xfs: fix variable set but not used warnings
[linux-2.6.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                     error;
465
466         if (!q)
467                 return 0;
468 again:
469         mutex_lock(&q->qi_dqlist_lock);
470         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
471                 xfs_dqlock(dqp);
472                 if (! XFS_DQ_IS_DIRTY(dqp)) {
473                         xfs_dqunlock(dqp);
474                         continue;
475                 }
476
477                 /* XXX a sentinel would be better */
478                 recl = q->qi_dqreclaims;
479                 if (!xfs_dqflock_nowait(dqp)) {
480                         /*
481                          * If we can't grab the flush lock then check
482                          * to see if the dquot has been flushed delayed
483                          * write.  If so, grab its buffer and send it
484                          * out immediately.  We'll be able to acquire
485                          * the flush lock when the I/O completes.
486                          */
487                         xfs_qm_dqflock_pushbuf_wait(dqp);
488                 }
489                 /*
490                  * Let go of the mplist lock. We don't want to hold it
491                  * across a disk write.
492                  */
493                 mutex_unlock(&q->qi_dqlist_lock);
494                 error = xfs_qm_dqflush(dqp, sync_mode);
495                 xfs_dqunlock(dqp);
496                 if (error)
497                         return error;
498
499                 mutex_lock(&q->qi_dqlist_lock);
500                 if (recl != q->qi_dqreclaims) {
501                         mutex_unlock(&q->qi_dqlist_lock);
502                         /* XXX restart limit */
503                         goto again;
504                 }
505         }
506
507         mutex_unlock(&q->qi_dqlist_lock);
508         /* return ! busy */
509         return 0;
510 }
511 /*
512  * Release the group dquot pointers the user dquots may be
513  * carrying around as a hint. mplist is locked on entry and exit.
514  */
515 STATIC void
516 xfs_qm_detach_gdquots(
517         struct xfs_mount        *mp)
518 {
519         struct xfs_quotainfo    *q = mp->m_quotainfo;
520         struct xfs_dquot        *dqp, *gdqp;
521         int                     nrecl;
522
523  again:
524         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
525         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
526                 xfs_dqlock(dqp);
527                 if ((gdqp = dqp->q_gdquot)) {
528                         xfs_dqlock(gdqp);
529                         dqp->q_gdquot = NULL;
530                 }
531                 xfs_dqunlock(dqp);
532
533                 if (gdqp) {
534                         /*
535                          * Can't hold the mplist lock across a dqput.
536                          * XXXmust convert to marker based iterations here.
537                          */
538                         nrecl = q->qi_dqreclaims;
539                         mutex_unlock(&q->qi_dqlist_lock);
540                         xfs_qm_dqput(gdqp);
541
542                         mutex_lock(&q->qi_dqlist_lock);
543                         if (nrecl != q->qi_dqreclaims)
544                                 goto again;
545                 }
546         }
547 }
548
549 /*
550  * Go through all the incore dquots of this file system and take them
551  * off the mplist and hashlist, if the dquot type matches the dqtype
552  * parameter. This is used when turning off quota accounting for
553  * users and/or groups, as well as when the filesystem is unmounting.
554  */
555 STATIC int
556 xfs_qm_dqpurge_int(
557         struct xfs_mount        *mp,
558         uint                    flags)
559 {
560         struct xfs_quotainfo    *q = mp->m_quotainfo;
561         struct xfs_dquot        *dqp, *n;
562         uint                    dqtype;
563         int                     nrecl;
564         int                     nmisses;
565
566         if (!q)
567                 return 0;
568
569         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
570         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
571         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
572
573         mutex_lock(&q->qi_dqlist_lock);
574
575         /*
576          * In the first pass through all incore dquots of this filesystem,
577          * we release the group dquot pointers the user dquots may be
578          * carrying around as a hint. We need to do this irrespective of
579          * what's being turned off.
580          */
581         xfs_qm_detach_gdquots(mp);
582
583       again:
584         nmisses = 0;
585         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
586         /*
587          * Try to get rid of all of the unwanted dquots. The idea is to
588          * get them off mplist and hashlist, but leave them on freelist.
589          */
590         list_for_each_entry_safe(dqp, n, &q->qi_dqlist, q_mplist) {
591                 /*
592                  * It's OK to look at the type without taking dqlock here.
593                  * We're holding the mplist lock here, and that's needed for
594                  * a dqreclaim.
595                  */
596                 if ((dqp->dq_flags & dqtype) == 0)
597                         continue;
598
599                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
600                         nrecl = q->qi_dqreclaims;
601                         mutex_unlock(&q->qi_dqlist_lock);
602                         mutex_lock(&dqp->q_hash->qh_lock);
603                         mutex_lock(&q->qi_dqlist_lock);
604
605                         /*
606                          * XXXTheoretically, we can get into a very long
607                          * ping pong game here.
608                          * No one can be adding dquots to the mplist at
609                          * this point, but somebody might be taking things off.
610                          */
611                         if (nrecl != q->qi_dqreclaims) {
612                                 mutex_unlock(&dqp->q_hash->qh_lock);
613                                 goto again;
614                         }
615                 }
616
617                 /*
618                  * Take the dquot off the mplist and hashlist. It may remain on
619                  * freelist in INACTIVE state.
620                  */
621                 nmisses += xfs_qm_dqpurge(dqp);
622         }
623         mutex_unlock(&q->qi_dqlist_lock);
624         return nmisses;
625 }
626
627 int
628 xfs_qm_dqpurge_all(
629         xfs_mount_t     *mp,
630         uint            flags)
631 {
632         int             ndquots;
633
634         /*
635          * Purge the dquot cache.
636          * None of the dquots should really be busy at this point.
637          */
638         if (mp->m_quotainfo) {
639                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
640                         delay(ndquots * 10);
641                 }
642         }
643         return 0;
644 }
645
646 STATIC int
647 xfs_qm_dqattach_one(
648         xfs_inode_t     *ip,
649         xfs_dqid_t      id,
650         uint            type,
651         uint            doalloc,
652         xfs_dquot_t     *udqhint, /* hint */
653         xfs_dquot_t     **IO_idqpp)
654 {
655         xfs_dquot_t     *dqp;
656         int             error;
657
658         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
659         error = 0;
660
661         /*
662          * See if we already have it in the inode itself. IO_idqpp is
663          * &i_udquot or &i_gdquot. This made the code look weird, but
664          * made the logic a lot simpler.
665          */
666         dqp = *IO_idqpp;
667         if (dqp) {
668                 trace_xfs_dqattach_found(dqp);
669                 return 0;
670         }
671
672         /*
673          * udqhint is the i_udquot field in inode, and is non-NULL only
674          * when the type arg is group/project. Its purpose is to save a
675          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
676          * the user dquot.
677          */
678         if (udqhint) {
679                 ASSERT(type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
680                 xfs_dqlock(udqhint);
681
682                 /*
683                  * No need to take dqlock to look at the id.
684                  *
685                  * The ID can't change until it gets reclaimed, and it won't
686                  * be reclaimed as long as we have a ref from inode and we
687                  * hold the ilock.
688                  */
689                 dqp = udqhint->q_gdquot;
690                 if (dqp && be32_to_cpu(dqp->q_core.d_id) == id) {
691                         xfs_dqlock(dqp);
692                         XFS_DQHOLD(dqp);
693                         ASSERT(*IO_idqpp == NULL);
694                         *IO_idqpp = dqp;
695
696                         xfs_dqunlock(dqp);
697                         xfs_dqunlock(udqhint);
698                         return 0;
699                 }
700
701                 /*
702                  * We can't hold a dquot lock when we call the dqget code.
703                  * We'll deadlock in no time, because of (not conforming to)
704                  * lock ordering - the inodelock comes before any dquot lock,
705                  * and we may drop and reacquire the ilock in xfs_qm_dqget().
706                  */
707                 xfs_dqunlock(udqhint);
708         }
709
710         /*
711          * Find the dquot from somewhere. This bumps the
712          * reference count of dquot and returns it locked.
