GNU Linux-libre 5.19-rc6-gnu
[releases.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);
681         if (error)
682                 goto out_free_inos;
683
684         return 0;
685
686 out_free_inos:
687         mutex_destroy(&qinf->qi_quotaofflock);
688         mutex_destroy(&qinf->qi_tree_lock);
689         xfs_qm_destroy_quotainos(qinf);
690 out_free_lru:
691         list_lru_destroy(&qinf->qi_lru);
692 out_free_qinf:
693         kmem_free(qinf);
694         mp->m_quotainfo = NULL;
695         return error;
696 }
697
698 /*
699  * Gets called when unmounting a filesystem or when all quotas get
700  * turned off.
701  * This purges the quota inodes, destroys locks and frees itself.
702  */
703 void
704 xfs_qm_destroy_quotainfo(
705         struct xfs_mount        *mp)
706 {
707         struct xfs_quotainfo    *qi;
708
709         qi = mp->m_quotainfo;
710         ASSERT(qi != NULL);
711
712         unregister_shrinker(&qi->qi_shrinker);
713         list_lru_destroy(&qi->qi_lru);
714         xfs_qm_destroy_quotainos(qi);
715         mutex_destroy(&qi->qi_tree_lock);
716         mutex_destroy(&qi->qi_quotaofflock);
717         kmem_free(qi);
718         mp->m_quotainfo = NULL;
719 }
720
721 /*
722  * Create an inode and return with a reference already taken, but unlocked
723  * This is how we create quota inodes
724  */
725 STATIC int
726 xfs_qm_qino_alloc(
727         struct xfs_mount        *mp,
728         struct xfs_inode        **ipp,
729         unsigned int            flags)
730 {
731         struct xfs_trans        *tp;
732         int                     error;
733         bool                    need_alloc = true;
734
735         *ipp = NULL;
736         /*
737          * With superblock that doesn't have separate pquotino, we
738          * share an inode between gquota and pquota. If the on-disk
739          * superblock has GQUOTA and the filesystem is now mounted
740          * with PQUOTA, just use sb_gquotino for sb_pquotino and
741          * vice-versa.
742          */
743         if (!xfs_has_pquotino(mp) &&
744                         (flags & (XFS_QMOPT_PQUOTA|XFS_QMOPT_GQUOTA))) {
745                 xfs_ino_t ino = NULLFSINO;
746
747                 if ((flags & XFS_QMOPT_PQUOTA) &&
748                              (mp->m_sb.sb_gquotino != NULLFSINO)) {
749                         ino = mp->m_sb.sb_gquotino;
750                         if (XFS_IS_CORRUPT(mp,
751                                            mp->m_sb.sb_pquotino != NULLFSINO))
752                                 return -EFSCORRUPTED;
753                 } else if ((flags & XFS_QMOPT_GQUOTA) &&
754                              (mp->m_sb.sb_pquotino != NULLFSINO)) {
755                         ino = mp->m_sb.sb_pquotino;
756                         if (XFS_IS_CORRUPT(mp,
757                                            mp->m_sb.sb_gquotino != NULLFSINO))
758                                 return -EFSCORRUPTED;
759                 }
760                 if (ino != NULLFSINO) {
761                         error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
762                         if (error)
763                                 return error;
764                         mp->m_sb.sb_gquotino = NULLFSINO;
765                         mp->m_sb.sb_pquotino = NULLFSINO;
766                         need_alloc = false;
767                 }
768         }
769
770         error = xfs_trans_alloc(mp, &M_RES(mp)->tr_create,
771                         need_alloc ? XFS_QM_QINOCREATE_SPACE_RES(mp) : 0,
772                         0, 0, &tp);
773         if (error)
774                 return error;
775
776         if (need_alloc) {
777                 xfs_ino_t       ino;
778
779                 error = xfs_dialloc(&tp, 0, S_IFREG, &ino);
780                 if (!error)
781                         error = xfs_init_new_inode(&init_user_ns, tp, NULL, ino,
782                                         S_IFREG, 1, 0, 0, false, ipp);
783                 if (error) {
784                         xfs_trans_cancel(tp);
785                         return error;
786                 }
787         }
788
789         /*
790          * Make the changes in the superblock, and log those too.
791          * sbfields arg may contain fields other than *QUOTINO;
792          * VERSIONNUM for example.
793          */
794         spin_lock(&mp->m_sb_lock);
795         if (flags & XFS_QMOPT_SBVERSION) {
796                 ASSERT(!xfs_has_quota(mp));
797
798                 xfs_add_quota(mp);
799                 mp->m_sb.sb_uquotino = NULLFSINO;
800                 mp->m_sb.sb_gquotino = NULLFSINO;
801                 mp->m_sb.sb_pquotino = NULLFSINO;
802
803                 /* qflags will get updated fully _after_ quotacheck */
804                 mp->m_sb.sb_qflags = mp->m_qflags & XFS_ALL_QUOTA_ACCT;
805         }
806         if (flags & XFS_QMOPT_UQUOTA)
807                 mp->m_sb.sb_uquotino = (*ipp)->i_ino;
808         else if (flags & XFS_QMOPT_GQUOTA)
809                 mp->m_sb.sb_gquotino = (*ipp)->i_ino;
810         else
811                 mp->m_sb.sb_pquotino = (*ipp)->i_ino;
812         spin_unlock(&mp->m_sb_lock);
813         xfs_log_sb(tp);
814
815         error = xfs_trans_commit(tp);
816         if (error) {
817                 ASSERT(xfs_is_shutdown(mp));
818                 xfs_alert(mp, "%s failed (error %d)!", __func__, error);
819         }
820         if (need_alloc)
821                 xfs_finish_inode_setup(*ipp);
822         return error;
823 }
824
825
826 STATIC void
827 xfs_qm_reset_dqcounts(
828         struct xfs_mount        *mp,
829         struct xfs_buf          *bp,
830         xfs_dqid_t              id,
831         xfs_dqtype_t            type)
832 {
833         struct xfs_dqblk        *dqb;
834         int                     j;
835
836         trace_xfs_reset_dqcounts(bp, _RET_IP_);
837
838         /*
839          * Reset all counters and timers. They'll be
840          * started afresh by xfs_qm_quotacheck.
