]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - fs/xfs/xfs_dquot.c
xfs: change available ranges of softlimit and hardlimit in quota check
[~shefty/rdma-dev.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_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_trans.h"
24 #include "xfs_sb.h"
25 #include "xfs_ag.h"
26 #include "xfs_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_mount.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_bmap.h"
32 #include "xfs_rtalloc.h"
33 #include "xfs_error.h"
34 #include "xfs_itable.h"
35 #include "xfs_attr.h"
36 #include "xfs_buf_item.h"
37 #include "xfs_trans_space.h"
38 #include "xfs_trans_priv.h"
39 #include "xfs_qm.h"
40 #include "xfs_trace.h"
41
42 /*
43  * Lock order:
44  *
45  * ip->i_lock
46  *   qh->qh_lock
47  *     qi->qi_dqlist_lock
48  *       dquot->q_qlock (xfs_dqlock() and friends)
49  *         dquot->q_flush (xfs_dqflock() and friends)
50  *         xfs_Gqm->qm_dqfrlist_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 static struct lock_class_key xfs_dquot_other_class;
64
65 /*
66  * This is called to free all the memory associated with a dquot
67  */
68 void
69 xfs_qm_dqdestroy(
70         xfs_dquot_t     *dqp)
71 {
72         ASSERT(list_empty(&dqp->q_freelist));
73
74         mutex_destroy(&dqp->q_qlock);
75         kmem_zone_free(xfs_Gqm->qm_dqzone, dqp);
76
77         atomic_dec(&xfs_Gqm->qm_totaldquots);
78 }
79
80 /*
81  * If default limits are in force, push them into the dquot now.
82  * We overwrite the dquot limits only if they are zero and this
83  * is not the root dquot.
84  */
85 void
86 xfs_qm_adjust_dqlimits(
87         xfs_mount_t             *mp,
88         xfs_disk_dquot_t        *d)
89 {
90         xfs_quotainfo_t         *q = mp->m_quotainfo;
91
92         ASSERT(d->d_id);
93
94         if (q->qi_bsoftlimit && !d->d_blk_softlimit)
95                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
96         if (q->qi_bhardlimit && !d->d_blk_hardlimit)
97                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
98         if (q->qi_isoftlimit && !d->d_ino_softlimit)
99                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
100         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
101                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
102         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
103                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
104         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
105                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
106 }
107
108 /*
109  * Check the limits and timers of a dquot and start or reset timers
110  * if necessary.
111  * This gets called even when quota enforcement is OFF, which makes our
112  * life a little less complicated. (We just don't reject any quota
113  * reservations in that case, when enforcement is off).
114  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
115  * enforcement's off.
116  * In contrast, warnings are a little different in that they don't
117  * 'automatically' get started when limits get exceeded.  They do
118  * get reset to zero, however, when we find the count to be under
119  * the soft limit (they are only ever set non-zero via userspace).
120  */
121 void
122 xfs_qm_adjust_dqtimers(
123         xfs_mount_t             *mp,
124         xfs_disk_dquot_t        *d)
125 {
126         ASSERT(d->d_id);
127
128 #ifdef DEBUG
129         if (d->d_blk_hardlimit)
130                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
131                        be64_to_cpu(d->d_blk_hardlimit));
132         if (d->d_ino_hardlimit)
133                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
134                        be64_to_cpu(d->d_ino_hardlimit));
135         if (d->d_rtb_hardlimit)
136                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
137                        be64_to_cpu(d->d_rtb_hardlimit));
138 #endif
139
140         if (!d->d_btimer) {
141                 if ((d->d_blk_softlimit &&
142                      (be64_to_cpu(d->d_bcount) >
143                       be64_to_cpu(d->d_blk_softlimit))) ||
144                     (d->d_blk_hardlimit &&
145                      (be64_to_cpu(d->d_bcount) >
146                       be64_to_cpu(d->d_blk_hardlimit)))) {
147                         d->d_btimer = cpu_to_be32(get_seconds() +
148                                         mp->m_quotainfo->qi_btimelimit);
149                 } else {
150                         d->d_bwarns = 0;
151                 }
152         } else {
153                 if ((!d->d_blk_softlimit ||
154                      (be64_to_cpu(d->d_bcount) <=
155                       be64_to_cpu(d->d_blk_softlimit))) &&
156                     (!d->d_blk_hardlimit ||
157                     (be64_to_cpu(d->d_bcount) <=
158                      be64_to_cpu(d->d_blk_hardlimit)))) {
159                         d->d_btimer = 0;
160                 }
161         }
162
163         if (!d->d_itimer) {
164                 if ((d->d_ino_softlimit &&
165                      (be64_to_cpu(d->d_icount) >
166                       be64_to_cpu(d->d_ino_softlimit))) ||
167                     (d->d_ino_hardlimit &&
168                      (be64_to_cpu(d->d_icount) >
169                       be64_to_cpu(d->d_ino_hardlimit)))) {
170                         d->d_itimer = cpu_to_be32(get_seconds() +
171                                         mp->m_quotainfo->qi_itimelimit);
172                 } else {
173                         d->d_iwarns = 0;
174                 }
175         } else {
176                 if ((!d->d_ino_softlimit ||
