Merge tag 'please-pull-fix-ia64-warnings' of git://git.kernel.org/pub/scm/linux/kerne...
[cascardo/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_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_trans.h"
23 #include "xfs_sb.h"
24 #include "xfs_ag.h"
25 #include "xfs_alloc.h"
26 #include "xfs_quota.h"
27 #include "xfs_mount.h"
28 #include "xfs_bmap_btree.h"
29 #include "xfs_inode.h"
30 #include "xfs_bmap.h"
31 #include "xfs_rtalloc.h"
32 #include "xfs_error.h"
33 #include "xfs_itable.h"
34 #include "xfs_attr.h"
35 #include "xfs_buf_item.h"
36 #include "xfs_trans_space.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_qm.h"
39 #include "xfs_cksum.h"
40 #include "xfs_trace.h"
41
42 /*
43  * Lock order:
44  *
45  * ip->i_lock
46  *   qi->qi_tree_lock
47  *     dquot->q_qlock (xfs_dqlock() and friends)
48  *       dquot->q_flush (xfs_dqflock() and friends)
49  *       qi->qi_lru_lock
50  *
51  * If two dquots need to be locked the order is user before group/project,
52  * otherwise by the lowest id first, see xfs_dqlock2.
53  */
54
55 #ifdef DEBUG
56 xfs_buftarg_t *xfs_dqerror_target;
57 int xfs_do_dqerror;
58 int xfs_dqreq_num;
59 int xfs_dqerror_mod = 33;
60 #endif
61
62 struct kmem_zone                *xfs_qm_dqtrxzone;
63 static struct kmem_zone         *xfs_qm_dqzone;
64
65 static struct lock_class_key xfs_dquot_other_class;
66
67 /*
68  * This is called to free all the memory associated with a dquot
69  */
70 void
71 xfs_qm_dqdestroy(
72         xfs_dquot_t     *dqp)
73 {
74         ASSERT(list_empty(&dqp->q_lru));
75
76         mutex_destroy(&dqp->q_qlock);
77         kmem_zone_free(xfs_qm_dqzone, dqp);
78
79         XFS_STATS_DEC(xs_qm_dquot);
80 }
81
82 /*
83  * If default limits are in force, push them into the dquot now.
84  * We overwrite the dquot limits only if they are zero and this
85  * is not the root dquot.
86  */
87 void
88 xfs_qm_adjust_dqlimits(
89         struct xfs_mount        *mp,
90         struct xfs_dquot        *dq)
91 {
92         struct xfs_quotainfo    *q = mp->m_quotainfo;
93         struct xfs_disk_dquot   *d = &dq->q_core;
94         int                     prealloc = 0;
95
96         ASSERT(d->d_id);
97
98         if (q->qi_bsoftlimit && !d->d_blk_softlimit) {
99                 d->d_blk_softlimit = cpu_to_be64(q->qi_bsoftlimit);
100                 prealloc = 1;
101         }
102         if (q->qi_bhardlimit && !d->d_blk_hardlimit) {
103                 d->d_blk_hardlimit = cpu_to_be64(q->qi_bhardlimit);
104                 prealloc = 1;
105         }
106         if (q->qi_isoftlimit && !d->d_ino_softlimit)
107                 d->d_ino_softlimit = cpu_to_be64(q->qi_isoftlimit);
108         if (q->qi_ihardlimit && !d->d_ino_hardlimit)
109                 d->d_ino_hardlimit = cpu_to_be64(q->qi_ihardlimit);
110         if (q->qi_rtbsoftlimit && !d->d_rtb_softlimit)
111                 d->d_rtb_softlimit = cpu_to_be64(q->qi_rtbsoftlimit);
112         if (q->qi_rtbhardlimit && !d->d_rtb_hardlimit)
113                 d->d_rtb_hardlimit = cpu_to_be64(q->qi_rtbhardlimit);
114
115         if (prealloc)
116                 xfs_dquot_set_prealloc_limits(dq);
117 }
118
119 /*
120  * Check the limits and timers of a dquot and start or reset timers
121  * if necessary.
122  * This gets called even when quota enforcement is OFF, which makes our
123  * life a little less complicated. (We just don't reject any quota
124  * reservations in that case, when enforcement is off).
125  * We also return 0 as the values of the timers in Q_GETQUOTA calls, when
126  * enforcement's off.
127  * In contrast, warnings are a little different in that they don't
128  * 'automatically' get started when limits get exceeded.  They do
129  * get reset to zero, however, when we find the count to be under
130  * the soft limit (they are only ever set non-zero via userspace).
