFS_IS_OQUOTA_ON(mp))
990                         ASSERT(ip->i_gdquot);
991         }
992 #endif
993
994         if (! (flags & XFS_QMOPT_ILOCKED))
995                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
996
997 #ifdef QUOTADEBUG
998         else
999                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1000 #endif
1001         return error;
1002 }
1003
1004 /*
1005  * Release dquots (and their references) if any.
1006  * The inode should be locked EXCL except when this's called by
1007  * xfs_ireclaim.
1008  */
1009 void
1010 xfs_qm_dqdetach(
1011         xfs_inode_t     *ip)
1012 {
1013         if (!(ip->i_udquot || ip->i_gdquot))
1014                 return;
1015
1016         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1017         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1018         if (ip->i_udquot) {
1019                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1020                 xfs_qm_dqrele(ip->i_udquot);
1021                 ip->i_udquot = NULL;
1022         }
1023         if (ip->i_gdquot) {
1024                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
1025                 xfs_qm_dqrele(ip->i_gdquot);
1026                 ip->i_gdquot = NULL;
1027         }
1028 }
1029
1030 /*
1031  * This is called by VFS_SYNC and flags arg determines the caller,
1032  * and its motives, as done in xfs_sync.
1033  *
1034  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1035  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1036  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1037  */
1038
1039 int
1040 xfs_qm_sync(
1041         xfs_mount_t     *mp,
1042         short           flags)
1043 {
1044         int             recl, restarts;
1045         xfs_dquot_t     *dqp;
1046         uint            flush_flags;
1047         boolean_t       nowait;
1048         int             error;
1049
1050         restarts = 0;
1051         /*
1052          * We won't block unless we are asked to.
1053          */
1054         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1055
1056   again:
1057         xfs_qm_mplist_lock(mp);
1058         /*
1059          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1060          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1061          * when we have the mplist lock, we know that dquots will be consistent
1062          * as long as we have it locked.
1063          */
1064         if (! XFS_IS_QUOTA_ON(mp)) {
1065                 xfs_qm_mplist_unlock(mp);
1066                 return 0;
1067         }
1068         FOREACH_DQUOT_IN_MP(dqp, mp) {
1069                 /*
1070                  * If this is vfs_sync calling, then skip the dquots that
1071                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1072                  * This is very similar to what xfs_sync does with inodes.
1073                  */
1074                 if (flags & SYNC_BDFLUSH) {
1075                         if (! XFS_DQ_IS_DIRTY(dqp))
1076                                 continue;
1077                 }
1078
1079                 if (nowait) {
1080                         /*
1081                          * Try to acquire the dquot lock. We are NOT out of
1082                          * lock order, but we just don't want to wait for this
1083                          * lock, unless somebody wanted us to.
1084                          */
1085                         if (! xfs_qm_dqlock_nowait(dqp))
1086                                 continue;
1087                 } else {
1088                         xfs_dqlock(dqp);
1089                 }
1090
1091                 /*
1092                  * Now, find out for sure if this dquot is dirty or not.
1093                  */
1094                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1095                         xfs_dqunlock(dqp);
1096                         continue;
1097                 }
1098
1099                 /* XXX a sentinel would be better */
1100                 recl = XFS_QI_MPLRECLAIMS(mp);
1101                 if (! xfs_qm_dqflock_nowait(dqp)) {
1102                         if (nowait) {
1103                                 xfs_dqunlock(dqp);
1104                                 continue;
1105                         }
1106                         /*
1107                          * If we can't grab the flush lock then if the caller
1108                          * really wanted us to give this our best shot, so
1109                          * see if we can give a push to the buffer before we wait
1110                          * on the flush lock. At this point, we know that
1111                          * even though the dquot is being flushed,
1112                          * it has (new) dirty data.
1113                          */
1114                         xfs_qm_dqflock_pushbuf_wait(dqp);
1115                 }
1116                 /*
1117                  * Let go of the mplist lock. We don't want to hold it
1118                  * across a disk write
1119                  */
1120                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1121                 xfs_qm_mplist_unlock(mp);
1122                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1123                 error = xfs_qm_dqflush(dqp, flush_flags);
1124                 xfs_dqunlock(dqp);
1125                 if (error && XFS_FORCED_SHUTDOWN(mp))
1126                         return 0;       /* Need to prevent umount failure */
1127                 else if (error)
1128                         return error;
1129
1130                 xfs_qm_mplist_lock(mp);
1131                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1132                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1133                                 break;
1134
1135                         xfs_qm_mplist_unlock(mp);
1136                         goto again;
1137                 }
1138         }
1139
1140         xfs_qm_mplist_unlock(mp);
1141         return 0;
1142 }
1143
1144
1145 /*
1146  * This initializes all the quota information that's kept in the
1147  * mount structure
1148  */
1149 STATIC int
1150 xfs_qm_init_quotainfo(
1151         xfs_mount_t     *mp)
1152 {
1153         xfs_quotainfo_t *qinf;
1154         int             error;
1155         xfs_dquot_t     *dqp;
1156
1157         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1158
1159         /*
1160          * Tell XQM that we exist as soon as possible.
1161          */
1162         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1163                 return error;
1164         }
1165
1166         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1167
1168         /*
1169          * See if quotainodes are setup, and if not, allocate them,
1170          * and change the superblock accordingly.
1171          */
1172         if ((error = xfs_qm_init_quotainos(mp))) {
1173                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1174                 mp->m_quotainfo = NULL;
1175                 return error;
1176         }
1177
1178         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1179         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1180         qinf->qi_dqreclaims = 0;
1181
1182         /* mutex used to serialize quotaoffs */
1183         mutex_init(&qinf->qi_quotaofflock);
1184
1185         /* Precalc some constants */
1186         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1187         ASSERT(qinf->qi_dqchunklen);
1188         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1189         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1190
1191         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1192
1193         /*
1194          * We try to get the limits from the superuser's limits fields.
1195          * This is quite hacky, but it is standard quota practice.
1196          * We look at the USR dquot with id == 0 first, but if user quotas
1197          * are not enabled we goto the GRP dquot with id == 0.
1198          * We don't really care to keep separate default limits for user
1199          * and group quotas, at least not at this point.
1200          */
1201         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1202                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1203                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1204                                 XFS_DQ_PROJ),
1205                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1206                              &dqp);
1207         if (! error) {
1208                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1209
1210                 /*
1211                  * The warnings and timers set the grace period given to
1212                  * a user or group before he or she can not perform any
1213                  * more writing. If it is zero, a default is used.
1214                  */
1215                 qinf->qi_btimelimit = ddqp->d_btimer ?
1216                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1217                 qinf->qi_itimelimit = ddqp->d_itimer ?
1218                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1219                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1220                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1221                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1222                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1223                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1224                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1225                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1226                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1227                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1228                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1229                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1230                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1231                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1232                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1233  
1234                 /*
1235                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1236                  * we don't want this dquot cached. We haven't done a
1237                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1238                  */
1239                 xfs_qm_dqdestroy(dqp);
1240         } else {
1241                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1242                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1243                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1244                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1245                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1246                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1247         }
1248
1249         return 0;
1250 }
1251
1252
1253 /*
1254  * Gets called when unmounting a filesystem or when all quotas get
1255  * turned off.
