Merge tag 'xfs-5.20-merge-8' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
[linux-2.6-block.git] / fs / xfs / xfs_qm.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_bit.h"
13 #include "xfs_sb.h"
14 #include "xfs_mount.h"
15 #include "xfs_inode.h"
16 #include "xfs_iwalk.h"
17 #include "xfs_quota.h"
18 #include "xfs_bmap.h"
19 #include "xfs_bmap_util.h"
20 #include "xfs_trans.h"
21 #include "xfs_trans_space.h"
22 #include "xfs_qm.h"
23 #include "xfs_trace.h"
24 #include "xfs_icache.h"
25 #include "xfs_error.h"
26 #include "xfs_ag.h"
27 #include "xfs_ialloc.h"
28 #include "xfs_log_priv.h"
29
30 /*
31  * The global quota manager. There is only one of these for the entire
32  * system, _not_ one per file system. XQM keeps track of the overall
33  * quota functionality, including maintaining the freelist and hash
34  * tables of dquots.
35  */
36 STATIC int      xfs_qm_init_quotainos(struct xfs_mount *mp);
37 STATIC int      xfs_qm_init_quotainfo(struct xfs_mount *mp);
38
39 STATIC void     xfs_qm_destroy_quotainos(struct xfs_quotainfo *qi);
40 STATIC void     xfs_qm_dqfree_one(struct xfs_dquot *dqp);
41 /*
42  * We use the batch lookup interface to iterate over the dquots as it
43  * currently is the only interface into the radix tree code that allows
44  * fuzzy lookups instead of exact matches.  Holding the lock over multiple
45  * operations is fine as all callers are used either during mount/umount
46  * or quotaoff.
47  */
48 #define XFS_DQ_LOOKUP_BATCH     32
49
50 STATIC int
51 xfs_qm_dquot_walk(
52         struct xfs_mount        *mp,
53         xfs_dqtype_t            type,
54         int                     (*execute)(struct xfs_dquot *dqp, void *data),
55         void                    *data)
56 {
57         struct xfs_quotainfo    *qi = mp->m_quotainfo;
58         struct radix_tree_root  *tree = xfs_dquot_tree(qi, type);
59         uint32_t                next_index;
60         int                     last_error = 0;
61         int                     skipped;
62         int                     nr_found;
63
64 restart:
65         skipped = 0;
66         next_index = 0;
67         nr_found = 0;
68
69         while (1) {
70                 struct xfs_dquot *batch[XFS_DQ_LOOKUP_BATCH];
71                 int             error = 0;
72                 int             i;
73
74                 mutex_lock(&qi->qi_tree_lock);
75                 nr_found = radix_tree_gang_lookup(tree, (void **)batch,
76                                         next_index, XFS_DQ_LOOKUP_BATCH);
77                 if (!nr_found) {
78                         mutex_unlock(&qi->qi_tree_lock);
79                         break;
80                 }
81
82                 for (i = 0; i < nr_found; i++) {
83                         struct xfs_dquot *dqp = batch[i];
84
85                         next_index = dqp->q_id + 1;
86
87                         error = execute(batch[i], data);
88                         if (error == -EAGAIN) {
89                                 skipped++;
90                                 continue;
91                         }
92                         if (error && last_error != -EFSCORRUPTED)
93                                 last_error = error;
94                 }
95
96                 mutex_unlock(&qi->qi_tree_lock);
97
98                 /* bail out if the filesystem is corrupted.  */
99                 if (last_error == -EFSCORRUPTED) {
100                         skipped = 0;
101                         break;
102                 }
103                 /* we're done if id overflows back to zero */
104                 if (!next_index)
105                         break;
106         }
107
108         if (skipped) {
109                 delay(1);
110                 goto restart;
111         }
112
113         return last_error;
114 }
115
116
117 /*
118  * Purge a dquot from all tracking data structures and free it.
119  */
120 STATIC int
121 xfs_qm_dqpurge(
122         struct xfs_dquot        *dqp,
123         void                    *data)
124 {
125         struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
126         int                     error = -EAGAIN;
127
128         xfs_dqlock(dqp);
129         if ((dqp->q_flags & XFS_DQFLAG_FREEING) || dqp->q_nrefs != 0)
130                 goto out_unlock;
131
132         dqp->q_flags |= XFS_DQFLAG_FREEING;
133
134         xfs_dqflock(dqp);
135
136         /*
137          * If we are turning this type of quotas off, we don't care
138          * about the dirty metadata sitting in this dquot. OTOH, if
139          * we're unmounting, we do care, so we flush it and wait.
140          */
141         if (XFS_DQ_IS_DIRTY(dqp)) {
142                 struct xfs_buf  *bp = NULL;
143
144                 /*
145                  * We don't care about getting disk errors here. We need
146                  * to purge this dquot anyway, so we go ahead regardless.
147                  */
148                 error = xfs_qm_dqflush(dqp, &bp);
149                 if (!error) {
150                         error = xfs_bwrite(bp);
151                         xfs_buf_relse(bp);
152                 } else if (error == -EAGAIN) {
153                         dqp->q_flags &= ~XFS_DQFLAG_FREEING;
154                         goto out_unlock;
155                 }
156                 xfs_dqflock(dqp);
157         }
158
159         ASSERT(atomic_read(&dqp->q_pincount) == 0);
160         ASSERT(xlog_is_shutdown(dqp->q_logitem.qli_item.li_log) ||
161                 !test_bit(XFS_LI_IN_AIL, &dqp->q_logitem.qli_item.li_flags));
162
163         xfs_dqfunlock(dqp);
164         xfs_dqunlock(dqp);
165
166         radix_tree_delete(xfs_dquot_tree(qi, xfs_dquot_type(dqp)), dqp->q_id);
167         qi->qi_dquots--;
168
169         /*
170          * We move dquots to the freelist as soon as their reference count
171          * hits zero, so it really should be on the freelist here.
172          */
173         ASSERT(!list_empty(&dqp->q_lru));
174         list_lru_del(&qi->qi_lru, &dqp->q_lru);
175         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot_unused);
176
177         xfs_qm_dqdestroy(dqp);
178         return 0;
179
180 out_unlock:
181         xfs_dqunlock(dqp);
182         return error;
183 }
184
185 /*
186  * Purge the dquot cache.
187  */
188 static void
189 xfs_qm_dqpurge_all(
190         struct xfs_mount        *mp)
191 {
192         xfs_qm_dquot_walk(mp, XFS_DQTYPE_USER, xfs_qm_dqpurge, NULL);
193         xfs_qm_dquot_walk(mp, XFS_DQTYPE_GROUP, xfs_qm_dqpurge, NULL);
194         xfs_qm_dquot_walk(mp, XFS_DQTYPE_PROJ, xfs_qm_dqpurge, NULL);
195 }
196
197 /*
198  * Just destroy the quotainfo structure.
199  */
200 void
201 xfs_qm_unmount(
202         struct xfs_mount        *mp)
203 {
204         if (mp->m_quotainfo) {
205                 xfs_qm_dqpurge_all(mp);
206                 xfs_qm_destroy_quotainfo(mp);
207         }
208 }
209
210 /*
211  * Called from the vfsops layer.
212  */
213 void
214 xfs_qm_unmount_quotas(
215         xfs_mount_t     *mp)
216 {
217         /*
218          * Release the dquots that root inode, et al might be holding,
219          * before we flush quotas and blow away the quotainfo structure.
220          */
221         ASSERT(mp->m_rootip);
222         xfs_qm_dqdetach(mp->m_rootip);
223         if (mp->m_rbmip)
224                 xfs_qm_dqdetach(mp->m_rbmip);
225         if (mp->m_rsumip)
226                 xfs_qm_dqdetach(mp->m_rsumip);
227
228         /*
229          * Release the quota inodes.
230          */
231         if (mp->m_quotainfo) {
232                 if (mp->m_quotainfo->qi_uquotaip) {
233                         xfs_irele(mp->m_quotainfo->qi_uquotaip);
234                         mp->m_quotainfo->qi_uquotaip = NULL;
235                 }
236                 if (mp->m_quotainfo->qi_gquotaip) {
237                         xfs_irele(mp->m_quotainfo->qi_gquotaip);
238                         mp->m_quotainfo->qi_gquotaip = NULL;
239                 }
240                 if (mp->m_quotainfo->qi_pquotaip) {
241                         xfs_irele(mp->m_quotainfo->qi_pquotaip);
242                         mp->m_quotainfo->qi_pquotaip = NULL;
243                 }
244         }
245 }
246
247 STATIC int
248 xfs_qm_dqattach_one(
249         struct xfs_inode        *ip,
250         xfs_dqtype_t            type,
251         bool                    doalloc,
252         struct xfs_dquot        **IO_idqpp)
253 {
254         struct xfs_dquot        *dqp;
255         int                     error;
256
257         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
258         error = 0;
259
260         /*
261          * See if we already have it in the inode itself. IO_idqpp is &i_udquot
262          * or &i_gdquot. This made the code look weird, but made the logic a lot
263          * simpler.
