f89f7b5241e68537e0231112a295d17b353f3141
[muen/linux.git] / fs / xfs / xfs_dquot.c
1 /*
2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_shared.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_bit.h"
25 #include "xfs_mount.h"
26 #include "xfs_defer.h"
27 #include "xfs_inode.h"
28 #include "xfs_bmap.h"
29 #include "xfs_bmap_util.h"
30 #include "xfs_alloc.h"
31 #include "xfs_quota.h"
32 #include "xfs_error.h"
33 #include "xfs_trans.h"
34 #include "xfs_buf_item.h"
35 #include "xfs_trans_space.h"
36 #include "xfs_trans_priv.h"
37 #include "xfs_qm.h"
38 #include "xfs_cksum.h"
39 #include "xfs_trace.h"
40 #include "xfs_log.h"
41 #include "xfs_bmap_btree.h"
42
43 /*
44  * Lock order:
45  *
46  * ip->i_lock
47  *   qi->qi_tree_lock
48  *     dquot->q_qlock (xfs_dqlock() and friends)
49  *       dquot->q_flush (xfs_dqflock() and friends)
50  *       qi->qi_lru_lock
51  *
52  * If two dquots need to be locked the order is user before group/project,
53  * otherwise by the lowest id first, see xfs_dqlock2.
54  */
55
56 #ifdef DEBUG
57 xfs_buftarg_t *xfs_dqerror_target;
58 int xfs_do_dqerror;
59 int xfs_dqreq_num;
60 int xfs_dqerror_mod = 33;
61 #endif
62
63 struct kmem_zone                *xfs_qm_dqtrxzone;
64 static struct kmem_zone         *xfs_qm_dqzone;
65
66 static struct lock_class_key xfs_dquot_group_class;
67 static struct lock_class_key xfs_dquot_project_class;
68
69 /*
70  * This is called to free all the memory associated with a dquot
71  */
72 void
73 xfs_qm_dqdestroy(
74         xfs_dquot_t     *dqp)
75 {
76         ASSERT(list_empty(&dqp->q_lru));
77
78         kmem_free(dqp->q_logitem.qli_item.li_lv_shadow);
79         mutex_destroy(&dqp->q_qlock);
80
81         XFS_STATS_DEC(dqp->q_mount, xs_qm_dquot);
82         kmem_zone_free(xfs_qm_dqzone, dqp);
83 }
84
85 /*
86  * If default limits are in force, push them into the dquot now.
87  * We overwrite the dquot limits only if they are zero and this
88  * is not the root dquot.
89  */
90 void
91 xfs_qm_adjust_dqlimits(
92         struct xfs_mount        *mp,
93         struct xfs_dquot        *dq)
94 {
95         struct xfs_quotainfo    *q = mp->m_quotainfo;
96         struct xfs_disk_dquot   *d = &dq->q_core;
97         struct xfs_def_quota    *defq;
98         int                     prealloc = 0;
99
100         ASSERT(d->d_id);
101         defq = xfs_get_defquota(dq, q);
102
103         if (defq->bsoftlimit && !d->d_blk_softlimit) {
104                 d->d_blk_softlimit = cpu_to_be64(defq->bsoftlimit);
105                 prealloc = 1;
106         }
107         if (defq->bhardlimit && !d->d_blk_hardlimit) {
108                 d->d_blk_hardlimit = cpu_to_be64(defq->bhardlimit);
109                 prealloc = 1;
110         }
111         if (defq->isoftlimit && !d->d_ino_softlimit)
112                 d->d_ino_softlimit = cpu_to_be64(defq->isoftlimit);
113         if (defq->ihardlimit && !d->d_ino_hardlimit)
114                 d->d_ino_hardlimit = cpu_to_be64(defq->ihardlimit);
115         if (defq->rtbsoftlimit && !d->d_rtb_softlimit)
116                 d->d_rtb_softlimit = cpu_to_be64(defq->rtbsoftlimit);
117         if (defq->rtbhardlimit && !d->d_rtb_hardlimit)
118                 d->d_rtb_hardlimit = cpu_to_be64(defq->rtbhardlimit);
119
120         if (prealloc)
121                 xfs_dquot_set_prealloc_limits(dq);
122 }
123
124 /*
125  * Check the limits and timers of a dquot and start or reset timers
126  * if necessary.
127  * This gets called even when quota enforcement is OFF, which makes our
128  * life a little less complicated. (We just don't reject any quota
129  * reservations in that case, when enforcement is off).
130  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
131  * enforcement's off.
132  * In contrast, warnings are a little different in that they don't
133  * 'automatically' get started when limits get exceeded.  They do
134  * get reset to zero, however, when we find the count to be under
135  * the soft limit (they are only ever set non-zero via userspace).