1256  * This purges the quota inodes, destroys locks and frees itself.
1257  */
1258 void
1259 xfs_qm_destroy_quotainfo(
1260         xfs_mount_t     *mp)
1261 {
1262         xfs_quotainfo_t *qi;
1263
1264         qi = mp->m_quotainfo;
1265         ASSERT(qi != NULL);
1266         ASSERT(xfs_Gqm != NULL);
1267
1268         /*
1269          * Release the reference that XQM kept, so that we know
1270          * when the XQM structure should be freed. We cannot assume
1271          * that xfs_Gqm is non-null after this point.
1272          */
1273         xfs_qm_rele_quotafs_ref(mp);
1274
1275         spinlock_destroy(&qi->qi_pinlock);
1276         xfs_qm_list_destroy(&qi->qi_dqlist);
1277
1278         if (qi->qi_uquotaip) {
1279                 XFS_PURGE_INODE(qi->qi_uquotaip);
1280                 qi->qi_uquotaip = NULL; /* paranoia */
1281         }
1282         if (qi->qi_gquotaip) {
1283                 XFS_PURGE_INODE(qi->qi_gquotaip);
1284                 qi->qi_gquotaip = NULL;
1285         }
1286         mutex_destroy(&qi->qi_quotaofflock);
1287         kmem_free(qi, sizeof(xfs_quotainfo_t));
1288         mp->m_quotainfo = NULL;
1289 }
1290
1291
1292
1293 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1294
1295 /* ARGSUSED */
1296 STATIC void
1297 xfs_qm_list_init(
1298         xfs_dqlist_t    *list,
1299         char            *str,
1300         int             n)
1301 {
1302         mutex_init(&list->qh_lock);
1303         list->qh_next = NULL;
1304         list->qh_version = 0;
1305         list->qh_nelems = 0;
1306 }
1307
1308 STATIC void
1309 xfs_qm_list_destroy(
1310         xfs_dqlist_t    *list)
1311 {
1312         mutex_destroy(&(list->qh_lock));
1313 }
1314
1315
1316 /*
1317  * Stripped down version of dqattach. This doesn't attach, or even look at the
1318  * dquots attached to the inode. The rationale is that there won't be any
1319  * attached at the time this is called from quotacheck.
1320  */
1321 STATIC int
1322 xfs_qm_dqget_noattach(
1323         xfs_inode_t     *ip,
1324         xfs_dquot_t     **O_udqpp,
1325         xfs_dquot_t     **O_gdqpp)
1326 {
1327         int             error;
1328         xfs_mount_t     *mp;
1329         xfs_dquot_t     *udqp, *gdqp;
1330
1331         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1332         mp = ip->i_mount;
1333         udqp = NULL;
1334         gdqp = NULL;
1335
1336         if (XFS_IS_UQUOTA_ON(mp)) {
1337                 ASSERT(ip->i_udquot == NULL);
1338                 /*
1339                  * We want the dquot allocated if it doesn't exist.
1340                  */
1341                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1342                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1343                                          &udqp))) {
1344                         /*
1345                          * Shouldn't be able to turn off quotas here.
1346                          */
1347                         ASSERT(error != ESRCH);
1348                         ASSERT(error != ENOENT);
1349                         return error;
1350                 }
1351                 ASSERT(udqp);
1352         }
1353
1354         if (XFS_IS_OQUOTA_ON(mp)) {
1355                 ASSERT(ip->i_gdquot == NULL);
1356                 if (udqp)
1357                         xfs_dqunlock(udqp);
1358                 error = XFS_IS_GQUOTA_ON(mp) ?
1359                                 xfs_qm_dqget(mp, ip,
1360                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1361                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1362                                              &gdqp) :
1363                                 xfs_qm_dqget(mp, ip,
1364                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1365                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1366                                              &gdqp);
1367                 if (error) {
1368                         if (udqp)
1369                                 xfs_qm_dqrele(udqp);
1370                         ASSERT(error != ESRCH);
1371                         ASSERT(error != ENOENT);
1372                         return error;
1373                 }
1374                 ASSERT(gdqp);
1375
1376                 /* Reacquire the locks in the right order */
1377                 if (udqp) {
1378                         if (! xfs_qm_dqlock_nowait(udqp)) {
1379                                 xfs_dqunlock(gdqp);
1380                                 xfs_dqlock(udqp);
1381                                 xfs_dqlock(gdqp);
1382                         }
1383                 }
1384         }
1385
1386         *O_udqpp = udqp;
1387         *O_gdqpp = gdqp;
1388
1389 #ifdef QUOTADEBUG
1390         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1391         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1392 #endif
1393         return 0;
1394 }
1395
1396 /*
1397  * Create an inode and return with a reference already taken, but unlocked
1398  * This is how we create quota inodes
1399  */
1400 STATIC int
1401 xfs_qm_qino_alloc(
1402         xfs_mount_t     *mp,
1403         xfs_inode_t     **ip,
1404         __int64_t       sbfields,
1405         uint            flags)
1406 {
1407         xfs_trans_t     *tp;
1408         int             error;
1409         unsigned long   s;
1410         int             committed;
1411
1412         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1413         if ((error = xfs_trans_reserve(tp,
1414                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1415                                       XFS_CREATE_LOG_RES(mp), 0,
1416                                       XFS_TRANS_PERM_LOG_RES,
1417                                       XFS_CREATE_LOG_COUNT))) {
1418                 xfs_trans_cancel(tp, 0);
1419                 return error;
1420         }
1421
1422         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1423                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1424                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1425                                  XFS_TRANS_ABORT);
1426                 return error;
1427         }
1428
1429         /*
1430          * Keep an extra reference to this quota inode. This inode is
1431          * locked exclusively and joined to the transaction already.
1432          */
1433         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1434         VN_HOLD(XFS_ITOV((*ip)));
1435
1436         /*
1437          * Make the changes in the superblock, and log those too.
1438          * sbfields arg may contain fields other than *QUOTINO;
1439          * VERSIONNUM for example.
1440          */
1441         s = XFS_SB_LOCK(mp);
1442         if (flags & XFS_QMOPT_SBVERSION) {
1443 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1444                 unsigned oldv = mp->m_sb.sb_versionnum;
1445 #endif
1446                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1447                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1448                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1449                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1450                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1451
1452                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1453                 mp->m_sb.sb_uquotino = NULLFSINO;
1454                 mp->m_sb.sb_gquotino = NULLFSINO;
1455
1456                 /* qflags will get updated _after_ quotacheck */
1457                 mp->m_sb.sb_qflags = 0;
1458 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1459                 cmn_err(CE_NOTE,
1460                         "Old superblock version %x, converting to %x.",
1461                         oldv, mp->m_sb.sb_versionnum);
1462 #endif
1463         }
1464         if (flags & XFS_QMOPT_UQUOTA)
1465                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1466         else
1467                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1468         XFS_SB_UNLOCK(mp, s);
1469         xfs_mod_sb(tp, sbfields);
1470
1471         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1472                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1473                 return error;
1474         }
1475         return 0;
1476 }
1477
1478
1479 STATIC int
1480 xfs_qm_reset_dqcounts(
1481         xfs_mount_t     *mp,
1482         xfs_buf_t       *bp,
1483         xfs_dqid_t      id,
1484         uint            type)
1485 {
1486         xfs_disk_dquot_t        *ddq;
1487         int                     j;
1488
1489         xfs_buftrace("RESET DQUOTS", bp);
1490         /*
1491          * Reset all counters and timers. They'll be
1492          * started afresh by xfs_qm_quotacheck.