264          */
265         dqp = *IO_idqpp;
266         if (dqp) {
267                 trace_xfs_dqattach_found(dqp);
268                 return 0;
269         }
270
271         /*
272          * Find the dquot from somewhere. This bumps the reference count of
273          * dquot and returns it locked.  This can return ENOENT if dquot didn't
274          * exist on disk and we didn't ask it to allocate; ESRCH if quotas got
275          * turned off suddenly.
276          */
277         error = xfs_qm_dqget_inode(ip, type, doalloc, &dqp);
278         if (error)
279                 return error;
280
281         trace_xfs_dqattach_get(dqp);
282
283         /*
284          * dqget may have dropped and re-acquired the ilock, but it guarantees
285          * that the dquot returned is the one that should go in the inode.
286          */
287         *IO_idqpp = dqp;
288         xfs_dqunlock(dqp);
289         return 0;
290 }
291
292 static bool
293 xfs_qm_need_dqattach(
294         struct xfs_inode        *ip)
295 {
296         struct xfs_mount        *mp = ip->i_mount;
297
298         if (!XFS_IS_QUOTA_ON(mp))
299                 return false;
300         if (!XFS_NOT_DQATTACHED(mp, ip))
301                 return false;
302         if (xfs_is_quota_inode(&mp->m_sb, ip->i_ino))
303                 return false;
304         return true;
305 }
306
307 /*
308  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
309  * into account.
310  * If @doalloc is true, the dquot(s) will be allocated if needed.
311  * Inode may get unlocked and relocked in here, and the caller must deal with
312  * the consequences.
313  */
314 int
315 xfs_qm_dqattach_locked(
316         xfs_inode_t     *ip,
317         bool            doalloc)
318 {
319         xfs_mount_t     *mp = ip->i_mount;
320         int             error = 0;
321
322         if (!xfs_qm_need_dqattach(ip))
323                 return 0;
324
325         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
326
327         if (XFS_IS_UQUOTA_ON(mp) && !ip->i_udquot) {
328                 error = xfs_qm_dqattach_one(ip, XFS_DQTYPE_USER,
329                                 doalloc, &ip->i_udquot);
330                 if (error)
331                         goto done;
332                 ASSERT(ip->i_udquot);
333         }
334
335         if (XFS_IS_GQUOTA_ON(mp) && !ip->i_gdquot) {
336                 error = xfs_qm_dqattach_one(ip, XFS_DQTYPE_GROUP,
337                                 doalloc, &ip->i_gdquot);
338                 if (error)
339                         goto done;
340                 ASSERT(ip->i_gdquot);
341         }
342
343         if (XFS_IS_PQUOTA_ON(mp) && !ip->i_pdquot) {
344                 error = xfs_qm_dqattach_one(ip, XFS_DQTYPE_PROJ,
345                                 doalloc, &ip->i_pdquot);
346                 if (error)
347                         goto done;
348                 ASSERT(ip->i_pdquot);
349         }
350
351 done:
352         /*
353          * Don't worry about the dquots that we may have attached before any
354          * error - they'll get detached later if it has not already been done.
355          */
356         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
357         return error;
358 }
359
360 int
361 xfs_qm_dqattach(
362         struct xfs_inode        *ip)
363 {
364         int                     error;
365
366         if (!xfs_qm_need_dqattach(ip))
367                 return 0;
368
369         xfs_ilock(ip, XFS_ILOCK_EXCL);
370         error = xfs_qm_dqattach_locked(ip, false);
371         xfs_iunlock(ip, XFS_ILOCK_EXCL);
372
373         return error;
374 }
375
376 /*
377  * Release dquots (and their references) if any.
378  * The inode should be locked EXCL except when this's called by
379  * xfs_ireclaim.
380  */
381 void
382 xfs_qm_dqdetach(
383         xfs_inode_t     *ip)
384 {
385         if (!(ip->i_udquot || ip->i_gdquot || ip->i_pdquot))
386                 return;
387
388         trace_xfs_dquot_dqdetach(ip);
389
390         ASSERT(!xfs_is_quota_inode(&ip->i_mount->m_sb, ip->i_ino));
391         if (ip->i_udquot) {
392                 xfs_qm_dqrele(ip->i_udquot);
393                 ip->i_udquot = NULL;
394         }
395         if (ip->i_gdquot) {
396                 xfs_qm_dqrele(ip->i_gdquot);
397                 ip->i_gdquot = NULL;
398         }
399         if (ip->i_pdquot) {
400                 xfs_qm_dqrele(ip->i_pdquot);
401                 ip->i_pdquot = NULL;
402         }
403 }
404
405 struct xfs_qm_isolate {
406         struct list_head        buffers;
407         struct list_head        dispose;
408 };
409
410 static enum lru_status
411 xfs_qm_dquot_isolate(
412         struct list_head        *item,
413         struct list_lru_one     *lru,
414         spinlock_t              *lru_lock,
415         void                    *arg)
416                 __releases(lru_lock) __acquires(lru_lock)
417 {
418         struct xfs_dquot        *dqp = container_of(item,
419                                                 struct xfs_dquot, q_lru);
420         struct xfs_qm_isolate   *isol = arg;
421
422         if (!xfs_dqlock_nowait(dqp))
423                 goto out_miss_busy;
424
425         /*
426          * This dquot has acquired a reference in the meantime remove it from
427          * the freelist and try again.
428          */
429         if (dqp->q_nrefs) {
430                 xfs_dqunlock(dqp);
431                 XFS_STATS_INC(dqp->q_mount, xs_qm_dqwants);
432
433                 trace_xfs_dqreclaim_want(dqp);
434                 list_lru_isolate(lru, &dqp->q_lru);
435                 XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot_unused);
436                 return LRU_REMOVED;
437         }
438
439         /*
440          * If the dquot is dirty, flush it. If it's already being flushed, just
441          * skip it so there is time for the IO to complete before we try to
442          * reclaim it again on the next LRU pass.
443          */
444         if (!xfs_dqflock_nowait(dqp)) {
445                 xfs_dqunlock(dqp);
446                 goto out_miss_busy;
447         }
448
449         if (XFS_DQ_IS_DIRTY(dqp)) {
450                 struct xfs_buf  *bp = NULL;
451                 int             error;
452
453                 trace_xfs_dqreclaim_dirty(dqp);
454
455                 /* we have to drop the LRU lock to flush the dquot */
456                 spin_unlock(lru_lock);
457
458                 error = xfs_qm_dqflush(dqp, &bp);
459                 if (error)
460                         goto out_unlock_dirty;
461
462                 xfs_buf_delwri_queue(bp, &isol->buffers);
463                 xfs_buf_relse(bp);
464                 goto out_unlock_dirty;
465         }
466         xfs_dqfunlock(dqp);
467
468         /*
469          * Prevent lookups now that we are past the point of no return.