136  */
137 void
138 xfs_qm_adjust_dqtimers(
139         xfs_mount_t             *mp,
140         xfs_disk_dquot_t        *d)
141 {
142         ASSERT(d->d_id);
143
144 #ifdef DEBUG
145         if (d->d_blk_hardlimit)
146                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
147                        be64_to_cpu(d->d_blk_hardlimit));
148         if (d->d_ino_hardlimit)
149                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
150                        be64_to_cpu(d->d_ino_hardlimit));
151         if (d->d_rtb_hardlimit)
152                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
153                        be64_to_cpu(d->d_rtb_hardlimit));
154 #endif
155
156         if (!d->d_btimer) {
157                 if ((d->d_blk_softlimit &&
158                      (be64_to_cpu(d->d_bcount) >
159                       be64_to_cpu(d->d_blk_softlimit))) ||
160                     (d->d_blk_hardlimit &&
161                      (be64_to_cpu(d->d_bcount) >
162                       be64_to_cpu(d->d_blk_hardlimit)))) {
163                         d->d_btimer = cpu_to_be32(get_seconds() +
164                                         mp->m_quotainfo->qi_btimelimit);
165                 } else {
166                         d->d_bwarns = 0;
167                 }
168         } else {
169                 if ((!d->d_blk_softlimit ||
170                      (be64_to_cpu(d->d_bcount) <=
171                       be64_to_cpu(d->d_blk_softlimit))) &&
172                     (!d->d_blk_hardlimit ||
173                     (be64_to_cpu(d->d_bcount) <=
174                      be64_to_cpu(d->d_blk_hardlimit)))) {
175                         d->d_btimer = 0;
176                 }
177         }
178
179         if (!d->d_itimer) {
180                 if ((d->d_ino_softlimit &&
181                      (be64_to_cpu(d->d_icount) >
182                       be64_to_cpu(d->d_ino_softlimit))) ||
183                     (d->d_ino_hardlimit &&
184                      (be64_to_cpu(d->d_icount) >
185                       be64_to_cpu(d->d_ino_hardlimit)))) {
186                         d->d_itimer = cpu_to_be32(get_seconds() +
187                                         mp->m_quotainfo->qi_itimelimit);
188                 } else {
189                         d->d_iwarns = 0;
190                 }
191         } else {
192                 if ((!d->d_ino_softlimit ||
193                      (be64_to_cpu(d->d_icount) <=
194                       be64_to_cpu(d->d_ino_softlimit)))  &&
195                     (!d->d_ino_hardlimit ||
196                      (be64_to_cpu(d->d_icount) <=
197                       be64_to_cpu(d->d_ino_hardlimit)))) {
198                         d->d_itimer = 0;
199                 }
200         }
201
202         if (!d->d_rtbtimer) {
203                 if ((d->d_rtb_softlimit &&
204                      (be64_to_cpu(d->d_rtbcount) >
205                       be64_to_cpu(d->d_rtb_softlimit))) ||
206                     (d->d_rtb_hardlimit &&
207                      (be64_to_cpu(d->d_rtbcount) >
208                       be64_to_cpu(d->d_rtb_hardlimit)))) {
209                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
210                                         mp->m_quotainfo->qi_rtbtimelimit);
211                 } else {
212                         d->d_rtbwarns = 0;
213                 }
214         } else {
215                 if ((!d->d_rtb_softlimit ||
216                      (be64_to_cpu(d->d_rtbcount) <=
217                       be64_to_cpu(d->d_rtb_softlimit))) &&
218                     (!d->d_rtb_hardlimit ||
219                      (be64_to_cpu(d->d_rtbcount) <=
220                       be64_to_cpu(d->d_rtb_hardlimit)))) {
221                         d->d_rtbtimer = 0;
222                 }
223         }
224 }
225
226 /*
227  * initialize a buffer full of dquots and log the whole thing
228  */
229 STATIC void
230 xfs_qm_init_dquot_blk(
231         xfs_trans_t     *tp,
232         xfs_mount_t     *mp,
233         xfs_dqid_t      id,
234         uint            type,
235         xfs_buf_t       *bp)
236 {
237         struct xfs_quotainfo    *q = mp->m_quotainfo;
238         xfs_dqblk_t     *d;
239         xfs_dqid_t      curid;
240         int             i;
241
242         ASSERT(tp);
243         ASSERT(xfs_buf_islocked(bp));
244
245         d = bp->b_addr;
246
247         /*
248          * ID of the first dquot in the block - id's are zero based.