131  */
132 void
133 xfs_qm_adjust_dqtimers(
134         xfs_mount_t             *mp,
135         xfs_disk_dquot_t        *d)
136 {
137         ASSERT(d->d_id);
138
139 #ifdef DEBUG
140         if (d->d_blk_hardlimit)
141                 ASSERT(be64_to_cpu(d->d_blk_softlimit) <=
142                        be64_to_cpu(d->d_blk_hardlimit));
143         if (d->d_ino_hardlimit)
144                 ASSERT(be64_to_cpu(d->d_ino_softlimit) <=
145                        be64_to_cpu(d->d_ino_hardlimit));
146         if (d->d_rtb_hardlimit)
147                 ASSERT(be64_to_cpu(d->d_rtb_softlimit) <=
148                        be64_to_cpu(d->d_rtb_hardlimit));
149 #endif
150
151         if (!d->d_btimer) {
152                 if ((d->d_blk_softlimit &&
153                      (be64_to_cpu(d->d_bcount) >
154                       be64_to_cpu(d->d_blk_softlimit))) ||
155                     (d->d_blk_hardlimit &&
156                      (be64_to_cpu(d->d_bcount) >
157                       be64_to_cpu(d->d_blk_hardlimit)))) {
158                         d->d_btimer = cpu_to_be32(get_seconds() +
159                                         mp->m_quotainfo->qi_btimelimit);
160                 } else {
161                         d->d_bwarns = 0;
162                 }
163         } else {
164                 if ((!d->d_blk_softlimit ||
165                      (be64_to_cpu(d->d_bcount) <=
166                       be64_to_cpu(d->d_blk_softlimit))) &&
167                     (!d->d_blk_hardlimit ||
168                     (be64_to_cpu(d->d_bcount) <=
169                      be64_to_cpu(d->d_blk_hardlimit)))) {
170                         d->d_btimer = 0;
171                 }
172         }
173
174         if (!d->d_itimer) {
175                 if ((d->d_ino_softlimit &&
176                      (be64_to_cpu(d->d_icount) >
177                       be64_to_cpu(d->d_ino_softlimit))) ||
178                     (d->d_ino_hardlimit &&
179                      (be64_to_cpu(d->d_icount) >
180                       be64_to_cpu(d->d_ino_hardlimit)))) {
181                         d->d_itimer = cpu_to_be32(get_seconds() +
182                                         mp->m_quotainfo->qi_itimelimit);
183                 } else {
184                         d->d_iwarns = 0;
185                 }
186         } else {
187                 if ((!d->d_ino_softlimit ||
188                      (be64_to_cpu(d->d_icount) <=
189                       be64_to_cpu(d->d_ino_softlimit)))  &&
190                     (!d->d_ino_hardlimit ||
191                      (be64_to_cpu(d->d_icount) <=
192                       be64_to_cpu(d->d_ino_hardlimit)))) {
193                         d->d_itimer = 0;
194                 }
195         }
196
197         if (!d->d_rtbtimer) {
198                 if ((d->d_rtb_softlimit &&
199                      (be64_to_cpu(d->d_rtbcount) >
200                       be64_to_cpu(d->d_rtb_softlimit))) ||
201                     (d->d_rtb_hardlimit &&
202                      (be64_to_cpu(d->d_rtbcount) >
203                       be64_to_cpu(d->d_rtb_hardlimit)))) {
204                         d->d_rtbtimer = cpu_to_be32(get_seconds() +
205                                         mp->m_quotainfo->qi_rtbtimelimit);
206                 } else {
207                         d->d_rtbwarns = 0;
208                 }
209         } else {
210                 if ((!d->d_rtb_softlimit ||
211                      (be64_to_cpu(d->d_rtbcount) <=
212                       be64_to_cpu(d->d_rtb_softlimit))) &&
213                     (!d->d_rtb_hardlimit ||
214                      (be64_to_cpu(d->d_rtbcount) <=
215                       be64_to_cpu(d->d_rtb_hardlimit)))) {
216                         d->d_rtbtimer = 0;
217                 }
218         }
219 }
220
221 /*
222  * initialize a buffer full of dquots and log the whole thing
223  */
224 STATIC void
225 xfs_qm_init_dquot_blk(
226         xfs_trans_t     *tp,
227         xfs_mount_t     *mp,
228         xfs_dqid_t      id,
229         uint            type,
230         xfs_buf_t       *bp)
231 {
232         struct xfs_quotainfo    *q = mp->m_quotainfo;
233         xfs_dqblk_t     *d;
234         int             curid, i;
235
236         ASSERT(tp);
237         ASSERT(xfs_buf_islocked(bp));
238
239         d = bp->b_addr;
240
241         /*
242          * ID of the first dquot in the block - id's are zero based.