470          */
471         dqp->q_flags |= XFS_DQFLAG_FREEING;
472         xfs_dqunlock(dqp);
473
474         ASSERT(dqp->q_nrefs == 0);
475         list_lru_isolate_move(lru, &dqp->q_lru, &isol->dispose);
476         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot_unused);
477         trace_xfs_dqreclaim_done(dqp);
478         XFS_STATS_INC(dqp->q_mount, xs_qm_dqreclaims);
479         return LRU_REMOVED;
480
481 out_miss_busy:
482         trace_xfs_dqreclaim_busy(dqp);
483         XFS_STATS_INC(dqp->q_mount, xs_qm_dqreclaim_misses);
484         return LRU_SKIP;
485
486 out_unlock_dirty:
487         trace_xfs_dqreclaim_busy(dqp);
488         XFS_STATS_INC(dqp->q_mount, xs_qm_dqreclaim_misses);
489         xfs_dqunlock(dqp);
490         spin_lock(lru_lock);
491         return LRU_RETRY;
492 }
493
494 static unsigned long
495 xfs_qm_shrink_scan(
496         struct shrinker         *shrink,
497         struct shrink_control   *sc)
498 {
499         struct xfs_quotainfo    *qi = container_of(shrink,
500                                         struct xfs_quotainfo, qi_shrinker);
501         struct xfs_qm_isolate   isol;
502         unsigned long           freed;
503         int                     error;
504
505         if ((sc->gfp_mask & (__GFP_FS|__GFP_DIRECT_RECLAIM)) != (__GFP_FS|__GFP_DIRECT_RECLAIM))
506                 return 0;
507
508         INIT_LIST_HEAD(&isol.buffers);
509         INIT_LIST_HEAD(&isol.dispose);
510
511         freed = list_lru_shrink_walk(&qi->qi_lru, sc,
512                                      xfs_qm_dquot_isolate, &isol);
513
514         error = xfs_buf_delwri_submit(&isol.buffers);
515         if (error)
516                 xfs_warn(NULL, "%s: dquot reclaim failed", __func__);
517
518         while (!list_empty(&isol.dispose)) {
519                 struct xfs_dquot        *dqp;
520
521                 dqp = list_first_entry(&isol.dispose, struct xfs_dquot, q_lru);
522                 list_del_init(&dqp->q_lru);
523                 xfs_qm_dqfree_one(dqp);
524         }
525
526         return freed;
527 }
528
529 static unsigned long
530 xfs_qm_shrink_count(
531         struct shrinker         *shrink,
532         struct shrink_control   *sc)
533 {
534         struct xfs_quotainfo    *qi = container_of(shrink,
535                                         struct xfs_quotainfo, qi_shrinker);
536
537         return list_lru_shrink_count(&qi->qi_lru, sc);
538 }
539
540 STATIC void
541 xfs_qm_set_defquota(
542         struct xfs_mount        *mp,
543         xfs_dqtype_t            type,
544         struct xfs_quotainfo    *qinf)
545 {
546         struct xfs_dquot        *dqp;
547         struct xfs_def_quota    *defq;
548         int                     error;
549
550         error = xfs_qm_dqget_uncached(mp, 0, type, &dqp);
551         if (error)
552                 return;
553
554         defq = xfs_get_defquota(qinf, xfs_dquot_type(dqp));
555
556         /*
557          * Timers and warnings have been already set, let's just set the
558          * default limits for this quota type
559          */
560         defq->blk.hard = dqp->q_blk.hardlimit;
561         defq->blk.soft = dqp->q_blk.softlimit;
562         defq->ino.hard = dqp->q_ino.hardlimit;
563         defq->ino.soft = dqp->q_ino.softlimit;
564         defq->rtb.hard = dqp->q_rtb.hardlimit;
565         defq->rtb.soft = dqp->q_rtb.softlimit;
566         xfs_qm_dqdestroy(dqp);
567 }
568
569 /* Initialize quota time limits from the root dquot. */
570 static void
571 xfs_qm_init_timelimits(
572         struct xfs_mount        *mp,
573         xfs_dqtype_t            type)
574 {
575         struct xfs_quotainfo    *qinf = mp->m_quotainfo;
576         struct xfs_def_quota    *defq;
577         struct xfs_dquot        *dqp;
578         int                     error;
579
580         defq = xfs_get_defquota(qinf, type);
581
582         defq->blk.time = XFS_QM_BTIMELIMIT;
583         defq->ino.time = XFS_QM_ITIMELIMIT;
584         defq->rtb.time = XFS_QM_RTBTIMELIMIT;
585
586         /*
587          * We try to get the limits from the superuser's limits fields.
588          * This is quite hacky, but it is standard quota practice.
589          *
590          * Since we may not have done a quotacheck by this point, just read
591          * the dquot without attaching it to any hashtables or lists.
592          */
593         error = xfs_qm_dqget_uncached(mp, 0, type, &dqp);
594         if (error)
595                 return;
596
597         /*
598          * The warnings and timers set the grace period given to
599          * a user or group before he or she can not perform any
600          * more writing. If it is zero, a default is used.
601          */
602         if (dqp->q_blk.timer)
603                 defq->blk.time = dqp->q_blk.timer;
604         if (dqp->q_ino.timer)
605                 defq->ino.time = dqp->q_ino.timer;
606         if (dqp->q_rtb.timer)
607                 defq->rtb.time = dqp->q_rtb.timer;
608
609         xfs_qm_dqdestroy(dqp);
610 }
611
612 /*
613  * This initializes all the quota information that's kept in the
614  * mount structure
615  */
616 STATIC int
617 xfs_qm_init_quotainfo(
618         struct xfs_mount        *mp)
619 {
620         struct xfs_quotainfo    *qinf;
621         int                     error;
622
623         ASSERT(XFS_IS_QUOTA_ON(mp));
624
625         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(struct xfs_quotainfo), 0);
626
627         error = list_lru_init(&qinf->qi_lru);
628         if (error)
629                 goto out_free_qinf;
630
631         /*
632          * See if quotainodes are setup, and if not, allocate them,
633          * and change the superblock accordingly.
634          */
635         error = xfs_qm_init_quotainos(mp);
636         if (error)
637                 goto out_free_lru;
638
639         INIT_RADIX_TREE(&qinf->qi_uquota_tree, GFP_NOFS);
640         INIT_RADIX_TREE(&qinf->qi_gquota_tree, GFP_NOFS);
641         INIT_RADIX_TREE(&qinf->qi_pquota_tree, GFP_NOFS);
642         mutex_init(&qinf->qi_tree_lock);
643
644         /* mutex used to serialize quotaoffs */
645         mutex_init(&qinf->qi_quotaofflock);
646
647         /* Precalc some constants */
648         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
649         qinf->qi_dqperchunk = xfs_calc_dquots_per_chunk(qinf->qi_dqchunklen);
650         if (xfs_has_bigtime(mp)) {
651                 qinf->qi_expiry_min =
652                         xfs_dq_bigtime_to_unix(XFS_DQ_BIGTIME_EXPIRY_MIN);
653                 qinf->qi_expiry_max =
654                         xfs_dq_bigtime_to_unix(XFS_DQ_BIGTIME_EXPIRY_MAX);
655         } else {
656                 qinf->qi_expiry_min = XFS_DQ_LEGACY_EXPIRY_MIN;
657                 qinf->qi_expiry_max = XFS_DQ_LEGACY_EXPIRY_MAX;
658         }
659         trace_xfs_quota_expiry_range(mp, qinf->qi_expiry_min,
660                         qinf->qi_expiry_max);
661
662         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
663
664         xfs_qm_init_timelimits(mp, XFS_DQTYPE_USER);
665         xfs_qm_init_timelimits(mp, XFS_DQTYPE_GROUP);
666         xfs_qm_init_timelimits(mp, XFS_DQTYPE_PROJ);
667
668         if (XFS_IS_UQUOTA_ON(mp))
669                 xfs_qm_set_defquota(mp, XFS_DQTYPE_USER, qinf);
670         if (XFS_IS_GQUOTA_ON(mp))
671                 xfs_qm_set_defquota(mp, XFS_DQTYPE_GROUP, qinf);
672         if (XFS_IS_PQUOTA_ON(mp))
673                 xfs_qm_set_defquota(mp, XFS_DQTYPE_PROJ, qinf);
674
675         qinf->qi_shrinker.count_objects = xfs_qm_shrink_count;
676         qinf->qi_shrinker.scan_objects = xfs_qm_shrink_scan;
677         qinf->qi_shrinker.seeks = DEFAULT_SEEKS;
678         qinf->qi_shrinker.flags = SHRINKER_NUMA_AWARE;
679
680         error = register_shrinker(&qinf->qi_shrinker, "xfs-qm:%s",
681                                   mp->m_super->s_id);
682         if (error)
683                 goto out_free_inos;
684
685         return 0;
686
687 out_free_inos:
688         mutex_destroy(&qinf->qi_quotaofflock);
689         mutex_destroy(&qinf->qi_tree_lock);
690         xfs_qm_destroy_quotainos(qinf);
691 out_free_lru:
692         list_lru_destroy(&qinf->qi_lru);
693 out_free_qinf:
694         kmem_free(qinf);
695         mp->m_quotainfo = NULL;
696         return error;
697 }
698
699 /*
700  * Gets called when unmounting a filesystem or when all quotas get
701  * turned off.
702  * This purges the quota inodes, destroys locks and frees itself.
703  */
704 void
705 xfs_qm_destroy_quotainfo(
706         struct xfs_mount        *mp)
707 {
708         struct xfs_quotainfo    *qi;
709
710         qi = mp->m_quotainfo;
711         ASSERT(qi != NULL);
712
713         unregister_shrinker(&qi->qi_shrinker);
714         list_lru_destroy(&qi->qi_lru);
715         xfs_qm_destroy_quotainos(qi);
716         mutex_destroy(&qi->qi_tree_lock);
717         mutex_destroy(&qi->qi_quotaofflock);
718         kmem_free(qi);
719         mp->m_quotainfo = NULL;
720 }
721
722 /*
723  * Create an inode and return with a reference already taken, but unlocked
724  * This is how we create quota inodes
725  */
726 STATIC int
727 xfs_qm_qino_alloc(
728         struct xfs_mount        *mp,
729         struct xfs_inode        **ipp,
730         unsigned int            flags)
731 {
732         struct xfs_trans        *tp;
733         int                     error;
734         bool                    need_alloc = true;
735
736         *ipp = NULL;
737         /*
738          * With superblock that doesn't have separate pquotino, we
739          * share an inode between gquota and pquota. If the on-disk
740          * superblock has GQUOTA and the filesystem is now mounted
741          * with PQUOTA, just use sb_gquotino for sb_pquotino and
742          * vice-versa.
