fix XEN_SAVE_RESTORE Kconfig dependencies
[pandora-kernel.git] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_trans.h"
24 #include "xfs_sb.h"
25 #include "xfs_ag.h"
26 #include "xfs_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_mount.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_ialloc_btree.h"
31 #include "xfs_dinode.h"
32 #include "xfs_inode.h"
33 #include "xfs_ialloc.h"
34 #include "xfs_itable.h"
35 #include "xfs_rtalloc.h"
36 #include "xfs_error.h"
37 #include "xfs_bmap.h"
38 #include "xfs_attr.h"
39 #include "xfs_buf_item.h"
40 #include "xfs_trans_space.h"
41 #include "xfs_utils.h"
42 #include "xfs_qm.h"
43 #include "xfs_trace.h"
44
45 /*
46  * The global quota manager. There is only one of these for the entire
47  * system, _not_ one per file system. XQM keeps track of the overall
48  * quota functionality, including maintaining the freelist and hash
49  * tables of dquots.
50  */
51 struct mutex    xfs_Gqm_lock;
52 struct xfs_qm   *xfs_Gqm;
53 uint            ndquot;
54
55 kmem_zone_t     *qm_dqzone;
56 kmem_zone_t     *qm_dqtrxzone;
57
58 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
59 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
60
61 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
62 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
63 STATIC int      xfs_qm_shake(struct shrinker *, int, gfp_t);
64
65 static struct shrinker xfs_qm_shaker = {
66         .shrink = xfs_qm_shake,
67         .seeks = DEFAULT_SEEKS,
68 };
69
70 #ifdef DEBUG
71 extern struct mutex     qcheck_lock;
72 #endif
73
74 #ifdef QUOTADEBUG
75 static void
76 xfs_qm_dquot_list_print(
77         struct xfs_mount *mp)
78 {
79         xfs_dquot_t     *dqp;
80         int             i = 0;
81
82         list_for_each_entry(dqp, &mp->m_quotainfo->qi_dqlist_lock, qi_mplist) {
83                 xfs_debug(mp, "   %d. \"%d (%s)\"   "
84                                   "bcnt = %lld, icnt = %lld, refs = %d",
85                         i++, be32_to_cpu(dqp->q_core.d_id),
86                         DQFLAGTO_TYPESTR(dqp),
87                         (long long)be64_to_cpu(dqp->q_core.d_bcount),
88                         (long long)be64_to_cpu(dqp->q_core.d_icount),
89                         dqp->q_nrefs);
90         }
91 }
92 #else
93 static void xfs_qm_dquot_list_print(struct xfs_mount *mp) { }
94 #endif
95
96 /*
97  * Initialize the XQM structure.
98  * Note that there is not one quota manager per file system.
99  */
100 STATIC struct xfs_qm *
101 xfs_Gqm_init(void)
102 {
103         xfs_dqhash_t    *udqhash, *gdqhash;
104         xfs_qm_t        *xqm;
105         size_t          hsize;
106         uint            i;
107
108         /*
109          * Initialize the dquot hash tables.
110          */
111         udqhash = kmem_zalloc_greedy(&hsize,
112                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
113                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t));
114         if (!udqhash)
115                 goto out;
116
117         gdqhash = kmem_zalloc_large(hsize);
118         if (!gdqhash)
119                 goto out_free_udqhash;
120
121         hsize /= sizeof(xfs_dqhash_t);
122         ndquot = hsize << 8;
123
124         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
125         xqm->qm_dqhashmask = hsize - 1;
126         xqm->qm_usr_dqhtable = udqhash;
127         xqm->qm_grp_dqhtable = gdqhash;
128         ASSERT(xqm->qm_usr_dqhtable != NULL);
129         ASSERT(xqm->qm_grp_dqhtable != NULL);
130
131         for (i = 0; i < hsize; i++) {
132                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
133                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
134         }
135
136         /*
137          * Freelist of all dquots of all file systems
138          */
139         INIT_LIST_HEAD(&xqm->qm_dqfrlist);
140         xqm->qm_dqfrlist_cnt = 0;
141         mutex_init(&xqm->qm_dqfrlist_lock);
142
143         /*
144          * dquot zone. we register our own low-memory callback.
145          */
146         if (!qm_dqzone) {
147                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
148                                                 "xfs_dquots");
149                 qm_dqzone = xqm->qm_dqzone;
150         } else
151                 xqm->qm_dqzone = qm_dqzone;
152
153         register_shrinker(&xfs_qm_shaker);
154
155         /*
156          * The t_dqinfo portion of transactions.
157          */
158         if (!qm_dqtrxzone) {
159                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
160                                                    "xfs_dqtrx");
161                 qm_dqtrxzone = xqm->qm_dqtrxzone;
162         } else
163                 xqm->qm_dqtrxzone = qm_dqtrxzone;
164
165         atomic_set(&xqm->qm_totaldquots, 0);
166         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
167         xqm->qm_nrefs = 0;
168 #ifdef DEBUG
169         mutex_init(&qcheck_lock);
170 #endif
171         return xqm;
172
173  out_free_udqhash:
174         kmem_free_large(udqhash);
175  out:
176         return NULL;
177 }
178
179 /*
180  * Destroy the global quota manager when its reference count goes to zero.
181  */
182 STATIC void
183 xfs_qm_destroy(
184         struct xfs_qm   *xqm)
185 {
186         struct xfs_dquot *dqp, *n;
187         int             hsize, i;
188
189         ASSERT(xqm != NULL);
190         ASSERT(xqm->qm_nrefs == 0);
191         unregister_shrinker(&xfs_qm_shaker);
192         hsize = xqm->qm_dqhashmask + 1;
193         for (i = 0; i < hsize; i++) {
194                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
195                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
196         }
197         kmem_free_large(xqm->qm_usr_dqhtable);
198         kmem_free_large(xqm->qm_grp_dqhtable);
199         xqm->qm_usr_dqhtable = NULL;
200         xqm->qm_grp_dqhtable = NULL;
201         xqm->qm_dqhashmask = 0;
202
203         /* frlist cleanup */
204         mutex_lock(&xqm->qm_dqfrlist_lock);
205         list_for_each_entry_safe(dqp, n, &xqm->qm_dqfrlist, q_freelist) {
206                 xfs_dqlock(dqp);
207 #ifdef QUOTADEBUG
208                 xfs_debug(dqp->q_mount, "FREELIST destroy 0x%p", dqp);
209 #endif
210                 list_del_init(&dqp->q_freelist);
211                 xfs_Gqm->qm_dqfrlist_cnt--;
212                 xfs_dqunlock(dqp);
213                 xfs_qm_dqdestroy(dqp);
214         }
215         mutex_unlock(&xqm->qm_dqfrlist_lock);
216         mutex_destroy(&xqm->qm_dqfrlist_lock);
217 #ifdef DEBUG
218         mutex_destroy(&qcheck_lock);
219 #endif
220         kmem_free(xqm);
221 }
222
223 /*
224  * Called at mount time to let XQM know that another file system is
225  * starting quotas. This isn't crucial information as the individual mount
226  * structures are pretty independent, but it helps the XQM keep a
227  * global view of what's going on.
228  */
229 /* ARGSUSED */
230 STATIC int
231 xfs_qm_hold_quotafs_ref(
232         struct xfs_mount *mp)
233 {
234         /*
235          * Need to lock the xfs_Gqm structure for things like this. For example,
236          * the structure could disappear between the entry to this routine and
237          * a HOLD operation if not locked.
238          */
239         mutex_lock(&xfs_Gqm_lock);
240
241         if (!xfs_Gqm) {
242                 xfs_Gqm = xfs_Gqm_init();
243                 if (!xfs_Gqm) {
244                         mutex_unlock(&xfs_Gqm_lock);
245                         return ENOMEM;
246                 }
247         }
248
249         /*
250          * We can keep a list of all filesystems with quotas mounted for
251          * debugging and statistical purposes, but ...
252          * Just take a reference and get out.
253          */
254         xfs_Gqm->qm_nrefs++;
255         mutex_unlock(&xfs_Gqm_lock);
256
257         return 0;
258 }
259
260
261 /*
262  * Release the reference that a filesystem took at mount time,
263  * so that we know when we need to destroy the entire quota manager.
264  */
265 /* ARGSUSED */
266 STATIC void
267 xfs_qm_rele_quotafs_ref(
268         struct xfs_mount *mp)
269 {
270         xfs_dquot_t     *dqp, *n;
271
272         ASSERT(xfs_Gqm);
273         ASSERT(xfs_Gqm->qm_nrefs > 0);
274
275         /*
276          * Go thru the freelist and destroy all inactive dquots.
277          */
278         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
279
280         list_for_each_entry_safe(dqp, n, &xfs_Gqm->qm_dqfrlist, q_freelist) {
281                 xfs_dqlock(dqp);
282                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
283                         ASSERT(dqp->q_mount == NULL);
284                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
285                         ASSERT(list_empty(&dqp->q_hashlist));
286                         ASSERT(list_empty(&dqp->q_mplist));
287                         list_del_init(&dqp->q_freelist);
288                         xfs_Gqm->qm_dqfrlist_cnt--;
289                         xfs_dqunlock(dqp);
290                         xfs_qm_dqdestroy(dqp);
291                 } else {
292                         xfs_dqunlock(dqp);
293                 }
294         }
295         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
296
297         /*
298          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
299          * be restarted.
300          */
301         mutex_lock(&xfs_Gqm_lock);
302         if (--xfs_Gqm->qm_nrefs == 0) {
303                 xfs_qm_destroy(xfs_Gqm);
304                 xfs_Gqm = NULL;
305         }
306         mutex_unlock(&xfs_Gqm_lock);
307 }
308
309 /*
310  * Just destroy the quotainfo structure.