243          */
244         curid = id - (id % q->qi_dqperchunk);
245         ASSERT(curid >= 0);
246         memset(d, 0, BBTOB(q->qi_dqchunklen));
247         for (i = 0; i < q->qi_dqperchunk; i++, d++, curid++) {
248                 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
249                 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
250                 d->dd_diskdq.d_id = cpu_to_be32(curid);
251                 d->dd_diskdq.d_flags = type;
252                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
253                         uuid_copy(&d->dd_uuid, &mp->m_sb.sb_uuid);
254                         xfs_update_cksum((char *)d, sizeof(struct xfs_dqblk),
255                                          XFS_DQUOT_CRC_OFF);
256                 }
257         }
258
259         xfs_trans_dquot_buf(tp, bp,
260                             (type & XFS_DQ_USER ? XFS_BLF_UDQUOT_BUF :
261                             ((type & XFS_DQ_PROJ) ? XFS_BLF_PDQUOT_BUF :
262                              XFS_BLF_GDQUOT_BUF)));
263         xfs_trans_log_buf(tp, bp, 0, BBTOB(q->qi_dqchunklen) - 1);
264 }
265
266 /*
267  * Initialize the dynamic speculative preallocation thresholds. The lo/hi
268  * watermarks correspond to the soft and hard limits by default. If a soft limit
269  * is not specified, we use 95% of the hard limit.
270  */
271 void
272 xfs_dquot_set_prealloc_limits(struct xfs_dquot *dqp)
273 {
274         __uint64_t space;
275
276         dqp->q_prealloc_hi_wmark = be64_to_cpu(dqp->q_core.d_blk_hardlimit);
277         dqp->q_prealloc_lo_wmark = be64_to_cpu(dqp->q_core.d_blk_softlimit);
278         if (!dqp->q_prealloc_lo_wmark) {
279                 dqp->q_prealloc_lo_wmark = dqp->q_prealloc_hi_wmark;
280                 do_div(dqp->q_prealloc_lo_wmark, 100);
281                 dqp->q_prealloc_lo_wmark *= 95;
282         }
283
284         space = dqp->q_prealloc_hi_wmark;
285
286         do_div(space, 100);
287         dqp->q_low_space[XFS_QLOWSP_1_PCNT] = space;
288         dqp->q_low_space[XFS_QLOWSP_3_PCNT] = space * 3;
289         dqp->q_low_space[XFS_QLOWSP_5_PCNT] = space * 5;
290 }
291
292 STATIC bool
293 xfs_dquot_buf_verify_crc(
294         struct xfs_mount        *mp,
295         struct xfs_buf          *bp)
296 {
297         struct xfs_dqblk        *d = (struct xfs_dqblk *)bp->b_addr;
298         int                     ndquots;
299         int                     i;
300
301         if (!xfs_sb_version_hascrc(&mp->m_sb))
302                 return true;
303
304         /*
305          * if we are in log recovery, the quota subsystem has not been
306          * initialised so we have no quotainfo structure. In that case, we need
307          * to manually calculate the number of dquots in the buffer.
308          */
309         if (mp->m_quotainfo)
310                 ndquots = mp->m_quotainfo->qi_dqperchunk;
311         else
312                 ndquots = xfs_qm_calc_dquots_per_chunk(mp,
313                                         XFS_BB_TO_FSB(mp, bp->b_length));
314
315         for (i = 0; i < ndquots; i++, d++) {
316                 if (!xfs_verify_cksum((char *)d, sizeof(struct xfs_dqblk),
317                                  XFS_DQUOT_CRC_OFF))
318                         return false;
319                 if (!uuid_equal(&d->dd_uuid, &mp->m_sb.sb_uuid))
320                         return false;
321         }
322         return true;
323 }
324
325 STATIC bool
326 xfs_dquot_buf_verify(
327         struct xfs_mount        *mp,
328         struct xfs_buf          *bp)
329 {
330         struct xfs_dqblk        *d = (struct xfs_dqblk *)bp->b_addr;
331         xfs_dqid_t              id = 0;
332         int                     ndquots;
333         int                     i;
334
335         /*
336          * if we are in log recovery, the quota subsystem has not been
337          * initialised so we have no quotainfo structure. In that case, we need
338          * to manually calculate the number of dquots in the buffer.