743          */
744         if (!xfs_has_pquotino(mp) &&
745                         (flags & (XFS_QMOPT_PQUOTA|XFS_QMOPT_GQUOTA))) {
746                 xfs_ino_t ino = NULLFSINO;
747
748                 if ((flags & XFS_QMOPT_PQUOTA) &&
749                              (mp->m_sb.sb_gquotino != NULLFSINO)) {
750                         ino = mp->m_sb.sb_gquotino;
751                         if (XFS_IS_CORRUPT(mp,
752                                            mp->m_sb.sb_pquotino != NULLFSINO))
753                                 return -EFSCORRUPTED;
754                 } else if ((flags & XFS_QMOPT_GQUOTA) &&
755                              (mp->m_sb.sb_pquotino != NULLFSINO)) {
756                         ino = mp->m_sb.sb_pquotino;
757                         if (XFS_IS_CORRUPT(mp,
758                                            mp->m_sb.sb_gquotino != NULLFSINO))
759                                 return -EFSCORRUPTED;
760                 }
761                 if (ino != NULLFSINO) {
762                         error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
763                         if (error)
764                                 return error;
765                         mp->m_sb.sb_gquotino = NULLFSINO;
766                         mp->m_sb.sb_pquotino = NULLFSINO;
767                         need_alloc = false;
768                 }
769         }
770
771         error = xfs_trans_alloc(mp, &M_RES(mp)->tr_create,
772                         need_alloc ? XFS_QM_QINOCREATE_SPACE_RES(mp) : 0,
773                         0, 0, &tp);
774         if (error)
775                 return error;
776
777         if (need_alloc) {
778                 xfs_ino_t       ino;
779
780                 error = xfs_dialloc(&tp, 0, S_IFREG, &ino);
781                 if (!error)
782                         error = xfs_init_new_inode(&init_user_ns, tp, NULL, ino,
783                                         S_IFREG, 1, 0, 0, false, ipp);
784                 if (error) {
785                         xfs_trans_cancel(tp);
786                         return error;
787                 }
788         }
789
790         /*
791          * Make the changes in the superblock, and log those too.
792          * sbfields arg may contain fields other than *QUOTINO;
793          * VERSIONNUM for example.
794          */
795         spin_lock(&mp->m_sb_lock);
796         if (flags & XFS_QMOPT_SBVERSION) {
797                 ASSERT(!xfs_has_quota(mp));
798
799                 xfs_add_quota(mp);
800                 mp->m_sb.sb_uquotino = NULLFSINO;
801                 mp->m_sb.sb_gquotino = NULLFSINO;
802                 mp->m_sb.sb_pquotino = NULLFSINO;
803
804                 /* qflags will get updated fully _after_ quotacheck */
805                 mp->m_sb.sb_qflags = mp->m_qflags & XFS_ALL_QUOTA_ACCT;
806         }
807         if (flags & XFS_QMOPT_UQUOTA)
808                 mp->m_sb.sb_uquotino = (*ipp)->i_ino;
809         else if (flags & XFS_QMOPT_GQUOTA)
810                 mp->m_sb.sb_gquotino = (*ipp)->i_ino;
811         else
812                 mp->m_sb.sb_pquotino = (*ipp)->i_ino;
813         spin_unlock(&mp->m_sb_lock);
814         xfs_log_sb(tp);
815
816         error = xfs_trans_commit(tp);
817         if (error) {
818                 ASSERT(xfs_is_shutdown(mp));
819                 xfs_alert(mp, "%s failed (error %d)!", __func__, error);
820         }
821         if (need_alloc)
822                 xfs_finish_inode_setup(*ipp);
823         return error;
824 }
825
826
827 STATIC void
828 xfs_qm_reset_dqcounts(
829         struct xfs_mount        *mp,
830         struct xfs_buf          *bp,
831         xfs_dqid_t              id,
832         xfs_dqtype_t            type)
833 {
834         struct xfs_dqblk        *dqb;
835         int                     j;
836
837         trace_xfs_reset_dqcounts(bp, _RET_IP_);
838
839         /*
840          * Reset all counters and timers. They'll be
841          * started afresh by xfs_qm_quotacheck.
842          */
843 #ifdef DEBUG
844         j = (int)XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB) /
845                 sizeof(struct xfs_dqblk);
846         ASSERT(mp->m_quotainfo->qi_dqperchunk == j);
847 #endif
848         dqb = bp->b_addr;
849         for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) {
850                 struct xfs_disk_dquot   *ddq;
851
852                 ddq = (struct xfs_disk_dquot *)&dqb[j];
853
854                 /*
855                  * Do a sanity check, and if needed, repair the dqblk. Don't
856                  * output any warnings because it's perfectly possible to
857                  * find uninitialised dquot blks. See comment in
858                  * xfs_dquot_verify.
859                  */
860                 if (xfs_dqblk_verify(mp, &dqb[j], id + j) ||
861                     (dqb[j].dd_diskdq.d_type & XFS_DQTYPE_REC_MASK) != type)
862                         xfs_dqblk_repair(mp, &dqb[j], id + j, type);
863
864                 /*
865                  * Reset type in case we are reusing group quota file for
866                  * project quotas or vice versa
867                  */
868                 ddq->d_type = type;
869                 ddq->d_bcount = 0;
870                 ddq->d_icount = 0;
871                 ddq->d_rtbcount = 0;
872
873                 /*
874                  * dquot id 0 stores the default grace period and the maximum
875                  * warning limit that were set by the administrator, so we
876                  * should not reset them.
877                  */
878                 if (ddq->d_id != 0) {
879                         ddq->d_btimer = 0;
880                         ddq->d_itimer = 0;
881                         ddq->d_rtbtimer = 0;
882                         ddq->d_bwarns = 0;
883                         ddq->d_iwarns = 0;
884                         ddq->d_rtbwarns = 0;
885                         if (xfs_has_bigtime(mp))
886                                 ddq->d_type |= XFS_DQTYPE_BIGTIME;
887                 }
888
889                 if (xfs_has_crc(mp)) {
890                         xfs_update_cksum((char *)&dqb[j],
891                                          sizeof(struct xfs_dqblk),
892                                          XFS_DQUOT_CRC_OFF);
893                 }
894         }
895 }
896
897 STATIC int
898 xfs_qm_reset_dqcounts_all(
899         struct xfs_mount        *mp,
900         xfs_dqid_t              firstid,
901         xfs_fsblock_t           bno,
902         xfs_filblks_t           blkcnt,
903         xfs_dqtype_t            type,
904         struct list_head        *buffer_list)
905 {
906         struct xfs_buf          *bp;
907         int                     error = 0;
908
909         ASSERT(blkcnt > 0);
910
911         /*
912          * Blkcnt arg can be a very big number, and might even be
913          * larger than the log itself. So, we have to break it up into
914          * manageable-sized transactions.
915          * Note that we don't start a permanent transaction here; we might
916          * not be able to get a log reservation for the whole thing up front,
917          * and we don't really care to either, because we just discard
918          * everything if we were to crash in the middle of this loop.
919          */
920         while (blkcnt--) {
921                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
922                               XFS_FSB_TO_DADDR(mp, bno),
923                               mp->m_quotainfo->qi_dqchunklen, 0, &bp,
924                               &xfs_dquot_buf_ops);
925
926                 /*
927                  * CRC and validation errors will return a EFSCORRUPTED here. If
928                  * this occurs, re-read without CRC validation so that we can
929                  * repair the damage via xfs_qm_reset_dqcounts(). This process
930                  * will leave a trace in the log indicating corruption has
931                  * been detected.
932                  */
933                 if (error == -EFSCORRUPTED) {
934                         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
935                                       XFS_FSB_TO_DADDR(mp, bno),
936                                       mp->m_quotainfo->qi_dqchunklen, 0, &bp,
937                                       NULL);
938                 }
939
940                 if (error)
941                         break;
942
943                 /*
944                  * A corrupt buffer might not have a verifier attached, so
945                  * make sure we have the correct one attached before writeback
946                  * occurs.
947                  */
948                 bp->b_ops = &xfs_dquot_buf_ops;
949                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
950                 xfs_buf_delwri_queue(bp, buffer_list);
951                 xfs_buf_relse(bp);
952
953                 /* goto the next block. */
954                 bno++;
955                 firstid += mp->m_quotainfo->qi_dqperchunk;
956         }
957
958         return error;
959 }
960
961 /*
962  * Iterate over all allocated dquot blocks in this quota inode, zeroing all
963  * counters for every chunk of dquots that we find.