311  */
312 void
313 xfs_qm_unmount(
314         struct xfs_mount        *mp)
315 {
316         if (mp->m_quotainfo) {
317                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL);
318                 xfs_qm_destroy_quotainfo(mp);
319         }
320 }
321
322
323 /*
324  * This is called from xfs_mountfs to start quotas and initialize all
325  * necessary data structures like quotainfo.  This is also responsible for
326  * running a quotacheck as necessary.  We are guaranteed that the superblock
327  * is consistently read in at this point.
328  *
329  * If we fail here, the mount will continue with quota turned off. We don't
330  * need to inidicate success or failure at all.
331  */
332 void
333 xfs_qm_mount_quotas(
334         xfs_mount_t     *mp)
335 {
336         int             error = 0;
337         uint            sbf;
338
339         /*
340          * If quotas on realtime volumes is not supported, we disable
341          * quotas immediately.
342          */
343         if (mp->m_sb.sb_rextents) {
344                 xfs_notice(mp, "Cannot turn on quotas for realtime filesystem");
345                 mp->m_qflags = 0;
346                 goto write_changes;
347         }
348
349         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
350
351         /*
352          * Allocate the quotainfo structure inside the mount struct, and
353          * create quotainode(s), and change/rev superblock if necessary.
354          */
355         error = xfs_qm_init_quotainfo(mp);
356         if (error) {
357                 /*
358                  * We must turn off quotas.
359                  */
360                 ASSERT(mp->m_quotainfo == NULL);
361                 mp->m_qflags = 0;
362                 goto write_changes;
363         }
364         /*
365          * If any of the quotas are not consistent, do a quotacheck.
366          */
367         if (XFS_QM_NEED_QUOTACHECK(mp)) {
368                 error = xfs_qm_quotacheck(mp);
369                 if (error) {
370                         /* Quotacheck failed and disabled quotas. */
371                         return;
372                 }
373         }
374         /* 
375          * If one type of quotas is off, then it will lose its
376          * quotachecked status, since we won't be doing accounting for
377          * that type anymore.
378          */
379         if (!XFS_IS_UQUOTA_ON(mp))
380                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
381         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp)))
382                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
383
384  write_changes:
385         /*
386          * We actually don't have to acquire the m_sb_lock at all.
387          * This can only be called from mount, and that's single threaded. XXX
388          */
389         spin_lock(&mp->m_sb_lock);
390         sbf = mp->m_sb.sb_qflags;
391         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
392         spin_unlock(&mp->m_sb_lock);
393
394         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
395                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
396                         /*
397                          * We could only have been turning quotas off.
398                          * We aren't in very good shape actually because
399                          * the incore structures are convinced that quotas are
400                          * off, but the on disk superblock doesn't know that !
401                          */
402                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
403                         xfs_alert(mp, "%s: Superblock update failed!",
404                                 __func__);
405                 }
406         }
407
408         if (error) {
409                 xfs_warn(mp, "Failed to initialize disk quotas.");
410                 return;
411         }
412
413 #ifdef QUOTADEBUG
414         if (XFS_IS_QUOTA_ON(mp))
415                 xfs_qm_internalqcheck(mp);
416 #endif
417 }
418
419 /*
420  * Called from the vfsops layer.
421  */
422 void
423 xfs_qm_unmount_quotas(
424         xfs_mount_t     *mp)
425 {
426         /*
427          * Release the dquots that root inode, et al might be holding,
428          * before we flush quotas and blow away the quotainfo structure.
429          */
430         ASSERT(mp->m_rootip);
431         xfs_qm_dqdetach(mp->m_rootip);
432         if (mp->m_rbmip)
433                 xfs_qm_dqdetach(mp->m_rbmip);
434         if (mp->m_rsumip)
435                 xfs_qm_dqdetach(mp->m_rsumip);
436
437         /*
438          * Release the quota inodes.
439          */
440         if (mp->m_quotainfo) {
441                 if (mp->m_quotainfo->qi_uquotaip) {
442                         IRELE(mp->m_quotainfo->qi_uquotaip);
443                         mp->m_quotainfo->qi_uquotaip = NULL;
444                 }
445                 if (mp->m_quotainfo->qi_gquotaip) {
446                         IRELE(mp->m_quotainfo->qi_gquotaip);
447                         mp->m_quotainfo->qi_gquotaip = NULL;
448                 }
449         }
450 }
451
452 /*
453  * Flush all dquots of the given file system to disk. The dquots are
454  * _not_ purged from memory here, just their data written to disk.
455  */
456 STATIC int
457 xfs_qm_dqflush_all(
458         struct xfs_mount        *mp,
459         int                     sync_mode)
460 {
461         struct xfs_quotainfo    *q = mp->m_quotainfo;
462         int                     recl;
463         struct xfs_dquot        *dqp;
464         int                     niters;
465         int                     error;
466
467         if (!q)
468                 return 0;
469         niters = 0;
470 again:
471         mutex_lock(&q->qi_dqlist_lock);
472         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
473                 xfs_dqlock(dqp);
474                 if (! XFS_DQ_IS_DIRTY(dqp)) {
475                         xfs_dqunlock(dqp);
476                         continue;
477                 }
478
479                 /* XXX a sentinel would be better */
480                 recl = q->qi_dqreclaims;
481                 if (!xfs_dqflock_nowait(dqp)) {
482                         /*
483                          * If we can't grab the flush lock then check
484                          * to see if the dquot has been flushed delayed
485                          * write.  If so, grab its buffer and send it
486                          * out immediately.  We'll be able to acquire
487                          * the flush lock when the I/O completes.
488                          */
489                         xfs_qm_dqflock_pushbuf_wait(dqp);
490                 }
491                 /*
492                  * Let go of the mplist lock. We don't want to hold it
493                  * across a disk write.
494                  */
495                 mutex_unlock(&q->qi_dqlist_lock);
496                 error = xfs_qm_dqflush(dqp, sync_mode);
497                 xfs_dqunlock(dqp);
498                 if (error)
499                         return error;
500
501                 mutex_lock(&q->qi_dqlist_lock);
502                 if (recl != q->qi_dqreclaims) {
503                         mutex_unlock(&q->qi_dqlist_lock);
504                         /* XXX restart limit */
505                         goto again;
506                 }
507         }
508
509         mutex_unlock(&q->qi_dqlist_lock);
510         /* return ! busy */
511         return 0;
512 }
513 /*
514  * Release the group dquot pointers the user dquots may be
515  * carrying around as a hint. mplist is locked on entry and exit.
516  */
517 STATIC void
518 xfs_qm_detach_gdquots(
519         struct xfs_mount        *mp)
520 {
521         struct xfs_quotainfo    *q = mp->m_quotainfo;
522         struct xfs_dquot        *dqp, *gdqp;
523         int                     nrecl;
524
525  again:
526         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
527         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
528                 xfs_dqlock(dqp);
529                 if ((gdqp = dqp->q_gdquot)) {
530                         xfs_dqlock(gdqp);
531                         dqp->q_gdquot = NULL;
532                 }
533                 xfs_dqunlock(dqp);
534
535                 if (gdqp) {
536                         /*
537                          * Can't hold the mplist lock across a dqput.
538                          * XXXmust convert to marker based iterations here.
539                          */
540                         nrecl = q->qi_dqreclaims;
541                         mutex_unlock(&q->qi_dqlist_lock);
542                         xfs_qm_dqput(gdqp);
543
544                         mutex_lock(&q->qi_dqlist_lock);
545                         if (nrecl != q->qi_dqreclaims)
546                                 goto again;
547                 }
548         }
549 }
550
551 /*
552  * Go through all the incore dquots of this file system and take them
553  * off the mplist and hashlist, if the dquot type matches the dqtype
554  * parameter. This is used when turning off quota accounting for
555  * users and/or groups, as well as when the filesystem is unmounting.
556  */
557 STATIC int
558 xfs_qm_dqpurge_int(
559         struct xfs_mount        *mp,
560         uint                    flags)
561 {
562         struct xfs_quotainfo    *q = mp->m_quotainfo;
563         struct xfs_dquot        *dqp, *n;
564         uint                    dqtype;
565         int                     nrecl;
566         int                     nmisses;
567
568         if (!q)
569                 return 0;
570
571         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
572         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
573         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
574
575         mutex_lock(&q->qi_dqlist_lock);
576
577         /*
578          * In the first pass through all incore dquots of this filesystem,
579          * we release the group dquot pointers the user dquots may be
580          * carrying around as a hint. We need to do this irrespective of
581          * what's being turned off.
582          */
583         xfs_qm_detach_gdquots(mp);
584
585       again:
586         nmisses = 0;
587         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
588         /*
589          * Try to get rid of all of the unwanted dquots. The idea is to
590          * get them off mplist and hashlist, but leave them on freelist.
591          */
592         list_for_each_entry_safe(dqp, n, &q->qi_dqlist, q_mplist) {
593                 /*
594                  * It's OK to look at the type without taking dqlock here.
595                  * We're holding the mplist lock here, and that's needed for
596                  * a dqreclaim.
597                  */
598                 if ((dqp->dq_flags & dqtype) == 0)
599                         continue;
600
601                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
602                         nrecl = q->qi_dqreclaims;
603                         mutex_unlock(&q->qi_dqlist_lock);
604                         mutex_lock(&dqp->q_hash->qh_lock);
605                         mutex_lock(&q->qi_dqlist_lock);
606
607                         /*
608                          * XXXTheoretically, we can get into a very long
609                          * ping pong game here.
610                          * No one can be adding dquots to the mplist at
611                          * this point, but somebody might be taking things off.
612                          */
613                         if (nrecl != q->qi_dqreclaims) {
614                                 mutex_unlock(&dqp->q_hash->qh_lock);
615                                 goto again;
616                         }
617                 }
618
619                 /*
620                  * Take the dquot off the mplist and hashlist. It may remain on
621                  * freelist in INACTIVE state.