177                      (be64_to_cpu(d->d_icount) <=
178                       be64_to_cpu(d->d_ino_softlimit)))  &&
179                     (!d->d_ino_hardlimit ||
180                      (be64_to_cpu(d->d_icount) <=
181                       be64_to_cpu(d->d_ino_hardlimit)))) {
182                         d->d_itimer = 0;
183                 }
184         }
185
186         if (!d->d_rtbtimer) {
187                 if ((d->d_rtb_softlimit &&
188                      (be64_to_cpu(d->d_rtbcount) >
189                       be64_to_cpu(d->d_rtb_softlimit))) ||
190                     (d->d_rtb_hardlimit &&
191                      (be64_to_cpu(d->d_rtbcount) >
192                       be64_to_cpu(d->d_rtb_hardlimit)))) {
193                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
194                                         mp->m_quotainfo->qi_rtbtimelimit);
195                 } else {
196                         d->d_rtbwarns = 0;
197                 }
198         } else {
199                 if ((!d->d_rtb_softlimit ||
200                      (be64_to_cpu(d->d_rtbcount) <=
201                       be64_to_cpu(d->d_rtb_softlimit))) &&
202                     (!d->d_rtb_hardlimit ||
203                      (be64_to_cpu(d->d_rtbcount) <=
204                       be64_to_cpu(d->d_rtb_hardlimit)))) {
205                         d->d_rtbtimer = 0;
206                 }
207         }
208 }
209
210 /*
211  * initialize a buffer full of dquots and log the whole thing
212  */
213 STATIC void
214 xfs_qm_init_dquot_blk(
215         xfs_trans_t     *tp,
216         xfs_mount_t     *mp,
217         xfs_dqid_t      id,
218         uint            type,
219         xfs_buf_t       *bp)
220 {
221         struct xfs_quotainfo    *q = mp->m_quotainfo;
222         xfs_dqblk_t     *d;
223         int             curid, i;
224
225         ASSERT(tp);
226         ASSERT(xfs_buf_islocked(bp));
227
228         d = bp->b_addr;
229
230         /*
231          * ID of the first dquot in the block - id's are zero based.
232          */
233         curid = id - (id % q->qi_dqperchunk);
234         ASSERT(curid >= 0);
235         memset(d, 0, BBTOB(q->qi_dqchunklen));
236         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
237                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
238                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
239                 d->dd_diskdq.d_id = cpu_to_be32(curid);
240                 d->dd_diskdq.d_flags = type;
241         }
242
243         xfs_trans_dquot_buf(tp, bp,
244                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
245                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
246                              XFS_BLF_GDQUOT_BUF)));
247         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
248 }
249
250
251
252 /*
253  * Allocate a block and fill it with dquots.
254  * This is called when the bmapi finds a hole.
255  */
256 STATIC int
257 xfs_qm_dqalloc(
258         xfs_trans_t     **tpp,
259         xfs_mount_t     *mp,
260         xfs_dquot_t     *dqp,
261         xfs_inode_t     *quotip,
262         xfs_fileoff_t   offset_fsb,
263         xfs_buf_t       **O_bpp)
264 {
265         xfs_fsblock_t   firstblock;
266         xfs_bmap_free_t flist;
267         xfs_bmbt_irec_t map;
268         int             nmaps, error, committed;
269         xfs_buf_t       *bp;
270         xfs_trans_t     *tp = *tpp;
271
272         ASSERT(tp != NULL);
273
274         trace_xfs_dqalloc(dqp);
275
276         /*
277          * Initialize the bmap freelist prior to calling bmapi code.
278          */
279         xfs_bmap_init(&flist, &firstblock);
280         xfs_ilock(quotip, XFS_ILOCK_EXCL);
281         /*
282          * Return if this type of quotas is turned off while we didn't
283          * have an inode lock
284          */
285         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
286                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
287                 return (ESRCH);
288         }
289
290         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
291         nmaps = 1;
292         error = xfs_bmapi_write(tp, quotip, offset_fsb,
293                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
294                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
295                                 &map, &nmaps, &flist);
296         if (error)
297                 goto error0;
298         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
299         ASSERT(nmaps == 1);
300         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
301                (map.br_startblock != HOLESTARTBLOCK));
302
303         /*
304          * Keep track of the blkno to save a lookup later
305          */
306         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
307
308         /* now we can just get the buffer (there's nothing to read yet) */
309         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
310                                dqp->q_blkno,
311                                mp->m_quotainfo->qi_dqchunklen,
312                                0);
313
314         error = xfs_buf_geterror(bp);
315         if (error)
316                 goto error1;
317
318         /*
319          * Make a chunk of dquots out of this buffer and log
320          * the entire thing.
321          */
322         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
323                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
324
325         /*
326          * xfs_bmap_finish() may commit the current transaction and
327          * start a second transaction if the freelist is not empty.