1493          */
1494 #ifdef DEBUG
1495         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1496         do_div(j, sizeof(xfs_dqblk_t));
1497         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1498 #endif
1499         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1500         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1501                 /*
1502                  * Do a sanity check, and if needed, repair the dqblk. Don't
1503                  * output any warnings because it's perfectly possible to
1504                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1505                  */
1506                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1507                                       "xfs_quotacheck");
1508                 ddq->d_bcount = 0;
1509                 ddq->d_icount = 0;
1510                 ddq->d_rtbcount = 0;
1511                 ddq->d_btimer = 0;
1512                 ddq->d_itimer = 0;
1513                 ddq->d_rtbtimer = 0;
1514                 ddq->d_bwarns = 0;
1515                 ddq->d_iwarns = 0;
1516                 ddq->d_rtbwarns = 0;
1517                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1518         }
1519
1520         return 0;
1521 }
1522
1523 STATIC int
1524 xfs_qm_dqiter_bufs(
1525         xfs_mount_t     *mp,
1526         xfs_dqid_t      firstid,
1527         xfs_fsblock_t   bno,
1528         xfs_filblks_t   blkcnt,
1529         uint            flags)
1530 {
1531         xfs_buf_t       *bp;
1532         int             error;
1533         int             notcommitted;
1534         int             incr;
1535         int             type;
1536
1537         ASSERT(blkcnt > 0);
1538         notcommitted = 0;
1539         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1540                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1541         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1542                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1543         error = 0;
1544
1545         /*
1546          * Blkcnt arg can be a very big number, and might even be
1547          * larger than the log itself. So, we have to break it up into
1548          * manageable-sized transactions.
1549          * Note that we don't start a permanent transaction here; we might
1550          * not be able to get a log reservation for the whole thing up front,
1551          * and we don't really care to either, because we just discard
1552          * everything if we were to crash in the middle of this loop.
1553          */
1554         while (blkcnt--) {
1555                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1556                               XFS_FSB_TO_DADDR(mp, bno),
1557                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1558                 if (error)
1559                         break;
1560
1561                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1562                 xfs_bdwrite(mp, bp);
1563                 /*
1564                  * goto the next block.
1565                  */
1566                 bno++;
1567                 firstid += XFS_QM_DQPERBLK(mp);
1568         }
1569         return error;
1570 }
1571
1572 /*
1573  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1574  * caller supplied function for every chunk of dquots that we find.
1575  */
1576 STATIC int
1577 xfs_qm_dqiterate(
1578         xfs_mount_t     *mp,
1579         xfs_inode_t     *qip,
1580         uint            flags)
1581 {
1582         xfs_bmbt_irec_t         *map;
1583         int                     i, nmaps;       /* number of map entries */
1584         int                     error;          /* return value */
1585         xfs_fileoff_t           lblkno;
1586         xfs_filblks_t           maxlblkcnt;
1587         xfs_dqid_t              firstid;
1588         xfs_fsblock_t           rablkno;
1589         xfs_filblks_t           rablkcnt;
1590
1591         error = 0;
1592         /*
1593          * This looks racy, but we can't keep an inode lock across a
1594          * trans_reserve. But, this gets called during quotacheck, and that
1595          * happens only at mount time which is single threaded.
1596          */
1597         if (qip->i_d.di_nblocks == 0)
1598                 return 0;
1599
1600         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1601
1602         lblkno = 0;
1603         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1604         do {
1605                 nmaps = XFS_DQITER_MAP_SIZE;
1606                 /*
1607                  * We aren't changing the inode itself. Just changing
1608                  * some of its data. No new blocks are added here, and
1609                  * the inode is never added to the transaction.
1610                  */
1611                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1612                 error = xfs_bmapi(NULL, qip, lblkno,
1613                                   maxlblkcnt - lblkno,
1614                                   XFS_BMAPI_METADATA,
1615                                   NULL,
1616                                   0, map, &nmaps, NULL, NULL);
1617                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1618                 if (error)
1619                         break;
1620
1621                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1622                 for (i = 0; i < nmaps; i++) {
1623                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1624                         ASSERT(map[i].br_blockcount);
1625
1626
1627                         lblkno += map[i].br_blockcount;
1628
1629                         if (map[i].br_startblock == HOLESTARTBLOCK)
1630                                 continue;
1631
1632                         firstid = (xfs_dqid_t) map[i].br_startoff *
1633                                 XFS_QM_DQPERBLK(mp);
1634                         /*
1635                          * Do a read-ahead on the next extent.
1636                          */
1637                         if ((i+1 < nmaps) &&
1638                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1639                                 rablkcnt =  map[i+1].br_blockcount;
1640                                 rablkno = map[i+1].br_startblock;
1641                                 while (rablkcnt--) {
1642                                         xfs_baread(mp->m_ddev_targp,
1643                                                XFS_FSB_TO_DADDR(mp, rablkno),
1644                                                (int)XFS_QI_DQCHUNKLEN(mp));
1645                                         rablkno++;
1646                                 }
1647                         }
1648                         /*
1649                          * Iterate thru all the blks in the extent and
1650                          * reset the counters of all the dquots inside them.
1651                          */
1652                         if ((error = xfs_qm_dqiter_bufs(mp,
1653                                                        firstid,
1654                                                        map[i].br_startblock,
1655                                                        map[i].br_blockcount,
1656                                                        flags))) {
1657                                 break;
1658                         }
1659                 }
1660
1661                 if (error)
1662                         break;
1663         } while (nmaps > 0);
1664
1665         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1666
1667         return error;
1668 }
1669
1670 /*
1671  * Called by dqusage_adjust in doing a quotacheck.
1672  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1673  * this updates its incore copy as well as the buffer copy. This is
1674  * so that once the quotacheck is done, we can just log all the buffers,
1675  * as opposed to logging numerous updates to individual dquots.
1676  */
1677 STATIC void
1678 xfs_qm_quotacheck_dqadjust(
1679         xfs_dquot_t             *dqp,
1680         xfs_qcnt_t              nblks,
1681         xfs_qcnt_t              rtblks)
1682 {
1683         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1684         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1685         /*
1686          * Adjust the inode count and the block count to reflect this inode's
1687          * resource usage.