249          */
250         curid = id - (id % q->qi_dqperchunk);
251         memset(d, 0, BBTOB(q->qi_dqchunklen));
252         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
253                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
254                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
255                 d->dd_diskdq.d_id = cpu_to_be32(curid);
256                 d->dd_diskdq.d_flags = type;
257                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
258                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_meta_uuid);
259                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
260                                          XFS_DQUOT_CRC_OFF);
261                 }
262         }
263
264         xfs_trans_dquot_buf(tp, bp,
265                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
266                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
267                              XFS_BLF_GDQUOT_BUF)));
268         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
269 }
270
271 /*
272  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
273  * watermarks correspond to the soft and hard limits by default. If a soft limit
274  * is not specified, we use 95% of the hard limit.
275  */
276 void
277 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
278 {
279         uint64_t space;
280
281         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
282         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
283         if (!dqp->q_prealloc_lo_wmark) {
284                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
285                 do_div(dqp->q_prealloc_lo_wmark, 100);
286                 dqp->q_prealloc_lo_wmark *= 95;
287         }
288
289         space = dqp->q_prealloc_hi_wmark;
290
291         do_div(space, 100);
292         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
293         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
294         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
295 }
296
297 /*
298  * Allocate a block and fill it with dquots.
299  * This is called when the bmapi finds a hole.
300  */
301 STATIC int
302 xfs_qm_dqalloc(
303         xfs_trans_t     **tpp,
304         xfs_mount_t     *mp,
305         xfs_dquot_t     *dqp,
306         xfs_inode_t     *quotip,
307         xfs_fileoff_t   offset_fsb,
308         xfs_buf_t       **O_bpp)
309 {
310         xfs_fsblock_t   firstblock;
311         struct xfs_defer_ops dfops;
312         xfs_bmbt_irec_t map;
313         int             nmaps, error;
314         xfs_buf_t       *bp;
315         xfs_trans_t     *tp = *tpp;
316
317         ASSERT(tp != NULL);
318
319         trace_xfs_dqalloc(dqp);
320
321         /*
322          * Initialize the bmap freelist prior to calling bmapi code.
323          */
324         xfs_defer_init(&dfops, &firstblock);
325         xfs_ilock(quotip, XFS_ILOCK_EXCL);
326         /*
327          * Return if this type of quotas is turned off while we didn't
328          * have an inode lock
329          */
330         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
331                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
332                 return -ESRCH;
333         }
334
335         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
336         nmaps = 1;
337         error = xfs_bmapi_write(tp, quotip, offset_fsb,
338                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
339                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
340                                 &map, &nmaps, &dfops);
341         if (error)
342                 goto error0;
343         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
344         ASSERT(nmaps == 1);
345         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
346                (map.br_startblock != HOLESTARTBLOCK));
347
348         /*
349          * Keep track of the blkno to save a lookup later
350          */
351         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
352
353         /* now we can just get the buffer (there's nothing to read yet) */
354         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
355                                dqp->q_blkno,
356                                mp->m_quotainfo->qi_dqchunklen,
357                                0);
358         if (!bp) {
359                 error = -ENOMEM;
360                 goto error1;
361         }
362         bp->b_ops = &xfs_dquot_buf_ops;
363
364         /*
365          * Make a chunk of dquots out of this buffer and log
366          * the entire thing.
367          */
368         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
369                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
370
371         /*
372          * xfs_defer_finish() may commit the current transaction and
373          * start a second transaction if the freelist is not empty.
374          *
375          * Since we still want to modify this buffer, we need to
376          * ensure that the buffer is not released on commit of
377          * the first transaction and ensure the buffer is added to the
378          * second transaction.
379          *
380          * If there is only one transaction then don't stop the buffer
381          * from being released when it commits later on.
382          */
383
384         xfs_trans_bhold(tp, bp);
385
386         error = xfs_defer_finish(tpp, &dfops, NULL);
387         if (error)
388                 goto error1;
389
390         /* Transaction was committed? */
391         if (*tpp != tp) {
392                 tp = *tpp;
393                 xfs_trans_bjoin(tp, bp);
394         } else {
395                 xfs_trans_bhold_release(tp, bp);
396         }
397
398         *O_bpp = bp;
399         return 0;
400
401 error1:
402         xfs_defer_cancel(&dfops);
403 error0:
404         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
405
406         return error;
407 }
408
409 STATIC int
410 xfs_qm_dqrepair(
411         struct xfs_mount        *mp,
412         struct xfs_trans        *tp,
413         struct xfs_dquot        *dqp,
414         xfs_dqid_t              firstid,
415         struct xfs_buf          **bpp)
416 {
417         int                     error;
418         struct xfs_disk_dquot   *ddq;
419         struct xfs_dqblk        *d;
420         int                     i;
421
422         /*
423          * Read the buffer without verification so we get the corrupted
424          * buffer returned to us. make sure we verify it on write, though.