841          */
842 #ifdef DEBUG
843         j = (int)XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB) /
844                 sizeof(struct xfs_dqblk);
845         ASSERT(mp->m_quotainfo->qi_dqperchunk == j);
846 #endif
847         dqb = bp->b_addr;
848         for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) {
849                 struct xfs_disk_dquot   *ddq;
850
851                 ddq = (struct xfs_disk_dquot *)&dqb[j];
852
853                 /*
854                  * Do a sanity check, and if needed, repair the dqblk. Don't
855                  * output any warnings because it's perfectly possible to
856                  * find uninitialised dquot blks. See comment in
857                  * xfs_dquot_verify.
858                  */
859                 if (xfs_dqblk_verify(mp, &dqb[j], id + j) ||
860                     (dqb[j].dd_diskdq.d_type & XFS_DQTYPE_REC_MASK) != type)
861                         xfs_dqblk_repair(mp, &dqb[j], id + j, type);
862
863                 /*
864                  * Reset type in case we are reusing group quota file for
865                  * project quotas or vice versa
866                  */
867                 ddq->d_type = type;
868                 ddq->d_bcount = 0;
869                 ddq->d_icount = 0;
870                 ddq->d_rtbcount = 0;
871
872                 /*
873                  * dquot id 0 stores the default grace period and the maximum
874                  * warning limit that were set by the administrator, so we
875                  * should not reset them.
876                  */
877                 if (ddq->d_id != 0) {
878                         ddq->d_btimer = 0;
879                         ddq->d_itimer = 0;
880                         ddq->d_rtbtimer = 0;
881                         ddq->d_bwarns = 0;
882                         ddq->d_iwarns = 0;
883                         ddq->d_rtbwarns = 0;
884                         if (xfs_has_bigtime(mp))
885                                 ddq->d_type |= XFS_DQTYPE_BIGTIME;
886                 }
887
888                 if (xfs_has_crc(mp)) {
889                         xfs_update_cksum((char *)&dqb[j],
890                                          sizeof(struct xfs_dqblk),
891                                          XFS_DQUOT_CRC_OFF);
892                 }
893         }
894 }
895
896 STATIC int
897 xfs_qm_reset_dqcounts_all(
898         struct xfs_mount        *mp,
899         xfs_dqid_t              firstid,
900         xfs_fsblock_t           bno,
901         xfs_filblks_t           blkcnt,
902         xfs_dqtype_t            type,
903         struct list_head        *buffer_list)
904 {
905         struct xfs_buf          *bp;
906         int                     error = 0;
907
908         ASSERT(blkcnt > 0);
909
910         /*
911          * Blkcnt arg can be a very big number, and might even be
912          * larger than the log itself. So, we have to break it up into
913          * manageable-sized transactions.
914          * Note that we don't start a permanent transaction here; we might
915          * not be able to get a log reservation for the whole thing up front,
916          * and we don't really care to either, because we just discard
917          * everything if we were to crash in the middle of this loop.
918          */
919         while (blkcnt--) {
920                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
921                               XFS_FSB_TO_DADDR(mp, bno),
922                               mp->m_quotainfo->qi_dqchunklen, 0, &bp,
923                               &xfs_dquot_buf_ops);
924
925                 /*
926                  * CRC and validation errors will return a EFSCORRUPTED here. If
927                  * this occurs, re-read without CRC validation so that we can
928                  * repair the damage via xfs_qm_reset_dqcounts(). This process
929                  * will leave a trace in the log indicating corruption has
930                  * been detected.
931                  */
932                 if (error == -EFSCORRUPTED) {
933                         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
934                                       XFS_FSB_TO_DADDR(mp, bno),
935                                       mp->m_quotainfo->qi_dqchunklen, 0, &bp,
936                                       NULL);
937                 }
938
939                 if (error)
940                         break;
941
942                 /*
943                  * A corrupt buffer might not have a verifier attached, so
944                  * make sure we have the correct one attached before writeback
945                  * occurs.
946                  */
947                 bp->b_ops = &xfs_dquot_buf_ops;
948                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
949                 xfs_buf_delwri_queue(bp, buffer_list);
950                 xfs_buf_relse(bp);
951
952                 /* goto the next block. */
953                 bno++;
954                 firstid += mp->m_quotainfo->qi_dqperchunk;
955         }
956
957         return error;
958 }
959
960 /*
961  * Iterate over all allocated dquot blocks in this quota inode, zeroing all
962  * counters for every chunk of dquots that we find.
963  */
964 STATIC int
965 xfs_qm_reset_dqcounts_buf(
966         struct xfs_mount        *mp,
967         struct xfs_inode        *qip,
968         xfs_dqtype_t            type,
969         struct list_head        *buffer_list)
970 {
971         struct xfs_bmbt_irec    *map;
972         int                     i, nmaps;       /* number of map entries */
973         int                     error;          /* return value */
974         xfs_fileoff_t           lblkno;
975         xfs_filblks_t           maxlblkcnt;
976         xfs_dqid_t              firstid;
977         xfs_fsblock_t           rablkno;
978         xfs_filblks_t           rablkcnt;
979
980         error = 0;
981         /*
982          * This looks racy, but we can't keep an inode lock across a
983          * trans_reserve. But, this gets called during quotacheck, and that
984          * happens only at mount time which is single threaded.