964  */
965 STATIC int
966 xfs_qm_reset_dqcounts_buf(
967         struct xfs_mount        *mp,
968         struct xfs_inode        *qip,
969         xfs_dqtype_t            type,
970         struct list_head        *buffer_list)
971 {
972         struct xfs_bmbt_irec    *map;
973         int                     i, nmaps;       /* number of map entries */
974         int                     error;          /* return value */
975         xfs_fileoff_t           lblkno;
976         xfs_filblks_t           maxlblkcnt;
977         xfs_dqid_t              firstid;
978         xfs_fsblock_t           rablkno;
979         xfs_filblks_t           rablkcnt;
980
981         error = 0;
982         /*
983          * This looks racy, but we can't keep an inode lock across a
984          * trans_reserve. But, this gets called during quotacheck, and that
985          * happens only at mount time which is single threaded.
986          */
987         if (qip->i_nblocks == 0)
988                 return 0;
989
990         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), 0);
991
992         lblkno = 0;
993         maxlblkcnt = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
994         do {
995                 uint            lock_mode;
996
997                 nmaps = XFS_DQITER_MAP_SIZE;
998                 /*
999                  * We aren't changing the inode itself. Just changing
1000                  * some of its data. No new blocks are added here, and
1001                  * the inode is never added to the transaction.
1002                  */
1003                 lock_mode = xfs_ilock_data_map_shared(qip);
1004                 error = xfs_bmapi_read(qip, lblkno, maxlblkcnt - lblkno,
1005                                        map, &nmaps, 0);
1006                 xfs_iunlock(qip, lock_mode);
1007                 if (error)
1008                         break;
1009
1010                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1011                 for (i = 0; i < nmaps; i++) {
1012                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1013                         ASSERT(map[i].br_blockcount);
1014
1015
1016                         lblkno += map[i].br_blockcount;
1017
1018                         if (map[i].br_startblock == HOLESTARTBLOCK)
1019                                 continue;
1020
1021                         firstid = (xfs_dqid_t) map[i].br_startoff *
1022                                 mp->m_quotainfo->qi_dqperchunk;
1023                         /*
1024                          * Do a read-ahead on the next extent.
1025                          */
1026                         if ((i+1 < nmaps) &&
1027                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1028                                 rablkcnt =  map[i+1].br_blockcount;
1029                                 rablkno = map[i+1].br_startblock;
1030                                 while (rablkcnt--) {
1031                                         xfs_buf_readahead(mp->m_ddev_targp,
1032                                                XFS_FSB_TO_DADDR(mp, rablkno),
1033                                                mp->m_quotainfo->qi_dqchunklen,
1034                                                &xfs_dquot_buf_ops);
1035                                         rablkno++;
1036                                 }
1037                         }
1038                         /*
1039                          * Iterate thru all the blks in the extent and
1040                          * reset the counters of all the dquots inside them.
1041                          */
1042                         error = xfs_qm_reset_dqcounts_all(mp, firstid,
1043                                                    map[i].br_startblock,
1044                                                    map[i].br_blockcount,
1045                                                    type, buffer_list);
1046                         if (error)
1047                                 goto out;
1048                 }
1049         } while (nmaps > 0);
1050
1051 out:
1052         kmem_free(map);
1053         return error;
1054 }
1055
1056 /*
1057  * Called by dqusage_adjust in doing a quotacheck.
1058  *
1059  * Given the inode, and a dquot id this updates both the incore dqout as well
1060  * as the buffer copy. This is so that once the quotacheck is done, we can
1061  * just log all the buffers, as opposed to logging numerous updates to
1062  * individual dquots.
1063  */
1064 STATIC int
1065 xfs_qm_quotacheck_dqadjust(
1066         struct xfs_inode        *ip,
1067         xfs_dqtype_t            type,
1068         xfs_qcnt_t              nblks,
1069         xfs_qcnt_t              rtblks)
1070 {
1071         struct xfs_mount        *mp = ip->i_mount;
1072         struct xfs_dquot        *dqp;
1073         xfs_dqid_t              id;
1074         int                     error;
1075
1076         id = xfs_qm_id_for_quotatype(ip, type);
1077         error = xfs_qm_dqget(mp, id, type, true, &dqp);
1078         if (error) {
1079                 /*
1080                  * Shouldn't be able to turn off quotas here.
1081                  */
1082                 ASSERT(error != -ESRCH);
1083                 ASSERT(error != -ENOENT);
1084                 return error;
1085         }
1086
1087         trace_xfs_dqadjust(dqp);
1088
1089         /*
1090          * Adjust the inode count and the block count to reflect this inode's
1091          * resource usage.
1092          */
1093         dqp->q_ino.count++;
1094         dqp->q_ino.reserved++;
1095         if (nblks) {
1096                 dqp->q_blk.count += nblks;
1097                 dqp->q_blk.reserved += nblks;
1098         }
1099         if (rtblks) {
1100                 dqp->q_rtb.count += rtblks;
1101                 dqp->q_rtb.reserved += rtblks;
1102         }
1103
1104         /*
1105          * Set default limits, adjust timers (since we changed usages)
1106          *
1107          * There are no timers for the default values set in the root dquot.
1108          */
1109         if (dqp->q_id) {
1110                 xfs_qm_adjust_dqlimits(dqp);
1111                 xfs_qm_adjust_dqtimers(dqp);
1112         }
1113
1114         dqp->q_flags |= XFS_DQFLAG_DIRTY;
1115         xfs_qm_dqput(dqp);
1116         return 0;
1117 }
1118
1119 /*
1120  * callback routine supplied to bulkstat(). Given an inumber, find its
1121  * dquots and update them to account for resources taken by that inode.
1122  */
1123 /* ARGSUSED */
1124 STATIC int
1125 xfs_qm_dqusage_adjust(
1126         struct xfs_mount        *mp,
1127         struct xfs_trans        *tp,
1128         xfs_ino_t               ino,
1129         void                    *data)
1130 {
1131         struct xfs_inode        *ip;
1132         xfs_qcnt_t              nblks;
1133         xfs_filblks_t           rtblks = 0;     /* total rt blks */
1134         int                     error;
1135
1136         ASSERT(XFS_IS_QUOTA_ON(mp));
1137
1138         /*
1139          * rootino must have its resources accounted for, not so with the quota
1140          * inodes.
1141          */
1142         if (xfs_is_quota_inode(&mp->m_sb, ino))
1143                 return 0;
1144
1145         /*
1146          * We don't _need_ to take the ilock EXCL here because quotacheck runs
1147          * at mount time and therefore nobody will be racing chown/chproj.
1148          */
1149         error = xfs_iget(mp, tp, ino, XFS_IGET_DONTCACHE, 0, &ip);
1150         if (error == -EINVAL || error == -ENOENT)
1151                 return 0;
1152         if (error)
1153                 return error;
1154
1155         ASSERT(ip->i_delayed_blks == 0);
1156
1157         if (XFS_IS_REALTIME_INODE(ip)) {
1158                 struct xfs_ifork        *ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1159
1160                 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1161                 if (error)
1162                         goto error0;
1163
1164                 xfs_bmap_count_leaves(ifp, &rtblks);
1165         }
1166
1167         nblks = (xfs_qcnt_t)ip->i_nblocks - rtblks;
1168
1169         /*
1170          * Add the (disk blocks and inode) resources occupied by this
1171          * inode to its dquots. We do this adjustment in the incore dquot,
1172          * and also copy the changes to its buffer.
1173          * We don't care about putting these changes in a transaction
1174          * envelope because if we crash in the middle of a 'quotacheck'
1175          * we have to start from the beginning anyway.
1176          * Once we're done, we'll log all the dquot bufs.
1177          *
1178          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1179          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1180          */
1181         if (XFS_IS_UQUOTA_ON(mp)) {
1182                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_USER, nblks,
1183                                 rtblks);
1184                 if (error)
1185                         goto error0;
1186         }
1187
1188         if (XFS_IS_GQUOTA_ON(mp)) {
1189                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_GROUP, nblks,
1190                                 rtblks);
1191                 if (error)
1192                         goto error0;
1193         }
1194
1195         if (XFS_IS_PQUOTA_ON(mp)) {
1196                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_PROJ, nblks,
1197                                 rtblks);
1198                 if (error)
1199                         goto error0;
1200         }
1201
1202 error0:
1203         xfs_irele(ip);
1204         return error;
1205 }
1206
1207 STATIC int
1208 xfs_qm_flush_one(
1209         struct xfs_dquot        *dqp,
1210         void                    *data)
1211 {
1212         struct xfs_mount        *mp = dqp->q_mount;
1213         struct list_head        *buffer_list = data;
1214         struct xfs_buf          *bp = NULL;
1215         int                     error = 0;
1216
1217         xfs_dqlock(dqp);
1218         if (dqp->q_flags & XFS_DQFLAG_FREEING)
1219                 goto out_unlock;
1220         if (!XFS_DQ_IS_DIRTY(dqp))
1221                 goto out_unlock;
1222
1223         /*
1224          * The only way the dquot is already flush locked by the time quotacheck
1225          * gets here is if reclaim flushed it before the dqadjust walk dirtied
1226          * it for the final time. Quotacheck collects all dquot bufs in the
1227          * local delwri queue before dquots are dirtied, so reclaim can't have
1228          * possibly queued it for I/O. The only way out is to push the buffer to
1229          * cycle the flush lock.