425          */
426         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
427                                    mp->m_quotainfo->qi_dqchunklen,
428                                    0, bpp, NULL);
429
430         if (error) {
431                 ASSERT(*bpp == NULL);
432                 return error;
433         }
434         (*bpp)->b_ops = &xfs_dquot_buf_ops;
435
436         ASSERT(xfs_buf_islocked(*bpp));
437         d = (struct xfs_dqblk *)(*bpp)->b_addr;
438
439         /* Do the actual repair of dquots in this buffer */
440         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
441                 ddq = &d[i].dd_diskdq;
442                 error = xfs_dqcheck(mp, ddq, firstid + i,
443                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
444                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
445                 if (error) {
446                         /* repair failed, we're screwed */
447                         xfs_trans_brelse(tp, *bpp);
448                         return -EIO;
449                 }
450         }
451
452         return 0;
453 }
454
455 /*
456  * Maps a dquot to the buffer containing its on-disk version.
457  * This returns a ptr to the buffer containing the on-disk dquot
458  * in the bpp param, and a ptr to the on-disk dquot within that buffer
459  */
460 STATIC int
461 xfs_qm_dqtobp(
462         xfs_trans_t             **tpp,
463         xfs_dquot_t             *dqp,
464         xfs_disk_dquot_t        **O_ddpp,
465         xfs_buf_t               **O_bpp,
466         uint                    flags)
467 {
468         struct xfs_bmbt_irec    map;
469         int                     nmaps = 1, error;
470         struct xfs_buf          *bp;
471         struct xfs_inode        *quotip;
472         struct xfs_mount        *mp = dqp->q_mount;
473         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
474         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
475         uint                    lock_mode = 0;
476
477         quotip = xfs_quota_inode(dqp->q_mount, dqp->dq_flags);
478         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
479
480         ASSERT(!(flags & XFS_QMOPT_NOLOCK) ||
481                 xfs_isilocked(quotip, XFS_ILOCK_SHARED) ||
482                 xfs_isilocked(quotip, XFS_ILOCK_EXCL));
483         if (!(flags & XFS_QMOPT_NOLOCK))
484                 lock_mode = xfs_ilock_data_map_shared(quotip);
485         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
486                 /*
487                  * Return if this type of quotas is turned off while we
488                  * didn't have the quota inode lock.
489                  */
490                 if (lock_mode)
491                         xfs_iunlock(quotip, lock_mode);
492                 return -ESRCH;
493         }
494
495         /*
496          * Find the block map; no allocations yet
497          */
498         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
499                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
500
501         if (lock_mode)
502                 xfs_iunlock(quotip, lock_mode);
503         if (error)
504                 return error;
505
506         ASSERT(nmaps == 1);
507         ASSERT(map.br_blockcount == 1);
508
509         /*
510          * Offset of dquot in the (fixed sized) dquot chunk.
511          */
512         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
513                 sizeof(xfs_dqblk_t);
514
515         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
516         if (map.br_startblock == HOLESTARTBLOCK) {
517                 /*
518                  * We don't allocate unless we're asked to
519                  */
520                 if (!(flags & XFS_QMOPT_DQALLOC))
521                         return -ENOENT;
522
523                 ASSERT(tp);
524                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
525                                         dqp->q_fileoffset, &bp);
526                 if (error)
527                         return error;
528                 tp = *tpp;
529         } else {
530                 trace_xfs_dqtobp_read(dqp);
531
532                 /*
533                  * store the blkno etc so that we don't have to do the
534                  * mapping all the time
535                  */
536                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
537
538                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
539                                            dqp->q_blkno,
540                                            mp->m_quotainfo->qi_dqchunklen,
541                                            0, &bp, &xfs_dquot_buf_ops);
542
543                 if (error == -EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
544                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
545                                                 mp->m_quotainfo->qi_dqperchunk;
546                         ASSERT(bp == NULL);
547                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
548                 }
549
550                 if (error) {
551                         ASSERT(bp == NULL);
552                         return error;
553                 }
554         }
555
556         ASSERT(xfs_buf_islocked(bp));
557         *O_bpp = bp;
558         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
559
560         return 0;
561 }
562
563
564 /*
565  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
566  * and release the buffer immediately.
567  *
568  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
569  */
570 int
571 xfs_qm_dqread(
572         struct xfs_mount        *mp,
573         xfs_dqid_t              id,
574         uint                    type,
575         uint                    flags,
576         struct xfs_dquot        **O_dqpp)
577 {
578         struct xfs_dquot        *dqp;
579         struct xfs_disk_dquot   *ddqp;
580         struct xfs_buf          *bp;
581         struct xfs_trans        *tp = NULL;
582         int                     error;
583
584         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
585
586         dqp->dq_flags = type;
587         dqp->q_core.d_id = cpu_to_be32(id);
588         dqp->q_mount = mp;
589         INIT_LIST_HEAD(&dqp->q_lru);
590         mutex_init(&dqp->q_qlock);
591         init_waitqueue_head(&dqp->q_pinwait);
592
593         /*
594          * Because we want to use a counting completion, complete
595          * the flush completion once to allow a single access to
596          * the flush completion without blocking.