985          */
986         if (qip->i_nblocks == 0)
987                 return 0;
988
989         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), 0);
990
991         lblkno = 0;
992         maxlblkcnt = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
993         do {
994                 uint            lock_mode;
995
996                 nmaps = XFS_DQITER_MAP_SIZE;
997                 /*
998                  * We aren't changing the inode itself. Just changing
999                  * some of its data. No new blocks are added here, and
1000                  * the inode is never added to the transaction.
1001                  */
1002                 lock_mode = xfs_ilock_data_map_shared(qip);
1003                 error = xfs_bmapi_read(qip, lblkno, maxlblkcnt - lblkno,
1004                                        map, &nmaps, 0);
1005                 xfs_iunlock(qip, lock_mode);
1006                 if (error)
1007                         break;
1008
1009                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1010                 for (i = 0; i < nmaps; i++) {
1011                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1012                         ASSERT(map[i].br_blockcount);
1013
1014
1015                         lblkno += map[i].br_blockcount;
1016
1017                         if (map[i].br_startblock == HOLESTARTBLOCK)
1018                                 continue;
1019
1020                         firstid = (xfs_dqid_t) map[i].br_startoff *
1021                                 mp->m_quotainfo->qi_dqperchunk;
1022                         /*
1023                          * Do a read-ahead on the next extent.
1024                          */
1025                         if ((i+1 < nmaps) &&
1026                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1027                                 rablkcnt =  map[i+1].br_blockcount;
1028                                 rablkno = map[i+1].br_startblock;
1029                                 while (rablkcnt--) {
1030                                         xfs_buf_readahead(mp->m_ddev_targp,
1031                                                XFS_FSB_TO_DADDR(mp, rablkno),
1032                                                mp->m_quotainfo->qi_dqchunklen,
1033                                                &xfs_dquot_buf_ops);
1034                                         rablkno++;
1035                                 }
1036                         }
1037                         /*
1038                          * Iterate thru all the blks in the extent and
1039                          * reset the counters of all the dquots inside them.
1040                          */
1041                         error = xfs_qm_reset_dqcounts_all(mp, firstid,
1042                                                    map[i].br_startblock,
1043                                                    map[i].br_blockcount,
1044                                                    type, buffer_list);
1045                         if (error)
1046                                 goto out;
1047                 }
1048         } while (nmaps > 0);
1049
1050 out:
1051         kmem_free(map);
1052         return error;
1053 }
1054
1055 /*
1056  * Called by dqusage_adjust in doing a quotacheck.
1057  *
1058  * Given the inode, and a dquot id this updates both the incore dqout as well
1059  * as the buffer copy. This is so that once the quotacheck is done, we can
1060  * just log all the buffers, as opposed to logging numerous updates to
1061  * individual dquots.
1062  */
1063 STATIC int
1064 xfs_qm_quotacheck_dqadjust(
1065         struct xfs_inode        *ip,
1066         xfs_dqtype_t            type,
1067         xfs_qcnt_t              nblks,
1068         xfs_qcnt_t              rtblks)
1069 {
1070         struct xfs_mount        *mp = ip->i_mount;
1071         struct xfs_dquot        *dqp;
1072         xfs_dqid_t              id;
1073         int                     error;
1074
1075         id = xfs_qm_id_for_quotatype(ip, type);
1076         error = xfs_qm_dqget(mp, id, type, true, &dqp);
1077         if (error) {
1078                 /*
1079                  * Shouldn't be able to turn off quotas here.
1080                  */
1081                 ASSERT(error != -ESRCH);
1082                 ASSERT(error != -ENOENT);
1083                 return error;
1084         }
1085
1086         trace_xfs_dqadjust(dqp);
1087
1088         /*
1089          * Adjust the inode count and the block count to reflect this inode's
1090          * resource usage.
1091          */
1092         dqp->q_ino.count++;
1093         dqp->q_ino.reserved++;
1094         if (nblks) {
1095                 dqp->q_blk.count += nblks;
1096                 dqp->q_blk.reserved += nblks;
1097         }
1098         if (rtblks) {
1099                 dqp->q_rtb.count += rtblks;
1100                 dqp->q_rtb.reserved += rtblks;
1101         }
1102
1103         /*
1104          * Set default limits, adjust timers (since we changed usages)
1105          *
1106          * There are no timers for the default values set in the root dquot.
1107          */
1108         if (dqp->q_id) {
1109                 xfs_qm_adjust_dqlimits(dqp);
1110                 xfs_qm_adjust_dqtimers(dqp);
1111         }
1112
1113         dqp->q_flags |= XFS_DQFLAG_DIRTY;
1114         xfs_qm_dqput(dqp);
1115         return 0;
1116 }
1117
1118 /*
1119  * callback routine supplied to bulkstat(). Given an inumber, find its
1120  * dquots and update them to account for resources taken by that inode.
1121  */
1122 /* ARGSUSED */
1123 STATIC int
1124 xfs_qm_dqusage_adjust(
1125         struct xfs_mount        *mp,
1126         struct xfs_trans        *tp,
1127         xfs_ino_t               ino,
1128         void                    *data)
1129 {
1130         struct xfs_inode        *ip;
1131         xfs_qcnt_t              nblks;
1132         xfs_filblks_t           rtblks = 0;     /* total rt blks */
1133         int                     error;
1134
1135         ASSERT(XFS_IS_QUOTA_ON(mp));
1136
1137         /*
1138          * rootino must have its resources accounted for, not so with the quota
1139          * inodes.
1140          */
1141         if (xfs_is_quota_inode(&mp->m_sb, ino))
1142                 return 0;
1143
1144         /*
1145          * We don't _need_ to take the ilock EXCL here because quotacheck runs
1146          * at mount time and therefore nobody will be racing chown/chproj.