622                  */
623                 nmisses += xfs_qm_dqpurge(dqp);
624         }
625         mutex_unlock(&q->qi_dqlist_lock);
626         return nmisses;
627 }
628
629 int
630 xfs_qm_dqpurge_all(
631         xfs_mount_t     *mp,
632         uint            flags)
633 {
634         int             ndquots;
635
636         /*
637          * Purge the dquot cache.
638          * None of the dquots should really be busy at this point.
639          */
640         if (mp->m_quotainfo) {
641                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
642                         delay(ndquots * 10);
643                 }
644         }
645         return 0;
646 }
647
648 STATIC int
649 xfs_qm_dqattach_one(
650         xfs_inode_t     *ip,
651         xfs_dqid_t      id,
652         uint            type,
653         uint            doalloc,
654         xfs_dquot_t     *udqhint, /* hint */
655         xfs_dquot_t     **IO_idqpp)
656 {
657         xfs_dquot_t     *dqp;
658         int             error;
659
660         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
661         error = 0;
662
663         /*
664          * See if we already have it in the inode itself. IO_idqpp is
665          * &i_udquot or &i_gdquot. This made the code look weird, but
666          * made the logic a lot simpler.
667          */
668         dqp = *IO_idqpp;
669         if (dqp) {
670                 trace_xfs_dqattach_found(dqp);
671                 return 0;
672         }
673
674         /*
675          * udqhint is the i_udquot field in inode, and is non-NULL only
676          * when the type arg is group/project. Its purpose is to save a
677          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
678          * the user dquot.
679          */
680         if (udqhint) {
681                 ASSERT(type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
682                 xfs_dqlock(udqhint);
683
684                 /*
685                  * No need to take dqlock to look at the id.
686                  *
687                  * The ID can't change until it gets reclaimed, and it won't
688                  * be reclaimed as long as we have a ref from inode and we
689                  * hold the ilock.
690                  */
691                 dqp = udqhint->q_gdquot;
692                 if (dqp && be32_to_cpu(dqp->q_core.d_id) == id) {
693                         xfs_dqlock(dqp);
694                         XFS_DQHOLD(dqp);
695                         ASSERT(*IO_idqpp == NULL);
696                         *IO_idqpp = dqp;
697
698                         xfs_dqunlock(dqp);
699                         xfs_dqunlock(udqhint);
700                         return 0;
701                 }
702
703                 /*
704                  * We can't hold a dquot lock when we call the dqget code.
705                  * We'll deadlock in no time, because of (not conforming to)
706                  * lock ordering - the inodelock comes before any dquot lock,
707                  * and we may drop and reacquire the ilock in xfs_qm_dqget().
708                  */
709                 xfs_dqunlock(udqhint);
710         }
711
712         /*
713          * Find the dquot from somewhere. This bumps the
714          * reference count of dquot and returns it locked.
715          * This can return ENOENT if dquot didn't exist on
716          * disk and we didn't ask it to allocate;
717          * ESRCH if quotas got turned off suddenly.
718          */
719         error = xfs_qm_dqget(ip->i_mount, ip, id, type, XFS_QMOPT_DOWARN, &dqp);
720         if (error)
721                 return error;
722
723         trace_xfs_dqattach_get(dqp);
724
725         /*
726          * dqget may have dropped and re-acquired the ilock, but it guarantees
727          * that the dquot returned is the one that should go in the inode.
728          */
729         *IO_idqpp = dqp;
730         xfs_dqunlock(dqp);
731         return 0;
732 }
733
734
735 /*
736  * Given a udquot and gdquot, attach a ptr to the group dquot in the
737  * udquot as a hint for future lookups. The idea sounds simple, but the
738  * execution isn't, because the udquot might have a group dquot attached
739  * already and getting rid of that gets us into lock ordering constraints.
740  * The process is complicated more by the fact that the dquots may or may not
741  * be locked on entry.
742  */
743 STATIC void
744 xfs_qm_dqattach_grouphint(
745         xfs_dquot_t     *udq,
746         xfs_dquot_t     *gdq)
747 {
748         xfs_dquot_t     *tmp;
749
750         xfs_dqlock(udq);
751
752         if ((tmp = udq->q_gdquot)) {
753                 if (tmp == gdq) {
754                         xfs_dqunlock(udq);
755                         return;
756                 }
757
758                 udq->q_gdquot = NULL;
759                 /*
760                  * We can't keep any dqlocks when calling dqrele,
761                  * because the freelist lock comes before dqlocks.
762                  */
763                 xfs_dqunlock(udq);
764                 /*
765                  * we took a hard reference once upon a time in dqget,
766                  * so give it back when the udquot no longer points at it
767                  * dqput() does the unlocking of the dquot.
768                  */
769                 xfs_qm_dqrele(tmp);
770
771                 xfs_dqlock(udq);
772                 xfs_dqlock(gdq);
773
774         } else {
775                 ASSERT(XFS_DQ_IS_LOCKED(udq));
776                 xfs_dqlock(gdq);
777         }
778
779         ASSERT(XFS_DQ_IS_LOCKED(udq));
780         ASSERT(XFS_DQ_IS_LOCKED(gdq));
781         /*
782          * Somebody could have attached a gdquot here,
783          * when we dropped the uqlock. If so, just do nothing.
784          */
785         if (udq->q_gdquot == NULL) {
786                 XFS_DQHOLD(gdq);
787                 udq->q_gdquot = gdq;
788         }
789
790         xfs_dqunlock(gdq);
791         xfs_dqunlock(udq);
792 }
793
794
795 /*
796  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
797  * into account.
798  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
799  * Inode may get unlocked and relocked in here, and the caller must deal with
800  * the consequences.
801  */
802 int
803 xfs_qm_dqattach_locked(
804         xfs_inode_t     *ip,
805         uint            flags)
806 {
807         xfs_mount_t     *mp = ip->i_mount;
808         uint            nquotas = 0;
809         int             error = 0;
810
811         if (!XFS_IS_QUOTA_RUNNING(mp) ||
812             !XFS_IS_QUOTA_ON(mp) ||
813             !XFS_NOT_DQATTACHED(mp, ip) ||
814             ip->i_ino == mp->m_sb.sb_uquotino ||
815             ip->i_ino == mp->m_sb.sb_gquotino)
816                 return 0;
817
818         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
819
820         if (XFS_IS_UQUOTA_ON(mp)) {
821                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
822                                                 flags & XFS_QMOPT_DQALLOC,
823                                                 NULL, &ip->i_udquot);
824                 if (error)
825                         goto done;
826                 nquotas++;
827         }
828
829         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
830         if (XFS_IS_OQUOTA_ON(mp)) {
831                 error = XFS_IS_GQUOTA_ON(mp) ?
832                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
833                                                 flags & XFS_QMOPT_DQALLOC,
834                                                 ip->i_udquot, &ip->i_gdquot) :
835                         xfs_qm_dqattach_one(ip, xfs_get_projid(ip), XFS_DQ_PROJ,
836                                                 flags & XFS_QMOPT_DQALLOC,
837                                                 ip->i_udquot, &ip->i_gdquot);
838                 /*
839                  * Don't worry about the udquot that we may have
840                  * attached above. It'll get detached, if not already.
841                  */
842                 if (error)
843                         goto done;
844                 nquotas++;
845         }
846
847         /*
848          * Attach this group quota to the user quota as a hint.
849          * This WON'T, in general, result in a thrash.
850          */
851         if (nquotas == 2) {
852                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
853                 ASSERT(ip->i_udquot);
854                 ASSERT(ip->i_gdquot);
855
856                 /*
857                  * We may or may not have the i_udquot locked at this point,
858                  * but this check is OK since we don't depend on the i_gdquot to
859                  * be accurate 100% all the time. It is just a hint, and this
860                  * will succeed in general.
861                  */
862                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
863                         goto done;
864                 /*
865                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
866                  */
867                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot);
868         }
869
870  done:
871 #ifdef QUOTADEBUG
872         if (! error) {
873                 if (XFS_IS_UQUOTA_ON(mp))
874                         ASSERT(ip->i_udquot);
875                 if (XFS_IS_OQUOTA_ON(mp))
876                         ASSERT(ip->i_gdquot);
877         }
878         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
879 #endif
880         return error;
881 }
882
883 int
884 xfs_qm_dqattach(
885         struct xfs_inode        *ip,
886         uint                    flags)
887 {
888         int                     error;
889
890         xfs_ilock(ip, XFS_ILOCK_EXCL);
891         error = xfs_qm_dqattach_locked(ip, flags);
892         xfs_iunlock(ip, XFS_ILOCK_EXCL);
893
894         return error;
895 }
896
897 /*
898  * Release dquots (and their references) if any.
899  * The inode should be locked EXCL except when this's called by
900  * xfs_ireclaim.
901  */
902 void
903 xfs_qm_dqdetach(
904         xfs_inode_t     *ip)
905 {
906         if (!(ip->i_udquot || ip->i_gdquot))
907                 return;
908
909         trace_xfs_dquot_dqdetach(ip);
910
911         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
912         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
913         if (ip->i_udquot) {
914                 xfs_qm_dqrele(ip->i_udquot);
915                 ip->i_udquot = NULL;
916         }
917         if (ip->i_gdquot) {
918                 xfs_qm_dqrele(ip->i_gdquot);
919                 ip->i_gdquot = NULL;
920         }
921 }
922
923 int
924 xfs_qm_sync(
925         struct xfs_mount        *mp,
926         int                     flags)
927 {
928         struct xfs_quotainfo    *q = mp->m_quotainfo;
929         int                     recl, restarts;
930         struct xfs_dquot        *dqp;
931         int                     error;
932
933         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
934                 return 0;
935
936         restarts = 0;
937
938   again:
939         mutex_lock(&q->qi_dqlist_lock);
940         /*
941          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
942          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
943          * when we have the mplist lock, we know that dquots will be consistent
944          * as long as we have it locked.