713          * This can return ENOENT if dquot didn't exist on
714          * disk and we didn't ask it to allocate;
715          * ESRCH if quotas got turned off suddenly.
716          */
717         error = xfs_qm_dqget(ip->i_mount, ip, id, type, XFS_QMOPT_DOWARN, &dqp);
718         if (error)
719                 return error;
720
721         trace_xfs_dqattach_get(dqp);
722
723         /*
724          * dqget may have dropped and re-acquired the ilock, but it guarantees
725          * that the dquot returned is the one that should go in the inode.
726          */
727         *IO_idqpp = dqp;
728         xfs_dqunlock(dqp);
729         return 0;
730 }
731
732
733 /*
734  * Given a udquot and gdquot, attach a ptr to the group dquot in the
735  * udquot as a hint for future lookups. The idea sounds simple, but the
736  * execution isn't, because the udquot might have a group dquot attached
737  * already and getting rid of that gets us into lock ordering constraints.
738  * The process is complicated more by the fact that the dquots may or may not
739  * be locked on entry.
740  */
741 STATIC void
742 xfs_qm_dqattach_grouphint(
743         xfs_dquot_t     *udq,
744         xfs_dquot_t     *gdq)
745 {
746         xfs_dquot_t     *tmp;
747
748         xfs_dqlock(udq);
749
750         if ((tmp = udq->q_gdquot)) {
751                 if (tmp == gdq) {
752                         xfs_dqunlock(udq);
753                         return;
754                 }
755
756                 udq->q_gdquot = NULL;
757                 /*
758                  * We can't keep any dqlocks when calling dqrele,
759                  * because the freelist lock comes before dqlocks.
760                  */
761                 xfs_dqunlock(udq);
762                 /*
763                  * we took a hard reference once upon a time in dqget,
764                  * so give it back when the udquot no longer points at it
765                  * dqput() does the unlocking of the dquot.
766                  */
767                 xfs_qm_dqrele(tmp);
768
769                 xfs_dqlock(udq);
770                 xfs_dqlock(gdq);
771
772         } else {
773                 ASSERT(XFS_DQ_IS_LOCKED(udq));
774                 xfs_dqlock(gdq);
775         }
776
777         ASSERT(XFS_DQ_IS_LOCKED(udq));
778         ASSERT(XFS_DQ_IS_LOCKED(gdq));
779         /*
780          * Somebody could have attached a gdquot here,
781          * when we dropped the uqlock. If so, just do nothing.
782          */
783         if (udq->q_gdquot == NULL) {
784                 XFS_DQHOLD(gdq);
785                 udq->q_gdquot = gdq;
786         }
787
788         xfs_dqunlock(gdq);
789         xfs_dqunlock(udq);
790 }
791
792
793 /*
794  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
795  * into account.
796  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
797  * Inode may get unlocked and relocked in here, and the caller must deal with
798  * the consequences.
799  */
800 int
801 xfs_qm_dqattach_locked(
802         xfs_inode_t     *ip,
803         uint            flags)
804 {
805         xfs_mount_t     *mp = ip->i_mount;
806         uint            nquotas = 0;
807         int             error = 0;
808
809         if (!XFS_IS_QUOTA_RUNNING(mp) ||
810             !XFS_IS_QUOTA_ON(mp) ||
811             !XFS_NOT_DQATTACHED(mp, ip) ||
812             ip->i_ino == mp->m_sb.sb_uquotino ||
813             ip->i_ino == mp->m_sb.sb_gquotino)
814                 return 0;
815
816         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
817
818         if (XFS_IS_UQUOTA_ON(mp)) {
819                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
820                                                 flags & XFS_QMOPT_DQALLOC,
821                                                 NULL, &ip->i_udquot);
822                 if (error)
823                         goto done;
824                 nquotas++;
825         }
826
827         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
828         if (XFS_IS_OQUOTA_ON(mp)) {
829                 error = XFS_IS_GQUOTA_ON(mp) ?
830                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
831                                                 flags & XFS_QMOPT_DQALLOC,
832                                                 ip->i_udquot, &ip->i_gdquot) :
833                         xfs_qm_dqattach_one(ip, xfs_get_projid(ip), XFS_DQ_PROJ,
834                                                 flags & XFS_QMOPT_DQALLOC,
835                                                 ip->i_udquot, &ip->i_gdquot);
836                 /*
837                  * Don't worry about the udquot that we may have
838                  * attached above. It'll get detached, if not already.
839                  */
840                 if (error)
841                         goto done;
842                 nquotas++;
843         }
844
845         /*
846          * Attach this group quota to the user quota as a hint.
847          * This WON'T, in general, result in a thrash.
848          */
849         if (nquotas == 2) {
850                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
851                 ASSERT(ip->i_udquot);
852                 ASSERT(ip->i_gdquot);
853
854                 /*
855                  * We may or may not have the i_udquot locked at this point,
856                  * but this check is OK since we don't depend on the i_gdquot to
857                  * be accurate 100% all the time. It is just a hint, and this
858                  * will succeed in general.
859                  */
860                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
861                         goto done;
862                 /*
863                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
864                  */
865                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot);
866         }
867
868  done:
869 #ifdef QUOTADEBUG
870         if (! error) {
871                 if (XFS_IS_UQUOTA_ON(mp))
872                         ASSERT(ip->i_udquot);
873                 if (XFS_IS_OQUOTA_ON(mp))
874                         ASSERT(ip->i_gdquot);
875         }
876         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
877 #endif
878         return error;
879 }
880
881 int
882 xfs_qm_dqattach(
883         struct xfs_inode        *ip,
884         uint                    flags)
885 {
886         int                     error;
887
888         xfs_ilock(ip, XFS_ILOCK_EXCL);
889         error = xfs_qm_dqattach_locked(ip, flags);
890         xfs_iunlock(ip, XFS_ILOCK_EXCL);
891
892         return error;
893 }
894
895 /*
896  * Release dquots (and their references) if any.
897  * The inode should be locked EXCL except when this's called by
898  * xfs_ireclaim.
899  */
900 void
901 xfs_qm_dqdetach(
902         xfs_inode_t     *ip)
903 {
904         if (!(ip->i_udquot || ip->i_gdquot))
905                 return;
906
907         trace_xfs_dquot_dqdetach(ip);
908
909         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
910         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
911         if (ip->i_udquot) {
912                 xfs_qm_dqrele(ip->i_udquot);
913                 ip->i_udquot = NULL;
914         }
915         if (ip->i_gdquot) {
916                 xfs_qm_dqrele(ip->i_gdquot);
917                 ip->i_gdquot = NULL;
918         }
919 }
920
921 int
922 xfs_qm_sync(
923         struct xfs_mount        *mp,
924         int                     flags)
925 {
926         struct xfs_quotainfo    *q = mp->m_quotainfo;
927         int                     recl, restarts;
928         struct xfs_dquot        *dqp;
929         int                     error;
930
931         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
932                 return 0;
933
934         restarts = 0;
935
936   again:
937         mutex_lock(&q->qi_dqlist_lock);
938         /*
939          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
940          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
941          * when we have the mplist lock, we know that dquots will be consistent
942          * as long as we have it locked.
943          */
944         if (!XFS_IS_QUOTA_ON(mp)) {
945                 mutex_unlock(&q->qi_dqlist_lock);
946                 return 0;
947         }
948         ASSERT(mutex_is_locked(&q->qi_dqlist_lock));
949         list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) {
950                 /*
951                  * If this is vfs_sync calling, then skip the dquots that
952                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
953                  * This is very similar to what xfs_sync does with inodes.
954                  */
955                 if (flags & SYNC_TRYLOCK) {
956                         if (!XFS_DQ_IS_DIRTY(dqp))
957                                 continue;
958                         if (!xfs_qm_dqlock_nowait(dqp))
959                                 continue;
960                 } else {
961                         xfs_dqlock(dqp);
962                 }
963
964                 /*
965                  * Now, find out for sure if this dquot is dirty or not.