328          *
329          * Since we still want to modify this buffer, we need to
330          * ensure that the buffer is not released on commit of
331          * the first transaction and ensure the buffer is added to the
332          * second transaction.
333          *
334          * If there is only one transaction then don't stop the buffer
335          * from being released when it commits later on.
336          */
337
338         xfs_trans_bhold(tp, bp);
339
340         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
341                 goto error1;
342         }
343
344         if (committed) {
345                 tp = *tpp;
346                 xfs_trans_bjoin(tp, bp);
347         } else {
348                 xfs_trans_bhold_release(tp, bp);
349         }
350
351         *O_bpp = bp;
352         return 0;
353
354       error1:
355         xfs_bmap_cancel(&flist);
356       error0:
357         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
358
359         return (error);
360 }
361
362 /*
363  * Maps a dquot to the buffer containing its on-disk version.
364  * This returns a ptr to the buffer containing the on-disk dquot
365  * in the bpp param, and a ptr to the on-disk dquot within that buffer
366  */
367 STATIC int
368 xfs_qm_dqtobp(
369         xfs_trans_t             **tpp,
370         xfs_dquot_t             *dqp,
371         xfs_disk_dquot_t        **O_ddpp,
372         xfs_buf_t               **O_bpp,
373         uint                    flags)
374 {
375         xfs_bmbt_irec_t map;
376         int             nmaps = 1, error;
377         xfs_buf_t       *bp;
378         xfs_inode_t     *quotip = XFS_DQ_TO_QIP(dqp);
379         xfs_mount_t     *mp = dqp->q_mount;
380         xfs_disk_dquot_t *ddq;
381         xfs_dqid_t      id = be32_to_cpu(dqp->q_core.d_id);
382         xfs_trans_t     *tp = (tpp ? *tpp : NULL);
383
384         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
385
386         xfs_ilock(quotip, XFS_ILOCK_SHARED);
387         if (XFS_IS_THIS_QUOTA_OFF(dqp)) {
388                 /*
389                  * Return if this type of quotas is turned off while we
390                  * didn't have the quota inode lock.
391                  */
392                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
393                 return ESRCH;
394         }
395
396         /*
397          * Find the block map; no allocations yet
398          */
399         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
400                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
401
402         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
403         if (error)
404                 return error;
405
406         ASSERT(nmaps == 1);
407         ASSERT(map.br_blockcount == 1);
408
409         /*
410          * Offset of dquot in the (fixed sized) dquot chunk.
411          */
412         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
413                 sizeof(xfs_dqblk_t);
414
415         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
416         if (map.br_startblock == HOLESTARTBLOCK) {
417                 /*
418                  * We don't allocate unless we're asked to
419                  */
420                 if (!(flags & XFS_QMOPT_DQALLOC))
421                         return ENOENT;
422
423                 ASSERT(tp);
424                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
425                                         dqp->q_fileoffset, &bp);
426                 if (error)
427                         return error;
428                 tp = *tpp;
429         } else {
430                 trace_xfs_dqtobp_read(dqp);
431
432                 /*
433                  * store the blkno etc so that we don't have to do the
434                  * mapping all the time
435                  */
436                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
437
438                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
439                                            dqp->q_blkno,
440                                            mp->m_quotainfo->qi_dqchunklen,
441                                            0, &bp);
442                 if (error || !bp)
443                         return XFS_ERROR(error);
444         }
445
446         ASSERT(xfs_buf_islocked(bp));
447
448         /*
449          * calculate the location of the dquot inside the buffer.
450          */
451         ddq = bp->b_addr + dqp->q_bufoffset;
452
453         /*
454          * A simple sanity check in case we got a corrupted dquot...
455          */
456         error = xfs_qm_dqcheck(mp, ddq, id, dqp->dq_flags & XFS_DQ_ALLTYPES,
457                            flags & (XFS_QMOPT_DQREPAIR|XFS_QMOPT_DOWARN),
458                            "dqtobp");
459         if (error) {
460                 if (!(flags & XFS_QMOPT_DQREPAIR)) {
461                         xfs_trans_brelse(tp, bp);
462                         return XFS_ERROR(EIO);
463                 }
464         }
465
466         *O_bpp = bp;
467         *O_ddpp = ddq;
468
469         return (0);
470 }
471
472
473 /*
474  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
475  * and release the buffer immediately.
476  *
477  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
478  */
479 int
480 xfs_qm_dqread(
481         struct xfs_mount        *mp,
482         xfs_dqid_t              id,
483         uint                    type,
484         uint                    flags,
485         struct xfs_dquot        **O_dqpp)
486 {
487         struct xfs_dquot        *dqp;
488         struct xfs_disk_dquot   *ddqp;
489         struct xfs_buf          *bp;
490         struct xfs_trans        *tp = NULL;
491         int                     error;
492         int                     cancelflags = 0;
493
494
495         dqp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
496
497         dqp->dq_flags = type;
498         dqp->q_core.d_id = cpu_to_be32(id);
499         dqp->q_mount = mp;
500         INIT_LIST_HEAD(&dqp->q_freelist);
501         mutex_init(&dqp->q_qlock);
502         init_waitqueue_head(&dqp->q_pinwait);
503
504         /*
505          * Because we want to use a counting completion, complete
506          * the flush completion once to allow a single access to
507          * the flush completion without blocking.