597          */
598         init_completion(&dqp->q_flush);
599         complete(&dqp->q_flush);
600
601         /*
602          * Make sure group quotas have a different lock class than user
603          * quotas.
604          */
605         switch (type) {
606         case XFS_DQ_USER:
607                 /* uses the default lock class */
608                 break;
609         case XFS_DQ_GROUP:
610                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_group_class);
611                 break;
612         case XFS_DQ_PROJ:
613                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_project_class);
614                 break;
615         default:
616                 ASSERT(0);
617                 break;
618         }
619
620         XFS_STATS_INC(mp, xs_qm_dquot);
621
622         trace_xfs_dqread(dqp);
623
624         if (flags & XFS_QMOPT_DQALLOC) {
625                 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_dqalloc,
626                                 XFS_QM_DQALLOC_SPACE_RES(mp), 0, 0, &tp);
627                 if (error)
628                         goto error0;
629         }
630
631         /*
632          * get a pointer to the on-disk dquot and the buffer containing it
633          * dqp already knows its own type (GROUP/USER).
634          */
635         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
636         if (error) {
637                 /*
638                  * This can happen if quotas got turned off (ESRCH),
639                  * or if the dquot didn't exist on disk and we ask to
640                  * allocate (ENOENT).
641                  */
642                 trace_xfs_dqread_fail(dqp);
643                 goto error1;
644         }
645
646         /* copy everything from disk dquot to the incore dquot */
647         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
648         xfs_qm_dquot_logitem_init(dqp);
649
650         /*
651          * Reservation counters are defined as reservation plus current usage
652          * to avoid having to add every time.
653          */
654         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
655         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
656         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
657
658         /* initialize the dquot speculative prealloc thresholds */
659         xfs_dquot_set_prealloc_limits(dqp);
660
661         /* Mark the buf so that this will stay incore a little longer */
662         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
663
664         /*
665          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
666          * So we need to release with xfs_trans_brelse().
667          * The strategy here is identical to that of inodes; we lock
668          * the dquot in xfs_qm_dqget() before making it accessible to
669          * others. This is because dquots, like inodes, need a good level of
670          * concurrency, and we don't want to take locks on the entire buffers
671          * for dquot accesses.
672          * Note also that the dquot buffer may even be dirty at this point, if
673          * this particular dquot was repaired. We still aren't afraid to
674          * brelse it because we have the changes incore.
675          */
676         ASSERT(xfs_buf_islocked(bp));
677         xfs_trans_brelse(tp, bp);
678
679         if (tp) {
680                 error = xfs_trans_commit(tp);
681                 if (error)
682                         goto error0;
683         }
684
685         *O_dqpp = dqp;
686         return error;
687
688 error1:
689         if (tp)
690                 xfs_trans_cancel(tp);
691 error0:
692         xfs_qm_dqdestroy(dqp);
693         *O_dqpp = NULL;
694         return error;
695 }
696
697 /*
698  * Advance to the next id in the current chunk, or if at the
699  * end of the chunk, skip ahead to first id in next allocated chunk
700  * using the SEEK_DATA interface.
701  */
702 static int
703 xfs_dq_get_next_id(
704         struct xfs_mount        *mp,
705         uint                    type,
706         xfs_dqid_t              *id)
707 {
708         struct xfs_inode        *quotip = xfs_quota_inode(mp, type);
709         xfs_dqid_t              next_id = *id + 1; /* simple advance */
710         uint                    lock_flags;
711         struct xfs_bmbt_irec    got;
712         xfs_extnum_t            idx;
713         xfs_fsblock_t           start;
714         int                     error = 0;
715
716         /* If we'd wrap past the max ID, stop */
717         if (next_id < *id)
718                 return -ENOENT;
719
720         /* If new ID is within the current chunk, advancing it sufficed */
721         if (next_id % mp->m_quotainfo->qi_dqperchunk) {
722                 *id = next_id;
723                 return 0;
724         }
725
726         /* Nope, next_id is now past the current chunk, so find the next one */
727         start = (xfs_fsblock_t)next_id / mp->m_quotainfo->qi_dqperchunk;
728
729         lock_flags = xfs_ilock_data_map_shared(quotip);
730         if (!(quotip->i_df.if_flags & XFS_IFEXTENTS)) {
731                 error = xfs_iread_extents(NULL, quotip, XFS_DATA_FORK);
732                 if (error)
733                         return error;
734         }
735
736         if (xfs_iext_lookup_extent(quotip, &quotip->i_df, start, &idx, &got)) {
737                 /* contiguous chunk, bump startoff for the id calculation */
738                 if (got.br_startoff < start)
739                         got.br_startoff = start;
740                 *id = got.br_startoff * mp->m_quotainfo->qi_dqperchunk;
741         } else {
742                 error = -ENOENT;
743         }
744
745         xfs_iunlock(quotip, lock_flags);
746
747         return error;
748 }
749
750 /*
751  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
752  * a locked dquot, doing an allocation (if requested) as needed.