1147          */
1148         error = xfs_iget(mp, tp, ino, XFS_IGET_DONTCACHE, 0, &ip);
1149         if (error == -EINVAL || error == -ENOENT)
1150                 return 0;
1151         if (error)
1152                 return error;
1153
1154         ASSERT(ip->i_delayed_blks == 0);
1155
1156         if (XFS_IS_REALTIME_INODE(ip)) {
1157                 struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1158
1159                 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1160                 if (error)
1161                         goto error0;
1162
1163                 xfs_bmap_count_leaves(ifp, &rtblks);
1164         }
1165
1166         nblks = (xfs_qcnt_t)ip->i_nblocks - rtblks;
1167
1168         /*
1169          * Add the (disk blocks and inode) resources occupied by this
1170          * inode to its dquots. We do this adjustment in the incore dquot,
1171          * and also copy the changes to its buffer.
1172          * We don't care about putting these changes in a transaction
1173          * envelope because if we crash in the middle of a 'quotacheck'
1174          * we have to start from the beginning anyway.
1175          * Once we're done, we'll log all the dquot bufs.
1176          *
1177          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1178          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1179          */
1180         if (XFS_IS_UQUOTA_ON(mp)) {
1181                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_USER, nblks,
1182                                 rtblks);
1183                 if (error)
1184                         goto error0;
1185         }
1186
1187         if (XFS_IS_GQUOTA_ON(mp)) {
1188                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_GROUP, nblks,
1189                                 rtblks);
1190                 if (error)
1191                         goto error0;
1192         }
1193
1194         if (XFS_IS_PQUOTA_ON(mp)) {
1195                 error = xfs_qm_quotacheck_dqadjust(ip, XFS_DQTYPE_PROJ, nblks,
1196                                 rtblks);
1197                 if (error)
1198                         goto error0;
1199         }
1200
1201 error0:
1202         xfs_irele(ip);
1203         return error;
1204 }
1205
1206 STATIC int
1207 xfs_qm_flush_one(
1208         struct xfs_dquot        *dqp,
1209         void                    *data)
1210 {
1211         struct xfs_mount        *mp = dqp->q_mount;
1212         struct list_head        *buffer_list = data;
1213         struct xfs_buf          *bp = NULL;
1214         int                     error = 0;
1215
1216         xfs_dqlock(dqp);
1217         if (dqp->q_flags & XFS_DQFLAG_FREEING)
1218                 goto out_unlock;
1219         if (!XFS_DQ_IS_DIRTY(dqp))
1220                 goto out_unlock;
1221
1222         /*
1223          * The only way the dquot is already flush locked by the time quotacheck
1224          * gets here is if reclaim flushed it before the dqadjust walk dirtied
1225          * it for the final time. Quotacheck collects all dquot bufs in the
1226          * local delwri queue before dquots are dirtied, so reclaim can't have
1227          * possibly queued it for I/O. The only way out is to push the buffer to
1228          * cycle the flush lock.
1229          */
1230         if (!xfs_dqflock_nowait(dqp)) {
1231                 /* buf is pinned in-core by delwri list */
1232                 bp = xfs_buf_incore(mp->m_ddev_targp, dqp->q_blkno,
1233                                 mp->m_quotainfo->qi_dqchunklen, 0);
1234                 if (!bp) {
1235                         error = -EINVAL;
1236                         goto out_unlock;
1237                 }
1238                 xfs_buf_unlock(bp);
1239
1240                 xfs_buf_delwri_pushbuf(bp, buffer_list);
1241                 xfs_buf_rele(bp);
1242
1243                 error = -EAGAIN;
1244                 goto out_unlock;
1245         }
1246
1247         error = xfs_qm_dqflush(dqp, &bp);
1248         if (error)
1249                 goto out_unlock;
1250
1251         xfs_buf_delwri_queue(bp, buffer_list);
1252         xfs_buf_relse(bp);
1253 out_unlock:
1254         xfs_dqunlock(dqp);
1255         return error;
1256 }
1257
1258 /*
1259  * Walk thru all the filesystem inodes and construct a consistent view
1260  * of the disk quota world. If the quotacheck fails, disable quotas.
1261  */
1262 STATIC int
1263 xfs_qm_quotacheck(
1264         xfs_mount_t     *mp)
1265 {
1266         int                     error, error2;
1267         uint                    flags;
1268         LIST_HEAD               (buffer_list);
1269         struct xfs_inode        *uip = mp->m_quotainfo->qi_uquotaip;
1270         struct xfs_inode        *gip = mp->m_quotainfo->qi_gquotaip;
1271         struct xfs_inode        *pip = mp->m_quotainfo->qi_pquotaip;
1272
1273         flags = 0;
1274
1275         ASSERT(uip || gip || pip);
1276         ASSERT(XFS_IS_QUOTA_ON(mp));
1277
1278         xfs_notice(mp, "Quotacheck needed: Please wait.");
1279
1280         /*
1281          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1282          * their counters to zero. We need a clean slate.
1283          * We don't log our changes till later.
1284          */
1285         if (uip) {
1286                 error = xfs_qm_reset_dqcounts_buf(mp, uip, XFS_DQTYPE_USER,
1287                                          &buffer_list);
1288                 if (error)
1289                         goto error_return;
1290                 flags |= XFS_UQUOTA_CHKD;
1291         }
1292
1293         if (gip) {
1294                 error = xfs_qm_reset_dqcounts_buf(mp, gip, XFS_DQTYPE_GROUP,
1295                                          &buffer_list);
1296                 if (error)
1297                         goto error_return;
1298                 flags |= XFS_GQUOTA_CHKD;
1299         }
1300
1301         if (pip) {
1302                 error = xfs_qm_reset_dqcounts_buf(mp, pip, XFS_DQTYPE_PROJ,
1303                                          &buffer_list);
1304                 if (error)
1305                         goto error_return;
1306                 flags |= XFS_PQUOTA_CHKD;
1307         }
1308
1309         error = xfs_iwalk_threaded(mp, 0, 0, xfs_qm_dqusage_adjust, 0, true,
1310                         NULL);
1311         if (error) {
1312                 /*
1313                  * The inode walk may have partially populated the dquot
1314                  * caches.  We must purge them before disabling quota and
1315                  * tearing down the quotainfo, or else the dquots will leak.