1230          */
1231         if (!xfs_dqflock_nowait(dqp)) {
1232                 /* buf is pinned in-core by delwri list */
1233                 error = xfs_buf_incore(mp->m_ddev_targp, dqp->q_blkno,
1234                                 mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1235                 if (error)
1236                         goto out_unlock;
1237
1238                 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1239                         error = -EAGAIN;
1240                         xfs_buf_relse(bp);
1241                         goto out_unlock;
1242                 }
1243                 xfs_buf_unlock(bp);
1244
1245                 xfs_buf_delwri_pushbuf(bp, buffer_list);
1246                 xfs_buf_rele(bp);
1247
1248                 error = -EAGAIN;
1249                 goto out_unlock;
1250         }
1251
1252         error = xfs_qm_dqflush(dqp, &bp);
1253         if (error)
1254                 goto out_unlock;
1255
1256         xfs_buf_delwri_queue(bp, buffer_list);
1257         xfs_buf_relse(bp);
1258 out_unlock:
1259         xfs_dqunlock(dqp);
1260         return error;
1261 }
1262
1263 /*
1264  * Walk thru all the filesystem inodes and construct a consistent view
1265  * of the disk quota world. If the quotacheck fails, disable quotas.
1266  */
1267 STATIC int
1268 xfs_qm_quotacheck(
1269         xfs_mount_t     *mp)
1270 {
1271         int                     error, error2;
1272         uint                    flags;
1273         LIST_HEAD               (buffer_list);
1274         struct xfs_inode        *uip = mp->m_quotainfo->qi_uquotaip;
1275         struct xfs_inode        *gip = mp->m_quotainfo->qi_gquotaip;
1276         struct xfs_inode        *pip = mp->m_quotainfo->qi_pquotaip;
1277
1278         flags = 0;
1279
1280         ASSERT(uip || gip || pip);
1281         ASSERT(XFS_IS_QUOTA_ON(mp));
1282
1283         xfs_notice(mp, "Quotacheck needed: Please wait.");
1284
1285         /*
1286          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1287          * their counters to zero. We need a clean slate.
1288          * We don't log our changes till later.
1289          */
1290         if (uip) {
1291                 error = xfs_qm_reset_dqcounts_buf(mp, uip, XFS_DQTYPE_USER,
1292                                          &buffer_list);
1293                 if (error)
1294                         goto error_return;
1295                 flags |= XFS_UQUOTA_CHKD;
1296         }
1297
1298         if (gip) {
1299                 error = xfs_qm_reset_dqcounts_buf(mp, gip, XFS_DQTYPE_GROUP,
1300                                          &buffer_list);
1301                 if (error)
1302                         goto error_return;
1303                 flags |= XFS_GQUOTA_CHKD;
1304         }
1305
1306         if (pip) {
1307                 error = xfs_qm_reset_dqcounts_buf(mp, pip, XFS_DQTYPE_PROJ,
1308                                          &buffer_list);
1309                 if (error)
1310                         goto error_return;
1311                 flags |= XFS_PQUOTA_CHKD;
1312         }
1313
1314         error = xfs_iwalk_threaded(mp, 0, 0, xfs_qm_dqusage_adjust, 0, true,
1315                         NULL);
1316         if (error) {
1317                 /*
1318                  * The inode walk may have partially populated the dquot
1319                  * caches.  We must purge them before disabling quota and
1320                  * tearing down the quotainfo, or else the dquots will leak.
1321                  */
1322                 xfs_qm_dqpurge_all(mp);
1323                 goto error_return;
1324         }
1325
1326         /*
1327          * We've made all the changes that we need to make incore.  Flush them
1328          * down to disk buffers if everything was updated successfully.
1329          */
1330         if (XFS_IS_UQUOTA_ON(mp)) {
1331                 error = xfs_qm_dquot_walk(mp, XFS_DQTYPE_USER, xfs_qm_flush_one,
1332                                           &buffer_list);
1333         }
1334         if (XFS_IS_GQUOTA_ON(mp)) {
1335                 error2 = xfs_qm_dquot_walk(mp, XFS_DQTYPE_GROUP, xfs_qm_flush_one,
1336                                            &buffer_list);
1337                 if (!error)
1338                         error = error2;
1339         }
1340         if (XFS_IS_PQUOTA_ON(mp)) {
1341                 error2 = xfs_qm_dquot_walk(mp, XFS_DQTYPE_PROJ, xfs_qm_flush_one,
1342                                            &buffer_list);
1343                 if (!error)
1344                         error = error2;
1345         }
1346
1347         error2 = xfs_buf_delwri_submit(&buffer_list);
1348         if (!error)
1349                 error = error2;
1350
1351         /*
1352          * We can get this error if we couldn't do a dquot allocation inside
1353          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1354          * dirty dquots that might be cached, we just want to get rid of them
1355          * and turn quotaoff. The dquots won't be attached to any of the inodes
1356          * at this point (because we intentionally didn't in dqget_noattach).
1357          */
1358         if (error) {
1359                 xfs_qm_dqpurge_all(mp);
1360                 goto error_return;
1361         }
1362
1363         /*
1364          * If one type of quotas is off, then it will lose its
1365          * quotachecked status, since we won't be doing accounting for
1366          * that type anymore.
1367          */
1368         mp->m_qflags &= ~XFS_ALL_QUOTA_CHKD;
1369         mp->m_qflags |= flags;
1370
1371  error_return:
1372         xfs_buf_delwri_cancel(&buffer_list);
1373
1374         if (error) {
1375                 xfs_warn(mp,
1376         "Quotacheck: Unsuccessful (Error %d): Disabling quotas.",
1377                         error);
1378                 /*
1379                  * We must turn off quotas.
1380                  */
1381                 ASSERT(mp->m_quotainfo != NULL);
1382                 xfs_qm_destroy_quotainfo(mp);
1383                 if (xfs_mount_reset_sbqflags(mp)) {
1384                         xfs_warn(mp,
1385                                 "Quotacheck: Failed to reset quota flags.");
1386                 }
1387         } else
1388                 xfs_notice(mp, "Quotacheck: Done.");
1389         return error;
1390 }
1391
1392 /*
1393  * This is called from xfs_mountfs to start quotas and initialize all
1394  * necessary data structures like quotainfo.  This is also responsible for
1395  * running a quotacheck as necessary.  We are guaranteed that the superblock
1396  * is consistently read in at this point.
1397  *
1398  * If we fail here, the mount will continue with quota turned off. We don't
1399  * need to inidicate success or failure at all.
1400  */
1401 void
1402 xfs_qm_mount_quotas(
1403         struct xfs_mount        *mp)
1404 {
1405         int                     error = 0;
1406         uint                    sbf;
1407
1408         /*
1409          * If quotas on realtime volumes is not supported, we disable
1410          * quotas immediately.
1411          */
1412         if (mp->m_sb.sb_rextents) {
1413                 xfs_notice(mp, "Cannot turn on quotas for realtime filesystem");
1414                 mp->m_qflags = 0;
1415                 goto write_changes;
1416         }
1417
1418         ASSERT(XFS_IS_QUOTA_ON(mp));
1419
1420         /*
1421          * Allocate the quotainfo structure inside the mount struct, and
1422          * create quotainode(s), and change/rev superblock if necessary.
1423          */
1424         error = xfs_qm_init_quotainfo(mp);
1425         if (error) {
1426                 /*
1427                  * We must turn off quotas.
1428                  */
1429                 ASSERT(mp->m_quotainfo == NULL);
1430                 mp->m_qflags = 0;
1431                 goto write_changes;
1432         }
1433         /*
1434          * If any of the quotas are not consistent, do a quotacheck.
1435          */
1436         if (XFS_QM_NEED_QUOTACHECK(mp)) {
1437                 error = xfs_qm_quotacheck(mp);
1438                 if (error) {
1439                         /* Quotacheck failed and disabled quotas. */
1440                         return;
1441                 }
1442         }
1443         /*
1444          * If one type of quotas is off, then it will lose its
1445          * quotachecked status, since we won't be doing accounting for
1446          * that type anymore.
1447          */
1448         if (!XFS_IS_UQUOTA_ON(mp))
1449                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
1450         if (!XFS_IS_GQUOTA_ON(mp))
1451                 mp->m_qflags &= ~XFS_GQUOTA_CHKD;
1452         if (!XFS_IS_PQUOTA_ON(mp))
1453                 mp->m_qflags &= ~XFS_PQUOTA_CHKD;
1454
1455  write_changes:
1456         /*
1457          * We actually don't have to acquire the m_sb_lock at all.