339          */
340         if (mp->m_quotainfo)
341                 ndquots = mp->m_quotainfo->qi_dqperchunk;
342         else
343                 ndquots = xfs_qm_calc_dquots_per_chunk(mp, bp->b_length);
344
345         /*
346          * On the first read of the buffer, verify that each dquot is valid.
347          * We don't know what the id of the dquot is supposed to be, just that
348          * they should be increasing monotonically within the buffer. If the
349          * first id is corrupt, then it will fail on the second dquot in the
350          * buffer so corruptions could point to the wrong dquot in this case.
351          */
352         for (i = 0; i < ndquots; i++) {
353                 struct xfs_disk_dquot   *ddq;
354                 int                     error;
355
356                 ddq = &d[i].dd_diskdq;
357
358                 if (i == 0)
359                         id = be32_to_cpu(ddq->d_id);
360
361                 error = xfs_qm_dqcheck(mp, ddq, id + i, 0, XFS_QMOPT_DOWARN,
362                                        "xfs_dquot_buf_verify");
363                 if (error)
364                         return false;
365         }
366         return true;
367 }
368
369 static void
370 xfs_dquot_buf_read_verify(
371         struct xfs_buf  *bp)
372 {
373         struct xfs_mount        *mp = bp->b_target->bt_mount;
374
375         if (!xfs_dquot_buf_verify_crc(mp, bp) || !xfs_dquot_buf_verify(mp, bp)) {
376                 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
377                 xfs_buf_ioerror(bp, EFSCORRUPTED);
378         }
379 }
380
381 /*
382  * we don't calculate the CRC here as that is done when the dquot is flushed to
383  * the buffer after the update is done. This ensures that the dquot in the
384  * buffer always has an up-to-date CRC value.
385  */
386 void
387 xfs_dquot_buf_write_verify(
388         struct xfs_buf  *bp)
389 {
390         struct xfs_mount        *mp = bp->b_target->bt_mount;
391
392         if (!xfs_dquot_buf_verify(mp, bp)) {
393                 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp, bp->b_addr);
394                 xfs_buf_ioerror(bp, EFSCORRUPTED);
395                 return;
396         }
397 }
398
399 const struct xfs_buf_ops xfs_dquot_buf_ops = {
400         .verify_read = xfs_dquot_buf_read_verify,
401         .verify_write = xfs_dquot_buf_write_verify,
402 };
403
404 /*
405  * Allocate a block and fill it with dquots.
406  * This is called when the bmapi finds a hole.
407  */
408 STATIC int
409 xfs_qm_dqalloc(
410         xfs_trans_t     **tpp,
411         xfs_mount_t     *mp,
412         xfs_dquot_t     *dqp,
413         xfs_inode_t     *quotip,
414         xfs_fileoff_t   offset_fsb,
415         xfs_buf_t       **O_bpp)
416 {
417         xfs_fsblock_t   firstblock;
418         xfs_bmap_free_t flist;
419         xfs_bmbt_irec_t map;
420         int             nmaps, error, committed;
421         xfs_buf_t       *bp;
422         xfs_trans_t     *tp = *tpp;
423
424         ASSERT(tp != NULL);
425
426         trace_xfs_dqalloc(dqp);
427
428         /*
429          * Initialize the bmap freelist prior to calling bmapi code.
430          */
431         xfs_bmap_init(&flist, &firstblock);
432         xfs_ilock(quotip, XFS_ILOCK_EXCL);
433         /*
434          * Return if this type of quotas is turned off while we didn't
435          * have an inode lock
436          */
437         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
438                 xfs_iunlock(quotip, XFS_ILOCK_EXCL);
439                 return (ESRCH);
440         }
441
442         xfs_trans_ijoin(tp, quotip, XFS_ILOCK_EXCL);
443         nmaps = 1;
444         error = xfs_bmapi_write(tp, quotip, offset_fsb,
445                                 XFS_DQUOT_CLUSTER_SIZE_FSB, XFS_BMAPI_METADATA,
446                                 &firstblock, XFS_QM_DQALLOC_SPACE_RES(mp),
447                                 &map, &nmaps, &flist);
448         if (error)
449                 goto error0;
450         ASSERT(map.br_blockcount == XFS_DQUOT_CLUSTER_SIZE_FSB);
451         ASSERT(nmaps == 1);
452         ASSERT((map.br_startblock != DELAYSTARTBLOCK) &&
453                (map.br_startblock != HOLESTARTBLOCK));
454
455         /*
456          * Keep track of the blkno to save a lookup later
457          */
458         dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
459
460         /* now we can just get the buffer (there's nothing to read yet) */
461         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
462                                dqp->q_blkno,
463                                mp->m_quotainfo->qi_dqchunklen,
464                                0);
465
466         error = xfs_buf_geterror(bp);
467         if (error)
468                 goto error1;
469         bp->b_ops = &xfs_dquot_buf_ops;
470
471         /*
472          * Make a chunk of dquots out of this buffer and log
473          * the entire thing.