966                  */
967                 if (! XFS_DQ_IS_DIRTY(dqp)) {
968                         xfs_dqunlock(dqp);
969                         continue;
970                 }
971
972                 /* XXX a sentinel would be better */
973                 recl = q->qi_dqreclaims;
974                 if (!xfs_dqflock_nowait(dqp)) {
975                         if (flags & SYNC_TRYLOCK) {
976                                 xfs_dqunlock(dqp);
977                                 continue;
978                         }
979                         /*
980                          * If we can't grab the flush lock then if the caller
981                          * really wanted us to give this our best shot, so
982                          * see if we can give a push to the buffer before we wait
983                          * on the flush lock. At this point, we know that
984                          * even though the dquot is being flushed,
985                          * it has (new) dirty data.
986                          */
987                         xfs_qm_dqflock_pushbuf_wait(dqp);
988                 }
989                 /*
990                  * Let go of the mplist lock. We don't want to hold it
991                  * across a disk write
992                  */
993                 mutex_unlock(&q->qi_dqlist_lock);
994                 error = xfs_qm_dqflush(dqp, flags);
995                 xfs_dqunlock(dqp);
996                 if (error && XFS_FORCED_SHUTDOWN(mp))
997                         return 0;       /* Need to prevent umount failure */
998                 else if (error)
999                         return error;
1000
1001                 mutex_lock(&q->qi_dqlist_lock);
1002                 if (recl != q->qi_dqreclaims) {
1003                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1004                                 break;
1005
1006                         mutex_unlock(&q->qi_dqlist_lock);
1007                         goto again;
1008                 }
1009         }
1010
1011         mutex_unlock(&q->qi_dqlist_lock);
1012         return 0;
1013 }
1014
1015 /*
1016  * The hash chains and the mplist use the same xfs_dqhash structure as
1017  * their list head, but we can take the mplist qh_lock and one of the
1018  * hash qh_locks at the same time without any problem as they aren't
1019  * related.
1020  */
1021 static struct lock_class_key xfs_quota_mplist_class;
1022
1023 /*
1024  * This initializes all the quota information that's kept in the
1025  * mount structure
1026  */
1027 STATIC int
1028 xfs_qm_init_quotainfo(
1029         xfs_mount_t     *mp)
1030 {
1031         xfs_quotainfo_t *qinf;
1032         int             error;
1033         xfs_dquot_t     *dqp;
1034
1035         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1036
1037         /*
1038          * Tell XQM that we exist as soon as possible.
1039          */
1040         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1041                 return error;
1042         }
1043
1044         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1045
1046         /*
1047          * See if quotainodes are setup, and if not, allocate them,
1048          * and change the superblock accordingly.
1049          */
1050         if ((error = xfs_qm_init_quotainos(mp))) {
1051                 kmem_free(qinf);
1052                 mp->m_quotainfo = NULL;
1053                 return error;
1054         }
1055
1056         INIT_LIST_HEAD(&qinf->qi_dqlist);
1057         mutex_init(&qinf->qi_dqlist_lock);
1058         lockdep_set_class(&qinf->qi_dqlist_lock, &xfs_quota_mplist_class);
1059
1060         qinf->qi_dqreclaims = 0;
1061
1062         /* mutex used to serialize quotaoffs */
1063         mutex_init(&qinf->qi_quotaofflock);
1064
1065         /* Precalc some constants */
1066         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1067         ASSERT(qinf->qi_dqchunklen);
1068         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1069         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1070
1071         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1072
1073         /*
1074          * We try to get the limits from the superuser's limits fields.
1075          * This is quite hacky, but it is standard quota practice.
1076          * We look at the USR dquot with id == 0 first, but if user quotas
1077          * are not enabled we goto the GRP dquot with id == 0.
1078          * We don't really care to keep separate default limits for user
1079          * and group quotas, at least not at this point.
1080          */
1081         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1082                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1083                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1084                                 XFS_DQ_PROJ),
1085                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1086                              &dqp);
1087         if (! error) {
1088                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1089
1090                 /*
1091                  * The warnings and timers set the grace period given to
1092                  * a user or group before he or she can not perform any
1093                  * more writing. If it is zero, a default is used.
1094                  */
1095                 qinf->qi_btimelimit = ddqp->d_btimer ?
1096                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1097                 qinf->qi_itimelimit = ddqp->d_itimer ?
1098                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1099                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1100                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1101                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1102                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1103                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1104                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1105                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1106                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1107                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1108                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1109                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1110                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1111                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1112                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1113  
1114                 /*
1115                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1116                  * we don't want this dquot cached. We haven't done a
1117                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1118                  */
1119                 xfs_qm_dqdestroy(dqp);
1120         } else {
1121                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1122                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1123                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1124                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1125                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1126                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1127         }
1128
1129         return 0;
1130 }
1131
1132
1133 /*
1134  * Gets called when unmounting a filesystem or when all quotas get
1135  * turned off.
1136  * This purges the quota inodes, destroys locks and frees itself.
1137  */
1138 void
1139 xfs_qm_destroy_quotainfo(
1140         xfs_mount_t     *mp)
1141 {
1142         xfs_quotainfo_t *qi;
1143
1144         qi = mp->m_quotainfo;
1145         ASSERT(qi != NULL);
1146         ASSERT(xfs_Gqm != NULL);
1147
1148         /*
1149          * Release the reference that XQM kept, so that we know
1150          * when the XQM structure should be freed. We cannot assume
1151          * that xfs_Gqm is non-null after this point.
1152          */
1153         xfs_qm_rele_quotafs_ref(mp);
1154
1155         ASSERT(list_empty(&qi->qi_dqlist));
1156         mutex_destroy(&qi->qi_dqlist_lock);
1157
1158         if (qi->qi_uquotaip) {
1159                 IRELE(qi->qi_uquotaip);
1160                 qi->qi_uquotaip = NULL; /* paranoia */
1161         }
1162         if (qi->qi_gquotaip) {
1163                 IRELE(qi->qi_gquotaip);
1164                 qi->qi_gquotaip = NULL;
1165         }
1166         mutex_destroy(&qi->qi_quotaofflock);
1167         kmem_free(qi);
1168         mp->m_quotainfo = NULL;
1169 }
1170
1171
1172
1173 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1174
1175 /* ARGSUSED */
1176 STATIC void
1177 xfs_qm_list_init(
1178         xfs_dqlist_t    *list,
1179         char            *str,
1180         int             n)
1181 {
1182         mutex_init(&list->qh_lock);
1183         INIT_LIST_HEAD(&list->qh_list);
1184         list->qh_version = 0;
1185         list->qh_nelems = 0;
1186 }
1187
1188 STATIC void
1189 xfs_qm_list_destroy(
1190         xfs_dqlist_t    *list)
1191 {
1192         mutex_destroy(&(list->qh_lock));
1193 }
1194
1195 /*
1196  * Create an inode and return with a reference already taken, but unlocked
1197  * This is how we create quota inodes
1198  */
1199 STATIC int
1200 xfs_qm_qino_alloc(
1201         xfs_mount_t     *mp,
1202         xfs_inode_t     **ip,
1203         __int64_t       sbfields,
1204         uint            flags)
1205 {
1206         xfs_trans_t     *tp;
1207         int             error;
1208         int             committed;
1209
1210         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1211         if ((error = xfs_trans_reserve(tp,
1212                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1213                                       XFS_CREATE_LOG_RES(mp), 0,
1214                                       XFS_TRANS_PERM_LOG_RES,
1215                                       XFS_CREATE_LOG_COUNT))) {
1216                 xfs_trans_cancel(tp, 0);
1217                 return error;
1218         }
1219
1220         error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0, 0, 1, ip, &committed);
1221         if (error) {
1222                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1223                                  XFS_TRANS_ABORT);
1224                 return error;
1225         }
1226
1227         /*
1228          * Make the changes in the superblock, and log those too.
1229          * sbfields arg may contain fields other than *QUOTINO;
1230          * VERSIONNUM for example.