1458          * This can only be called from mount, and that's single threaded. XXX
1459          */
1460         spin_lock(&mp->m_sb_lock);
1461         sbf = mp->m_sb.sb_qflags;
1462         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
1463         spin_unlock(&mp->m_sb_lock);
1464
1465         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
1466                 if (xfs_sync_sb(mp, false)) {
1467                         /*
1468                          * We could only have been turning quotas off.
1469                          * We aren't in very good shape actually because
1470                          * the incore structures are convinced that quotas are
1471                          * off, but the on disk superblock doesn't know that !
1472                          */
1473                         ASSERT(!(XFS_IS_QUOTA_ON(mp)));
1474                         xfs_alert(mp, "%s: Superblock update failed!",
1475                                 __func__);
1476                 }
1477         }
1478
1479         if (error) {
1480                 xfs_warn(mp, "Failed to initialize disk quotas.");
1481                 return;
1482         }
1483 }
1484
1485 /*
1486  * This is called after the superblock has been read in and we're ready to
1487  * iget the quota inodes.
1488  */
1489 STATIC int
1490 xfs_qm_init_quotainos(
1491         xfs_mount_t     *mp)
1492 {
1493         struct xfs_inode        *uip = NULL;
1494         struct xfs_inode        *gip = NULL;
1495         struct xfs_inode        *pip = NULL;
1496         int                     error;
1497         uint                    flags = 0;
1498
1499         ASSERT(mp->m_quotainfo);
1500
1501         /*
1502          * Get the uquota and gquota inodes
1503          */
1504         if (xfs_has_quota(mp)) {
1505                 if (XFS_IS_UQUOTA_ON(mp) &&
1506                     mp->m_sb.sb_uquotino != NULLFSINO) {
1507                         ASSERT(mp->m_sb.sb_uquotino > 0);
1508                         error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1509                                              0, 0, &uip);
1510                         if (error)
1511                                 return error;
1512                 }
1513                 if (XFS_IS_GQUOTA_ON(mp) &&
1514                     mp->m_sb.sb_gquotino != NULLFSINO) {
1515                         ASSERT(mp->m_sb.sb_gquotino > 0);
1516                         error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1517                                              0, 0, &gip);
1518                         if (error)
1519                                 goto error_rele;
1520                 }
1521                 if (XFS_IS_PQUOTA_ON(mp) &&
1522                     mp->m_sb.sb_pquotino != NULLFSINO) {
1523                         ASSERT(mp->m_sb.sb_pquotino > 0);
1524                         error = xfs_iget(mp, NULL, mp->m_sb.sb_pquotino,
1525                                              0, 0, &pip);
1526                         if (error)
1527                                 goto error_rele;
1528                 }
1529         } else {
1530                 flags |= XFS_QMOPT_SBVERSION;
1531         }
1532
1533         /*
1534          * Create the three inodes, if they don't exist already. The changes
1535          * made above will get added to a transaction and logged in one of
1536          * the qino_alloc calls below.  If the device is readonly,
1537          * temporarily switch to read-write to do this.
1538          */
1539         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1540                 error = xfs_qm_qino_alloc(mp, &uip,
1541                                               flags | XFS_QMOPT_UQUOTA);
1542                 if (error)
1543                         goto error_rele;
1544
1545                 flags &= ~XFS_QMOPT_SBVERSION;
1546         }
1547         if (XFS_IS_GQUOTA_ON(mp) && gip == NULL) {
1548                 error = xfs_qm_qino_alloc(mp, &gip,
1549                                           flags | XFS_QMOPT_GQUOTA);
1550                 if (error)
1551                         goto error_rele;
1552
1553                 flags &= ~XFS_QMOPT_SBVERSION;
1554         }
1555         if (XFS_IS_PQUOTA_ON(mp) && pip == NULL) {
1556                 error = xfs_qm_qino_alloc(mp, &pip,
1557                                           flags | XFS_QMOPT_PQUOTA);
1558                 if (error)
1559                         goto error_rele;
1560         }
1561
1562         mp->m_quotainfo->qi_uquotaip = uip;
1563         mp->m_quotainfo->qi_gquotaip = gip;
1564         mp->m_quotainfo->qi_pquotaip = pip;
1565
1566         return 0;
1567
1568 error_rele:
1569         if (uip)
1570                 xfs_irele(uip);
1571         if (gip)
1572                 xfs_irele(gip);
1573         if (pip)
1574                 xfs_irele(pip);
1575         return error;
1576 }
1577
1578 STATIC void
1579 xfs_qm_destroy_quotainos(
1580         struct xfs_quotainfo    *qi)
1581 {
1582         if (qi->qi_uquotaip) {
1583                 xfs_irele(qi->qi_uquotaip);
1584                 qi->qi_uquotaip = NULL; /* paranoia */
1585         }
1586         if (qi->qi_gquotaip) {
1587                 xfs_irele(qi->qi_gquotaip);
1588                 qi->qi_gquotaip = NULL;
1589         }
1590         if (qi->qi_pquotaip) {
1591                 xfs_irele(qi->qi_pquotaip);
1592                 qi->qi_pquotaip = NULL;
1593         }
1594 }
1595
1596 STATIC void
1597 xfs_qm_dqfree_one(
1598         struct xfs_dquot        *dqp)
1599 {
1600         struct xfs_mount        *mp = dqp->q_mount;
1601         struct xfs_quotainfo    *qi = mp->m_quotainfo;
1602
1603         mutex_lock(&qi->qi_tree_lock);
1604         radix_tree_delete(xfs_dquot_tree(qi, xfs_dquot_type(dqp)), dqp->q_id);
1605
1606         qi->qi_dquots--;
1607         mutex_unlock(&qi->qi_tree_lock);
1608
1609         xfs_qm_dqdestroy(dqp);
1610 }
1611
1612 /* --------------- utility functions for vnodeops ---------------- */
1613
1614
1615 /*
1616  * Given an inode, a uid, gid and prid make sure that we have
1617  * allocated relevant dquot(s) on disk, and that we won't exceed inode
1618  * quotas by creating this file.
1619  * This also attaches dquot(s) to the given inode after locking it,
1620  * and returns the dquots corresponding to the uid and/or gid.
1621  *
1622  * in   : inode (unlocked)
1623  * out  : udquot, gdquot with references taken and unlocked
1624  */
1625 int
1626 xfs_qm_vop_dqalloc(
1627         struct xfs_inode        *ip,
1628         kuid_t                  uid,
1629         kgid_t                  gid,
1630         prid_t                  prid,
1631         uint                    flags,
1632         struct xfs_dquot        **O_udqpp,
1633         struct xfs_dquot        **O_gdqpp,
1634         struct xfs_dquot        **O_pdqpp)
1635 {
1636         struct xfs_mount        *mp = ip->i_mount;
1637         struct inode            *inode = VFS_I(ip);
1638         struct user_namespace   *user_ns = inode->i_sb->s_user_ns;
1639         struct xfs_dquot        *uq = NULL;
1640         struct xfs_dquot        *gq = NULL;
1641         struct xfs_dquot        *pq = NULL;
1642         int                     error;
1643         uint                    lockflags;
1644
1645         if (!XFS_IS_QUOTA_ON(mp))
1646                 return 0;
1647
1648         lockflags = XFS_ILOCK_EXCL;
1649         xfs_ilock(ip, lockflags);
1650
1651         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
1652                 gid = inode->i_gid;
1653
1654         /*
1655          * Attach the dquot(s) to this inode, doing a dquot allocation
1656          * if necessary. The dquot(s) will not be locked.
1657          */
1658         if (XFS_NOT_DQATTACHED(mp, ip)) {
1659                 error = xfs_qm_dqattach_locked(ip, true);
1660                 if (error) {
1661                         xfs_iunlock(ip, lockflags);
1662                         return error;
1663                 }
1664         }
1665
1666         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
1667                 ASSERT(O_udqpp);
1668                 if (!uid_eq(inode->i_uid, uid)) {
1669                         /*
1670                          * What we need is the dquot that has this uid, and
1671                          * if we send the inode to dqget, the uid of the inode
1672                          * takes priority over what's sent in the uid argument.
1673                          * We must unlock inode here before calling dqget if
1674                          * we're not sending the inode, because otherwise
1675                          * we'll deadlock by doing trans_reserve while
1676                          * holding ilock.
1677                          */
1678                         xfs_iunlock(ip, lockflags);
1679                         error = xfs_qm_dqget(mp, from_kuid(user_ns, uid),
1680                                         XFS_DQTYPE_USER, true, &uq);
1681                         if (error) {
1682                                 ASSERT(error != -ENOENT);
1683                                 return error;
1684                         }
1685                         /*
1686                          * Get the ilock in the right order.