1316                  */
1317                 xfs_qm_dqpurge_all(mp);
1318                 goto error_return;
1319         }
1320
1321         /*
1322          * We've made all the changes that we need to make incore.  Flush them
1323          * down to disk buffers if everything was updated successfully.
1324          */
1325         if (XFS_IS_UQUOTA_ON(mp)) {
1326                 error = xfs_qm_dquot_walk(mp, XFS_DQTYPE_USER, xfs_qm_flush_one,
1327                                           &buffer_list);
1328         }
1329         if (XFS_IS_GQUOTA_ON(mp)) {
1330                 error2 = xfs_qm_dquot_walk(mp, XFS_DQTYPE_GROUP, xfs_qm_flush_one,
1331                                            &buffer_list);
1332                 if (!error)
1333                         error = error2;
1334         }
1335         if (XFS_IS_PQUOTA_ON(mp)) {
1336                 error2 = xfs_qm_dquot_walk(mp, XFS_DQTYPE_PROJ, xfs_qm_flush_one,
1337                                            &buffer_list);
1338                 if (!error)
1339                         error = error2;
1340         }
1341
1342         error2 = xfs_buf_delwri_submit(&buffer_list);
1343         if (!error)
1344                 error = error2;
1345
1346         /*
1347          * We can get this error if we couldn't do a dquot allocation inside
1348          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1349          * dirty dquots that might be cached, we just want to get rid of them
1350          * and turn quotaoff. The dquots won't be attached to any of the inodes
1351          * at this point (because we intentionally didn't in dqget_noattach).
1352          */
1353         if (error) {
1354                 xfs_qm_dqpurge_all(mp);
1355                 goto error_return;
1356         }
1357
1358         /*
1359          * If one type of quotas is off, then it will lose its
1360          * quotachecked status, since we won't be doing accounting for
1361          * that type anymore.
1362          */
1363         mp->m_qflags &= ~XFS_ALL_QUOTA_CHKD;
1364         mp->m_qflags |= flags;
1365
1366  error_return:
1367         xfs_buf_delwri_cancel(&buffer_list);
1368
1369         if (error) {
1370                 xfs_warn(mp,
1371         "Quotacheck: Unsuccessful (Error %d): Disabling quotas.",
1372                         error);
1373                 /*
1374                  * We must turn off quotas.
1375                  */
1376                 ASSERT(mp->m_quotainfo != NULL);
1377                 xfs_qm_destroy_quotainfo(mp);
1378                 if (xfs_mount_reset_sbqflags(mp)) {
1379                         xfs_warn(mp,
1380                                 "Quotacheck: Failed to reset quota flags.");
1381                 }
1382         } else
1383                 xfs_notice(mp, "Quotacheck: Done.");
1384         return error;
1385 }
1386
1387 /*
1388  * This is called from xfs_mountfs to start quotas and initialize all
1389  * necessary data structures like quotainfo.  This is also responsible for
1390  * running a quotacheck as necessary.  We are guaranteed that the superblock
1391  * is consistently read in at this point.
1392  *
1393  * If we fail here, the mount will continue with quota turned off. We don't
1394  * need to inidicate success or failure at all.
1395  */
1396 void
1397 xfs_qm_mount_quotas(
1398         struct xfs_mount        *mp)
1399 {
1400         int                     error = 0;
1401         uint                    sbf;
1402
1403         /*
1404          * If quotas on realtime volumes is not supported, we disable
1405          * quotas immediately.
1406          */
1407         if (mp->m_sb.sb_rextents) {
1408                 xfs_notice(mp, "Cannot turn on quotas for realtime filesystem");
1409                 mp->m_qflags = 0;
1410                 goto write_changes;
1411         }
1412
1413         ASSERT(XFS_IS_QUOTA_ON(mp));
1414
1415         /*
1416          * Allocate the quotainfo structure inside the mount struct, and
1417          * create quotainode(s), and change/rev superblock if necessary.
1418          */
1419         error = xfs_qm_init_quotainfo(mp);
1420         if (error) {
1421                 /*
1422                  * We must turn off quotas.
1423                  */
1424                 ASSERT(mp->m_quotainfo == NULL);
1425                 mp->m_qflags = 0;
1426                 goto write_changes;
1427         }
1428         /*
1429          * If any of the quotas are not consistent, do a quotacheck.
1430          */
1431         if (XFS_QM_NEED_QUOTACHECK(mp)) {
1432                 error = xfs_qm_quotacheck(mp);
1433                 if (error) {
1434                         /* Quotacheck failed and disabled quotas. */
1435                         return;
1436                 }
1437         }
1438         /*
1439          * If one type of quotas is off, then it will lose its
1440          * quotachecked status, since we won't be doing accounting for
1441          * that type anymore.
1442          */
1443         if (!XFS_IS_UQUOTA_ON(mp))
1444                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
1445         if (!XFS_IS_GQUOTA_ON(mp))
1446                 mp->m_qflags &= ~XFS_GQUOTA_CHKD;
1447         if (!XFS_IS_PQUOTA_ON(mp))
1448                 mp->m_qflags &= ~XFS_PQUOTA_CHKD;
1449
1450  write_changes:
1451         /*
1452          * We actually don't have to acquire the m_sb_lock at all.
1453          * This can only be called from mount, and that's single threaded. XXX
1454          */
1455         spin_lock(&mp->m_sb_lock);
1456         sbf = mp->m_sb.sb_qflags;
1457         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
1458         spin_unlock(&mp->m_sb_lock);
1459
1460         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
1461                 if (xfs_sync_sb(mp, false)) {
1462                         /*
1463                          * We could only have been turning quotas off.