945          */
946         if (!XFS_IS_QUOTA_ON(mp)) {
947                 mutex_unlock(&q->qi_dqlist_lock);
948                 return 0;
949         }
950         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
951         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
952                 /*
953                  * If this is vfs_sync calling, then skip the dquots that
954                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
955                  * This is very similar to what xfs_sync does with inodes.
956                  */
957                 if (flags & SYNC_TRYLOCK) {
958                         if (!XFS_DQ_IS_DIRTY(dqp))
959                                 continue;
960                         if (!xfs_qm_dqlock_nowait(dqp))
961                                 continue;
962                 } else {
963                         xfs_dqlock(dqp);
964                 }
965
966                 /*
967                  * Now, find out for sure if this dquot is dirty or not.
968                  */
969                 if (! XFS_DQ_IS_DIRTY(dqp)) {
970                         xfs_dqunlock(dqp);
971                         continue;
972                 }
973
974                 /* XXX a sentinel would be better */
975                 recl = q->qi_dqreclaims;
976                 if (!xfs_dqflock_nowait(dqp)) {
977                         if (flags & SYNC_TRYLOCK) {
978                                 xfs_dqunlock(dqp);
979                                 continue;
980                         }
981                         /*
982                          * If we can't grab the flush lock then if the caller
983                          * really wanted us to give this our best shot, so
984                          * see if we can give a push to the buffer before we wait
985                          * on the flush lock. At this point, we know that
986                          * even though the dquot is being flushed,
987                          * it has (new) dirty data.
988                          */
989                         xfs_qm_dqflock_pushbuf_wait(dqp);
990                 }
991                 /*
992                  * Let go of the mplist lock. We don't want to hold it
993                  * across a disk write
994                  */
995                 mutex_unlock(&q->qi_dqlist_lock);
996                 error = xfs_qm_dqflush(dqp, flags);
997                 xfs_dqunlock(dqp);
998                 if (error && XFS_FORCED_SHUTDOWN(mp))
999                         return 0;       /* Need to prevent umount failure */
1000                 else if (error)
1001                         return error;
1002
1003                 mutex_lock(&q->qi_dqlist_lock);
1004                 if (recl != q->qi_dqreclaims) {
1005                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1006                                 break;
1007
1008                         mutex_unlock(&q->qi_dqlist_lock);
1009                         goto again;
1010                 }
1011         }
1012
1013         mutex_unlock(&q->qi_dqlist_lock);
1014         return 0;
1015 }
1016
1017 /*
1018  * The hash chains and the mplist use the same xfs_dqhash structure as
1019  * their list head, but we can take the mplist qh_lock and one of the
1020  * hash qh_locks at the same time without any problem as they aren't
1021  * related.
1022  */
1023 static struct lock_class_key xfs_quota_mplist_class;
1024
1025 /*
1026  * This initializes all the quota information that's kept in the
1027  * mount structure
1028  */
1029 STATIC int
1030 xfs_qm_init_quotainfo(
1031         xfs_mount_t     *mp)
1032 {
1033         xfs_quotainfo_t *qinf;
1034         int             error;
1035         xfs_dquot_t     *dqp;
1036
1037         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1038
1039         /*
1040          * Tell XQM that we exist as soon as possible.
1041          */
1042         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1043                 return error;
1044         }
1045
1046         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1047
1048         /*
1049          * See if quotainodes are setup, and if not, allocate them,
1050          * and change the superblock accordingly.
1051          */
1052         if ((error = xfs_qm_init_quotainos(mp))) {
1053                 kmem_free(qinf);
1054                 mp->m_quotainfo = NULL;
1055                 return error;
1056         }
1057
1058         INIT_LIST_HEAD(&qinf->qi_dqlist);
1059         mutex_init(&qinf->qi_dqlist_lock);
1060         lockdep_set_class(&qinf->qi_dqlist_lock, &xfs_quota_mplist_class);
1061
1062         qinf->qi_dqreclaims = 0;
1063
1064         /* mutex used to serialize quotaoffs */
1065         mutex_init(&qinf->qi_quotaofflock);
1066
1067         /* Precalc some constants */
1068         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1069         ASSERT(qinf->qi_dqchunklen);
1070         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1071         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1072
1073         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1074
1075         /*
1076          * We try to get the limits from the superuser's limits fields.
1077          * This is quite hacky, but it is standard quota practice.
1078          * We look at the USR dquot with id == 0 first, but if user quotas
1079          * are not enabled we goto the GRP dquot with id == 0.
1080          * We don't really care to keep separate default limits for user
1081          * and group quotas, at least not at this point.
1082          */
1083         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1084                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1085                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1086                                 XFS_DQ_PROJ),
1087                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1088                              &dqp);
1089         if (! error) {
1090                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1091
1092                 /*
1093                  * The warnings and timers set the grace period given to
1094                  * a user or group before he or she can not perform any
1095                  * more writing. If it is zero, a default is used.
1096                  */
1097                 qinf->qi_btimelimit = ddqp->d_btimer ?
1098                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1099                 qinf->qi_itimelimit = ddqp->d_itimer ?
1100                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1101                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1102                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1103                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1104                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1105                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1106                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1107                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1108                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1109                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1110                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1111                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1112                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1113                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1114                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1115  
1116                 /*
1117                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1118                  * we don't want this dquot cached. We haven't done a
1119                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1120                  */
1121                 xfs_qm_dqdestroy(dqp);
1122         } else {
1123                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1124                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1125                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1126                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1127                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1128                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1129         }
1130
1131         return 0;
1132 }
1133
1134
1135 /*
1136  * Gets called when unmounting a filesystem or when all quotas get
1137  * turned off.
1138  * This purges the quota inodes, destroys locks and frees itself.
1139  */
1140 void
1141 xfs_qm_destroy_quotainfo(
1142         xfs_mount_t     *mp)
1143 {
1144         xfs_quotainfo_t *qi;
1145
1146         qi = mp->m_quotainfo;
1147         ASSERT(qi != NULL);
1148         ASSERT(xfs_Gqm != NULL);
1149
1150         /*
1151          * Release the reference that XQM kept, so that we know
1152          * when the XQM structure should be freed. We cannot assume
1153          * that xfs_Gqm is non-null after this point.
1154          */
1155         xfs_qm_rele_quotafs_ref(mp);
1156
1157         ASSERT(list_empty(&qi->qi_dqlist));
1158         mutex_destroy(&qi->qi_dqlist_lock);
1159
1160         if (qi->qi_uquotaip) {
1161                 IRELE(qi->qi_uquotaip);
1162                 qi->qi_uquotaip = NULL; /* paranoia */
1163         }
1164         if (qi->qi_gquotaip) {
1165                 IRELE(qi->qi_gquotaip);
1166                 qi->qi_gquotaip = NULL;
1167         }
1168         mutex_destroy(&qi->qi_quotaofflock);
1169         kmem_free(qi);
1170         mp->m_quotainfo = NULL;
1171 }
1172
1173
1174
1175 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1176
1177 /* ARGSUSED */
1178 STATIC void
1179 xfs_qm_list_init(
1180         xfs_dqlist_t    *list,
1181         char            *str,
1182         int             n)
1183 {
1184         mutex_init(&list->qh_lock);
1185         INIT_LIST_HEAD(&list->qh_list);
1186         list->qh_version = 0;
1187         list->qh_nelems = 0;
1188 }
1189
1190 STATIC void
1191 xfs_qm_list_destroy(
1192         xfs_dqlist_t    *list)
1193 {
1194         mutex_destroy(&(list->qh_lock));
1195 }
1196
1197 /*
1198  * Create an inode and return with a reference already taken, but unlocked
1199  * This is how we create quota inodes
1200  */
1201 STATIC int
1202 xfs_qm_qino_alloc(
1203         xfs_mount_t     *mp,
1204         xfs_inode_t     **ip,
1205         __int64_t       sbfields,
1206         uint            flags)
1207 {
1208         xfs_trans_t     *tp;
1209         int             error;
1210         int             committed;
1211
1212         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1213         if ((error = xfs_trans_reserve(tp,
1214                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1215                                       XFS_CREATE_LOG_RES(mp), 0,
1216                                       XFS_TRANS_PERM_LOG_RES,
1217                                       XFS_CREATE_LOG_COUNT))) {
1218                 xfs_trans_cancel(tp, 0);
1219                 return error;
1220         }
1221
1222         error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0, 0, 1, ip, &committed);
1223         if (error) {
1224                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1225                                  XFS_TRANS_ABORT);
1226                 return error;
1227         }
1228
1229         /*
1230          * Make the changes in the superblock, and log those too.
1231          * sbfields arg may contain fields other than *QUOTINO;
1232          * VERSIONNUM for example.
1233          */
1234         spin_lock(&mp->m_sb_lock);
1235         if (flags & XFS_QMOPT_SBVERSION) {
1236                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1237                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1238                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1239                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1240                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1241
1242                 xfs_sb_version_addquota(&mp->m_sb);
1243                 mp->m_sb.sb_uquotino = NULLFSINO;
1244                 mp->m_sb.sb_gquotino = NULLFSINO;
1245
1246                 /* qflags will get updated _after_ quotacheck */
1247                 mp->m_sb.sb_qflags = 0;
1248         }
1249         if (flags & XFS_QMOPT_UQUOTA)
1250                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1251         else
1252                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1253         spin_unlock(&mp->m_sb_lock);
1254         xfs_mod_sb(tp, sbfields);
1255
1256         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1257                 xfs_alert(mp, "%s failed (error %d)!", __func__, error);
1258                 return error;
1259         }
1260         return 0;
1261 }
1262
1263
1264 STATIC void
1265 xfs_qm_reset_dqcounts(
1266         xfs_mount_t     *mp,
1267         xfs_buf_t       *bp,
1268         xfs_dqid_t      id,
1269         uint            type)
1270 {
1271         xfs_disk_dquot_t        *ddq;
1272         int                     j;
1273
1274         trace_xfs_reset_dqcounts(bp, _RET_IP_);
1275
1276         /*
1277          * Reset all counters and timers. They'll be
1278          * started afresh by xfs_qm_quotacheck.