1688          */
1689         be64_add(&dqp->q_core.d_icount, 1);
1690         dqp->q_res_icount++;
1691         if (nblks) {
1692                 be64_add(&dqp->q_core.d_bcount, nblks);
1693                 dqp->q_res_bcount += nblks;
1694         }
1695         if (rtblks) {
1696                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1697                 dqp->q_res_rtbcount += rtblks;
1698         }
1699
1700         /*
1701          * Set default limits, adjust timers (since we changed usages)
1702          */
1703         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1704                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1705                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1706         }
1707
1708         dqp->dq_flags |= XFS_DQ_DIRTY;
1709 }
1710
1711 STATIC int
1712 xfs_qm_get_rtblks(
1713         xfs_inode_t     *ip,
1714         xfs_qcnt_t      *O_rtblks)
1715 {
1716         xfs_filblks_t   rtblks;                 /* total rt blks */
1717         xfs_extnum_t    idx;                    /* extent record index */
1718         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1719         xfs_extnum_t    nextents;               /* number of extent entries */
1720         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1721         int             error;
1722
1723         ASSERT(XFS_IS_REALTIME_INODE(ip));
1724         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1725         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1726                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1727                         return error;
1728         }
1729         rtblks = 0;
1730         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1731         for (idx = 0; idx < nextents; idx++) {
1732                 ep = xfs_iext_get_ext(ifp, idx);
1733                 rtblks += xfs_bmbt_get_blockcount(ep);
1734         }
1735         *O_rtblks = (xfs_qcnt_t)rtblks;
1736         return 0;
1737 }
1738
1739 /*
1740  * callback routine supplied to bulkstat(). Given an inumber, find its
1741  * dquots and update them to account for resources taken by that inode.
1742  */
1743 /* ARGSUSED */
1744 STATIC int
1745 xfs_qm_dqusage_adjust(
1746         xfs_mount_t     *mp,            /* mount point for filesystem */
1747         xfs_ino_t       ino,            /* inode number to get data for */
1748         void            __user *buffer, /* not used */
1749         int             ubsize,         /* not used */
1750         void            *private_data,  /* not used */
1751         xfs_daddr_t     bno,            /* starting block of inode cluster */
1752         int             *ubused,        /* not used */
1753         void            *dip,           /* on-disk inode pointer (not used) */
1754         int             *res)           /* result code value */
1755 {
1756         xfs_inode_t     *ip;
1757         xfs_dquot_t     *udqp, *gdqp;
1758         xfs_qcnt_t      nblks, rtblks;
1759         int             error;
1760
1761         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1762
1763         /*
1764          * rootino must have its resources accounted for, not so with the quota
1765          * inodes.
1766          */
1767         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1768                 *res = BULKSTAT_RV_NOTHING;
1769                 return XFS_ERROR(EINVAL);
1770         }
1771
1772         /*
1773          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1774          * interface expects the inode to be exclusively locked because that's
1775          * the case in all other instances. It's OK that we do this because
1776          * quotacheck is done only at mount time.
1777          */
1778         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1779                 *res = BULKSTAT_RV_NOTHING;
1780                 return error;
1781         }
1782
1783         if (ip->i_d.di_mode == 0) {
1784                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1785                 *res = BULKSTAT_RV_NOTHING;
1786                 return XFS_ERROR(ENOENT);
1787         }
1788
1789         /*
1790          * Obtain the locked dquots. In case of an error (eg. allocation
1791          * fails for ENOSPC), we return the negative of the error number
1792          * to bulkstat, so that it can get propagated to quotacheck() and
1793          * making us disable quotas for the file system.
1794          */
1795         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1796                 xfs_iput(ip, XFS_ILOCK_EXCL);
1797                 *res = BULKSTAT_RV_GIVEUP;
1798                 return error;
1799         }
1800
1801         rtblks = 0;
1802         if (! XFS_IS_REALTIME_INODE(ip)) {
1803                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1804         } else {
1805                 /*
1806                  * Walk thru the extent list and count the realtime blocks.
1807                  */
1808                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1809                         xfs_iput(ip, XFS_ILOCK_EXCL);
1810                         if (udqp)
1811                                 xfs_qm_dqput(udqp);
1812                         if (gdqp)
1813                                 xfs_qm_dqput(gdqp);
1814                         *res = BULKSTAT_RV_GIVEUP;
1815                         return error;
1816                 }
1817                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1818         }
1819         ASSERT(ip->i_delayed_blks == 0);
1820
1821         /*
1822          * We can't release the inode while holding its dquot locks.
1823          * The inode can go into inactive and might try to acquire the dquotlocks.
1824          * So, just unlock here and do a vn_rele at the end.
1825          */
1826         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1827
1828         /*
1829          * Add the (disk blocks and inode) resources occupied by this
1830          * inode to its dquots. We do this adjustment in the incore dquot,
1831          * and also copy the changes to its buffer.
1832          * We don't care about putting these changes in a transaction
1833          * envelope because if we crash in the middle of a 'quotacheck'
1834          * we have to start from the beginning anyway.
1835          * Once we're done, we'll log all the dquot bufs.
1836          *
1837          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1838          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1839          */
1840         if (XFS_IS_UQUOTA_ON(mp)) {
1841                 ASSERT(udqp);
1842                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1843                 xfs_qm_dqput(udqp);
1844         }
1845         if (XFS_IS_OQUOTA_ON(mp)) {
1846                 ASSERT(gdqp);
1847                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1848                 xfs_qm_dqput(gdqp);
1849         }
1850         /*
1851          * Now release the inode. This will send it to 'inactive', and
1852          * possibly even free blocks.
1853          */
1854         VN_RELE(XFS_ITOV(ip));
1855
1856         /*
1857          * Goto next inode.
1858          */
1859         *res = BULKSTAT_RV_DIDONE;
1860         return 0;
1861 }
1862
1863 /*
1864  * Walk thru all the filesystem inodes and construct a consistent view
1865  * of the disk quota world. If the quotacheck fails, disable quotas.
1866  */
1867 int
1868 xfs_qm_quotacheck(
1869         xfs_mount_t     *mp)
1870 {
1871         int             done, count, error;
1872         xfs_ino_t       lastino;
1873         size_t          structsz;
1874         xfs_inode_t     *uip, *gip;
1875         uint            flags;
1876
1877         count = INT_MAX;
1878         structsz = 1;
1879         lastino = 0;
1880         flags = 0;
1881
1882         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1883         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1884
1885         /*
1886          * There should be no cached dquots. The (simplistic) quotacheck
1887          * algorithm doesn't like that.
1888          */
1889         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1890
1891         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1892
1893         /*
1894          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1895          * their counters to zero. We need a clean slate.
1896          * We don't log our changes till later.
1897          */
1898         if ((uip = XFS_QI_UQIP(mp))) {
1899                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1900                         goto error_return;
1901                 flags |= XFS_UQUOTA_CHKD;
1902         }
1903
1904         if ((gip = XFS_QI_GQIP(mp))) {
1905                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1906                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1907                         goto error_return;
1908                 flags |= XFS_OQUOTA_CHKD;
1909         }
1910
1911         do {
1912                 /*
1913                  * Iterate thru all the inodes in the file system,
1914                  * adjusting the corresponding dquot counters in core.
1915                  */
1916                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1917                                      xfs_qm_dqusage_adjust, NULL,
1918                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1919                         break;
1920
1921         } while (! done);
1922
1923         /*
1924          * We can get this error if we couldn't do a dquot allocation inside
1925          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1926          * dirty dquots that might be cached, we just want to get rid of them
1927          * and turn quotaoff. The dquots won't be attached to any of the inodes
1928          * at this point (because we intentionally didn't in dqget_noattach).
1929          */
1930         if (error) {
1931                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1932                 goto error_return;
1933         }
1934         /*
1935          * We've made all the changes that we need to make incore.
1936          * Now flush_them down to disk buffers.