1464                          * We aren't in very good shape actually because
1465                          * the incore structures are convinced that quotas are
1466                          * off, but the on disk superblock doesn't know that !
1467                          */
1468                         ASSERT(!(XFS_IS_QUOTA_ON(mp)));
1469                         xfs_alert(mp, "%s: Superblock update failed!",
1470                                 __func__);
1471                 }
1472         }
1473
1474         if (error) {
1475                 xfs_warn(mp, "Failed to initialize disk quotas.");
1476                 return;
1477         }
1478 }
1479
1480 /*
1481  * This is called after the superblock has been read in and we're ready to
1482  * iget the quota inodes.
1483  */
1484 STATIC int
1485 xfs_qm_init_quotainos(
1486         xfs_mount_t     *mp)
1487 {
1488         struct xfs_inode        *uip = NULL;
1489         struct xfs_inode        *gip = NULL;
1490         struct xfs_inode        *pip = NULL;
1491         int                     error;
1492         uint                    flags = 0;
1493
1494         ASSERT(mp->m_quotainfo);
1495
1496         /*
1497          * Get the uquota and gquota inodes
1498          */
1499         if (xfs_has_quota(mp)) {
1500                 if (XFS_IS_UQUOTA_ON(mp) &&
1501                     mp->m_sb.sb_uquotino != NULLFSINO) {
1502                         ASSERT(mp->m_sb.sb_uquotino > 0);
1503                         error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1504                                              0, 0, &uip);
1505                         if (error)
1506                                 return error;
1507                 }
1508                 if (XFS_IS_GQUOTA_ON(mp) &&
1509                     mp->m_sb.sb_gquotino != NULLFSINO) {
1510                         ASSERT(mp->m_sb.sb_gquotino > 0);
1511                         error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1512                                              0, 0, &gip);
1513                         if (error)
1514                                 goto error_rele;
1515                 }
1516                 if (XFS_IS_PQUOTA_ON(mp) &&
1517                     mp->m_sb.sb_pquotino != NULLFSINO) {
1518                         ASSERT(mp->m_sb.sb_pquotino > 0);
1519                         error = xfs_iget(mp, NULL, mp->m_sb.sb_pquotino,
1520                                              0, 0, &pip);
1521                         if (error)
1522                                 goto error_rele;
1523                 }
1524         } else {
1525                 flags |= XFS_QMOPT_SBVERSION;
1526         }
1527
1528         /*
1529          * Create the three inodes, if they don't exist already. The changes
1530          * made above will get added to a transaction and logged in one of
1531          * the qino_alloc calls below.  If the device is readonly,
1532          * temporarily switch to read-write to do this.
1533          */
1534         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1535                 error = xfs_qm_qino_alloc(mp, &uip,
1536                                               flags | XFS_QMOPT_UQUOTA);
1537                 if (error)
1538                         goto error_rele;
1539
1540                 flags &= ~XFS_QMOPT_SBVERSION;
1541         }
1542         if (XFS_IS_GQUOTA_ON(mp) && gip == NULL) {
1543                 error = xfs_qm_qino_alloc(mp, &gip,
1544                                           flags | XFS_QMOPT_GQUOTA);
1545                 if (error)
1546                         goto error_rele;
1547
1548                 flags &= ~XFS_QMOPT_SBVERSION;
1549         }
1550         if (XFS_IS_PQUOTA_ON(mp) && pip == NULL) {
1551                 error = xfs_qm_qino_alloc(mp, &pip,
1552                                           flags | XFS_QMOPT_PQUOTA);
1553                 if (error)
1554                         goto error_rele;
1555         }
1556
1557         mp->m_quotainfo->qi_uquotaip = uip;
1558         mp->m_quotainfo->qi_gquotaip = gip;
1559         mp->m_quotainfo->qi_pquotaip = pip;
1560
1561         return 0;
1562
1563 error_rele:
1564         if (uip)
1565                 xfs_irele(uip);
1566         if (gip)
1567                 xfs_irele(gip);
1568         if (pip)
1569                 xfs_irele(pip);
1570         return error;
1571 }
1572
1573 STATIC void
1574 xfs_qm_destroy_quotainos(
1575         struct xfs_quotainfo    *qi)
1576 {
1577         if (qi->qi_uquotaip) {
1578                 xfs_irele(qi->qi_uquotaip);
1579                 qi->qi_uquotaip = NULL; /* paranoia */
1580         }
1581         if (qi->qi_gquotaip) {
1582                 xfs_irele(qi->qi_gquotaip);
1583                 qi->qi_gquotaip = NULL;
1584         }
1585         if (qi->qi_pquotaip) {
1586                 xfs_irele(qi->qi_pquotaip);
1587                 qi->qi_pquotaip = NULL;
1588         }
1589 }
1590
1591 STATIC void
1592 xfs_qm_dqfree_one(
1593         struct xfs_dquot        *dqp)
1594 {
1595         struct xfs_mount        *mp = dqp->q_mount;
1596         struct xfs_quotainfo    *qi = mp->m_quotainfo;
1597
1598         mutex_lock(&qi->qi_tree_lock);
1599         radix_tree_delete(xfs_dquot_tree(qi, xfs_dquot_type(dqp)), dqp->q_id);
1600
1601         qi->qi_dquots--;
1602         mutex_unlock(&qi->qi_tree_lock);
1603
1604         xfs_qm_dqdestroy(dqp);
1605 }
1606
1607 /* --------------- utility functions for vnodeops ---------------- */
1608
1609
1610 /*
1611  * Given an inode, a uid, gid and prid make sure that we have
1612  * allocated relevant dquot(s) on disk, and that we won't exceed inode
1613  * quotas by creating this file.