1279          */
1280 #ifdef DEBUG
1281         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1282         do_div(j, sizeof(xfs_dqblk_t));
1283         ASSERT(mp->m_quotainfo->qi_dqperchunk == j);
1284 #endif
1285         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1286         for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) {
1287                 /*
1288                  * Do a sanity check, and if needed, repair the dqblk. Don't
1289                  * output any warnings because it's perfectly possible to
1290                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1291                  */
1292                 (void) xfs_qm_dqcheck(mp, ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1293                                       "xfs_quotacheck");
1294                 ddq->d_bcount = 0;
1295                 ddq->d_icount = 0;
1296                 ddq->d_rtbcount = 0;
1297                 ddq->d_btimer = 0;
1298                 ddq->d_itimer = 0;
1299                 ddq->d_rtbtimer = 0;
1300                 ddq->d_bwarns = 0;
1301                 ddq->d_iwarns = 0;
1302                 ddq->d_rtbwarns = 0;
1303                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1304         }
1305 }
1306
1307 STATIC int
1308 xfs_qm_dqiter_bufs(
1309         xfs_mount_t     *mp,
1310         xfs_dqid_t      firstid,
1311         xfs_fsblock_t   bno,
1312         xfs_filblks_t   blkcnt,
1313         uint            flags)
1314 {
1315         xfs_buf_t       *bp;
1316         int             error;
1317         int             notcommitted;
1318         int             incr;
1319         int             type;
1320
1321         ASSERT(blkcnt > 0);
1322         notcommitted = 0;
1323         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1324                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1325         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1326                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1327         error = 0;
1328
1329         /*
1330          * Blkcnt arg can be a very big number, and might even be
1331          * larger than the log itself. So, we have to break it up into
1332          * manageable-sized transactions.
1333          * Note that we don't start a permanent transaction here; we might
1334          * not be able to get a log reservation for the whole thing up front,
1335          * and we don't really care to either, because we just discard
1336          * everything if we were to crash in the middle of this loop.
1337          */
1338         while (blkcnt--) {
1339                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1340                               XFS_FSB_TO_DADDR(mp, bno),
1341                               mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1342                 if (error)
1343                         break;
1344
1345                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1346                 xfs_bdwrite(mp, bp);
1347                 /*
1348                  * goto the next block.
1349                  */
1350                 bno++;
1351                 firstid += mp->m_quotainfo->qi_dqperchunk;
1352         }
1353         return error;
1354 }
1355
1356 /*
1357  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1358  * caller supplied function for every chunk of dquots that we find.
1359  */
1360 STATIC int
1361 xfs_qm_dqiterate(
1362         xfs_mount_t     *mp,
1363         xfs_inode_t     *qip,
1364         uint            flags)
1365 {
1366         xfs_bmbt_irec_t         *map;
1367         int                     i, nmaps;       /* number of map entries */
1368         int                     error;          /* return value */
1369         xfs_fileoff_t           lblkno;
1370         xfs_filblks_t           maxlblkcnt;
1371         xfs_dqid_t              firstid;
1372         xfs_fsblock_t           rablkno;
1373         xfs_filblks_t           rablkcnt;
1374
1375         error = 0;
1376         /*
1377          * This looks racy, but we can't keep an inode lock across a
1378          * trans_reserve. But, this gets called during quotacheck, and that
1379          * happens only at mount time which is single threaded.
1380          */
1381         if (qip->i_d.di_nblocks == 0)
1382                 return 0;
1383
1384         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1385
1386         lblkno = 0;
1387         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1388         do {
1389                 nmaps = XFS_DQITER_MAP_SIZE;
1390                 /*
1391                  * We aren't changing the inode itself. Just changing
1392                  * some of its data. No new blocks are added here, and
1393                  * the inode is never added to the transaction.
1394                  */
1395                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1396                 error = xfs_bmapi(NULL, qip, lblkno,
1397                                   maxlblkcnt - lblkno,
1398                                   XFS_BMAPI_METADATA,
1399                                   NULL,
1400                                   0, map, &nmaps, NULL);
1401                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1402                 if (error)
1403                         break;
1404
1405                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1406                 for (i = 0; i < nmaps; i++) {
1407                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1408                         ASSERT(map[i].br_blockcount);
1409
1410
1411                         lblkno += map[i].br_blockcount;
1412
1413                         if (map[i].br_startblock == HOLESTARTBLOCK)
1414                                 continue;
1415
1416                         firstid = (xfs_dqid_t) map[i].br_startoff *
1417                                 mp->m_quotainfo->qi_dqperchunk;
1418                         /*
1419                          * Do a read-ahead on the next extent.
1420                          */
1421                         if ((i+1 < nmaps) &&
1422                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1423                                 rablkcnt =  map[i+1].br_blockcount;
1424                                 rablkno = map[i+1].br_startblock;
1425                                 while (rablkcnt--) {
1426                                         xfs_buf_readahead(mp->m_ddev_targp,
1427                                                XFS_FSB_TO_DADDR(mp, rablkno),
1428                                                mp->m_quotainfo->qi_dqchunklen);
1429                                         rablkno++;
1430                                 }
1431                         }
1432                         /*
1433                          * Iterate thru all the blks in the extent and
1434                          * reset the counters of all the dquots inside them.
1435                          */
1436                         if ((error = xfs_qm_dqiter_bufs(mp,
1437                                                        firstid,
1438                                                        map[i].br_startblock,
1439                                                        map[i].br_blockcount,
1440                                                        flags))) {
1441                                 break;
1442                         }
1443                 }
1444
1445                 if (error)
1446                         break;
1447         } while (nmaps > 0);
1448
1449         kmem_free(map);
1450
1451         return error;
1452 }
1453
1454 /*
1455  * Called by dqusage_adjust in doing a quotacheck.
1456  *
1457  * Given the inode, and a dquot id this updates both the incore dqout as well
1458  * as the buffer copy. This is so that once the quotacheck is done, we can
1459  * just log all the buffers, as opposed to logging numerous updates to
1460  * individual dquots.
1461  */
1462 STATIC int
1463 xfs_qm_quotacheck_dqadjust(
1464         struct xfs_inode        *ip,
1465         xfs_dqid_t              id,
1466         uint                    type,
1467         xfs_qcnt_t              nblks,
1468         xfs_qcnt_t              rtblks)
1469 {
1470         struct xfs_mount        *mp = ip->i_mount;
1471         struct xfs_dquot        *dqp;
1472         int                     error;
1473
1474         error = xfs_qm_dqget(mp, ip, id, type,
1475                              XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN, &dqp);
1476         if (error) {
1477                 /*
1478                  * Shouldn't be able to turn off quotas here.
1479                  */
1480                 ASSERT(error != ESRCH);
1481                 ASSERT(error != ENOENT);
1482                 return error;
1483         }
1484
1485         trace_xfs_dqadjust(dqp);
1486
1487         /*
1488          * Adjust the inode count and the block count to reflect this inode's
1489          * resource usage.
1490          */
1491         be64_add_cpu(&dqp->q_core.d_icount, 1);
1492         dqp->q_res_icount++;
1493         if (nblks) {
1494                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1495                 dqp->q_res_bcount += nblks;
1496         }
1497         if (rtblks) {
1498                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1499                 dqp->q_res_rtbcount += rtblks;
1500         }
1501
1502         /*
1503          * Set default limits, adjust timers (since we changed usages)
1504          *
1505          * There are no timers for the default values set in the root dquot.
1506          */
1507         if (dqp->q_core.d_id) {
1508                 xfs_qm_adjust_dqlimits(mp, &dqp->q_core);
1509                 xfs_qm_adjust_dqtimers(mp, &dqp->q_core);
1510         }
1511
1512         dqp->dq_flags |= XFS_DQ_DIRTY;
1513         xfs_qm_dqput(dqp);
1514         return 0;
1515 }
1516
1517 STATIC int
1518 xfs_qm_get_rtblks(
1519         xfs_inode_t     *ip,
1520         xfs_qcnt_t      *O_rtblks)
1521 {
1522         xfs_filblks_t   rtblks;                 /* total rt blks */
1523         xfs_extnum_t    idx;                    /* extent record index */
1524         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1525         xfs_extnum_t    nextents;               /* number of extent entries */
1526         int             error;
1527
1528         ASSERT(XFS_IS_REALTIME_INODE(ip));
1529         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1530         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1531                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1532                         return error;
1533         }
1534         rtblks = 0;
1535         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1536         for (idx = 0; idx < nextents; idx++)
1537                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1538         *O_rtblks = (xfs_qcnt_t)rtblks;
1539         return 0;
1540 }
1541
1542 /*
1543  * callback routine supplied to bulkstat(). Given an inumber, find its
1544  * dquots and update them to account for resources taken by that inode.
1545  */
1546 /* ARGSUSED */
1547 STATIC int
1548 xfs_qm_dqusage_adjust(
1549         xfs_mount_t     *mp,            /* mount point for filesystem */
1550         xfs_ino_t       ino,            /* inode number to get data for */
1551         void            __user *buffer, /* not used */
1552         int             ubsize,         /* not used */
1553         int             *ubused,        /* not used */
1554         int             *res)           /* result code value */
1555 {
1556         xfs_inode_t     *ip;
1557         xfs_qcnt_t      nblks, rtblks = 0;
1558         int             error;
1559
1560         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1561
1562         /*
1563          * rootino must have its resources accounted for, not so with the quota
1564          * inodes.