753  * When both an inode and an id are given, the inode's id takes precedence.
754  * That is, if the id changes while we don't hold the ilock inside this
755  * function, the new dquot is returned, not necessarily the one requested
756  * in the id argument.
757  */
758 int
759 xfs_qm_dqget(
760         xfs_mount_t     *mp,
761         xfs_inode_t     *ip,      /* locked inode (optional) */
762         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
763         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
764         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
765         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
766 {
767         struct xfs_quotainfo    *qi = mp->m_quotainfo;
768         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
769         struct xfs_dquot        *dqp;
770         int                     error;
771
772         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
773         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
774             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
775             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
776                 return -ESRCH;
777         }
778
779 #ifdef DEBUG
780         if (xfs_do_dqerror) {
781                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
782                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
783                         xfs_debug(mp, "Returning error in dqget");
784                         return -EIO;
785                 }
786         }
787
788         ASSERT(type == XFS_DQ_USER ||
789                type == XFS_DQ_PROJ ||
790                type == XFS_DQ_GROUP);
791         if (ip) {
792                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
793                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
794         }
795 #endif
796
797 restart:
798         mutex_lock(&qi->qi_tree_lock);
799         dqp = radix_tree_lookup(tree, id);
800         if (dqp) {
801                 xfs_dqlock(dqp);
802                 if (dqp->dq_flags & XFS_DQ_FREEING) {
803                         xfs_dqunlock(dqp);
804                         mutex_unlock(&qi->qi_tree_lock);
805                         trace_xfs_dqget_freeing(dqp);
806                         delay(1);
807                         goto restart;
808                 }
809
810                 /* uninit / unused quota found in radix tree, keep looking  */
811                 if (flags & XFS_QMOPT_DQNEXT) {
812                         if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
813                                 xfs_dqunlock(dqp);
814                                 mutex_unlock(&qi->qi_tree_lock);
815                                 error = xfs_dq_get_next_id(mp, type, &id);
816                                 if (error)
817                                         return error;
818                                 goto restart;
819                         }
820                 }
821
822                 dqp->q_nrefs++;
823                 mutex_unlock(&qi->qi_tree_lock);
824
825                 trace_xfs_dqget_hit(dqp);
826                 XFS_STATS_INC(mp, xs_qm_dqcachehits);
827                 *O_dqpp = dqp;
828                 return 0;
829         }
830         mutex_unlock(&qi->qi_tree_lock);
831         XFS_STATS_INC(mp, xs_qm_dqcachemisses);
832
833         /*
834          * Dquot cache miss. We don't want to keep the inode lock across
835          * a (potential) disk read. Also we don't want to deal with the lock
836          * ordering between quotainode and this inode. OTOH, dropping the inode
837          * lock here means dealing with a chown that can happen before
838          * we re-acquire the lock.
839          */
840         if (ip)
841                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
842
843         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
844
845         if (ip)
846                 xfs_ilock(ip, XFS_ILOCK_EXCL);
847
848         /* If we are asked to find next active id, keep looking */
849         if (error == -ENOENT && (flags & XFS_QMOPT_DQNEXT)) {
850                 error = xfs_dq_get_next_id(mp, type, &id);
851                 if (!error)
852                         goto restart;
853         }
854
855         if (error)
856                 return error;
857
858         if (ip) {
859                 /*
860                  * A dquot could be attached to this inode by now, since
861                  * we had dropped the ilock.
862                  */
863                 if (xfs_this_quota_on(mp, type)) {
864                         struct xfs_dquot        *dqp1;
865
866                         dqp1 = xfs_inode_dquot(ip, type);
867                         if (dqp1) {
868                                 xfs_qm_dqdestroy(dqp);
869                                 dqp = dqp1;
870                                 xfs_dqlock(dqp);
871                                 goto dqret;
872                         }
873                 } else {
874                         /* inode stays locked on return */
875                         xfs_qm_dqdestroy(dqp);
876                         return -ESRCH;
877                 }
878         }
879
880         mutex_lock(&qi->qi_tree_lock);
881         error = radix_tree_insert(tree, id, dqp);
882         if (unlikely(error)) {
883                 WARN_ON(error != -EEXIST);
884
885                 /*
886                  * Duplicate found. Just throw away the new dquot and start
887                  * over.