1687                          */
1688                         xfs_dqunlock(uq);
1689                         lockflags = XFS_ILOCK_SHARED;
1690                         xfs_ilock(ip, lockflags);
1691                 } else {
1692                         /*
1693                          * Take an extra reference, because we'll return
1694                          * this to caller
1695                          */
1696                         ASSERT(ip->i_udquot);
1697                         uq = xfs_qm_dqhold(ip->i_udquot);
1698                 }
1699         }
1700         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
1701                 ASSERT(O_gdqpp);
1702                 if (!gid_eq(inode->i_gid, gid)) {
1703                         xfs_iunlock(ip, lockflags);
1704                         error = xfs_qm_dqget(mp, from_kgid(user_ns, gid),
1705                                         XFS_DQTYPE_GROUP, true, &gq);
1706                         if (error) {
1707                                 ASSERT(error != -ENOENT);
1708                                 goto error_rele;
1709                         }
1710                         xfs_dqunlock(gq);
1711                         lockflags = XFS_ILOCK_SHARED;
1712                         xfs_ilock(ip, lockflags);
1713                 } else {
1714                         ASSERT(ip->i_gdquot);
1715                         gq = xfs_qm_dqhold(ip->i_gdquot);
1716                 }
1717         }
1718         if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
1719                 ASSERT(O_pdqpp);
1720                 if (ip->i_projid != prid) {
1721                         xfs_iunlock(ip, lockflags);
1722                         error = xfs_qm_dqget(mp, prid,
1723                                         XFS_DQTYPE_PROJ, true, &pq);
1724                         if (error) {
1725                                 ASSERT(error != -ENOENT);
1726                                 goto error_rele;
1727                         }
1728                         xfs_dqunlock(pq);
1729                         lockflags = XFS_ILOCK_SHARED;
1730                         xfs_ilock(ip, lockflags);
1731                 } else {
1732                         ASSERT(ip->i_pdquot);
1733                         pq = xfs_qm_dqhold(ip->i_pdquot);
1734                 }
1735         }
1736         trace_xfs_dquot_dqalloc(ip);
1737
1738         xfs_iunlock(ip, lockflags);
1739         if (O_udqpp)
1740                 *O_udqpp = uq;
1741         else
1742                 xfs_qm_dqrele(uq);
1743         if (O_gdqpp)
1744                 *O_gdqpp = gq;
1745         else
1746                 xfs_qm_dqrele(gq);
1747         if (O_pdqpp)
1748                 *O_pdqpp = pq;
1749         else
1750                 xfs_qm_dqrele(pq);
1751         return 0;
1752
1753 error_rele:
1754         xfs_qm_dqrele(gq);
1755         xfs_qm_dqrele(uq);
1756         return error;
1757 }
1758
1759 /*
1760  * Actually transfer ownership, and do dquot modifications.
1761  * These were already reserved.
1762  */
1763 struct xfs_dquot *
1764 xfs_qm_vop_chown(
1765         struct xfs_trans        *tp,
1766         struct xfs_inode        *ip,
1767         struct xfs_dquot        **IO_olddq,
1768         struct xfs_dquot        *newdq)
1769 {
1770         struct xfs_dquot        *prevdq;
1771         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
1772                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
1773
1774
1775         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1776         ASSERT(XFS_IS_QUOTA_ON(ip->i_mount));
1777
1778         /* old dquot */
1779         prevdq = *IO_olddq;
1780         ASSERT(prevdq);
1781         ASSERT(prevdq != newdq);
1782
1783         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_nblocks));
1784         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
1785
1786         /* the sparkling new dquot */
1787         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_nblocks);
1788         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
1789
1790         /*
1791          * Back when we made quota reservations for the chown, we reserved the
1792          * ondisk blocks + delalloc blocks with the new dquot.  Now that we've
1793          * switched the dquots, decrease the new dquot's block reservation
1794          * (having already bumped up the real counter) so that we don't have
1795          * any reservation to give back when we commit.
1796          */
1797         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_RES_BLKS,
1798                         -ip->i_delayed_blks);
1799
1800         /*
1801          * Give the incore reservation for delalloc blocks back to the old
1802          * dquot.  We don't normally handle delalloc quota reservations
1803          * transactionally, so just lock the dquot and subtract from the
1804          * reservation.  Dirty the transaction because it's too late to turn
1805          * back now.
1806          */
1807         tp->t_flags |= XFS_TRANS_DIRTY;
1808         xfs_dqlock(prevdq);
1809         ASSERT(prevdq->q_blk.reserved >= ip->i_delayed_blks);
1810         prevdq->q_blk.reserved -= ip->i_delayed_blks;
1811         xfs_dqunlock(prevdq);
1812
1813         /*
1814          * Take an extra reference, because the inode is going to keep
1815          * this dquot pointer even after the trans_commit.
1816          */
1817         *IO_olddq = xfs_qm_dqhold(newdq);
1818
1819         return prevdq;
1820 }
1821
1822 int
1823 xfs_qm_vop_rename_dqattach(
1824         struct xfs_inode        **i_tab)
1825 {
1826         struct xfs_mount        *mp = i_tab[0]->i_mount;
1827         int                     i;
1828
1829         if (!XFS_IS_QUOTA_ON(mp))
1830                 return 0;
1831
1832         for (i = 0; (i < 4 && i_tab[i]); i++) {
1833                 struct xfs_inode        *ip = i_tab[i];
1834                 int                     error;
1835
1836                 /*
1837                  * Watch out for duplicate entries in the table.
1838                  */
1839                 if (i == 0 || ip != i_tab[i-1]) {
1840                         if (XFS_NOT_DQATTACHED(mp, ip)) {
1841                                 error = xfs_qm_dqattach(ip);
1842                                 if (error)
1843                                         return error;
1844                         }
1845                 }
1846         }
1847         return 0;
1848 }
1849
1850 void
1851 xfs_qm_vop_create_dqattach(
1852         struct xfs_trans        *tp,
1853         struct xfs_inode        *ip,
1854         struct xfs_dquot        *udqp,
1855         struct xfs_dquot        *gdqp,
1856         struct xfs_dquot        *pdqp)
1857 {
1858         struct xfs_mount        *mp = tp->t_mountp;
1859
1860         if (!XFS_IS_QUOTA_ON(mp))
1861                 return;
1862
1863         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1864
1865         if (udqp && XFS_IS_UQUOTA_ON(mp)) {
1866                 ASSERT(ip->i_udquot == NULL);
1867                 ASSERT(i_uid_read(VFS_I(ip)) == udqp->q_id);
1868
1869                 ip->i_udquot = xfs_qm_dqhold(udqp);
1870                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
1871         }
1872         if (gdqp && XFS_IS_GQUOTA_ON(mp)) {
1873                 ASSERT(ip->i_gdquot == NULL);
1874                 ASSERT(i_gid_read(VFS_I(ip)) == gdqp->q_id);
1875
1876                 ip->i_gdquot = xfs_qm_dqhold(gdqp);
1877                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
1878         }
1879         if (pdqp && XFS_IS_PQUOTA_ON(mp)) {
1880                 ASSERT(ip->i_pdquot == NULL);
1881                 ASSERT(ip->i_projid == pdqp->q_id);
1882
1883                 ip->i_pdquot = xfs_qm_dqhold(pdqp);
1884                 xfs_trans_mod_dquot(tp, pdqp, XFS_TRANS_DQ_ICOUNT, 1);
1885         }
1886 }
1887
1888 /* Decide if this inode's dquot is near an enforcement boundary. */
1889 bool
1890 xfs_inode_near_dquot_enforcement(
1891         struct xfs_inode        *ip,
1892         xfs_dqtype_t            type)
1893 {
1894         struct xfs_dquot        *dqp;
1895         int64_t                 freesp;
1896
1897         /* We only care for quotas that are enabled and enforced. */
1898         dqp = xfs_inode_dquot(ip, type);
1899         if (!dqp || !xfs_dquot_is_enforced(dqp))
1900                 return false;
1901
1902         if (xfs_dquot_res_over_limits(&dqp->q_ino) ||
1903             xfs_dquot_res_over_limits(&dqp->q_rtb))
1904                 return true;
1905
1906         /* For space on the data device, check the various thresholds. */
1907         if (!dqp->q_prealloc_hi_wmark)
1908                 return false;
1909
1910         if (dqp->q_blk.reserved < dqp->q_prealloc_lo_wmark)
1911                 return false;
1912
1913         if (dqp->q_blk.reserved >= dqp->q_prealloc_hi_wmark)
1914                 return true;
1915
1916         freesp = dqp->q_prealloc_hi_wmark - dqp->q_blk.reserved;
1917         if (freesp < dqp->q_low_space[XFS_QLOWSP_5_PCNT])
1918                 return true;
1919
1920         return false;
1921 }