1937          */
1938         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1939
1940         /*
1941          * We didn't log anything, because if we crashed, we'll have to
1942          * start the quotacheck from scratch anyway. However, we must make
1943          * sure that our dquot changes are secure before we put the
1944          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1945          * flush.
1946          */
1947         XFS_bflush(mp->m_ddev_targp);
1948
1949         /*
1950          * If one type of quotas is off, then it will lose its
1951          * quotachecked status, since we won't be doing accounting for
1952          * that type anymore.
1953          */
1954         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1955         mp->m_qflags |= flags;
1956
1957         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1958
1959  error_return:
1960         if (error) {
1961                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1962                         "Disabling quotas.",
1963                         mp->m_fsname, error);
1964                 /*
1965                  * We must turn off quotas.
1966                  */
1967                 ASSERT(mp->m_quotainfo != NULL);
1968                 ASSERT(xfs_Gqm != NULL);
1969                 xfs_qm_destroy_quotainfo(mp);
1970                 (void)xfs_mount_reset_sbqflags(mp);
1971         } else {
1972                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1973         }
1974         return (error);
1975 }
1976
1977 /*
1978  * This is called after the superblock has been read in and we're ready to
1979  * iget the quota inodes.
1980  */
1981 STATIC int
1982 xfs_qm_init_quotainos(
1983         xfs_mount_t     *mp)
1984 {
1985         xfs_inode_t     *uip, *gip;
1986         int             error;
1987         __int64_t       sbflags;
1988         uint            flags;
1989
1990         ASSERT(mp->m_quotainfo);
1991         uip = gip = NULL;
1992         sbflags = 0;
1993         flags = 0;
1994
1995         /*
1996          * Get the uquota and gquota inodes
1997          */
1998         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1999                 if (XFS_IS_UQUOTA_ON(mp) &&
2000                     mp->m_sb.sb_uquotino != NULLFSINO) {
2001                         ASSERT(mp->m_sb.sb_uquotino > 0);
2002                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
2003                                              0, 0, &uip, 0)))
2004                                 return XFS_ERROR(error);
2005                 }
2006                 if (XFS_IS_OQUOTA_ON(mp) &&
2007                     mp->m_sb.sb_gquotino != NULLFSINO) {
2008                         ASSERT(mp->m_sb.sb_gquotino > 0);
2009                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2010                                              0, 0, &gip, 0))) {
2011                                 if (uip)
2012                                         VN_RELE(XFS_ITOV(uip));
2013                                 return XFS_ERROR(error);
2014                         }
2015                 }
2016         } else {
2017                 flags |= XFS_QMOPT_SBVERSION;
2018                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2019                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2020         }
2021
2022         /*
2023          * Create the two inodes, if they don't exist already. The changes
2024          * made above will get added to a transaction and logged in one of
2025          * the qino_alloc calls below.  If the device is readonly,
2026          * temporarily switch to read-write to do this.
2027          */
2028         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2029                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2030                                               sbflags | XFS_SB_UQUOTINO,
2031                                               flags | XFS_QMOPT_UQUOTA)))
2032                         return XFS_ERROR(error);
2033
2034                 flags &= ~XFS_QMOPT_SBVERSION;
2035         }
2036         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
2037                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
2038                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2039                 error = xfs_qm_qino_alloc(mp, &gip,
2040                                           sbflags | XFS_SB_GQUOTINO, flags);
2041                 if (error) {
2042                         if (uip)
2043                                 VN_RELE(XFS_ITOV(uip));
2044
2045                         return XFS_ERROR(error);
2046                 }
2047         }
2048
2049         XFS_QI_UQIP(mp) = uip;
2050         XFS_QI_GQIP(mp) = gip;
2051
2052         return 0;
2053 }
2054
2055
2056 /*
2057  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2058  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2059  * favor the lookup function ...
2060  * XXXsup merge this with qm_reclaim_one().
2061  */
2062 STATIC int
2063 xfs_qm_shake_freelist(
2064         int howmany)
2065 {
2066         int             nreclaimed;
2067         xfs_dqhash_t    *hash;
2068         xfs_dquot_t     *dqp, *nextdqp;
2069         int             restarts;
2070         int             nflushes;
2071
2072         if (howmany <= 0)
2073                 return 0;
2074
2075         nreclaimed = 0;
2076         restarts = 0;
2077         nflushes = 0;
2078
2079 #ifdef QUOTADEBUG
2080         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2081 #endif
2082         /* lock order is : hashchainlock, freelistlock, mplistlock */
2083  tryagain:
2084         xfs_qm_freelist_lock(xfs_Gqm);
2085
2086         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2087              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2088               nreclaimed < howmany); ) {
2089                 xfs_dqlock(dqp);
2090
2091                 /*
2092                  * We are racing with dqlookup here. Naturally we don't
2093                  * want to reclaim a dquot that lookup wants.
2094                  */
2095                 if (dqp->dq_flags & XFS_DQ_WANT) {
2096                         xfs_dqunlock(dqp);
2097                         xfs_qm_freelist_unlock(xfs_Gqm);
2098                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2099                                 return nreclaimed;
2100                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2101                         goto tryagain;
2102                 }
2103
2104                 /*
2105                  * If the dquot is inactive, we are assured that it is
2106                  * not on the mplist or the hashlist, and that makes our
2107                  * life easier.
2108                  */
2109                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2110                         ASSERT(dqp->q_mount == NULL);
2111                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2112                         ASSERT(dqp->HL_PREVP == NULL);
2113                         ASSERT(dqp->MPL_PREVP == NULL);
2114                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2115                         nextdqp = dqp->dq_flnext;
2116                         goto off_freelist;
2117                 }
2118
2119                 ASSERT(dqp->MPL_PREVP);
2120                 /*
2121                  * Try to grab the flush lock. If this dquot is in the process of
2122                  * getting flushed to disk, we don't want to reclaim it.
2123                  */
2124                 if (! xfs_qm_dqflock_nowait(dqp)) {
2125                         xfs_dqunlock(dqp);
2126                         dqp = dqp->dq_flnext;
2127                         continue;
2128                 }
2129
2130                 /*
2131                  * We have the flush lock so we know that this is not in the
2132                  * process of being flushed. So, if this is dirty, flush it
2133                  * DELWRI so that we don't get a freelist infested with
2134                  * dirty dquots.
2135                  */
2136                 if (XFS_DQ_IS_DIRTY(dqp)) {
2137                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2138                         /*
2139                          * We flush it delayed write, so don't bother
2140                          * releasing the mplock.
2141                          */
2142                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2143                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2144                         dqp = dqp->dq_flnext;
2145                         continue;
2146                 }
2147                 /*
2148                  * We're trying to get the hashlock out of order. This races
2149                  * with dqlookup; so, we giveup and goto the next dquot if
2150                  * we couldn't get the hashlock. This way, we won't starve
2151                  * a dqlookup process that holds the hashlock that is
2152                  * waiting for the freelist lock.
2153                  */
2154                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2155                         xfs_dqfunlock(dqp);
2156                         xfs_dqunlock(dqp);
2157                         dqp = dqp->dq_flnext;
2158                         continue;
2159                 }
2160                 /*
2161                  * This races with dquot allocation code as well as dqflush_all
2162                  * and reclaim code. So, if we failed to grab the mplist lock,
2163                  * giveup everything and start over.