888                  */
889                 mutex_unlock(&qi->qi_tree_lock);
890                 trace_xfs_dqget_dup(dqp);
891                 xfs_qm_dqdestroy(dqp);
892                 XFS_STATS_INC(mp, xs_qm_dquot_dups);
893                 goto restart;
894         }
895
896         /*
897          * We return a locked dquot to the caller, with a reference taken
898          */
899         xfs_dqlock(dqp);
900         dqp->q_nrefs = 1;
901
902         qi->qi_dquots++;
903         mutex_unlock(&qi->qi_tree_lock);
904
905         /* If we are asked to find next active id, keep looking */
906         if (flags & XFS_QMOPT_DQNEXT) {
907                 if (XFS_IS_DQUOT_UNINITIALIZED(dqp)) {
908                         xfs_qm_dqput(dqp);
909                         error = xfs_dq_get_next_id(mp, type, &id);
910                         if (error)
911                                 return error;
912                         goto restart;
913                 }
914         }
915
916  dqret:
917         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
918         trace_xfs_dqget_miss(dqp);
919         *O_dqpp = dqp;
920         return 0;
921 }
922
923 /*
924  * Release a reference to the dquot (decrement ref-count) and unlock it.
925  *
926  * If there is a group quota attached to this dquot, carefully release that
927  * too without tripping over deadlocks'n'stuff.
928  */
929 void
930 xfs_qm_dqput(
931         struct xfs_dquot        *dqp)
932 {
933         ASSERT(dqp->q_nrefs > 0);
934         ASSERT(XFS_DQ_IS_LOCKED(dqp));
935
936         trace_xfs_dqput(dqp);
937
938         if (--dqp->q_nrefs == 0) {
939                 struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
940                 trace_xfs_dqput_free(dqp);
941
942                 if (list_lru_add(&qi->qi_lru, &dqp->q_lru))
943                         XFS_STATS_INC(dqp->q_mount, xs_qm_dquot_unused);
944         }
945         xfs_dqunlock(dqp);
946 }
947
948 /*
949  * Release a dquot. Flush it if dirty, then dqput() it.
950  * dquot must not be locked.
951  */
952 void
953 xfs_qm_dqrele(
954         xfs_dquot_t     *dqp)
955 {
956         if (!dqp)
957                 return;
958
959         trace_xfs_dqrele(dqp);
960
961         xfs_dqlock(dqp);
962         /*
963          * We don't care to flush it if the dquot is dirty here.
964          * That will create stutters that we want to avoid.
965          * Instead we do a delayed write when we try to reclaim
966          * a dirty dquot. Also xfs_sync will take part of the burden...
967          */
968         xfs_qm_dqput(dqp);
969 }
970
971 /*
972  * This is the dquot flushing I/O completion routine.  It is called
973  * from interrupt level when the buffer containing the dquot is
974  * flushed to disk.  It is responsible for removing the dquot logitem
975  * from the AIL if it has not been re-logged, and unlocking the dquot's
976  * flush lock. This behavior is very similar to that of inodes..
977  */
978 STATIC void
979 xfs_qm_dqflush_done(
980         struct xfs_buf          *bp,
981         struct xfs_log_item     *lip)
982 {
983         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
984         xfs_dquot_t             *dqp = qip->qli_dquot;
985         struct xfs_ail          *ailp = lip->li_ailp;
986
987         /*
988          * We only want to pull the item from the AIL if its
989          * location in the log has not changed since we started the flush.
990          * Thus, we only bother if the dquot's lsn has
991          * not changed. First we check the lsn outside the lock
992          * since it's cheaper, and then we recheck while
993          * holding the lock before removing the dquot from the AIL.
994          */
995         if ((lip->li_flags & XFS_LI_IN_AIL) &&
996             lip->li_lsn == qip->qli_flush_lsn) {
997
998                 /* xfs_trans_ail_delete() drops the AIL lock. */
999                 spin_lock(&ailp->xa_lock);
1000                 if (lip->li_lsn == qip->qli_flush_lsn)
1001                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1002                 else
1003                         spin_unlock(&ailp->xa_lock);
1004         }
1005
1006         /*
1007          * Release the dq's flush lock since we're done with it.
1008          */
1009         xfs_dqfunlock(dqp);
1010 }
1011
1012 /*
1013  * Write a modified dquot to disk.
1014  * The dquot must be locked and the flush lock too taken by caller.
1015  * The flush lock will not be unlocked until the dquot reaches the disk,
1016  * but the dquot is free to be unlocked and modified by the caller
1017  * in the interim. Dquot is still locked on return. This behavior is
1018  * identical to that of inodes.
1019  */
1020 int
1021 xfs_qm_dqflush(
1022         struct xfs_dquot        *dqp,
1023         struct xfs_buf          **bpp)
1024 {
1025         struct xfs_mount        *mp = dqp->q_mount;
1026         struct xfs_buf          *bp;
1027         struct xfs_disk_dquot   *ddqp;
1028         int                     error;
1029
1030         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1031         ASSERT(!completion_done(&dqp->q_flush));
1032
1033         trace_xfs_dqflush(dqp);
1034
1035         *bpp = NULL;
1036
1037         xfs_qm_dqunpin_wait(dqp);
1038
1039         /*
1040          * This may have been unpinned because the filesystem is shutting
1041          * down forcibly. If that's the case we must not write this dquot
1042          * to disk, because the log record didn't make it to disk.