474          */
475         xfs_qm_init_dquot_blk(tp, mp, be32_to_cpu(dqp->q_core.d_id),
476                               dqp->dq_flags & XFS_DQ_ALLTYPES, bp);
477
478         /*
479          * xfs_bmap_finish() may commit the current transaction and
480          * start a second transaction if the freelist is not empty.
481          *
482          * Since we still want to modify this buffer, we need to
483          * ensure that the buffer is not released on commit of
484          * the first transaction and ensure the buffer is added to the
485          * second transaction.
486          *
487          * If there is only one transaction then don't stop the buffer
488          * from being released when it commits later on.
489          */
490
491         xfs_trans_bhold(tp, bp);
492
493         if ((error = xfs_bmap_finish(tpp, &flist, &committed))) {
494                 goto error1;
495         }
496
497         if (committed) {
498                 tp = *tpp;
499                 xfs_trans_bjoin(tp, bp);
500         } else {
501                 xfs_trans_bhold_release(tp, bp);
502         }
503
504         *O_bpp = bp;
505         return 0;
506
507       error1:
508         xfs_bmap_cancel(&flist);
509       error0:
510         xfs_iunlock(quotip, XFS_ILOCK_EXCL);
511
512         return (error);
513 }
514 STATIC int
515 xfs_qm_dqrepair(
516         struct xfs_mount        *mp,
517         struct xfs_trans        *tp,
518         struct xfs_dquot        *dqp,
519         xfs_dqid_t              firstid,
520         struct xfs_buf          **bpp)
521 {
522         int                     error;
523         struct xfs_disk_dquot   *ddq;
524         struct xfs_dqblk        *d;
525         int                     i;
526
527         /*
528          * Read the buffer without verification so we get the corrupted
529          * buffer returned to us. make sure we verify it on write, though.
530          */
531         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp, dqp->q_blkno,
532                                    mp->m_quotainfo->qi_dqchunklen,
533                                    0, bpp, NULL);
534
535         if (error) {
536                 ASSERT(*bpp == NULL);
537                 return XFS_ERROR(error);
538         }
539         (*bpp)->b_ops = &xfs_dquot_buf_ops;
540
541         ASSERT(xfs_buf_islocked(*bpp));
542         d = (struct xfs_dqblk *)(*bpp)->b_addr;
543
544         /* Do the actual repair of dquots in this buffer */
545         for (i = 0; i < mp->m_quotainfo->qi_dqperchunk; i++) {
546                 ddq = &d[i].dd_diskdq;
547                 error = xfs_qm_dqcheck(mp, ddq, firstid + i,
548                                        dqp->dq_flags & XFS_DQ_ALLTYPES,
549                                        XFS_QMOPT_DQREPAIR, "xfs_qm_dqrepair");
550                 if (error) {
551                         /* repair failed, we're screwed */
552                         xfs_trans_brelse(tp, *bpp);
553                         return XFS_ERROR(EIO);
554                 }
555         }
556
557         return 0;
558 }
559
560 /*
561  * Maps a dquot to the buffer containing its on-disk version.
562  * This returns a ptr to the buffer containing the on-disk dquot
563  * in the bpp param, and a ptr to the on-disk dquot within that buffer
564  */
565 STATIC int
566 xfs_qm_dqtobp(
567         xfs_trans_t             **tpp,
568         xfs_dquot_t             *dqp,
569         xfs_disk_dquot_t        **O_ddpp,
570         xfs_buf_t               **O_bpp,
571         uint                    flags)
572 {
573         struct xfs_bmbt_irec    map;
574         int                     nmaps = 1, error;
575         struct xfs_buf          *bp;
576         struct xfs_inode        *quotip = xfs_dq_to_quota_inode(dqp);
577         struct xfs_mount        *mp = dqp->q_mount;
578         xfs_dqid_t              id = be32_to_cpu(dqp->q_core.d_id);
579         struct xfs_trans        *tp = (tpp ? *tpp : NULL);
580
581         dqp->q_fileoffset = (xfs_fileoff_t)id / mp->m_quotainfo->qi_dqperchunk;
582
583         xfs_ilock(quotip, XFS_ILOCK_SHARED);
584         if (!xfs_this_quota_on(dqp->q_mount, dqp->dq_flags)) {
585                 /*
586                  * Return if this type of quotas is turned off while we
587                  * didn't have the quota inode lock.