2164                  */
2165                 hash = dqp->q_hash;
2166                 ASSERT(hash);
2167                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2168                         /* XXX put a sentinel so that we can come back here */
2169                         xfs_dqfunlock(dqp);
2170                         xfs_dqunlock(dqp);
2171                         XFS_DQ_HASH_UNLOCK(hash);
2172                         xfs_qm_freelist_unlock(xfs_Gqm);
2173                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2174                                 return nreclaimed;
2175                         goto tryagain;
2176                 }
2177                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2178 #ifdef QUOTADEBUG
2179                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2180                         dqp, be32_to_cpu(dqp->q_core.d_id));
2181 #endif
2182                 ASSERT(dqp->q_nrefs == 0);
2183                 nextdqp = dqp->dq_flnext;
2184                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2185                 XQM_HASHLIST_REMOVE(hash, dqp);
2186                 xfs_dqfunlock(dqp);
2187                 xfs_qm_mplist_unlock(dqp->q_mount);
2188                 XFS_DQ_HASH_UNLOCK(hash);
2189
2190  off_freelist:
2191                 XQM_FREELIST_REMOVE(dqp);
2192                 xfs_dqunlock(dqp);
2193                 nreclaimed++;
2194                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2195                 xfs_qm_dqdestroy(dqp);
2196                 dqp = nextdqp;
2197         }
2198         xfs_qm_freelist_unlock(xfs_Gqm);
2199         return nreclaimed;
2200 }
2201
2202
2203 /*
2204  * The kmem_shake interface is invoked when memory is running low.
2205  */
2206 /* ARGSUSED */
2207 STATIC int
2208 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2209 {
2210         int     ndqused, nfree, n;
2211
2212         if (!kmem_shake_allow(gfp_mask))
2213                 return 0;
2214         if (!xfs_Gqm)
2215                 return 0;
2216
2217         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2218         /* incore dquots in all f/s's */
2219         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2220
2221         ASSERT(ndqused >= 0);
2222
2223         if (nfree <= ndqused && nfree < ndquot)
2224                 return 0;
2225
2226         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2227         n = nfree - ndqused - ndquot;           /* # over target */
2228
2229         return xfs_qm_shake_freelist(MAX(nfree, n));
2230 }
2231
2232
2233 /*
2234  * Just pop the least recently used dquot off the freelist and
2235  * recycle it. The returned dquot is locked.
2236  */
2237 STATIC xfs_dquot_t *
2238 xfs_qm_dqreclaim_one(void)
2239 {
2240         xfs_dquot_t     *dqpout;
2241         xfs_dquot_t     *dqp;
2242         int             restarts;
2243         int             nflushes;
2244
2245         restarts = 0;
2246         dqpout = NULL;
2247         nflushes = 0;
2248
2249         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2250  startagain:
2251         xfs_qm_freelist_lock(xfs_Gqm);
2252
2253         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2254                 xfs_dqlock(dqp);
2255
2256                 /*
2257                  * We are racing with dqlookup here. Naturally we don't
2258                  * want to reclaim a dquot that lookup wants. We release the
2259                  * freelist lock and start over, so that lookup will grab
2260                  * both the dquot and the freelistlock.
2261                  */
2262                 if (dqp->dq_flags & XFS_DQ_WANT) {
2263                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2264                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2265                         xfs_dqunlock(dqp);
2266                         xfs_qm_freelist_unlock(xfs_Gqm);
2267                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2268                                 return NULL;
2269                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2270                         goto startagain;
2271                 }
2272
2273                 /*
2274                  * If the dquot is inactive, we are assured that it is
2275                  * not on the mplist or the hashlist, and that makes our
2276                  * life easier.
2277                  */
2278                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2279                         ASSERT(dqp->q_mount == NULL);
2280                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2281                         ASSERT(dqp->HL_PREVP == NULL);
2282                         ASSERT(dqp->MPL_PREVP == NULL);
2283                         XQM_FREELIST_REMOVE(dqp);
2284                         xfs_dqunlock(dqp);
2285                         dqpout = dqp;
2286                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2287                         break;
2288                 }
2289
2290                 ASSERT(dqp->q_hash);
2291                 ASSERT(dqp->MPL_PREVP);
2292
2293                 /*
2294                  * Try to grab the flush lock. If this dquot is in the process of
2295                  * getting flushed to disk, we don't want to reclaim it.
2296                  */
2297                 if (! xfs_qm_dqflock_nowait(dqp)) {
2298                         xfs_dqunlock(dqp);
2299                         continue;
2300                 }
2301
2302                 /*
2303                  * We have the flush lock so we know that this is not in the
2304                  * process of being flushed. So, if this is dirty, flush it
2305                  * DELWRI so that we don't get a freelist infested with
2306                  * dirty dquots.
2307                  */
2308                 if (XFS_DQ_IS_DIRTY(dqp)) {
2309                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2310                         /*
2311                          * We flush it delayed write, so don't bother
2312                          * releasing the freelist lock.
2313                          */
2314                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2315                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2316                         continue;
2317                 }
2318
2319                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2320                         xfs_dqfunlock(dqp);
2321                         xfs_dqunlock(dqp);
2322                         continue;
2323                 }
2324
2325                 if (! xfs_qm_dqhashlock_nowait(dqp))
2326                         goto mplistunlock;
2327
2328                 ASSERT(dqp->q_nrefs == 0);
2329                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2330                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2331                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2332                 XQM_FREELIST_REMOVE(dqp);
2333                 dqpout = dqp;
2334                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2335  mplistunlock:
2336                 xfs_qm_mplist_unlock(dqp->q_mount);
2337                 xfs_dqfunlock(dqp);
2338                 xfs_dqunlock(dqp);
2339                 if (dqpout)
2340                         break;
2341         }
2342
2343         xfs_qm_freelist_unlock(xfs_Gqm);
2344         return dqpout;
2345 }
2346
2347
2348 /*------------------------------------------------------------------*/
2349
2350 /*
2351  * Return a new incore dquot. Depending on the number of
2352  * dquots in the system, we either allocate a new one on the kernel heap,
2353  * or reclaim a free one.
2354  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2355  * to reclaim an existing one from the freelist.
2356  */
2357 boolean_t
2358 xfs_qm_dqalloc_incore(
2359         xfs_dquot_t **O_dqpp)
2360 {
2361         xfs_dquot_t     *dqp;
2362
2363         /*
2364          * Check against high water mark to see if we want to pop
2365          * a nincompoop dquot off the freelist.
2366          */
2367         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2368                 /*
2369                  * Try to recycle a dquot from the freelist.
2370                  */
2371                 if ((dqp = xfs_qm_dqreclaim_one())) {
2372                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2373                         /*
2374                          * Just zero the core here. The rest will get
2375                          * reinitialized by caller. XXX we shouldn't even
2376                          * do this zero ...
2377                          */
2378                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2379                         *O_dqpp = dqp;
2380                         return B_FALSE;
2381                 }
2382                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2383         }
2384
2385         /*
2386          * Allocate a brand new dquot on the kernel heap and return it
2387          * to the caller to initialize.