508          */
509         init_completion(&dqp->q_flush);
510         complete(&dqp->q_flush);
511
512         /*
513          * Make sure group quotas have a different lock class than user
514          * quotas.
515          */
516         if (!(type & XFS_DQ_USER))
517                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
518
519         atomic_inc(&xfs_Gqm->qm_totaldquots);
520
521         trace_xfs_dqread(dqp);
522
523         if (flags & XFS_QMOPT_DQALLOC) {
524                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
525                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
526                                 XFS_WRITE_LOG_RES(mp) +
527                                 /*
528                                  * Round the chunklen up to the next multiple
529                                  * of 128 (buf log item chunk size)).
530                                  */
531                                 BBTOB(mp->m_quotainfo->qi_dqchunklen) - 1 + 128,
532                                 0,
533                                 XFS_TRANS_PERM_LOG_RES,
534                                 XFS_WRITE_LOG_COUNT);
535                 if (error)
536                         goto error1;
537                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
538         }
539
540         /*
541          * get a pointer to the on-disk dquot and the buffer containing it
542          * dqp already knows its own type (GROUP/USER).
543          */
544         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
545         if (error) {
546                 /*
547                  * This can happen if quotas got turned off (ESRCH),
548                  * or if the dquot didn't exist on disk and we ask to
549                  * allocate (ENOENT).
550                  */
551                 trace_xfs_dqread_fail(dqp);
552                 cancelflags |= XFS_TRANS_ABORT;
553                 goto error1;
554         }
555
556         /* copy everything from disk dquot to the incore dquot */
557         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
558         xfs_qm_dquot_logitem_init(dqp);
559
560         /*
561          * Reservation counters are defined as reservation plus current usage
562          * to avoid having to add every time.
563          */
564         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
565         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
566         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
567
568         /* Mark the buf so that this will stay incore a little longer */
569         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
570
571         /*
572          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
573          * So we need to release with xfs_trans_brelse().
574          * The strategy here is identical to that of inodes; we lock
575          * the dquot in xfs_qm_dqget() before making it accessible to
576          * others. This is because dquots, like inodes, need a good level of
577          * concurrency, and we don't want to take locks on the entire buffers
578          * for dquot accesses.
579          * Note also that the dquot buffer may even be dirty at this point, if
580          * this particular dquot was repaired. We still aren't afraid to
581          * brelse it because we have the changes incore.
582          */
583         ASSERT(xfs_buf_islocked(bp));
584         xfs_trans_brelse(tp, bp);
585
586         if (tp) {
587                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
588                 if (error)
589                         goto error0;
590         }
591
592         *O_dqpp = dqp;
593         return error;
594
595 error1:
596         if (tp)
597                 xfs_trans_cancel(tp, cancelflags);
598 error0:
599         xfs_qm_dqdestroy(dqp);
600         *O_dqpp = NULL;
601         return error;
602 }
603
604 /*
605  * Lookup a dquot in the incore dquot hashtable. We keep two separate
606  * hashtables for user and group dquots; and, these are global tables
607  * inside the XQM, not per-filesystem tables.
608  * The hash chain must be locked by caller, and it is left locked
609  * on return. Returning dquot is locked.
610  */
611 STATIC int
612 xfs_qm_dqlookup(
613         xfs_mount_t             *mp,
614         xfs_dqid_t              id,
615         xfs_dqhash_t            *qh,
616         xfs_dquot_t             **O_dqpp)
617 {
618         xfs_dquot_t             *dqp;
619
620         ASSERT(mutex_is_locked(&qh->qh_lock));
621
622         /*
623          * Traverse the hashchain looking for a match
624          */
625         list_for_each_entry(dqp, &qh->qh_list, q_hashlist) {
626                 /*
627                  * We already have the hashlock. We don't need the
628                  * dqlock to look at the id field of the dquot, since the
629                  * id can't be modified without the hashlock anyway.
630                  */
631                 if (be32_to_cpu(dqp->q_core.d_id) != id || dqp->q_mount != mp)
632                         continue;
633
634                 trace_xfs_dqlookup_found(dqp);
635
636                 xfs_dqlock(dqp);
637                 if (dqp->dq_flags & XFS_DQ_FREEING) {
638                         *O_dqpp = NULL;
639                         xfs_dqunlock(dqp);
640                         return -1;
641                 }
642
643                 dqp->q_nrefs++;
644
645                 /*
646                  * move the dquot to the front of the hashchain
647                  */
648                 list_move(&dqp->q_hashlist, &qh->qh_list);
649                 trace_xfs_dqlookup_done(dqp);
650                 *O_dqpp = dqp;
651                 return 0;
652         }
653
654         *O_dqpp = NULL;
655         return 1;
656 }
657
658 /*
659  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
660  * a locked dquot, doing an allocation (if requested) as needed.