588                  */
589                 xfs_iunlock(quotip, XFS_ILOCK_SHARED);
590                 return ESRCH;
591         }
592
593         /*
594          * Find the block map; no allocations yet
595          */
596         error = xfs_bmapi_read(quotip, dqp->q_fileoffset,
597                                XFS_DQUOT_CLUSTER_SIZE_FSB, &map, &nmaps, 0);
598
599         xfs_iunlock(quotip, XFS_ILOCK_SHARED);
600         if (error)
601                 return error;
602
603         ASSERT(nmaps == 1);
604         ASSERT(map.br_blockcount == 1);
605
606         /*
607          * Offset of dquot in the (fixed sized) dquot chunk.
608          */
609         dqp->q_bufoffset = (id % mp->m_quotainfo->qi_dqperchunk) *
610                 sizeof(xfs_dqblk_t);
611
612         ASSERT(map.br_startblock != DELAYSTARTBLOCK);
613         if (map.br_startblock == HOLESTARTBLOCK) {
614                 /*
615                  * We don't allocate unless we're asked to
616                  */
617                 if (!(flags & XFS_QMOPT_DQALLOC))
618                         return ENOENT;
619
620                 ASSERT(tp);
621                 error = xfs_qm_dqalloc(tpp, mp, dqp, quotip,
622                                         dqp->q_fileoffset, &bp);
623                 if (error)
624                         return error;
625                 tp = *tpp;
626         } else {
627                 trace_xfs_dqtobp_read(dqp);
628
629                 /*
630                  * store the blkno etc so that we don't have to do the
631                  * mapping all the time
632                  */
633                 dqp->q_blkno = XFS_FSB_TO_DADDR(mp, map.br_startblock);
634
635                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
636                                            dqp->q_blkno,
637                                            mp->m_quotainfo->qi_dqchunklen,
638                                            0, &bp, &xfs_dquot_buf_ops);
639
640                 if (error == EFSCORRUPTED && (flags & XFS_QMOPT_DQREPAIR)) {
641                         xfs_dqid_t firstid = (xfs_dqid_t)map.br_startoff *
642                                                 mp->m_quotainfo->qi_dqperchunk;
643                         ASSERT(bp == NULL);
644                         error = xfs_qm_dqrepair(mp, tp, dqp, firstid, &bp);
645                 }
646
647                 if (error) {
648                         ASSERT(bp == NULL);
649                         return XFS_ERROR(error);
650                 }
651         }
652
653         ASSERT(xfs_buf_islocked(bp));
654         *O_bpp = bp;
655         *O_ddpp = bp->b_addr + dqp->q_bufoffset;
656
657         return (0);
658 }
659
660
661 /*
662  * Read in the ondisk dquot using dqtobp() then copy it to an incore version,
663  * and release the buffer immediately.
664  *
665  * If XFS_QMOPT_DQALLOC is set, allocate a dquot on disk if it needed.
666  */
667 int
668 xfs_qm_dqread(
669         struct xfs_mount        *mp,
670         xfs_dqid_t              id,
671         uint                    type,
672         uint                    flags,
673         struct xfs_dquot        **O_dqpp)
674 {
675         struct xfs_dquot        *dqp;
676         struct xfs_disk_dquot   *ddqp;
677         struct xfs_buf          *bp;
678         struct xfs_trans        *tp = NULL;
679         int                     error;
680         int                     cancelflags = 0;
681
682
683         dqp = kmem_zone_zalloc(xfs_qm_dqzone, KM_SLEEP);
684
685         dqp->dq_flags = type;
686         dqp->q_core.d_id = cpu_to_be32(id);
687         dqp->q_mount = mp;
688         INIT_LIST_HEAD(&dqp->q_lru);
689         mutex_init(&dqp->q_qlock);
690         init_waitqueue_head(&dqp->q_pinwait);
691
692         /*
693          * Because we want to use a counting completion, complete
694          * the flush completion once to allow a single access to
695          * the flush completion without blocking.
696          */
697         init_completion(&dqp->q_flush);
698         complete(&dqp->q_flush);
699
700         /*
701          * Make sure group quotas have a different lock class than user
702          * quotas.