1231          */
1232         spin_lock(&mp->m_sb_lock);
1233         if (flags & XFS_QMOPT_SBVERSION) {
1234                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1235                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1236                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1237                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1238                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1239
1240                 xfs_sb_version_addquota(&mp->m_sb);
1241                 mp->m_sb.sb_uquotino = NULLFSINO;
1242                 mp->m_sb.sb_gquotino = NULLFSINO;
1243
1244                 /* qflags will get updated _after_ quotacheck */
1245                 mp->m_sb.sb_qflags = 0;
1246         }
1247         if (flags & XFS_QMOPT_UQUOTA)
1248                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1249         else
1250                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1251         spin_unlock(&mp->m_sb_lock);
1252         xfs_mod_sb(tp, sbfields);
1253
1254         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1255                 xfs_alert(mp, "%s failed (error %d)!", __func__, error);
1256                 return error;
1257         }
1258         return 0;
1259 }
1260
1261
1262 STATIC void
1263 xfs_qm_reset_dqcounts(
1264         xfs_mount_t     *mp,
1265         xfs_buf_t       *bp,
1266         xfs_dqid_t      id,
1267         uint            type)
1268 {
1269         xfs_disk_dquot_t        *ddq;
1270         int                     j;
1271
1272         trace_xfs_reset_dqcounts(bp, _RET_IP_);
1273
1274         /*
1275          * Reset all counters and timers. They'll be
1276          * started afresh by xfs_qm_quotacheck.
1277          */
1278 #ifdef DEBUG
1279         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1280         do_div(j, sizeof(xfs_dqblk_t));
1281         ASSERT(mp->m_quotainfo->qi_dqperchunk == j);
1282 #endif
1283         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1284         for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) {
1285                 /*
1286                  * Do a sanity check, and if needed, repair the dqblk. Don't
1287                  * output any warnings because it's perfectly possible to
1288                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1289                  */
1290                 (void) xfs_qm_dqcheck(mp, ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1291                                       "xfs_quotacheck");
1292                 ddq->d_bcount = 0;
1293                 ddq->d_icount = 0;
1294                 ddq->d_rtbcount = 0;
1295                 ddq->d_btimer = 0;
1296                 ddq->d_itimer = 0;
1297                 ddq->d_rtbtimer = 0;
1298                 ddq->d_bwarns = 0;
1299                 ddq->d_iwarns = 0;
1300                 ddq->d_rtbwarns = 0;
1301                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1302         }
1303 }
1304
1305 STATIC int
1306 xfs_qm_dqiter_bufs(
1307         xfs_mount_t     *mp,
1308         xfs_dqid_t      firstid,
1309         xfs_fsblock_t   bno,
1310         xfs_filblks_t   blkcnt,
1311         uint            flags)
1312 {
1313         xfs_buf_t       *bp;
1314         int             error;
1315         int             type;
1316
1317         ASSERT(blkcnt > 0);
1318         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1319                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1320         error = 0;
1321
1322         /*
1323          * Blkcnt arg can be a very big number, and might even be
1324          * larger than the log itself. So, we have to break it up into
1325          * manageable-sized transactions.
1326          * Note that we don't start a permanent transaction here; we might
1327          * not be able to get a log reservation for the whole thing up front,
1328          * and we don't really care to either, because we just discard
1329          * everything if we were to crash in the middle of this loop.
1330          */
1331         while (blkcnt--) {
1332                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1333                               XFS_FSB_TO_DADDR(mp, bno),
1334                               mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1335                 if (error)
1336                         break;
1337
1338                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1339                 xfs_bdwrite(mp, bp);
1340                 /*
1341                  * goto the next block.
1342                  */
1343                 bno++;
1344                 firstid += mp->m_quotainfo->qi_dqperchunk;
1345         }
1346         return error;
1347 }
1348
1349 /*
1350  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1351  * caller supplied function for every chunk of dquots that we find.
1352  */
1353 STATIC int
1354 xfs_qm_dqiterate(
1355         xfs_mount_t     *mp,
1356         xfs_inode_t     *qip,
1357         uint            flags)
1358 {
1359         xfs_bmbt_irec_t         *map;
1360         int                     i, nmaps;       /* number of map entries */
1361         int                     error;          /* return value */
1362         xfs_fileoff_t           lblkno;
1363         xfs_filblks_t           maxlblkcnt;
1364         xfs_dqid_t              firstid;
1365         xfs_fsblock_t           rablkno;
1366         xfs_filblks_t           rablkcnt;
1367
1368         error = 0;
1369         /*
1370          * This looks racy, but we can't keep an inode lock across a
1371          * trans_reserve. But, this gets called during quotacheck, and that
1372          * happens only at mount time which is single threaded.
1373          */
1374         if (qip->i_d.di_nblocks == 0)
1375                 return 0;
1376
1377         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1378
1379         lblkno = 0;
1380         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1381         do {
1382                 nmaps = XFS_DQITER_MAP_SIZE;
1383                 /*
1384                  * We aren't changing the inode itself. Just changing
1385                  * some of its data. No new blocks are added here, and
1386                  * the inode is never added to the transaction.
1387                  */
1388                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1389                 error = xfs_bmapi(NULL, qip, lblkno,
1390                                   maxlblkcnt - lblkno,
1391                                   XFS_BMAPI_METADATA,
1392                                   NULL,
1393                                   0, map, &nmaps, NULL);
1394                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1395                 if (error)
1396                         break;
1397
1398                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1399                 for (i = 0; i < nmaps; i++) {
1400                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1401                         ASSERT(map[i].br_blockcount);
1402
1403
1404                         lblkno += map[i].br_blockcount;
1405
1406                         if (map[i].br_startblock == HOLESTARTBLOCK)
1407                                 continue;
1408
1409                         firstid = (xfs_dqid_t) map[i].br_startoff *
1410                                 mp->m_quotainfo->qi_dqperchunk;
1411                         /*
1412                          * Do a read-ahead on the next extent.
1413                          */
1414                         if ((i+1 < nmaps) &&
1415                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1416                                 rablkcnt =  map[i+1].br_blockcount;
1417                                 rablkno = map[i+1].br_startblock;
1418                                 while (rablkcnt--) {
1419                                         xfs_buf_readahead(mp->m_ddev_targp,
1420                                                XFS_FSB_TO_DADDR(mp, rablkno),
1421                                                mp->m_quotainfo->qi_dqchunklen);
1422                                         rablkno++;
1423                                 }
1424                         }
1425                         /*
1426                          * Iterate thru all the blks in the extent and
1427                          * reset the counters of all the dquots inside them.
1428                          */
1429                         if ((error = xfs_qm_dqiter_bufs(mp,
1430                                                        firstid,
1431                                                        map[i].br_startblock,
1432                                                        map[i].br_blockcount,
1433                                                        flags))) {
1434                                 break;
1435                         }
1436                 }
1437
1438                 if (error)
1439                         break;
1440         } while (nmaps > 0);
1441
1442         kmem_free(map);
1443
1444         return error;
1445 }
1446
1447 /*
1448  * Called by dqusage_adjust in doing a quotacheck.
1449  *
1450  * Given the inode, and a dquot id this updates both the incore dqout as well
1451  * as the buffer copy. This is so that once the quotacheck is done, we can
1452  * just log all the buffers, as opposed to logging numerous updates to
1453  * individual dquots.
1454  */
1455 STATIC int
1456 xfs_qm_quotacheck_dqadjust(
1457         struct xfs_inode        *ip,
1458         xfs_dqid_t              id,
1459         uint                    type,
1460         xfs_qcnt_t              nblks,
1461         xfs_qcnt_t              rtblks)
1462 {
1463         struct xfs_mount        *mp = ip->i_mount;
1464         struct xfs_dquot        *dqp;
1465         int                     error;
1466
1467         error = xfs_qm_dqget(mp, ip, id, type,
1468                              XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN, &dqp);
1469         if (error) {
1470                 /*
1471                  * Shouldn't be able to turn off quotas here.