661  * When both an inode and an id are given, the inode's id takes precedence.
662  * That is, if the id changes while we don't hold the ilock inside this
663  * function, the new dquot is returned, not necessarily the one requested
664  * in the id argument.
665  */
666 int
667 xfs_qm_dqget(
668         xfs_mount_t     *mp,
669         xfs_inode_t     *ip,      /* locked inode (optional) */
670         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
671         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
672         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
673         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
674 {
675         xfs_dquot_t     *dqp;
676         xfs_dqhash_t    *h;
677         uint            version;
678         int             error;
679
680         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
681         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
682             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
683             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
684                 return (ESRCH);
685         }
686         h = XFS_DQ_HASH(mp, id, type);
687
688 #ifdef DEBUG
689         if (xfs_do_dqerror) {
690                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
691                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
692                         xfs_debug(mp, "Returning error in dqget");
693                         return (EIO);
694                 }
695         }
696
697         ASSERT(type == XFS_DQ_USER ||
698                type == XFS_DQ_PROJ ||
699                type == XFS_DQ_GROUP);
700         if (ip) {
701                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
702                 if (type == XFS_DQ_USER)
703                         ASSERT(ip->i_udquot == NULL);
704                 else
705                         ASSERT(ip->i_gdquot == NULL);
706         }
707 #endif
708
709 restart:
710         mutex_lock(&h->qh_lock);
711
712         /*
713          * Look in the cache (hashtable).
714          * The chain is kept locked during lookup.
715          */
716         switch (xfs_qm_dqlookup(mp, id, h, O_dqpp)) {
717         case -1:
718                 XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
719                 mutex_unlock(&h->qh_lock);
720                 delay(1);
721                 goto restart;
722         case 0:
723                 XQM_STATS_INC(xqmstats.xs_qm_dqcachehits);
724                 /*
725                  * The dquot was found, moved to the front of the chain,
726                  * taken off the freelist if it was on it, and locked
727                  * at this point. Just unlock the hashchain and return.
728                  */
729                 ASSERT(*O_dqpp);
730                 ASSERT(XFS_DQ_IS_LOCKED(*O_dqpp));
731                 mutex_unlock(&h->qh_lock);
732                 trace_xfs_dqget_hit(*O_dqpp);
733                 return 0;       /* success */
734         default:
735                 XQM_STATS_INC(xqmstats.xs_qm_dqcachemisses);
736                 break;
737         }
738
739         /*
740          * Dquot cache miss. We don't want to keep the inode lock across
741          * a (potential) disk read. Also we don't want to deal with the lock
742          * ordering between quotainode and this inode. OTOH, dropping the inode
743          * lock here means dealing with a chown that can happen before
744          * we re-acquire the lock.
745          */
746         if (ip)
747                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
748         /*
749          * Save the hashchain version stamp, and unlock the chain, so that
750          * we don't keep the lock across a disk read
751          */
752         version = h->qh_version;
753         mutex_unlock(&h->qh_lock);
754
755         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
756
757         if (ip)
758                 xfs_ilock(ip, XFS_ILOCK_EXCL);
759
760         if (error)
761                 return error;
762
763         /*
764          * Dquot lock comes after hashlock in the lock ordering
765          */
766         if (ip) {
767                 /*
768                  * A dquot could be attached to this inode by now, since
769                  * we had dropped the ilock.
770                  */
771                 if (type == XFS_DQ_USER) {
772                         if (!XFS_IS_UQUOTA_ON(mp)) {
773                                 /* inode stays locked on return */
774                                 xfs_qm_dqdestroy(dqp);
775                                 return XFS_ERROR(ESRCH);
776                         }
777                         if (ip->i_udquot) {
778                                 xfs_qm_dqdestroy(dqp);
779                                 dqp = ip->i_udquot;
780                                 xfs_dqlock(dqp);
781                                 goto dqret;
782                         }
783                 } else {
784                         if (!XFS_IS_OQUOTA_ON(mp)) {
785                                 /* inode stays locked on return */
786                                 xfs_qm_dqdestroy(dqp);
787                                 return XFS_ERROR(ESRCH);
788                         }
789                         if (ip->i_gdquot) {
790                                 xfs_qm_dqdestroy(dqp);
791                                 dqp = ip->i_gdquot;
792                                 xfs_dqlock(dqp);
793                                 goto dqret;
794                         }
795                 }
796         }
797
798         /*
799          * Hashlock comes after ilock in lock order
800          */
801         mutex_lock(&h->qh_lock);
802         if (version != h->qh_version) {
803                 xfs_dquot_t *tmpdqp;
804                 /*
805                  * Now, see if somebody else put the dquot in the
806                  * hashtable before us. This can happen because we didn't
807                  * keep the hashchain lock. We don't have to worry about
808                  * lock order between the two dquots here since dqp isn't
809                  * on any findable lists yet.