703          */
704         if (!(type & XFS_DQ_USER))
705                 lockdep_set_class(&dqp->q_qlock, &xfs_dquot_other_class);
706
707         XFS_STATS_INC(xs_qm_dquot);
708
709         trace_xfs_dqread(dqp);
710
711         if (flags & XFS_QMOPT_DQALLOC) {
712                 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_DQALLOC);
713                 error = xfs_trans_reserve(tp, XFS_QM_DQALLOC_SPACE_RES(mp),
714                                           XFS_QM_DQALLOC_LOG_RES(mp), 0,
715                                           XFS_TRANS_PERM_LOG_RES,
716                                           XFS_WRITE_LOG_COUNT);
717                 if (error)
718                         goto error1;
719                 cancelflags = XFS_TRANS_RELEASE_LOG_RES;
720         }
721
722         /*
723          * get a pointer to the on-disk dquot and the buffer containing it
724          * dqp already knows its own type (GROUP/USER).
725          */
726         error = xfs_qm_dqtobp(&tp, dqp, &ddqp, &bp, flags);
727         if (error) {
728                 /*
729                  * This can happen if quotas got turned off (ESRCH),
730                  * or if the dquot didn't exist on disk and we ask to
731                  * allocate (ENOENT).
732                  */
733                 trace_xfs_dqread_fail(dqp);
734                 cancelflags |= XFS_TRANS_ABORT;
735                 goto error1;
736         }
737
738         /* copy everything from disk dquot to the incore dquot */
739         memcpy(&dqp->q_core, ddqp, sizeof(xfs_disk_dquot_t));
740         xfs_qm_dquot_logitem_init(dqp);
741
742         /*
743          * Reservation counters are defined as reservation plus current usage
744          * to avoid having to add every time.
745          */
746         dqp->q_res_bcount = be64_to_cpu(ddqp->d_bcount);
747         dqp->q_res_icount = be64_to_cpu(ddqp->d_icount);
748         dqp->q_res_rtbcount = be64_to_cpu(ddqp->d_rtbcount);
749
750         /* initialize the dquot speculative prealloc thresholds */
751         xfs_dquot_set_prealloc_limits(dqp);
752
753         /* Mark the buf so that this will stay incore a little longer */
754         xfs_buf_set_ref(bp, XFS_DQUOT_REF);
755
756         /*
757          * We got the buffer with a xfs_trans_read_buf() (in dqtobp())
758          * So we need to release with xfs_trans_brelse().
759          * The strategy here is identical to that of inodes; we lock
760          * the dquot in xfs_qm_dqget() before making it accessible to
761          * others. This is because dquots, like inodes, need a good level of
762          * concurrency, and we don't want to take locks on the entire buffers
763          * for dquot accesses.
764          * Note also that the dquot buffer may even be dirty at this point, if
765          * this particular dquot was repaired. We still aren't afraid to
766          * brelse it because we have the changes incore.
767          */
768         ASSERT(xfs_buf_islocked(bp));
769         xfs_trans_brelse(tp, bp);
770
771         if (tp) {
772                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
773                 if (error)
774                         goto error0;
775         }
776
777         *O_dqpp = dqp;
778         return error;
779
780 error1:
781         if (tp)
782                 xfs_trans_cancel(tp, cancelflags);
783 error0:
784         xfs_qm_dqdestroy(dqp);
785         *O_dqpp = NULL;
786         return error;
787 }
788
789 /*
790  * Given the file system, inode OR id, and type (UDQUOT/GDQUOT), return a
791  * a locked dquot, doing an allocation (if requested) as needed.
792  * When both an inode and an id are given, the inode's id takes precedence.
793  * That is, if the id changes while we don't hold the ilock inside this
794  * function, the new dquot is returned, not necessarily the one requested
795  * in the id argument.