1472                  */
1473                 ASSERT(error != ESRCH);
1474                 ASSERT(error != ENOENT);
1475                 return error;
1476         }
1477
1478         trace_xfs_dqadjust(dqp);
1479
1480         /*
1481          * Adjust the inode count and the block count to reflect this inode's
1482          * resource usage.
1483          */
1484         be64_add_cpu(&dqp->q_core.d_icount, 1);
1485         dqp->q_res_icount++;
1486         if (nblks) {
1487                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1488                 dqp->q_res_bcount += nblks;
1489         }
1490         if (rtblks) {
1491                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1492                 dqp->q_res_rtbcount += rtblks;
1493         }
1494
1495         /*
1496          * Set default limits, adjust timers (since we changed usages)
1497          *
1498          * There are no timers for the default values set in the root dquot.
1499          */
1500         if (dqp->q_core.d_id) {
1501                 xfs_qm_adjust_dqlimits(mp, &dqp->q_core);
1502                 xfs_qm_adjust_dqtimers(mp, &dqp->q_core);
1503         }
1504
1505         dqp->dq_flags |= XFS_DQ_DIRTY;
1506         xfs_qm_dqput(dqp);
1507         return 0;
1508 }
1509
1510 STATIC int
1511 xfs_qm_get_rtblks(
1512         xfs_inode_t     *ip,
1513         xfs_qcnt_t      *O_rtblks)
1514 {
1515         xfs_filblks_t   rtblks;                 /* total rt blks */
1516         xfs_extnum_t    idx;                    /* extent record index */
1517         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1518         xfs_extnum_t    nextents;               /* number of extent entries */
1519         int             error;
1520
1521         ASSERT(XFS_IS_REALTIME_INODE(ip));
1522         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1523         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1524                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1525                         return error;
1526         }
1527         rtblks = 0;
1528         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1529         for (idx = 0; idx < nextents; idx++)
1530                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1531         *O_rtblks = (xfs_qcnt_t)rtblks;
1532         return 0;
1533 }
1534
1535 /*
1536  * callback routine supplied to bulkstat(). Given an inumber, find its
1537  * dquots and update them to account for resources taken by that inode.
1538  */
1539 /* ARGSUSED */
1540 STATIC int
1541 xfs_qm_dqusage_adjust(
1542         xfs_mount_t     *mp,            /* mount point for filesystem */
1543         xfs_ino_t       ino,            /* inode number to get data for */
1544         void            __user *buffer, /* not used */
1545         int             ubsize,         /* not used */
1546         int             *ubused,        /* not used */
1547         int             *res)           /* result code value */
1548 {
1549         xfs_inode_t     *ip;
1550         xfs_qcnt_t      nblks, rtblks = 0;
1551         int             error;
1552
1553         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1554
1555         /*
1556          * rootino must have its resources accounted for, not so with the quota
1557          * inodes.
1558          */
1559         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1560                 *res = BULKSTAT_RV_NOTHING;
1561                 return XFS_ERROR(EINVAL);
1562         }
1563
1564         /*
1565          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1566          * interface expects the inode to be exclusively locked because that's
1567          * the case in all other instances. It's OK that we do this because
1568          * quotacheck is done only at mount time.
1569          */
1570         error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip);
1571         if (error) {
1572                 *res = BULKSTAT_RV_NOTHING;
1573                 return error;
1574         }
1575
1576         ASSERT(ip->i_delayed_blks == 0);
1577
1578         if (XFS_IS_REALTIME_INODE(ip)) {
1579                 /*
1580                  * Walk thru the extent list and count the realtime blocks.
1581                  */
1582                 error = xfs_qm_get_rtblks(ip, &rtblks);
1583                 if (error)
1584                         goto error0;
1585         }
1586
1587         nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1588
1589         /*
1590          * Add the (disk blocks and inode) resources occupied by this
1591          * inode to its dquots. We do this adjustment in the incore dquot,
1592          * and also copy the changes to its buffer.
1593          * We don't care about putting these changes in a transaction
1594          * envelope because if we crash in the middle of a 'quotacheck'
1595          * we have to start from the beginning anyway.
1596          * Once we're done, we'll log all the dquot bufs.
1597          *
1598          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1599          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1600          */
1601         if (XFS_IS_UQUOTA_ON(mp)) {
1602                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_uid,
1603                                                    XFS_DQ_USER, nblks, rtblks);
1604                 if (error)
1605                         goto error0;
1606         }
1607
1608         if (XFS_IS_GQUOTA_ON(mp)) {
1609                 error = xfs_qm_quotacheck_dqadjust(ip, ip->i_d.di_gid,
1610                                                    XFS_DQ_GROUP, nblks, rtblks);
1611                 if (error)
1612                         goto error0;
1613         }
1614
1615         if (XFS_IS_PQUOTA_ON(mp)) {
1616                 error = xfs_qm_quotacheck_dqadjust(ip, xfs_get_projid(ip),
1617                                                    XFS_DQ_PROJ, nblks, rtblks);
1618                 if (error)
1619                         goto error0;
1620         }
1621
1622         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1623         IRELE(ip);
1624         *res = BULKSTAT_RV_DIDONE;
1625         return 0;
1626
1627 error0:
1628         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1629         IRELE(ip);
1630         *res = BULKSTAT_RV_GIVEUP;
1631         return error;
1632 }
1633
1634 /*
1635  * Walk thru all the filesystem inodes and construct a consistent view
1636  * of the disk quota world. If the quotacheck fails, disable quotas.
1637  */
1638 int
1639 xfs_qm_quotacheck(
1640         xfs_mount_t     *mp)
1641 {
1642         int             done, count, error;
1643         xfs_ino_t       lastino;
1644         size_t          structsz;
1645         xfs_inode_t     *uip, *gip;
1646         uint            flags;
1647
1648         count = INT_MAX;
1649         structsz = 1;
1650         lastino = 0;
1651         flags = 0;
1652
1653         ASSERT(mp->m_quotainfo->qi_uquotaip || mp->m_quotainfo->qi_gquotaip);
1654         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1655
1656         /*
1657          * There should be no cached dquots. The (simplistic) quotacheck
1658          * algorithm doesn't like that.
1659          */
1660         ASSERT(list_empty(&mp->m_quotainfo->qi_dqlist));
1661
1662         xfs_notice(mp, "Quotacheck needed: Please wait.");
1663
1664         /*
1665          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1666          * their counters to zero. We need a clean slate.
1667          * We don't log our changes till later.
1668          */
1669         uip = mp->m_quotainfo->qi_uquotaip;
1670         if (uip) {
1671                 error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA);
1672                 if (error)
1673                         goto error_return;
1674                 flags |= XFS_UQUOTA_CHKD;
1675         }
1676
1677         gip = mp->m_quotainfo->qi_gquotaip;
1678         if (gip) {
1679                 error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1680                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1681                 if (error)
1682                         goto error_return;
1683                 flags |= XFS_OQUOTA_CHKD;
1684         }
1685
1686         do {
1687                 /*
1688                  * Iterate thru all the inodes in the file system,
1689                  * adjusting the corresponding dquot counters in core.
1690                  */
1691                 error = xfs_bulkstat(mp, &lastino, &count,
1692                                      xfs_qm_dqusage_adjust,
1693                                      structsz, NULL, &done);
1694                 if (error)
1695                         break;
1696
1697         } while (!done);
1698
1699         /*
1700          * We've made all the changes that we need to make incore.
1701          * Flush them down to disk buffers if everything was updated
1702          * successfully.
1703          */
1704         if (!error)
1705                 error = xfs_qm_dqflush_all(mp, 0);
1706
1707         /*
1708          * We can get this error if we couldn't do a dquot allocation inside
1709          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1710          * dirty dquots that might be cached, we just want to get rid of them
1711          * and turn quotaoff. The dquots won't be attached to any of the inodes
1712          * at this point (because we intentionally didn't in dqget_noattach).