1565          */
1566         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1567                 *res = BULKSTAT_RV_NOTHING;
1568                 return XFS_ERROR(EINVAL);
1569         }
1570
1571         /*
1572          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1573          * interface expects the inode to be exclusively locked because that's
1574          * the case in all other instances. It's OK that we do this because
1575          * quotacheck is done only at mount time.
1576          */
1577         error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip);
1578         if (error) {
1579                 *res = BULKSTAT_RV_NOTHING;
1580                 return error;
1581         }
1582
1583         ASSERT(ip->i_delayed_blks == 0);
1584
1585         if (XFS_IS_REALTIME_INODE(ip)) {
1586                 /*
1587                  * Walk thru the extent list and count the realtime blocks.
1588                  */
1589                 error = xfs_qm_get_rtblks(ip, &rtblks);
1590                 if (error)
1591                         goto error0;
1592         }
1593
1594         nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1595
1596         /*
1597          * Add the (disk blocks and inode) resources occupied by this
1598          * inode to its dquots. We do this adjustment in the incore dquot,
1599          * and also copy the changes to its buffer.
1600          * We don't care about putting these changes in a transaction
1601          * envelope because if we crash in the middle of a 'quotacheck'
1602          * we have to start from the beginning anyway.
1603          * Once we're done, we'll log all the dquot bufs.
1604          *
1605          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1606          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1607          */
1608         if (XFS_IS_UQUOTA_ON(mp)) {
1609                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_uid,
1610                                                    XFS_DQ_USER, nblks, rtblks);
1611                 if (error)
1612                         goto error0;
1613         }
1614
1615         if (XFS_IS_GQUOTA_ON(mp)) {
1616                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_gid,
1617                                                    XFS_DQ_GROUP, nblks, rtblks);
1618                 if (error)
1619                         goto error0;
1620         }
1621
1622         if (XFS_IS_PQUOTA_ON(mp)) {
1623                 error = xfs_qm_quotacheck_dqadjust(ip, xfs_get_projid(ip),
1624                                                    XFS_DQ_PROJ, nblks, rtblks);
1625                 if (error)
1626                         goto error0;
1627         }
1628
1629         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1630         IRELE(ip);
1631         *res = BULKSTAT_RV_DIDONE;
1632         return 0;
1633
1634 error0:
1635         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1636         IRELE(ip);
1637         *res = BULKSTAT_RV_GIVEUP;
1638         return error;
1639 }
1640
1641 /*
1642  * Walk thru all the filesystem inodes and construct a consistent view
1643  * of the disk quota world. If the quotacheck fails, disable quotas.
1644  */
1645 int
1646 xfs_qm_quotacheck(
1647         xfs_mount_t     *mp)
1648 {
1649         int             done, count, error;
1650         xfs_ino_t       lastino;
1651         size_t          structsz;
1652         xfs_inode_t     *uip, *gip;
1653         uint            flags;
1654
1655         count = INT_MAX;
1656         structsz = 1;
1657         lastino = 0;
1658         flags = 0;
1659
1660         ASSERT(mp->m_quotainfo->qi_uquotaip || mp->m_quotainfo->qi_gquotaip);
1661         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1662
1663         /*
1664          * There should be no cached dquots. The (simplistic) quotacheck
1665          * algorithm doesn't like that.
1666          */
1667         ASSERT(list_empty(&mp->m_quotainfo->qi_dqlist));
1668
1669         xfs_notice(mp, "Quotacheck needed: Please wait.");
1670
1671         /*
1672          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1673          * their counters to zero. We need a clean slate.
1674          * We don't log our changes till later.
1675          */
1676         uip = mp->m_quotainfo->qi_uquotaip;
1677         if (uip) {
1678                 error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA);
1679                 if (error)
1680                         goto error_return;
1681                 flags |= XFS_UQUOTA_CHKD;
1682         }
1683
1684         gip = mp->m_quotainfo->qi_gquotaip;
1685         if (gip) {
1686                 error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1687                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1688                 if (error)
1689                         goto error_return;
1690                 flags |= XFS_OQUOTA_CHKD;
1691         }
1692
1693         do {
1694                 /*
1695                  * Iterate thru all the inodes in the file system,
1696                  * adjusting the corresponding dquot counters in core.
1697                  */
1698                 error = xfs_bulkstat(mp, &lastino, &count,
1699                                      xfs_qm_dqusage_adjust,
1700                                      structsz, NULL, &done);
1701                 if (error)
1702                         break;
1703
1704         } while (!done);
1705
1706         /*
1707          * We've made all the changes that we need to make incore.
1708          * Flush them down to disk buffers if everything was updated
1709          * successfully.
1710          */
1711         if (!error)
1712                 error = xfs_qm_dqflush_all(mp, 0);
1713
1714         /*
1715          * We can get this error if we couldn't do a dquot allocation inside
1716          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1717          * dirty dquots that might be cached, we just want to get rid of them
1718          * and turn quotaoff. The dquots won't be attached to any of the inodes
1719          * at this point (because we intentionally didn't in dqget_noattach).
1720          */
1721         if (error) {
1722                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL);
1723                 goto error_return;
1724         }
1725
1726         /*
1727          * We didn't log anything, because if we crashed, we'll have to
1728          * start the quotacheck from scratch anyway. However, we must make
1729          * sure that our dquot changes are secure before we put the
1730          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1731          * flush.
1732          */
1733         XFS_bflush(mp->m_ddev_targp);
1734
1735         /*
1736          * If one type of quotas is off, then it will lose its
1737          * quotachecked status, since we won't be doing accounting for
1738          * that type anymore.
1739          */
1740         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1741         mp->m_qflags |= flags;
1742
1743         xfs_qm_dquot_list_print(mp);
1744
1745  error_return:
1746         if (error) {
1747                 xfs_warn(mp,
1748         "Quotacheck: Unsuccessful (Error %d): Disabling quotas.",
1749                         error);
1750                 /*
1751                  * We must turn off quotas.
1752                  */
1753                 ASSERT(mp->m_quotainfo != NULL);
1754                 ASSERT(xfs_Gqm != NULL);
1755                 xfs_qm_destroy_quotainfo(mp);
1756                 if (xfs_mount_reset_sbqflags(mp)) {
1757                         xfs_warn(mp,
1758                                 "Quotacheck: Failed to reset quota flags.");
1759                 }
1760         } else
1761                 xfs_notice(mp, "Quotacheck: Done.");
1762         return (error);
1763 }
1764
1765 /*
1766  * This is called after the superblock has been read in and we're ready to
1767  * iget the quota inodes.
1768  */
1769 STATIC int
1770 xfs_qm_init_quotainos(
1771         xfs_mount_t     *mp)
1772 {
1773         xfs_inode_t     *uip, *gip;
1774         int             error;
1775         __int64_t       sbflags;
1776         uint            flags;
1777
1778         ASSERT(mp->m_quotainfo);
1779         uip = gip = NULL;
1780         sbflags = 0;
1781         flags = 0;
1782
1783         /*
1784          * Get the uquota and gquota inodes
1785          */
1786         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1787                 if (XFS_IS_UQUOTA_ON(mp) &&
1788                     mp->m_sb.sb_uquotino != NULLFSINO) {
1789                         ASSERT(mp->m_sb.sb_uquotino > 0);
1790                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1791                                              0, 0, &uip)))
1792                                 return XFS_ERROR(error);
1793                 }
1794                 if (XFS_IS_OQUOTA_ON(mp) &&
1795                     mp->m_sb.sb_gquotino != NULLFSINO) {
1796                         ASSERT(mp->m_sb.sb_gquotino > 0);
1797                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1798                                              0, 0, &gip))) {
1799                                 if (uip)
1800                                         IRELE(uip);
1801                                 return XFS_ERROR(error);
1802                         }
1803                 }
1804         } else {
1805                 flags |= XFS_QMOPT_SBVERSION;
1806                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1807                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1808         }
1809
1810         /*
1811          * Create the two inodes, if they don't exist already. The changes
1812          * made above will get added to a transaction and logged in one of
1813          * the qino_alloc calls below.  If the device is readonly,
1814          * temporarily switch to read-write to do this.
1815          */
1816         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1817                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1818                                               sbflags | XFS_SB_UQUOTINO,
1819                                               flags | XFS_QMOPT_UQUOTA)))
1820                         return XFS_ERROR(error);
1821
1822                 flags &= ~XFS_QMOPT_SBVERSION;
1823         }
1824         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1825                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1826                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1827                 error = xfs_qm_qino_alloc(mp, &gip,
1828                                           sbflags | XFS_SB_GQUOTINO, flags);
1829                 if (error) {
1830                         if (uip)
1831                                 IRELE(uip);
1832
1833                         return XFS_ERROR(error);
1834                 }
1835         }
1836
1837         mp->m_quotainfo->qi_uquotaip = uip;
1838         mp->m_quotainfo->qi_gquotaip = gip;
1839
1840         return 0;
1841 }
1842
1843
1844
1845 /*
1846  * Just pop the least recently used dquot off the freelist and
1847  * recycle it. The returned dquot is locked.