1043          *
1044          * We also have to remove the log item from the AIL in this case,
1045          * as we wait for an emptry AIL as part of the unmount process.
1046          */
1047         if (XFS_FORCED_SHUTDOWN(mp)) {
1048                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
1049                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1050
1051                 xfs_trans_ail_remove(lip, SHUTDOWN_CORRUPT_INCORE);
1052
1053                 error = -EIO;
1054                 goto out_unlock;
1055         }
1056
1057         /*
1058          * Get the buffer containing the on-disk dquot
1059          */
1060         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1061                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp,
1062                                    &xfs_dquot_buf_ops);
1063         if (error)
1064                 goto out_unlock;
1065
1066         /*
1067          * Calculate the location of the dquot inside the buffer.
1068          */
1069         ddqp = bp->b_addr + dqp->q_bufoffset;
1070
1071         /*
1072          * A simple sanity check in case we got a corrupted dquot..
1073          */
1074         error = xfs_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1075                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1076         if (error) {
1077                 xfs_buf_relse(bp);
1078                 xfs_dqfunlock(dqp);
1079                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1080                 return -EIO;
1081         }
1082
1083         /* This is the only portion of data that needs to persist */
1084         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1085
1086         /*
1087          * Clear the dirty field and remember the flush lsn for later use.
1088          */
1089         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1090
1091         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1092                                         &dqp->q_logitem.qli_item.li_lsn);
1093
1094         /*
1095          * copy the lsn into the on-disk dquot now while we have the in memory
1096          * dquot here. This can't be done later in the write verifier as we
1097          * can't get access to the log item at that point in time.
1098          *
1099          * We also calculate the CRC here so that the on-disk dquot in the
1100          * buffer always has a valid CRC. This ensures there is no possibility
1101          * of a dquot without an up-to-date CRC getting to disk.
1102          */
1103         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1104                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1105
1106                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1107                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1108                                  XFS_DQUOT_CRC_OFF);
1109         }
1110
1111         /*
1112          * Attach an iodone routine so that we can remove this dquot from the
1113          * AIL and release the flush lock once the dquot is synced to disk.
1114          */
1115         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1116                                   &dqp->q_logitem.qli_item);
1117
1118         /*
1119          * If the buffer is pinned then push on the log so we won't
1120          * get stuck waiting in the write for too long.
1121          */
1122         if (xfs_buf_ispinned(bp)) {
1123                 trace_xfs_dqflush_force(dqp);
1124                 xfs_log_force(mp, 0);
1125         }
1126
1127         trace_xfs_dqflush_done(dqp);
1128         *bpp = bp;
1129         return 0;
1130
1131 out_unlock:
1132         xfs_dqfunlock(dqp);
1133         return -EIO;
1134 }
1135
1136 /*
1137  * Lock two xfs_dquot structures.
1138  *
1139  * To avoid deadlocks we always lock the quota structure with
1140  * the lowerd id first.
1141  */
1142 void
1143 xfs_dqlock2(
1144         xfs_dquot_t     *d1,
1145         xfs_dquot_t     *d2)
1146 {
1147         if (d1 && d2) {
1148                 ASSERT(d1 != d2);
1149                 if (be32_to_cpu(d1->q_core.d_id) >
1150                     be32_to_cpu(d2->q_core.d_id)) {
1151                         mutex_lock(&d2->q_qlock);
1152                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1153                 } else {
1154                         mutex_lock(&d1->q_qlock);
1155                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1156                 }
1157         } else if (d1) {
1158                 mutex_lock(&d1->q_qlock);
1159         } else if (d2) {
1160                 mutex_lock(&d2->q_qlock);
1161         }
1162 }
1163
1164 int __init
1165 xfs_qm_init(void)
1166 {
1167         xfs_qm_dqzone =
1168                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1169         if (!xfs_qm_dqzone)
1170                 goto out;
1171
1172         xfs_qm_dqtrxzone =
1173                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1174         if (!xfs_qm_dqtrxzone)
1175                 goto out_free_dqzone;
1176
1177         return 0;
1178
1179 out_free_dqzone:
1180         kmem_zone_destroy(xfs_qm_dqzone);
1181 out:
1182         return -ENOMEM;
1183 }
1184
1185 void
1186 xfs_qm_exit(void)
1187 {
1188         kmem_zone_destroy(xfs_qm_dqtrxzone);
1189         kmem_zone_destroy(xfs_qm_dqzone);
1190 }