1713          */
1714         if (error) {
1715                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL);
1716                 goto error_return;
1717         }
1718
1719         /*
1720          * We didn't log anything, because if we crashed, we'll have to
1721          * start the quotacheck from scratch anyway. However, we must make
1722          * sure that our dquot changes are secure before we put the
1723          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1724          * flush.
1725          */
1726         XFS_bflush(mp->m_ddev_targp);
1727
1728         /*
1729          * If one type of quotas is off, then it will lose its
1730          * quotachecked status, since we won't be doing accounting for
1731          * that type anymore.
1732          */
1733         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1734         mp->m_qflags |= flags;
1735
1736         xfs_qm_dquot_list_print(mp);
1737
1738  error_return:
1739         if (error) {
1740                 xfs_warn(mp,
1741         "Quotacheck: Unsuccessful (Error %d): Disabling quotas.",
1742                         error);
1743                 /*
1744                  * We must turn off quotas.
1745                  */
1746                 ASSERT(mp->m_quotainfo != NULL);
1747                 ASSERT(xfs_Gqm != NULL);
1748                 xfs_qm_destroy_quotainfo(mp);
1749                 if (xfs_mount_reset_sbqflags(mp)) {
1750                         xfs_warn(mp,
1751                                 "Quotacheck: Failed to reset quota flags.");
1752                 }
1753         } else
1754                 xfs_notice(mp, "Quotacheck: Done.");
1755         return (error);
1756 }
1757
1758 /*
1759  * This is called after the superblock has been read in and we're ready to
1760  * iget the quota inodes.
1761  */
1762 STATIC int
1763 xfs_qm_init_quotainos(
1764         xfs_mount_t     *mp)
1765 {
1766         xfs_inode_t     *uip, *gip;
1767         int             error;
1768         __int64_t       sbflags;
1769         uint            flags;
1770
1771         ASSERT(mp->m_quotainfo);
1772         uip = gip = NULL;
1773         sbflags = 0;
1774         flags = 0;
1775
1776         /*
1777          * Get the uquota and gquota inodes
1778          */
1779         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1780                 if (XFS_IS_UQUOTA_ON(mp) &&
1781                     mp->m_sb.sb_uquotino != NULLFSINO) {
1782                         ASSERT(mp->m_sb.sb_uquotino > 0);
1783                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1784                                              0, 0, &uip)))
1785                                 return XFS_ERROR(error);
1786                 }
1787                 if (XFS_IS_OQUOTA_ON(mp) &&
1788                     mp->m_sb.sb_gquotino != NULLFSINO) {
1789                         ASSERT(mp->m_sb.sb_gquotino > 0);
1790                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1791                                              0, 0, &gip))) {
1792                                 if (uip)
1793                                         IRELE(uip);
1794                                 return XFS_ERROR(error);
1795                         }
1796                 }
1797         } else {
1798                 flags |= XFS_QMOPT_SBVERSION;
1799                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1800                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1801         }
1802
1803         /*
1804          * Create the two inodes, if they don't exist already. The changes
1805          * made above will get added to a transaction and logged in one of
1806          * the qino_alloc calls below.  If the device is readonly,
1807          * temporarily switch to read-write to do this.
1808          */
1809         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1810                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1811                                               sbflags | XFS_SB_UQUOTINO,
1812                                               flags | XFS_QMOPT_UQUOTA)))
1813                         return XFS_ERROR(error);
1814
1815                 flags &= ~XFS_QMOPT_SBVERSION;
1816         }
1817         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1818                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1819                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1820                 error = xfs_qm_qino_alloc(mp, &gip,
1821                                           sbflags | XFS_SB_GQUOTINO, flags);
1822                 if (error) {
1823                         if (uip)
1824                                 IRELE(uip);
1825
1826                         return XFS_ERROR(error);
1827                 }
1828         }
1829
1830         mp->m_quotainfo->qi_uquotaip = uip;
1831         mp->m_quotainfo->qi_gquotaip = gip;
1832
1833         return 0;
1834 }
1835
1836
1837
1838 /*
1839  * Just pop the least recently used dquot off the freelist and
1840  * recycle it. The returned dquot is locked.
1841  */
1842 STATIC xfs_dquot_t *
1843 xfs_qm_dqreclaim_one(void)
1844 {
1845         xfs_dquot_t     *dqpout;
1846         xfs_dquot_t     *dqp;
1847         int             restarts;
1848         int             startagain;
1849
1850         restarts = 0;
1851         dqpout = NULL;
1852
1853         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
1854 again:
1855         startagain = 0;
1856         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1857
1858         list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) {
1859                 struct xfs_mount *mp = dqp->q_mount;
1860                 xfs_dqlock(dqp);
1861
1862                 /*
1863                  * We are racing with dqlookup here. Naturally we don't
1864                  * want to reclaim a dquot that lookup wants. We release the
1865                  * freelist lock and start over, so that lookup will grab
1866                  * both the dquot and the freelistlock.
1867                  */
1868                 if (dqp->dq_flags & XFS_DQ_WANT) {
1869                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
1870
1871                         trace_xfs_dqreclaim_want(dqp);
1872                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
1873                         restarts++;
1874                         startagain = 1;
1875                         goto dqunlock;
1876                 }
1877
1878                 /*
1879                  * If the dquot is inactive, we are assured that it is
1880                  * not on the mplist or the hashlist, and that makes our
1881                  * life easier.
1882                  */
1883                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
1884                         ASSERT(mp == NULL);
1885                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
1886                         ASSERT(list_empty(&dqp->q_hashlist));
1887                         ASSERT(list_empty(&dqp->q_mplist));
1888                         list_del_init(&dqp->q_freelist);
1889                         xfs_Gqm->qm_dqfrlist_cnt--;
1890                         dqpout = dqp;
1891                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
1892                         goto dqunlock;
1893                 }
1894
1895                 ASSERT(dqp->q_hash);
1896                 ASSERT(!list_empty(&dqp->q_mplist));
1897
1898                 /*
1899                  * Try to grab the flush lock. If this dquot is in the process
1900                  * of getting flushed to disk, we don't want to reclaim it.
1901                  */
1902                 if (!xfs_dqflock_nowait(dqp))
1903                         goto dqunlock;
1904
1905                 /*
1906                  * We have the flush lock so we know that this is not in the
1907                  * process of being flushed. So, if this is dirty, flush it
1908                  * DELWRI so that we don't get a freelist infested with
1909                  * dirty dquots.
1910                  */
1911                 if (XFS_DQ_IS_DIRTY(dqp)) {
1912                         int     error;
1913
1914                         trace_xfs_dqreclaim_dirty(dqp);
1915
1916                         /*
1917                          * We flush it delayed write, so don't bother
1918                          * releasing the freelist lock.
1919                          */
1920                         error = xfs_qm_dqflush(dqp, 0);
1921                         if (error) {
1922                                 xfs_warn(mp, "%s: dquot %p flush failed",
1923                                         __func__, dqp);
1924                         }
1925                         goto dqunlock;
1926                 }
1927
1928                 /*
1929                  * We're trying to get the hashlock out of order. This races
1930                  * with dqlookup; so, we giveup and goto the next dquot if
1931                  * we couldn't get the hashlock. This way, we won't starve
1932                  * a dqlookup process that holds the hashlock that is
1933                  * waiting for the freelist lock.
1934                  */
1935                 if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
1936                         restarts++;
1937                         goto dqfunlock;
1938                 }
1939
1940                 /*
1941                  * This races with dquot allocation code as well as dqflush_all
1942                  * and reclaim code. So, if we failed to grab the mplist lock,
1943                  * giveup everything and start over.