1614  * This also attaches dquot(s) to the given inode after locking it,
1615  * and returns the dquots corresponding to the uid and/or gid.
1616  *
1617  * in   : inode (unlocked)
1618  * out  : udquot, gdquot with references taken and unlocked
1619  */
1620 int
1621 xfs_qm_vop_dqalloc(
1622         struct xfs_inode        *ip,
1623         kuid_t                  uid,
1624         kgid_t                  gid,
1625         prid_t                  prid,
1626         uint                    flags,
1627         struct xfs_dquot        **O_udqpp,
1628         struct xfs_dquot        **O_gdqpp,
1629         struct xfs_dquot        **O_pdqpp)
1630 {
1631         struct xfs_mount        *mp = ip->i_mount;
1632         struct inode            *inode = VFS_I(ip);
1633         struct user_namespace   *user_ns = inode->i_sb->s_user_ns;
1634         struct xfs_dquot        *uq = NULL;
1635         struct xfs_dquot        *gq = NULL;
1636         struct xfs_dquot        *pq = NULL;
1637         int                     error;
1638         uint                    lockflags;
1639
1640         if (!XFS_IS_QUOTA_ON(mp))
1641                 return 0;
1642
1643         lockflags = XFS_ILOCK_EXCL;
1644         xfs_ilock(ip, lockflags);
1645
1646         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
1647                 gid = inode->i_gid;
1648
1649         /*
1650          * Attach the dquot(s) to this inode, doing a dquot allocation
1651          * if necessary. The dquot(s) will not be locked.
1652          */
1653         if (XFS_NOT_DQATTACHED(mp, ip)) {
1654                 error = xfs_qm_dqattach_locked(ip, true);
1655                 if (error) {
1656                         xfs_iunlock(ip, lockflags);
1657                         return error;
1658                 }
1659         }
1660
1661         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
1662                 ASSERT(O_udqpp);
1663                 if (!uid_eq(inode->i_uid, uid)) {
1664                         /*
1665                          * What we need is the dquot that has this uid, and
1666                          * if we send the inode to dqget, the uid of the inode
1667                          * takes priority over what's sent in the uid argument.
1668                          * We must unlock inode here before calling dqget if
1669                          * we're not sending the inode, because otherwise
1670                          * we'll deadlock by doing trans_reserve while
1671                          * holding ilock.
1672                          */
1673                         xfs_iunlock(ip, lockflags);
1674                         error = xfs_qm_dqget(mp, from_kuid(user_ns, uid),
1675                                         XFS_DQTYPE_USER, true, &uq);
1676                         if (error) {
1677                                 ASSERT(error != -ENOENT);
1678                                 return error;
1679                         }
1680                         /*
1681                          * Get the ilock in the right order.
1682                          */
1683                         xfs_dqunlock(uq);
1684                         lockflags = XFS_ILOCK_SHARED;
1685                         xfs_ilock(ip, lockflags);
1686                 } else {
1687                         /*
1688                          * Take an extra reference, because we'll return
1689                          * this to caller
1690                          */
1691                         ASSERT(ip->i_udquot);
1692                         uq = xfs_qm_dqhold(ip->i_udquot);
1693                 }
1694         }
1695         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
1696                 ASSERT(O_gdqpp);
1697                 if (!gid_eq(inode->i_gid, gid)) {
1698                         xfs_iunlock(ip, lockflags);
1699                         error = xfs_qm_dqget(mp, from_kgid(user_ns, gid),
1700                                         XFS_DQTYPE_GROUP, true, &gq);
1701                         if (error) {
1702                                 ASSERT(error != -ENOENT);
1703                                 goto error_rele;
1704                         }
1705                         xfs_dqunlock(gq);
1706                         lockflags = XFS_ILOCK_SHARED;
1707                         xfs_ilock(ip, lockflags);
1708                 } else {
1709                         ASSERT(ip->i_gdquot);
1710                         gq = xfs_qm_dqhold(ip->i_gdquot);
1711                 }
1712         }
1713         if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
1714                 ASSERT(O_pdqpp);
1715                 if (ip->i_projid != prid) {
1716                         xfs_iunlock(ip, lockflags);
1717                         error = xfs_qm_dqget(mp, prid,
1718                                         XFS_DQTYPE_PROJ, true, &pq);
1719                         if (error) {
1720                                 ASSERT(error != -ENOENT);
1721                                 goto error_rele;
1722                         }
1723                         xfs_dqunlock(pq);
1724                         lockflags = XFS_ILOCK_SHARED;
1725                         xfs_ilock(ip, lockflags);
1726                 } else {
1727                         ASSERT(ip->i_pdquot);
1728                         pq = xfs_qm_dqhold(ip->i_pdquot);
1729                 }
1730         }
1731         trace_xfs_dquot_dqalloc(ip);
1732
1733         xfs_iunlock(ip, lockflags);
1734         if (O_udqpp)
1735                 *O_udqpp = uq;
1736         else
1737                 xfs_qm_dqrele(uq);
1738         if (O_gdqpp)
1739                 *O_gdqpp = gq;
1740         else
1741                 xfs_qm_dqrele(gq);
1742         if (O_pdqpp)
1743                 *O_pdqpp = pq;
1744         else
1745                 xfs_qm_dqrele(pq);
1746         return 0;
1747
1748 error_rele:
1749         xfs_qm_dqrele(gq);
1750         xfs_qm_dqrele(uq);
1751         return error;
1752 }
1753
1754 /*
1755  * Actually transfer ownership, and do dquot modifications.
1756  * These were already reserved.