810                  */
811                 switch (xfs_qm_dqlookup(mp, id, h, &tmpdqp)) {
812                 case 0:
813                 case -1:
814                         /*
815                          * Duplicate found, either in cache or on its way out.
816                          * Just throw away the new dquot and start over.
817                          */
818                         if (tmpdqp)
819                                 xfs_qm_dqput(tmpdqp);
820                         mutex_unlock(&h->qh_lock);
821                         xfs_qm_dqdestroy(dqp);
822                         XQM_STATS_INC(xqmstats.xs_qm_dquot_dups);
823                         goto restart;
824                 default:
825                         break;
826                 }
827         }
828
829         /*
830          * Put the dquot at the beginning of the hash-chain and mp's list
831          * LOCK ORDER: hashlock, freelistlock, mplistlock, udqlock, gdqlock ..
832          */
833         ASSERT(mutex_is_locked(&h->qh_lock));
834         dqp->q_hash = h;
835         list_add(&dqp->q_hashlist, &h->qh_list);
836         h->qh_version++;
837
838         /*
839          * Attach this dquot to this filesystem's list of all dquots,
840          * kept inside the mount structure in m_quotainfo field
841          */
842         mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
843
844         /*
845          * We return a locked dquot to the caller, with a reference taken
846          */
847         xfs_dqlock(dqp);
848         dqp->q_nrefs = 1;
849
850         list_add(&dqp->q_mplist, &mp->m_quotainfo->qi_dqlist);
851         mp->m_quotainfo->qi_dquots++;
852         mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
853         mutex_unlock(&h->qh_lock);
854  dqret:
855         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
856         trace_xfs_dqget_miss(dqp);
857         *O_dqpp = dqp;
858         return (0);
859 }
860
861
862 /*
863  * Release a reference to the dquot (decrement ref-count)
864  * and unlock it. If there is a group quota attached to this
865  * dquot, carefully release that too without tripping over
866  * deadlocks'n'stuff.
867  */
868 void
869 xfs_qm_dqput(
870         struct xfs_dquot        *dqp)
871 {
872         struct xfs_dquot        *gdqp;
873
874         ASSERT(dqp->q_nrefs > 0);
875         ASSERT(XFS_DQ_IS_LOCKED(dqp));
876
877         trace_xfs_dqput(dqp);
878
879 recurse:
880         if (--dqp->q_nrefs > 0) {
881                 xfs_dqunlock(dqp);
882                 return;
883         }
884
885         trace_xfs_dqput_free(dqp);
886
887         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
888         if (list_empty(&dqp->q_freelist)) {
889                 list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
890                 xfs_Gqm->qm_dqfrlist_cnt++;
891         }
892         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
893
894         /*
895          * If we just added a udquot to the freelist, then we want to release
896          * the gdquot reference that it (probably) has. Otherwise it'll keep
897          * the gdquot from getting reclaimed.
898          */
899         gdqp = dqp->q_gdquot;
900         if (gdqp) {
901                 xfs_dqlock(gdqp);
902                 dqp->q_gdquot = NULL;
903         }
904         xfs_dqunlock(dqp);
905
906         /*
907          * If we had a group quota hint, release it now.
908          */
909         if (gdqp) {
910                 dqp = gdqp;
911                 goto recurse;
912         }
913 }
914
915 /*
916  * Release a dquot. Flush it if dirty, then dqput() it.
917  * dquot must not be locked.
918  */
919 void
920 xfs_qm_dqrele(
921         xfs_dquot_t     *dqp)
922 {
923         if (!dqp)
924                 return;
925
926         trace_xfs_dqrele(dqp);
927
928         xfs_dqlock(dqp);
929         /*
930          * We don't care to flush it if the dquot is dirty here.
931          * That will create stutters that we want to avoid.
932          * Instead we do a delayed write when we try to reclaim
933          * a dirty dquot. Also xfs_sync will take part of the burden...
934          */
935         xfs_qm_dqput(dqp);
936 }
937
938 /*
939  * This is the dquot flushing I/O completion routine.  It is called
940  * from interrupt level when the buffer containing the dquot is
941  * flushed to disk.  It is responsible for removing the dquot logitem
942  * from the AIL if it has not been re-logged, and unlocking the dquot's
943  * flush lock. This behavior is very similar to that of inodes..
944  */
945 STATIC void
946 xfs_qm_dqflush_done(
947         struct xfs_buf          *bp,
948         struct xfs_log_item     *lip)
949 {
950         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
951         xfs_dquot_t             *dqp = qip->qli_dquot;
952         struct xfs_ail          *ailp = lip->li_ailp;
953
954         /*
955          * We only want to pull the item from the AIL if its
956          * location in the log has not changed since we started the flush.
957          * Thus, we only bother if the dquot's lsn has
958          * not changed. First we check the lsn outside the lock
959          * since it's cheaper, and then we recheck while
960          * holding the lock before removing the dquot from the AIL.