1944                  */
1945                 if (!mutex_trylock(&mp->m_quotainfo->qi_dqlist_lock)) {
1946                         restarts++;
1947                         startagain = 1;
1948                         goto qhunlock;
1949                 }
1950
1951                 ASSERT(dqp->q_nrefs == 0);
1952                 list_del_init(&dqp->q_mplist);
1953                 mp->m_quotainfo->qi_dquots--;
1954                 mp->m_quotainfo->qi_dqreclaims++;
1955                 list_del_init(&dqp->q_hashlist);
1956                 dqp->q_hash->qh_version++;
1957                 list_del_init(&dqp->q_freelist);
1958                 xfs_Gqm->qm_dqfrlist_cnt--;
1959                 dqpout = dqp;
1960                 mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1961 qhunlock:
1962                 mutex_unlock(&dqp->q_hash->qh_lock);
1963 dqfunlock:
1964                 xfs_dqfunlock(dqp);
1965 dqunlock:
1966                 xfs_dqunlock(dqp);
1967                 if (dqpout)
1968                         break;
1969                 if (restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
1970                         break;
1971                 if (startagain) {
1972                         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1973                         goto again;
1974                 }
1975         }
1976         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1977         return dqpout;
1978 }
1979
1980 /*
1981  * Traverse the freelist of dquots and attempt to reclaim a maximum of
1982  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
1983  * favor the lookup function ...
1984  */
1985 STATIC int
1986 xfs_qm_shake_freelist(
1987         int     howmany)
1988 {
1989         int             nreclaimed = 0;
1990         xfs_dquot_t     *dqp;
1991
1992         if (howmany <= 0)
1993                 return 0;
1994
1995         while (nreclaimed < howmany) {
1996                 dqp = xfs_qm_dqreclaim_one();
1997                 if (!dqp)
1998                         return nreclaimed;
1999                 xfs_qm_dqdestroy(dqp);
2000                 nreclaimed++;
2001         }
2002         return nreclaimed;
2003 }
2004
2005 /*
2006  * The kmem_shake interface is invoked when memory is running low.
2007  */
2008 /* ARGSUSED */
2009 STATIC int
2010 xfs_qm_shake(
2011         struct shrinker *shrink,
2012         int             nr_to_scan,
2013         gfp_t           gfp_mask)
2014 {
2015         int     ndqused, nfree, n;
2016
2017         if (!kmem_shake_allow(gfp_mask))
2018                 return 0;
2019         if (!xfs_Gqm)
2020                 return 0;
2021
2022         nfree = xfs_Gqm->qm_dqfrlist_cnt; /* free dquots */
2023         /* incore dquots in all f/s's */
2024         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2025
2026         ASSERT(ndqused >= 0);
2027
2028         if (nfree <= ndqused && nfree < ndquot)
2029                 return 0;
2030
2031         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2032         n = nfree - ndqused - ndquot;           /* # over target */
2033
2034         return xfs_qm_shake_freelist(MAX(nfree, n));
2035 }
2036
2037
2038 /*------------------------------------------------------------------*/
2039
2040 /*
2041  * Return a new incore dquot. Depending on the number of
2042  * dquots in the system, we either allocate a new one on the kernel heap,
2043  * or reclaim a free one.
2044  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2045  * to reclaim an existing one from the freelist.
2046  */
2047 boolean_t
2048 xfs_qm_dqalloc_incore(
2049         xfs_dquot_t **O_dqpp)
2050 {
2051         xfs_dquot_t     *dqp;
2052
2053         /*
2054          * Check against high water mark to see if we want to pop
2055          * a nincompoop dquot off the freelist.
2056          */
2057         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2058                 /*
2059                  * Try to recycle a dquot from the freelist.
2060                  */
2061                 if ((dqp = xfs_qm_dqreclaim_one())) {
2062                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2063                         /*
2064                          * Just zero the core here. The rest will get
2065                          * reinitialized by caller. XXX we shouldn't even
2066                          * do this zero ...
2067                          */
2068                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2069                         *O_dqpp = dqp;
2070                         return B_FALSE;
2071                 }
2072                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2073         }
2074
2075         /*
2076          * Allocate a brand new dquot on the kernel heap and return it
2077          * to the caller to initialize.
2078          */
2079         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2080         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2081         atomic_inc(&xfs_Gqm->qm_totaldquots);
2082
2083         return B_TRUE;
2084 }
2085
2086
2087 /*
2088  * Start a transaction and write the incore superblock changes to
2089  * disk. flags parameter indicates which fields have changed.
2090  */
2091 int
2092 xfs_qm_write_sb_changes(
2093         xfs_mount_t     *mp,
2094         __int64_t       flags)
2095 {
2096         xfs_trans_t     *tp;
2097         int             error;
2098
2099 #ifdef QUOTADEBUG
2100         xfs_notice(mp, "Writing superblock quota changes");
2101 #endif
2102         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2103         if ((error = xfs_trans_reserve(tp, 0,
2104                                       mp->m_sb.sb_sectsize + 128, 0,
2105                                       0,
2106                                       XFS_DEFAULT_LOG_COUNT))) {
2107                 xfs_trans_cancel(tp, 0);
2108                 return error;
2109         }
2110
2111         xfs_mod_sb(tp, flags);
2112         error = xfs_trans_commit(tp, 0);
2113
2114         return error;
2115 }
2116
2117
2118 /* --------------- utility functions for vnodeops ---------------- */
2119
2120
2121 /*
2122  * Given an inode, a uid, gid and prid make sure that we have
2123  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2124  * quotas by creating this file.
2125  * This also attaches dquot(s) to the given inode after locking it,
2126  * and returns the dquots corresponding to the uid and/or gid.
2127  *
2128  * in   : inode (unlocked)
2129  * out  : udquot, gdquot with references taken and unlocked
2130  */
2131 int
2132 xfs_qm_vop_dqalloc(
2133         struct xfs_inode        *ip,
2134         uid_t                   uid,
2135         gid_t                   gid,
2136         prid_t                  prid,
2137         uint                    flags,
2138         struct xfs_dquot        **O_udqpp,
2139         struct xfs_dquot        **O_gdqpp)
2140 {
2141         struct xfs_mount        *mp = ip->i_mount;
2142         struct xfs_dquot        *uq, *gq;
2143         int                     error;
2144         uint                    lockflags;
2145
2146         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2147                 return 0;
2148
2149         lockflags = XFS_ILOCK_EXCL;
2150         xfs_ilock(ip, lockflags);
2151
2152         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2153                 gid = ip->i_d.di_gid;
2154
2155         /*
2156          * Attach the dquot(s) to this inode, doing a dquot allocation
2157          * if necessary. The dquot(s) will not be locked.
2158          */
2159         if (XFS_NOT_DQATTACHED(mp, ip)) {
2160                 error = xfs_qm_dqattach_locked(ip, XFS_QMOPT_DQALLOC);
2161                 if (error) {
2162                         xfs_iunlock(ip, lockflags);
2163                         return error;
2164                 }
2165         }
2166
2167         uq = gq = NULL;
2168         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2169                 if (ip->i_d.di_uid != uid) {
2170                         /*
2171                          * What we need is the dquot that has this uid, and
2172                          * if we send the inode to dqget, the uid of the inode
2173                          * takes priority over what's sent in the uid argument.
2174                          * We must unlock inode here before calling dqget if
2175                          * we're not sending the inode, because otherwise
2176                          * we'll deadlock by doing trans_reserve while
2177                          * holding ilock.
2178                          */
2179                         xfs_iunlock(ip, lockflags);
2180                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2181                                                  XFS_DQ_USER,
2182                                                  XFS_QMOPT_DQALLOC |
2183                                                  XFS_QMOPT_DOWARN,
2184                                                  &uq))) {
2185                                 ASSERT(error != ENOENT);
2186                                 return error;
2187                         }
2188                         /*
2189                          * Get the ilock in the right order.