1757  */
1758 struct xfs_dquot *
1759 xfs_qm_vop_chown(
1760         struct xfs_trans        *tp,
1761         struct xfs_inode        *ip,
1762         struct xfs_dquot        **IO_olddq,
1763         struct xfs_dquot        *newdq)
1764 {
1765         struct xfs_dquot        *prevdq;
1766         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
1767                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
1768
1769
1770         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1771         ASSERT(XFS_IS_QUOTA_ON(ip->i_mount));
1772
1773         /* old dquot */
1774         prevdq = *IO_olddq;
1775         ASSERT(prevdq);
1776         ASSERT(prevdq != newdq);
1777
1778         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_nblocks));
1779         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
1780
1781         /* the sparkling new dquot */
1782         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_nblocks);
1783         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
1784
1785         /*
1786          * Back when we made quota reservations for the chown, we reserved the
1787          * ondisk blocks + delalloc blocks with the new dquot.  Now that we've
1788          * switched the dquots, decrease the new dquot's block reservation
1789          * (having already bumped up the real counter) so that we don't have
1790          * any reservation to give back when we commit.
1791          */
1792         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_RES_BLKS,
1793                         -ip->i_delayed_blks);
1794
1795         /*
1796          * Give the incore reservation for delalloc blocks back to the old
1797          * dquot.  We don't normally handle delalloc quota reservations
1798          * transactionally, so just lock the dquot and subtract from the
1799          * reservation.  Dirty the transaction because it's too late to turn
1800          * back now.
1801          */
1802         tp->t_flags |= XFS_TRANS_DIRTY;
1803         xfs_dqlock(prevdq);
1804         ASSERT(prevdq->q_blk.reserved >= ip->i_delayed_blks);
1805         prevdq->q_blk.reserved -= ip->i_delayed_blks;
1806         xfs_dqunlock(prevdq);
1807
1808         /*
1809          * Take an extra reference, because the inode is going to keep
1810          * this dquot pointer even after the trans_commit.
1811          */
1812         *IO_olddq = xfs_qm_dqhold(newdq);
1813
1814         return prevdq;
1815 }
1816
1817 int
1818 xfs_qm_vop_rename_dqattach(
1819         struct xfs_inode        **i_tab)
1820 {
1821         struct xfs_mount        *mp = i_tab[0]->i_mount;
1822         int                     i;
1823
1824         if (!XFS_IS_QUOTA_ON(mp))
1825                 return 0;
1826
1827         for (i = 0; (i < 4 && i_tab[i]); i++) {
1828                 struct xfs_inode        *ip = i_tab[i];
1829                 int                     error;
1830
1831                 /*
1832                  * Watch out for duplicate entries in the table.
1833                  */
1834                 if (i == 0 || ip != i_tab[i-1]) {
1835                         if (XFS_NOT_DQATTACHED(mp, ip)) {
1836                                 error = xfs_qm_dqattach(ip);
1837                                 if (error)
1838                                         return error;
1839                         }
1840                 }
1841         }
1842         return 0;
1843 }
1844
1845 void
1846 xfs_qm_vop_create_dqattach(
1847         struct xfs_trans        *tp,
1848         struct xfs_inode        *ip,
1849         struct xfs_dquot        *udqp,
1850         struct xfs_dquot        *gdqp,
1851         struct xfs_dquot        *pdqp)
1852 {
1853         struct xfs_mount        *mp = tp->t_mountp;
1854
1855         if (!XFS_IS_QUOTA_ON(mp))
1856                 return;
1857
1858         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1859
1860         if (udqp && XFS_IS_UQUOTA_ON(mp)) {
1861                 ASSERT(ip->i_udquot == NULL);
1862                 ASSERT(i_uid_read(VFS_I(ip)) == udqp->q_id);
1863
1864                 ip->i_udquot = xfs_qm_dqhold(udqp);
1865                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
1866         }
1867         if (gdqp && XFS_IS_GQUOTA_ON(mp)) {
1868                 ASSERT(ip->i_gdquot == NULL);
1869                 ASSERT(i_gid_read(VFS_I(ip)) == gdqp->q_id);
1870
1871                 ip->i_gdquot = xfs_qm_dqhold(gdqp);
1872                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
1873         }
1874         if (pdqp && XFS_IS_PQUOTA_ON(mp)) {
1875                 ASSERT(ip->i_pdquot == NULL);
1876                 ASSERT(ip->i_projid == pdqp->q_id);
1877
1878                 ip->i_pdquot = xfs_qm_dqhold(pdqp);
1879                 xfs_trans_mod_dquot(tp, pdqp, XFS_TRANS_DQ_ICOUNT, 1);
1880         }
1881 }
1882
1883 /* Decide if this inode's dquot is near an enforcement boundary. */
1884 bool
1885 xfs_inode_near_dquot_enforcement(
1886         struct xfs_inode        *ip,
1887         xfs_dqtype_t            type)
1888 {
1889         struct xfs_dquot        *dqp;
1890         int64_t                 freesp;
1891
1892         /* We only care for quotas that are enabled and enforced. */
1893         dqp = xfs_inode_dquot(ip, type);
1894         if (!dqp || !xfs_dquot_is_enforced(dqp))
1895                 return false;
1896
1897         if (xfs_dquot_res_over_limits(&dqp->q_ino) ||
1898             xfs_dquot_res_over_limits(&dqp->q_rtb))
1899                 return true;
1900
1901         /* For space on the data device, check the various thresholds. */
1902         if (!dqp->q_prealloc_hi_wmark)
1903                 return false;
1904
1905         if (dqp->q_blk.reserved < dqp->q_prealloc_lo_wmark)
1906                 return false;
1907
1908         if (dqp->q_blk.reserved >= dqp->q_prealloc_hi_wmark)
1909                 return true;
1910
1911         freesp = dqp->q_prealloc_hi_wmark - dqp->q_blk.reserved;
1912         if (freesp < dqp->q_low_space[XFS_QLOWSP_5_PCNT])
1913                 return true;
1914
1915         return false;
1916 }