1848  */
1849 STATIC xfs_dquot_t *
1850 xfs_qm_dqreclaim_one(void)
1851 {
1852         xfs_dquot_t     *dqpout;
1853         xfs_dquot_t     *dqp;
1854         int             restarts;
1855         int             startagain;
1856
1857         restarts = 0;
1858         dqpout = NULL;
1859
1860         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
1861 again:
1862         startagain = 0;
1863         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1864
1865         list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) {
1866                 struct xfs_mount *mp = dqp->q_mount;
1867                 xfs_dqlock(dqp);
1868
1869                 /*
1870                  * We are racing with dqlookup here. Naturally we don't
1871                  * want to reclaim a dquot that lookup wants. We release the
1872                  * freelist lock and start over, so that lookup will grab
1873                  * both the dquot and the freelistlock.
1874                  */
1875                 if (dqp->dq_flags & XFS_DQ_WANT) {
1876                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
1877
1878                         trace_xfs_dqreclaim_want(dqp);
1879                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
1880                         restarts++;
1881                         startagain = 1;
1882                         goto dqunlock;
1883                 }
1884
1885                 /*
1886                  * If the dquot is inactive, we are assured that it is
1887                  * not on the mplist or the hashlist, and that makes our
1888                  * life easier.
1889                  */
1890                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
1891                         ASSERT(mp == NULL);
1892                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
1893                         ASSERT(list_empty(&dqp->q_hashlist));
1894                         ASSERT(list_empty(&dqp->q_mplist));
1895                         list_del_init(&dqp->q_freelist);
1896                         xfs_Gqm->qm_dqfrlist_cnt--;
1897                         dqpout = dqp;
1898                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
1899                         goto dqunlock;
1900                 }
1901
1902                 ASSERT(dqp->q_hash);
1903                 ASSERT(!list_empty(&dqp->q_mplist));
1904
1905                 /*
1906                  * Try to grab the flush lock. If this dquot is in the process
1907                  * of getting flushed to disk, we don't want to reclaim it.
1908                  */
1909                 if (!xfs_dqflock_nowait(dqp))
1910                         goto dqunlock;
1911
1912                 /*
1913                  * We have the flush lock so we know that this is not in the
1914                  * process of being flushed. So, if this is dirty, flush it
1915                  * DELWRI so that we don't get a freelist infested with
1916                  * dirty dquots.
1917                  */
1918                 if (XFS_DQ_IS_DIRTY(dqp)) {
1919                         int     error;
1920
1921                         trace_xfs_dqreclaim_dirty(dqp);
1922
1923                         /*
1924                          * We flush it delayed write, so don't bother
1925                          * releasing the freelist lock.
1926                          */
1927                         error = xfs_qm_dqflush(dqp, 0);
1928                         if (error) {
1929                                 xfs_warn(mp, "%s: dquot %p flush failed",
1930                                         __func__, dqp);
1931                         }
1932                         goto dqunlock;
1933                 }
1934
1935                 /*
1936                  * We're trying to get the hashlock out of order. This races
1937                  * with dqlookup; so, we giveup and goto the next dquot if
1938                  * we couldn't get the hashlock. This way, we won't starve
1939                  * a dqlookup process that holds the hashlock that is
1940                  * waiting for the freelist lock.
1941                  */
1942                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
1943                         restarts++;
1944                         goto dqfunlock;
1945                 }
1946
1947                 /*
1948                  * This races with dquot allocation code as well as dqflush_all
1949                  * and reclaim code. So, if we failed to grab the mplist lock,
1950                  * giveup everything and start over.
1951                  */
1952                 if (!mutex_trylock(&mp->m_quotainfo->qi_dqlist_lock)) {
1953                         restarts++;
1954                         startagain = 1;
1955                         goto qhunlock;
1956                 }
1957
1958                 ASSERT(dqp->q_nrefs == 0);
1959                 list_del_init(&dqp->q_mplist);
1960                 mp->m_quotainfo->qi_dquots--;
1961                 mp->m_quotainfo->qi_dqreclaims++;
1962                 list_del_init(&dqp->q_hashlist);
1963                 dqp->q_hash->qh_version++;
1964                 list_del_init(&dqp->q_freelist);
1965                 xfs_Gqm->qm_dqfrlist_cnt--;
1966                 dqpout = dqp;
1967                 mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1968 qhunlock:
1969                 mutex_unlock(&dqp->q_hash->qh_lock);
1970 dqfunlock:
1971                 xfs_dqfunlock(dqp);
1972 dqunlock:
1973                 xfs_dqunlock(dqp);
1974                 if (dqpout)
1975                         break;
1976                 if (restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
1977                         break;
1978                 if (startagain) {
1979                         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1980                         goto again;
1981                 }
1982         }
1983         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1984         return dqpout;
1985 }
1986
1987 /*
1988  * Traverse the freelist of dquots and attempt to reclaim a maximum of
1989  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
1990  * favor the lookup function ...
1991  */
1992 STATIC int
1993 xfs_qm_shake_freelist(
1994         int     howmany)
1995 {
1996         int             nreclaimed = 0;
1997         xfs_dquot_t     *dqp;
1998
1999         if (howmany <= 0)
2000                 return 0;
2001
2002         while (nreclaimed < howmany) {
2003                 dqp = xfs_qm_dqreclaim_one();
2004                 if (!dqp)
2005                         return nreclaimed;
2006                 xfs_qm_dqdestroy(dqp);
2007                 nreclaimed++;
2008         }
2009         return nreclaimed;
2010 }
2011
2012 /*
2013  * The kmem_shake interface is invoked when memory is running low.
2014  */
2015 /* ARGSUSED */
2016 STATIC int
2017 xfs_qm_shake(
2018         struct shrinker *shrink,
2019         int             nr_to_scan,
2020         gfp_t           gfp_mask)
2021 {
2022         int     ndqused, nfree, n;
2023
2024         if (!kmem_shake_allow(gfp_mask))
2025                 return 0;
2026         if (!xfs_Gqm)
2027                 return 0;
2028
2029         nfree = xfs_Gqm->qm_dqfrlist_cnt; /* free dquots */
2030         /* incore dquots in all f/s's */
2031         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2032
2033         ASSERT(ndqused >= 0);
2034
2035         if (nfree <= ndqused && nfree < ndquot)
2036                 return 0;
2037
2038         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2039         n = nfree - ndqused - ndquot;           /* # over target */
2040
2041         return xfs_qm_shake_freelist(MAX(nfree, n));
2042 }
2043
2044
2045 /*------------------------------------------------------------------*/
2046
2047 /*
2048  * Return a new incore dquot. Depending on the number of
2049  * dquots in the system, we either allocate a new one on the kernel heap,
2050  * or reclaim a free one.
2051  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2052  * to reclaim an existing one from the freelist.
2053  */
2054 boolean_t
2055 xfs_qm_dqalloc_incore(
2056         xfs_dquot_t **O_dqpp)
2057 {
2058         xfs_dquot_t     *dqp;
2059
2060         /*
2061          * Check against high water mark to see if we want to pop
2062          * a nincompoop dquot off the freelist.
2063          */
2064         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2065                 /*
2066                  * Try to recycle a dquot from the freelist.
2067                  */
2068                 if ((dqp = xfs_qm_dqreclaim_one())) {
2069                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2070                         /*
2071                          * Just zero the core here. The rest will get
2072                          * reinitialized by caller. XXX we shouldn't even
2073                          * do this zero ...
2074                          */
2075                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2076                         *O_dqpp = dqp;
2077                         return B_FALSE;
2078                 }
2079                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2080         }
2081
2082         /*
2083          * Allocate a brand new dquot on the kernel heap and return it
2084          * to the caller to initialize.
2085          */
2086         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2087         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2088         atomic_inc(&xfs_Gqm->qm_totaldquots);
2089
2090         return B_TRUE;
2091 }
2092
2093
2094 /*
2095  * Start a transaction and write the incore superblock changes to
2096  * disk. flags parameter indicates which fields have changed.
2097  */
2098 int
2099 xfs_qm_write_sb_changes(
2100         xfs_mount_t     *mp,
2101         __int64_t       flags)
2102 {
2103         xfs_trans_t     *tp;
2104         int             error;
2105
2106 #ifdef QUOTADEBUG
2107         xfs_notice(mp, "Writing superblock quota changes");
2108 #endif
2109         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2110         if ((error = xfs_trans_reserve(tp, 0,
2111                                       mp->m_sb.sb_sectsize + 128, 0,
2112                                       0,
2113                                       XFS_DEFAULT_LOG_COUNT))) {
2114                 xfs_trans_cancel(tp, 0);
2115                 return error;
2116         }
2117
2118         xfs_mod_sb(tp, flags);
2119         error = xfs_trans_commit(tp, 0);
2120
2121         return error;
2122 }
2123
2124
2125 /* --------------- utility functions for vnodeops ---------------- */
2126
2127
2128 /*
2129  * Given an inode, a uid, gid and prid make sure that we have
2130  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2131  * quotas by creating this file.
2132  * This also attaches dquot(s) to the given inode after locking it,
2133  * and returns the dquots corresponding to the uid and/or gid.
2134  *
2135  * in   : inode (unlocked)
2136  * out  : udquot, gdquot with references taken and unlocked
2137  */
2138 int
2139 xfs_qm_vop_dqalloc(
2140         struct xfs_inode        *ip,
2141         uid_t                   uid,
2142         gid_t                   gid,
2143         prid_t                  prid,
2144         uint                    flags,
2145         struct xfs_dquot        **O_udqpp,
2146         struct xfs_dquot        **O_gdqpp)
2147 {
2148         struct xfs_mount        *mp = ip->i_mount;
2149         struct xfs_dquot        *uq, *gq;
2150         int                     error;
2151         uint                    lockflags;
2152
2153         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2154                 return 0;
2155
2156         lockflags = XFS_ILOCK_EXCL;
2157         xfs_ilock(ip, lockflags);
2158
2159         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2160                 gid = ip->i_d.di_gid;
2161
2162         /*
2163          * Attach the dquot(s) to this inode, doing a dquot allocation
2164          * if necessary. The dquot(s) will not be locked.