2190                          */
2191                         xfs_dqunlock(uq);
2192                         lockflags = XFS_ILOCK_SHARED;
2193                         xfs_ilock(ip, lockflags);
2194                 } else {
2195                         /*
2196                          * Take an extra reference, because we'll return
2197                          * this to caller
2198                          */
2199                         ASSERT(ip->i_udquot);
2200                         uq = ip->i_udquot;
2201                         xfs_dqlock(uq);
2202                         XFS_DQHOLD(uq);
2203                         xfs_dqunlock(uq);
2204                 }
2205         }
2206         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2207                 if (ip->i_d.di_gid != gid) {
2208                         xfs_iunlock(ip, lockflags);
2209                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2210                                                  XFS_DQ_GROUP,
2211                                                  XFS_QMOPT_DQALLOC |
2212                                                  XFS_QMOPT_DOWARN,
2213                                                  &gq))) {
2214                                 if (uq)
2215                                         xfs_qm_dqrele(uq);
2216                                 ASSERT(error != ENOENT);
2217                                 return error;
2218                         }
2219                         xfs_dqunlock(gq);
2220                         lockflags = XFS_ILOCK_SHARED;
2221                         xfs_ilock(ip, lockflags);
2222                 } else {
2223                         ASSERT(ip->i_gdquot);
2224                         gq = ip->i_gdquot;
2225                         xfs_dqlock(gq);
2226                         XFS_DQHOLD(gq);
2227                         xfs_dqunlock(gq);
2228                 }
2229         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2230                 if (xfs_get_projid(ip) != prid) {
2231                         xfs_iunlock(ip, lockflags);
2232                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2233                                                  XFS_DQ_PROJ,
2234                                                  XFS_QMOPT_DQALLOC |
2235                                                  XFS_QMOPT_DOWARN,
2236                                                  &gq))) {
2237                                 if (uq)
2238                                         xfs_qm_dqrele(uq);
2239                                 ASSERT(error != ENOENT);
2240                                 return (error);
2241                         }
2242                         xfs_dqunlock(gq);
2243                         lockflags = XFS_ILOCK_SHARED;
2244                         xfs_ilock(ip, lockflags);
2245                 } else {
2246                         ASSERT(ip->i_gdquot);
2247                         gq = ip->i_gdquot;
2248                         xfs_dqlock(gq);
2249                         XFS_DQHOLD(gq);
2250                         xfs_dqunlock(gq);
2251                 }
2252         }
2253         if (uq)
2254                 trace_xfs_dquot_dqalloc(ip);
2255
2256         xfs_iunlock(ip, lockflags);
2257         if (O_udqpp)
2258                 *O_udqpp = uq;
2259         else if (uq)
2260                 xfs_qm_dqrele(uq);
2261         if (O_gdqpp)
2262                 *O_gdqpp = gq;
2263         else if (gq)
2264                 xfs_qm_dqrele(gq);
2265         return 0;
2266 }
2267
2268 /*
2269  * Actually transfer ownership, and do dquot modifications.
2270  * These were already reserved.
2271  */
2272 xfs_dquot_t *
2273 xfs_qm_vop_chown(
2274         xfs_trans_t     *tp,
2275         xfs_inode_t     *ip,
2276         xfs_dquot_t     **IO_olddq,
2277         xfs_dquot_t     *newdq)
2278 {
2279         xfs_dquot_t     *prevdq;
2280         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2281                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2282
2283
2284         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2285         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2286
2287         /* old dquot */
2288         prevdq = *IO_olddq;
2289         ASSERT(prevdq);
2290         ASSERT(prevdq != newdq);
2291
2292         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2293         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2294
2295         /* the sparkling new dquot */
2296         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2297         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2298
2299         /*
2300          * Take an extra reference, because the inode
2301          * is going to keep this dquot pointer even
2302          * after the trans_commit.
2303          */
2304         xfs_dqlock(newdq);
2305         XFS_DQHOLD(newdq);
2306         xfs_dqunlock(newdq);
2307         *IO_olddq = newdq;
2308
2309         return prevdq;
2310 }
2311
2312 /*
2313  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2314  */
2315 int
2316 xfs_qm_vop_chown_reserve(
2317         xfs_trans_t     *tp,
2318         xfs_inode_t     *ip,
2319         xfs_dquot_t     *udqp,
2320         xfs_dquot_t     *gdqp,
2321         uint            flags)
2322 {
2323         xfs_mount_t     *mp = ip->i_mount;
2324         uint            delblks, blkflags, prjflags = 0;
2325         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2326         int             error;
2327
2328
2329         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2330         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2331
2332         delblks = ip->i_delayed_blks;
2333         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2334         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2335                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2336
2337         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2338             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2339                 delblksudq = udqp;
2340                 /*
2341                  * If there are delayed allocation blocks, then we have to
2342                  * unreserve those from the old dquot, and add them to the
2343                  * new dquot.
2344                  */
2345                 if (delblks) {
2346                         ASSERT(ip->i_udquot);
2347                         unresudq = ip->i_udquot;
2348                 }
2349         }
2350         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2351                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2352                      xfs_get_projid(ip) != be32_to_cpu(gdqp->q_core.d_id))
2353                         prjflags = XFS_QMOPT_ENOSPC;
2354
2355                 if (prjflags ||
2356                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2357                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2358                         delblksgdq = gdqp;
2359                         if (delblks) {
2360                                 ASSERT(ip->i_gdquot);
2361                                 unresgdq = ip->i_gdquot;
2362                         }
2363                 }
2364         }
2365
2366         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2367                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2368                                 flags | blkflags | prjflags)))
2369                 return (error);
2370
2371         /*
2372          * Do the delayed blks reservations/unreservations now. Since, these
2373          * are done without the help of a transaction, if a reservation fails
2374          * its previous reservations won't be automatically undone by trans
2375          * code. So, we have to do it manually here.
2376          */
2377         if (delblks) {
2378                 /*
2379                  * Do the reservations first. Unreservation can't fail.
2380                  */
2381                 ASSERT(delblksudq || delblksgdq);
2382                 ASSERT(unresudq || unresgdq);
2383                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2384                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2385                                 flags | blkflags | prjflags)))
2386                         return (error);
2387                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2388                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2389                                 blkflags);
2390         }
2391
2392         return (0);
2393 }
2394
2395 int
2396 xfs_qm_vop_rename_dqattach(
2397         struct xfs_inode        **i_tab)
2398 {
2399         struct xfs_mount        *mp = i_tab[0]->i_mount;
2400         int                     i;
2401
2402         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2403                 return 0;
2404
2405         for (i = 0; (i < 4 && i_tab[i]); i++) {
2406                 struct xfs_inode        *ip = i_tab[i];
2407                 int                     error;
2408
2409                 /*
2410                  * Watch out for duplicate entries in the table.
2411                  */
2412                 if (i == 0 || ip != i_tab[i-1]) {
2413                         if (XFS_NOT_DQATTACHED(mp, ip)) {
2414                                 error = xfs_qm_dqattach(ip, 0);
2415                                 if (error)
2416                                         return error;
2417                         }
2418                 }
2419         }
2420         return 0;
2421 }
2422
2423 void
2424 xfs_qm_vop_create_dqattach(
2425         struct xfs_trans        *tp,
2426         struct xfs_inode        *ip,
2427         struct xfs_dquot        *udqp,
2428         struct xfs_dquot        *gdqp)
2429 {
2430         struct xfs_mount        *mp = tp->t_mountp;
2431
2432         if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp))
2433                 return;
2434
2435         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2436         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2437
2438         if (udqp) {
2439                 xfs_dqlock(udqp);
2440                 XFS_DQHOLD(udqp);
2441                 xfs_dqunlock(udqp);
2442                 ASSERT(ip->i_udquot == NULL);
2443                 ip->i_udquot = udqp;
2444                 ASSERT(XFS_IS_UQUOTA_ON(mp));
2445                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2446                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2447         }
2448         if (gdqp) {
2449                 xfs_dqlock(gdqp);
2450                 XFS_DQHOLD(gdqp);
2451                 xfs_dqunlock(gdqp);
2452                 ASSERT(ip->i_gdquot == NULL);
2453                 ip->i_gdquot = gdqp;
2454                 ASSERT(XFS_IS_OQUOTA_ON(mp));
2455                 ASSERT((XFS_IS_GQUOTA_ON(mp) ?
2456                         ip->i_d.di_gid : xfs_get_projid(ip)) ==
2457                                 be32_to_cpu(gdqp->q_core.d_id));
2458                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2459         }
2460 }
2461