961          */
962         if ((lip->li_flags & XFS_LI_IN_AIL) &&
963             lip->li_lsn == qip->qli_flush_lsn) {
964
965                 /* xfs_trans_ail_delete() drops the AIL lock. */
966                 spin_lock(&ailp->xa_lock);
967                 if (lip->li_lsn == qip->qli_flush_lsn)
968                         xfs_trans_ail_delete(ailp, lip);
969                 else
970                         spin_unlock(&ailp->xa_lock);
971         }
972
973         /*
974          * Release the dq's flush lock since we're done with it.
975          */
976         xfs_dqfunlock(dqp);
977 }
978
979 /*
980  * Write a modified dquot to disk.
981  * The dquot must be locked and the flush lock too taken by caller.
982  * The flush lock will not be unlocked until the dquot reaches the disk,
983  * but the dquot is free to be unlocked and modified by the caller
984  * in the interim. Dquot is still locked on return. This behavior is
985  * identical to that of inodes.
986  */
987 int
988 xfs_qm_dqflush(
989         xfs_dquot_t             *dqp,
990         uint                    flags)
991 {
992         struct xfs_mount        *mp = dqp->q_mount;
993         struct xfs_buf          *bp;
994         struct xfs_disk_dquot   *ddqp;
995         int                     error;
996
997         ASSERT(XFS_DQ_IS_LOCKED(dqp));
998         ASSERT(!completion_done(&dqp->q_flush));
999
1000         trace_xfs_dqflush(dqp);
1001
1002         /*
1003          * If not dirty, or it's pinned and we are not supposed to block, nada.
1004          */
1005         if (!XFS_DQ_IS_DIRTY(dqp) ||
1006             ((flags & SYNC_TRYLOCK) && atomic_read(&dqp->q_pincount) > 0)) {
1007                 xfs_dqfunlock(dqp);
1008                 return 0;
1009         }
1010         xfs_qm_dqunpin_wait(dqp);
1011
1012         /*
1013          * This may have been unpinned because the filesystem is shutting
1014          * down forcibly. If that's the case we must not write this dquot
1015          * to disk, because the log record didn't make it to disk!
1016          */
1017         if (XFS_FORCED_SHUTDOWN(mp)) {
1018                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1019                 xfs_dqfunlock(dqp);
1020                 return XFS_ERROR(EIO);
1021         }
1022
1023         /*
1024          * Get the buffer containing the on-disk dquot
1025          */
1026         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1027                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp);
1028         if (error) {
1029                 ASSERT(error != ENOENT);
1030                 xfs_dqfunlock(dqp);
1031                 return error;
1032         }
1033
1034         /*
1035          * Calculate the location of the dquot inside the buffer.
1036          */
1037         ddqp = bp->b_addr + dqp->q_bufoffset;
1038
1039         /*
1040          * A simple sanity check in case we got a corrupted dquot..
1041          */
1042         error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1043                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1044         if (error) {
1045                 xfs_buf_relse(bp);
1046                 xfs_dqfunlock(dqp);
1047                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1048                 return XFS_ERROR(EIO);
1049         }
1050
1051         /* This is the only portion of data that needs to persist */
1052         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1053
1054         /*
1055          * Clear the dirty field and remember the flush lsn for later use.
1056          */
1057         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1058
1059         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1060                                         &dqp->q_logitem.qli_item.li_lsn);
1061
1062         /*
1063          * Attach an iodone routine so that we can remove this dquot from the
1064          * AIL and release the flush lock once the dquot is synced to disk.
1065          */
1066         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1067                                   &dqp->q_logitem.qli_item);
1068
1069         /*
1070          * If the buffer is pinned then push on the log so we won't
1071          * get stuck waiting in the write for too long.
1072          */
1073         if (xfs_buf_ispinned(bp)) {
1074                 trace_xfs_dqflush_force(dqp);
1075                 xfs_log_force(mp, 0);
1076         }
1077
1078         if (flags & SYNC_WAIT)
1079                 error = xfs_bwrite(bp);
1080         else
1081                 xfs_buf_delwri_queue(bp);
1082
1083         xfs_buf_relse(bp);
1084
1085         trace_xfs_dqflush_done(dqp);
1086
1087         /*
1088          * dqp is still locked, but caller is free to unlock it now.
1089          */
1090         return error;
1091
1092 }
1093
1094 void
1095 xfs_dqunlock(
1096         xfs_dquot_t *dqp)
1097 {
1098         xfs_dqunlock_nonotify(dqp);
1099         if (dqp->q_logitem.qli_dquot == dqp) {
1100                 xfs_trans_unlocked_item(dqp->q_logitem.qli_item.li_ailp,
1101                                         &dqp->q_logitem.qli_item);
1102         }
1103 }
1104
1105 /*
1106  * Lock two xfs_dquot structures.
1107  *
1108  * To avoid deadlocks we always lock the quota structure with
1109  * the lowerd id first.