2165          */
2166         if (XFS_NOT_DQATTACHED(mp, ip)) {
2167                 error = xfs_qm_dqattach_locked(ip, XFS_QMOPT_DQALLOC);
2168                 if (error) {
2169                         xfs_iunlock(ip, lockflags);
2170                         return error;
2171                 }
2172         }
2173
2174         uq = gq = NULL;
2175         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2176                 if (ip->i_d.di_uid != uid) {
2177                         /*
2178                          * What we need is the dquot that has this uid, and
2179                          * if we send the inode to dqget, the uid of the inode
2180                          * takes priority over what's sent in the uid argument.
2181                          * We must unlock inode here before calling dqget if
2182                          * we're not sending the inode, because otherwise
2183                          * we'll deadlock by doing trans_reserve while
2184                          * holding ilock.
2185                          */
2186                         xfs_iunlock(ip, lockflags);
2187                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2188                                                  XFS_DQ_USER,
2189                                                  XFS_QMOPT_DQALLOC |
2190                                                  XFS_QMOPT_DOWARN,
2191                                                  &uq))) {
2192                                 ASSERT(error != ENOENT);
2193                                 return error;
2194                         }
2195                         /*
2196                          * Get the ilock in the right order.
2197                          */
2198                         xfs_dqunlock(uq);
2199                         lockflags = XFS_ILOCK_SHARED;
2200                         xfs_ilock(ip, lockflags);
2201                 } else {
2202                         /*
2203                          * Take an extra reference, because we'll return
2204                          * this to caller
2205                          */
2206                         ASSERT(ip->i_udquot);
2207                         uq = ip->i_udquot;
2208                         xfs_dqlock(uq);
2209                         XFS_DQHOLD(uq);
2210                         xfs_dqunlock(uq);
2211                 }
2212         }
2213         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2214                 if (ip->i_d.di_gid != gid) {
2215                         xfs_iunlock(ip, lockflags);
2216                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2217                                                  XFS_DQ_GROUP,
2218                                                  XFS_QMOPT_DQALLOC |
2219                                                  XFS_QMOPT_DOWARN,
2220                                                  &gq))) {
2221                                 if (uq)
2222                                         xfs_qm_dqrele(uq);
2223                                 ASSERT(error != ENOENT);
2224                                 return error;
2225                         }
2226                         xfs_dqunlock(gq);
2227                         lockflags = XFS_ILOCK_SHARED;
2228                         xfs_ilock(ip, lockflags);
2229                 } else {
2230                         ASSERT(ip->i_gdquot);
2231                         gq = ip->i_gdquot;
2232                         xfs_dqlock(gq);
2233                         XFS_DQHOLD(gq);
2234                         xfs_dqunlock(gq);
2235                 }
2236         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2237                 if (xfs_get_projid(ip) != prid) {
2238                         xfs_iunlock(ip, lockflags);
2239                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2240                                                  XFS_DQ_PROJ,
2241                                                  XFS_QMOPT_DQALLOC |
2242                                                  XFS_QMOPT_DOWARN,
2243                                                  &gq))) {
2244                                 if (uq)
2245                                         xfs_qm_dqrele(uq);
2246                                 ASSERT(error != ENOENT);
2247                                 return (error);
2248                         }
2249                         xfs_dqunlock(gq);
2250                         lockflags = XFS_ILOCK_SHARED;
2251                         xfs_ilock(ip, lockflags);
2252                 } else {
2253                         ASSERT(ip->i_gdquot);
2254                         gq = ip->i_gdquot;
2255                         xfs_dqlock(gq);
2256                         XFS_DQHOLD(gq);
2257                         xfs_dqunlock(gq);
2258                 }
2259         }
2260         if (uq)
2261                 trace_xfs_dquot_dqalloc(ip);
2262
2263         xfs_iunlock(ip, lockflags);
2264         if (O_udqpp)
2265                 *O_udqpp = uq;
2266         else if (uq)
2267                 xfs_qm_dqrele(uq);
2268         if (O_gdqpp)
2269                 *O_gdqpp = gq;
2270         else if (gq)
2271                 xfs_qm_dqrele(gq);
2272         return 0;
2273 }
2274
2275 /*
2276  * Actually transfer ownership, and do dquot modifications.
2277  * These were already reserved.
2278  */
2279 xfs_dquot_t *
2280 xfs_qm_vop_chown(
2281         xfs_trans_t     *tp,
2282         xfs_inode_t     *ip,
2283         xfs_dquot_t     **IO_olddq,
2284         xfs_dquot_t     *newdq)
2285 {
2286         xfs_dquot_t     *prevdq;
2287         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2288                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2289
2290
2291         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2292         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2293
2294         /* old dquot */
2295         prevdq = *IO_olddq;
2296         ASSERT(prevdq);
2297         ASSERT(prevdq != newdq);
2298
2299         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2300         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2301
2302         /* the sparkling new dquot */
2303         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2304         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2305
2306         /*
2307          * Take an extra reference, because the inode
2308          * is going to keep this dquot pointer even
2309          * after the trans_commit.
2310          */
2311         xfs_dqlock(newdq);
2312         XFS_DQHOLD(newdq);
2313         xfs_dqunlock(newdq);
2314         *IO_olddq = newdq;
2315
2316         return prevdq;
2317 }
2318
2319 /*
2320  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2321  */
2322 int
2323 xfs_qm_vop_chown_reserve(
2324         xfs_trans_t     *tp,
2325         xfs_inode_t     *ip,
2326         xfs_dquot_t     *udqp,
2327         xfs_dquot_t     *gdqp,
2328         uint            flags)
2329 {
2330         xfs_mount_t     *mp = ip->i_mount;
2331         uint            delblks, blkflags, prjflags = 0;
2332         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2333         int             error;
2334
2335
2336         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2337         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2338
2339         delblks = ip->i_delayed_blks;
2340         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2341         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2342                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2343
2344         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2345             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2346                 delblksudq = udqp;
2347                 /*
2348                  * If there are delayed allocation blocks, then we have to
2349                  * unreserve those from the old dquot, and add them to the
2350                  * new dquot.
2351                  */
2352                 if (delblks) {
2353                         ASSERT(ip->i_udquot);
2354                         unresudq = ip->i_udquot;
2355                 }
2356         }
2357         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2358                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2359                      xfs_get_projid(ip) != be32_to_cpu(gdqp->q_core.d_id))
2360                         prjflags = XFS_QMOPT_ENOSPC;
2361
2362                 if (prjflags ||
2363                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2364                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2365                         delblksgdq = gdqp;
2366                         if (delblks) {
2367                                 ASSERT(ip->i_gdquot);
2368                                 unresgdq = ip->i_gdquot;
2369                         }
2370                 }
2371         }
2372
2373         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2374                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2375                                 flags | blkflags | prjflags)))
2376                 return (error);
2377
2378         /*
2379          * Do the delayed blks reservations/unreservations now. Since, these
2380          * are done without the help of a transaction, if a reservation fails
2381          * its previous reservations won't be automatically undone by trans
2382          * code. So, we have to do it manually here.
2383          */
2384         if (delblks) {
2385                 /*
2386                  * Do the reservations first. Unreservation can't fail.
2387                  */
2388                 ASSERT(delblksudq || delblksgdq);
2389                 ASSERT(unresudq || unresgdq);
2390                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2391                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2392                                 flags | blkflags | prjflags)))
2393                         return (error);
2394                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2395                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2396                                 blkflags);
2397         }
2398
2399         return (0);
2400 }
2401
2402 int
2403 xfs_qm_vop_rename_dqattach(
2404         struct xfs_inode        **i_tab)
2405 {
2406         struct xfs_mount        *mp = i_tab[0]->i_mount;
2407         int                     i;
2408
2409         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2410                 return 0;
2411
2412         for (i = 0; (i < 4 && i_tab[i]); i++) {
2413                 struct xfs_inode        *ip = i_tab[i];
2414                 int                     error;
2415
2416                 /*
2417                  * Watch out for duplicate entries in the table.
2418                  */
2419                 if (i == 0 || ip != i_tab[i-1]) {
2420                         if (XFS_NOT_DQATTACHED(mp, ip)) {
2421                                 error = xfs_qm_dqattach(ip, 0);
2422                                 if (error)
2423                                         return error;
2424                         }
2425                 }
2426         }
2427         return 0;
2428 }
2429
2430 void
2431 xfs_qm_vop_create_dqattach(
2432         struct xfs_trans        *tp,
2433         struct xfs_inode        *ip,
2434         struct xfs_dquot        *udqp,
2435         struct xfs_dquot        *gdqp)
2436 {
2437         struct xfs_mount        *mp = tp->t_mountp;
2438
2439         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2440                 return;
2441
2442         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2443         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2444
2445         if (udqp) {
2446                 xfs_dqlock(udqp);
2447                 XFS_DQHOLD(udqp);
2448                 xfs_dqunlock(udqp);
2449                 ASSERT(ip->i_udquot == NULL);
2450                 ip->i_udquot = udqp;
2451                 ASSERT(XFS_IS_UQUOTA_ON(mp));
2452                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2453                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2454         }
2455         if (gdqp) {
2456                 xfs_dqlock(gdqp);
2457                 XFS_DQHOLD(gdqp);
2458                 xfs_dqunlock(gdqp);
2459                 ASSERT(ip->i_gdquot == NULL);
2460                 ip->i_gdquot = gdqp;
2461                 ASSERT(XFS_IS_OQUOTA_ON(mp));
2462                 ASSERT((XFS_IS_GQUOTA_ON(mp) ?
2463                         ip->i_d.di_gid : xfs_get_projid(ip)) ==
2464                                 be32_to_cpu(gdqp->q_core.d_id));
2465                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2466         }
2467 }
2468