2388          */
2389         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2390         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2391         atomic_inc(&xfs_Gqm->qm_totaldquots);
2392
2393         return B_TRUE;
2394 }
2395
2396
2397 /*
2398  * Start a transaction and write the incore superblock changes to
2399  * disk. flags parameter indicates which fields have changed.
2400  */
2401 int
2402 xfs_qm_write_sb_changes(
2403         xfs_mount_t     *mp,
2404         __int64_t       flags)
2405 {
2406         xfs_trans_t     *tp;
2407         int             error;
2408
2409 #ifdef QUOTADEBUG
2410         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2411 #endif
2412         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2413         if ((error = xfs_trans_reserve(tp, 0,
2414                                       mp->m_sb.sb_sectsize + 128, 0,
2415                                       0,
2416                                       XFS_DEFAULT_LOG_COUNT))) {
2417                 xfs_trans_cancel(tp, 0);
2418                 return error;
2419         }
2420
2421         xfs_mod_sb(tp, flags);
2422         (void) xfs_trans_commit(tp, 0);
2423
2424         return 0;
2425 }
2426
2427
2428 /* --------------- utility functions for vnodeops ---------------- */
2429
2430
2431 /*
2432  * Given an inode, a uid and gid (from cred_t) make sure that we have
2433  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2434  * quotas by creating this file.
2435  * This also attaches dquot(s) to the given inode after locking it,
2436  * and returns the dquots corresponding to the uid and/or gid.
2437  *
2438  * in   : inode (unlocked)
2439  * out  : udquot, gdquot with references taken and unlocked
2440  */
2441 int
2442 xfs_qm_vop_dqalloc(
2443         xfs_mount_t     *mp,
2444         xfs_inode_t     *ip,
2445         uid_t           uid,
2446         gid_t           gid,
2447         prid_t          prid,
2448         uint            flags,
2449         xfs_dquot_t     **O_udqpp,
2450         xfs_dquot_t     **O_gdqpp)
2451 {
2452         int             error;
2453         xfs_dquot_t     *uq, *gq;
2454         uint            lockflags;
2455
2456         if (!XFS_IS_QUOTA_ON(mp))
2457                 return 0;
2458
2459         lockflags = XFS_ILOCK_EXCL;
2460         xfs_ilock(ip, lockflags);
2461
2462         if ((flags & XFS_QMOPT_INHERIT) &&
2463             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2464                 gid = ip->i_d.di_gid;
2465
2466         /*
2467          * Attach the dquot(s) to this inode, doing a dquot allocation
2468          * if necessary. The dquot(s) will not be locked.
2469          */
2470         if (XFS_NOT_DQATTACHED(mp, ip)) {
2471                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2472                                             XFS_QMOPT_ILOCKED))) {
2473                         xfs_iunlock(ip, lockflags);
2474                         return error;
2475                 }
2476         }
2477
2478         uq = gq = NULL;
2479         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2480                 if (ip->i_d.di_uid != uid) {
2481                         /*
2482                          * What we need is the dquot that has this uid, and
2483                          * if we send the inode to dqget, the uid of the inode
2484                          * takes priority over what's sent in the uid argument.
2485                          * We must unlock inode here before calling dqget if
2486                          * we're not sending the inode, because otherwise
2487                          * we'll deadlock by doing trans_reserve while
2488                          * holding ilock.
2489                          */
2490                         xfs_iunlock(ip, lockflags);
2491                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2492                                                  XFS_DQ_USER,
2493                                                  XFS_QMOPT_DQALLOC |
2494                                                  XFS_QMOPT_DOWARN,
2495                                                  &uq))) {
2496                                 ASSERT(error != ENOENT);
2497                                 return error;
2498                         }
2499                         /*
2500                          * Get the ilock in the right order.
2501                          */
2502                         xfs_dqunlock(uq);
2503                         lockflags = XFS_ILOCK_SHARED;
2504                         xfs_ilock(ip, lockflags);
2505                 } else {
2506                         /*
2507                          * Take an extra reference, because we'll return
2508                          * this to caller
2509                          */
2510                         ASSERT(ip->i_udquot);
2511                         uq = ip->i_udquot;
2512                         xfs_dqlock(uq);
2513                         XFS_DQHOLD(uq);
2514                         xfs_dqunlock(uq);
2515                 }
2516         }
2517         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2518                 if (ip->i_d.di_gid != gid) {
2519                         xfs_iunlock(ip, lockflags);
2520                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2521                                                  XFS_DQ_GROUP,
2522                                                  XFS_QMOPT_DQALLOC |
2523                                                  XFS_QMOPT_DOWARN,
2524                                                  &gq))) {
2525                                 if (uq)
2526                                         xfs_qm_dqrele(uq);
2527                                 ASSERT(error != ENOENT);
2528                                 return error;
2529                         }
2530                         xfs_dqunlock(gq);
2531                         lockflags = XFS_ILOCK_SHARED;
2532                         xfs_ilock(ip, lockflags);
2533                 } else {
2534                         ASSERT(ip->i_gdquot);
2535                         gq = ip->i_gdquot;
2536                         xfs_dqlock(gq);
2537                         XFS_DQHOLD(gq);
2538                         xfs_dqunlock(gq);
2539                 }
2540         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2541                 if (ip->i_d.di_projid != prid) {
2542                         xfs_iunlock(ip, lockflags);
2543                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2544                                                  XFS_DQ_PROJ,
2545                                                  XFS_QMOPT_DQALLOC |
2546                                                  XFS_QMOPT_DOWARN,
2547                                                  &gq))) {
2548                                 if (uq)
2549                                         xfs_qm_dqrele(uq);
2550                                 ASSERT(error != ENOENT);
2551                                 return (error);
2552                         }
2553                         xfs_dqunlock(gq);
2554                         lockflags = XFS_ILOCK_SHARED;
2555                         xfs_ilock(ip, lockflags);
2556                 } else {
2557                         ASSERT(ip->i_gdquot);
2558                         gq = ip->i_gdquot;
2559                         xfs_dqlock(gq);
2560                         XFS_DQHOLD(gq);
2561                         xfs_dqunlock(gq);
2562                 }
2563         }
2564         if (uq)
2565                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2566
2567         xfs_iunlock(ip, lockflags);
2568         if (O_udqpp)
2569                 *O_udqpp = uq;
2570         else if (uq)
2571                 xfs_qm_dqrele(uq);
2572         if (O_gdqpp)
2573                 *O_gdqpp = gq;
2574         else if (gq)
2575                 xfs_qm_dqrele(gq);
2576         return 0;
2577 }
2578
2579 /*
2580  * Actually transfer ownership, and do dquot modifications.
2581  * These were already reserved.
2582  */
2583 xfs_dquot_t *
2584 xfs_qm_vop_chown(
2585         xfs_trans_t     *tp,
2586         xfs_inode_t     *ip,
2587         xfs_dquot_t     **IO_olddq,
2588         xfs_dquot_t     *newdq)
2589 {
2590         xfs_dquot_t     *prevdq;
2591         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2592                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2593
2594         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2595         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2596
2597         /* old dquot */
2598         prevdq = *IO_olddq;
2599         ASSERT(prevdq);
2600         ASSERT(prevdq != newdq);
2601
2602         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2603         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2604
2605         /* the sparkling new dquot */
2606         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2607         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2608
2609         /*
2610          * Take an extra reference, because the inode
2611          * is going to keep this dquot pointer even
2612          * after the trans_commit.