1110  */
1111 void
1112 xfs_dqlock2(
1113         xfs_dquot_t     *d1,
1114         xfs_dquot_t     *d2)
1115 {
1116         if (d1 && d2) {
1117                 ASSERT(d1 != d2);
1118                 if (be32_to_cpu(d1->q_core.d_id) >
1119                     be32_to_cpu(d2->q_core.d_id)) {
1120                         mutex_lock(&d2->q_qlock);
1121                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1122                 } else {
1123                         mutex_lock(&d1->q_qlock);
1124                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1125                 }
1126         } else if (d1) {
1127                 mutex_lock(&d1->q_qlock);
1128         } else if (d2) {
1129                 mutex_lock(&d2->q_qlock);
1130         }
1131 }
1132
1133 /*
1134  * Take a dquot out of the mount's dqlist as well as the hashlist.  This is
1135  * called via unmount as well as quotaoff, and the purge will always succeed.
1136  */
1137 void
1138 xfs_qm_dqpurge(
1139         struct xfs_dquot        *dqp)
1140 {
1141         struct xfs_mount        *mp = dqp->q_mount;
1142         struct xfs_dqhash       *qh = dqp->q_hash;
1143
1144         xfs_dqlock(dqp);
1145
1146         /*
1147          * If we're turning off quotas, we have to make sure that, for
1148          * example, we don't delete quota disk blocks while dquots are
1149          * in the process of getting written to those disk blocks.
1150          * This dquot might well be on AIL, and we can't leave it there
1151          * if we're turning off quotas. Basically, we need this flush
1152          * lock, and are willing to block on it.
1153          */
1154         if (!xfs_dqflock_nowait(dqp)) {
1155                 /*
1156                  * Block on the flush lock after nudging dquot buffer,
1157                  * if it is incore.
1158                  */
1159                 xfs_dqflock_pushbuf_wait(dqp);
1160         }
1161
1162         /*
1163          * If we are turning this type of quotas off, we don't care
1164          * about the dirty metadata sitting in this dquot. OTOH, if
1165          * we're unmounting, we do care, so we flush it and wait.
1166          */
1167         if (XFS_DQ_IS_DIRTY(dqp)) {
1168                 int     error;
1169
1170                 /*
1171                  * We don't care about getting disk errors here. We need
1172                  * to purge this dquot anyway, so we go ahead regardless.
1173                  */
1174                 error = xfs_qm_dqflush(dqp, SYNC_WAIT);
1175                 if (error)
1176                         xfs_warn(mp, "%s: dquot %p flush failed",
1177                                 __func__, dqp);
1178                 xfs_dqflock(dqp);
1179         }
1180
1181         ASSERT(atomic_read(&dqp->q_pincount) == 0);
1182         ASSERT(XFS_FORCED_SHUTDOWN(mp) ||
1183                !(dqp->q_logitem.qli_item.li_flags & XFS_LI_IN_AIL));
1184
1185         xfs_dqfunlock(dqp);
1186         xfs_dqunlock(dqp);
1187
1188         mutex_lock(&qh->qh_lock);
1189         list_del_init(&dqp->q_hashlist);
1190         qh->qh_version++;
1191         mutex_unlock(&qh->qh_lock);
1192
1193         mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
1194         list_del_init(&dqp->q_mplist);
1195         mp->m_quotainfo->qi_dqreclaims++;
1196         mp->m_quotainfo->qi_dquots--;
1197         mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
1198
1199         /*
1200          * We move dquots to the freelist as soon as their reference count
1201          * hits zero, so it really should be on the freelist here.
1202          */
1203         mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
1204         ASSERT(!list_empty(&dqp->q_freelist));
1205         list_del_init(&dqp->q_freelist);
1206         xfs_Gqm->qm_dqfrlist_cnt--;
1207         mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
1208
1209         xfs_qm_dqdestroy(dqp);
1210 }
1211
1212 /*
1213  * Give the buffer a little push if it is incore and
1214  * wait on the flush lock.
1215  */
1216 void
1217 xfs_dqflock_pushbuf_wait(
1218         xfs_dquot_t     *dqp)
1219 {
1220         xfs_mount_t     *mp = dqp->q_mount;
1221         xfs_buf_t       *bp;
1222
1223         /*
1224          * Check to see if the dquot has been flushed delayed
1225          * write.  If so, grab its buffer and send it
1226          * out immediately.  We'll be able to acquire
1227          * the flush lock when the I/O completes.
1228          */
1229         bp = xfs_incore(mp->m_ddev_targp, dqp->q_blkno,
1230                         mp->m_quotainfo->qi_dqchunklen, XBF_TRYLOCK);
1231         if (!bp)
1232                 goto out_lock;
1233
1234         if (XFS_BUF_ISDELAYWRITE(bp)) {
1235                 if (xfs_buf_ispinned(bp))
1236                         xfs_log_force(mp, 0);
1237                 xfs_buf_delwri_promote(bp);
1238                 wake_up_process(bp->b_target->bt_task);
1239         }
1240         xfs_buf_relse(bp);
1241 out_lock:
1242         xfs_dqflock(dqp);
1243 }