2613          */
2614         xfs_dqlock(newdq);
2615         XFS_DQHOLD(newdq);
2616         xfs_dqunlock(newdq);
2617         *IO_olddq = newdq;
2618
2619         return prevdq;
2620 }
2621
2622 /*
2623  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2624  */
2625 int
2626 xfs_qm_vop_chown_reserve(
2627         xfs_trans_t     *tp,
2628         xfs_inode_t     *ip,
2629         xfs_dquot_t     *udqp,
2630         xfs_dquot_t     *gdqp,
2631         uint            flags)
2632 {
2633         int             error;
2634         xfs_mount_t     *mp;
2635         uint            delblks, blkflags, prjflags = 0;
2636         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2637
2638         ASSERT(XFS_ISLOCKED_INODE(ip));
2639         mp = ip->i_mount;
2640         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2641
2642         delblks = ip->i_delayed_blks;
2643         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2644         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2645                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2646
2647         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2648             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2649                 delblksudq = udqp;
2650                 /*
2651                  * If there are delayed allocation blocks, then we have to
2652                  * unreserve those from the old dquot, and add them to the
2653                  * new dquot.
2654                  */
2655                 if (delblks) {
2656                         ASSERT(ip->i_udquot);
2657                         unresudq = ip->i_udquot;
2658                 }
2659         }
2660         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2661                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2662                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2663                         prjflags = XFS_QMOPT_ENOSPC;
2664
2665                 if (prjflags ||
2666                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2667                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2668                         delblksgdq = gdqp;
2669                         if (delblks) {
2670                                 ASSERT(ip->i_gdquot);
2671                                 unresgdq = ip->i_gdquot;
2672                         }
2673                 }
2674         }
2675
2676         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2677                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2678                                 flags | blkflags | prjflags)))
2679                 return (error);
2680
2681         /*
2682          * Do the delayed blks reservations/unreservations now. Since, these
2683          * are done without the help of a transaction, if a reservation fails
2684          * its previous reservations won't be automatically undone by trans
2685          * code. So, we have to do it manually here.
2686          */
2687         if (delblks) {
2688                 /*
2689                  * Do the reservations first. Unreservation can't fail.
2690                  */
2691                 ASSERT(delblksudq || delblksgdq);
2692                 ASSERT(unresudq || unresgdq);
2693                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2694                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2695                                 flags | blkflags | prjflags)))
2696                         return (error);
2697                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2698                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2699                                 blkflags);
2700         }
2701
2702         return (0);
2703 }
2704
2705 int
2706 xfs_qm_vop_rename_dqattach(
2707         xfs_inode_t     **i_tab)
2708 {
2709         xfs_inode_t     *ip;
2710         int             i;
2711         int             error;
2712
2713         ip = i_tab[0];
2714
2715         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2716                 return 0;
2717
2718         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2719                 error = xfs_qm_dqattach(ip, 0);
2720                 if (error)
2721                         return error;
2722         }
2723         for (i = 1; (i < 4 && i_tab[i]); i++) {
2724                 /*
2725                  * Watch out for duplicate entries in the table.
2726                  */
2727                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2728                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2729                                 error = xfs_qm_dqattach(ip, 0);
2730                                 if (error)
2731                                         return error;
2732                         }
2733                 }
2734         }
2735         return 0;
2736 }
2737
2738 void
2739 xfs_qm_vop_dqattach_and_dqmod_newinode(
2740         xfs_trans_t     *tp,
2741         xfs_inode_t     *ip,
2742         xfs_dquot_t     *udqp,
2743         xfs_dquot_t     *gdqp)
2744 {
2745         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2746                 return;
2747
2748         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2749         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2750
2751         if (udqp) {
2752                 xfs_dqlock(udqp);
2753                 XFS_DQHOLD(udqp);
2754                 xfs_dqunlock(udqp);
2755                 ASSERT(ip->i_udquot == NULL);
2756                 ip->i_udquot = udqp;
2757                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2758                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2759                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2760         }
2761         if (gdqp) {
2762                 xfs_dqlock(gdqp);
2763                 XFS_DQHOLD(gdqp);
2764                 xfs_dqunlock(gdqp);
2765                 ASSERT(ip->i_gdquot == NULL);
2766                 ip->i_gdquot = gdqp;
2767                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2768                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2769                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2770                                 be32_to_cpu(gdqp->q_core.d_id));
2771                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2772         }
2773 }
2774
2775 /* ------------- list stuff -----------------*/
2776 STATIC void
2777 xfs_qm_freelist_init(xfs_frlist_t *ql)
2778 {
2779         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2780         mutex_init(&ql->qh_lock);
2781         ql->qh_version = 0;
2782         ql->qh_nelems = 0;
2783 }
2784
2785 STATIC void
2786 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2787 {
2788         xfs_dquot_t     *dqp, *nextdqp;
2789
2790         mutex_lock(&ql->qh_lock);
2791         for (dqp = ql->qh_next;
2792              dqp != (xfs_dquot_t *)ql; ) {
2793                 xfs_dqlock(dqp);
2794                 nextdqp = dqp->dq_flnext;
2795 #ifdef QUOTADEBUG
2796                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2797 #endif
2798                 XQM_FREELIST_REMOVE(dqp);
2799                 xfs_dqunlock(dqp);
2800                 xfs_qm_dqdestroy(dqp);
2801                 dqp = nextdqp;
2802         }
2803         mutex_unlock(&ql->qh_lock);
2804         mutex_destroy(&ql->qh_lock);
2805
2806         ASSERT(ql->qh_nelems == 0);
2807 }
2808
2809 STATIC void
2810 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2811 {
2812         dq->dq_flnext = ql->qh_next;
2813         dq->dq_flprev = (xfs_dquot_t *)ql;
2814         ql->qh_next = dq;
2815         dq->dq_flnext->dq_flprev = dq;
2816         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2817         xfs_Gqm->qm_dqfreelist.qh_version++;
2818 }
2819
2820 void
2821 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2822 {
2823         xfs_dquot_t *next = dq->dq_flnext;
2824         xfs_dquot_t *prev = dq->dq_flprev;
2825
2826         next->dq_flprev = prev;
2827         prev->dq_flnext = next;
2828         dq->dq_flnext = dq->dq_flprev = dq;
2829         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2830         xfs_Gqm->qm_dqfreelist.qh_version++;
2831 }
2832
2833 void
2834 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2835 {
2836         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2837 }
2838
2839 STATIC int
2840 xfs_qm_dqhashlock_nowait(
2841         xfs_dquot_t *dqp)
2842 {
2843         int locked;
2844
2845         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2846         return locked;
2847 }
2848
2849 int
2850 xfs_qm_freelist_lock_nowait(
2851         xfs_qm_t *xqm)
2852 {
2853         int locked;
2854
2855         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2856         return locked;
2857 }
2858
2859 STATIC int
2860 xfs_qm_mplist_nowait(
2861         xfs_mount_t     *mp)
2862 {
2863         int locked;
2864
2865         ASSERT(mp->m_quotainfo);
2866         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2867         return locked;
2868 }