796  */
797 int
798 xfs_qm_dqget(
799         xfs_mount_t     *mp,
800         xfs_inode_t     *ip,      /* locked inode (optional) */
801         xfs_dqid_t      id,       /* uid/projid/gid depending on type */
802         uint            type,     /* XFS_DQ_USER/XFS_DQ_PROJ/XFS_DQ_GROUP */
803         uint            flags,    /* DQALLOC, DQSUSER, DQREPAIR, DOWARN */
804         xfs_dquot_t     **O_dqpp) /* OUT : locked incore dquot */
805 {
806         struct xfs_quotainfo    *qi = mp->m_quotainfo;
807         struct radix_tree_root *tree = xfs_dquot_tree(qi, type);
808         struct xfs_dquot        *dqp;
809         int                     error;
810
811         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
812         if ((! XFS_IS_UQUOTA_ON(mp) && type == XFS_DQ_USER) ||
813             (! XFS_IS_PQUOTA_ON(mp) && type == XFS_DQ_PROJ) ||
814             (! XFS_IS_GQUOTA_ON(mp) && type == XFS_DQ_GROUP)) {
815                 return (ESRCH);
816         }
817
818 #ifdef DEBUG
819         if (xfs_do_dqerror) {
820                 if ((xfs_dqerror_target == mp->m_ddev_targp) &&
821                     (xfs_dqreq_num++ % xfs_dqerror_mod) == 0) {
822                         xfs_debug(mp, "Returning error in dqget");
823                         return (EIO);
824                 }
825         }
826
827         ASSERT(type == XFS_DQ_USER ||
828                type == XFS_DQ_PROJ ||
829                type == XFS_DQ_GROUP);
830         if (ip) {
831                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
832                 ASSERT(xfs_inode_dquot(ip, type) == NULL);
833         }
834 #endif
835
836 restart:
837         mutex_lock(&qi->qi_tree_lock);
838         dqp = radix_tree_lookup(tree, id);
839         if (dqp) {
840                 xfs_dqlock(dqp);
841                 if (dqp->dq_flags & XFS_DQ_FREEING) {
842                         xfs_dqunlock(dqp);
843                         mutex_unlock(&qi->qi_tree_lock);
844                         trace_xfs_dqget_freeing(dqp);
845                         delay(1);
846                         goto restart;
847                 }
848
849                 dqp->q_nrefs++;
850                 mutex_unlock(&qi->qi_tree_lock);
851
852                 trace_xfs_dqget_hit(dqp);
853                 XFS_STATS_INC(xs_qm_dqcachehits);
854                 *O_dqpp = dqp;
855                 return 0;
856         }
857         mutex_unlock(&qi->qi_tree_lock);
858         XFS_STATS_INC(xs_qm_dqcachemisses);
859
860         /*
861          * Dquot cache miss. We don't want to keep the inode lock across
862          * a (potential) disk read. Also we don't want to deal with the lock
863          * ordering between quotainode and this inode. OTOH, dropping the inode
864          * lock here means dealing with a chown that can happen before
865          * we re-acquire the lock.
866          */
867         if (ip)
868                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
869
870         error = xfs_qm_dqread(mp, id, type, flags, &dqp);
871
872         if (ip)
873                 xfs_ilock(ip, XFS_ILOCK_EXCL);
874
875         if (error)
876                 return error;
877
878         if (ip) {
879                 /*
880                  * A dquot could be attached to this inode by now, since
881                  * we had dropped the ilock.
882                  */
883                 if (xfs_this_quota_on(mp, type)) {
884                         struct xfs_dquot        *dqp1;
885
886                         dqp1 = xfs_inode_dquot(ip, type);
887                         if (dqp1) {
888                                 xfs_qm_dqdestroy(dqp);
889                                 dqp = dqp1;
890                                 xfs_dqlock(dqp);
891                                 goto dqret;
892                         }
893                 } else {
894                         /* inode stays locked on return */
895                         xfs_qm_dqdestroy(dqp);
896                         return XFS_ERROR(ESRCH);
897                 }
898         }
899
900         mutex_lock(&qi->qi_tree_lock);
901         error = -radix_tree_insert(tree, id, dqp);
902         if (unlikely(error)) {
903                 WARN_ON(error != EEXIST);
904
905                 /*
906                  * Duplicate found. Just throw away the new dquot and start
907                  * over.
908                  */
909                 mutex_unlock(&qi->qi_tree_lock);
910                 trace_xfs_dqget_dup(dqp);
911                 xfs_qm_dqdestroy(dqp);
912                 XFS_STATS_INC(xs_qm_dquot_dups);
913                 goto restart;
914         }
915
916         /*
917          * We return a locked dquot to the caller, with a reference taken
918          */
919         xfs_dqlock(dqp);
920         dqp->q_nrefs = 1;
921
922         qi->qi_dquots++;
923         mutex_unlock(&qi->qi_tree_lock);
924
925  dqret:
926         ASSERT((ip == NULL) || xfs_isilocked(ip, XFS_ILOCK_EXCL));
927         trace_xfs_dqget_miss(dqp);
928         *O_dqpp = dqp;
929         return (0);
930 }
931
932
933 STATIC void
934 xfs_qm_dqput_final(
935         struct xfs_dquot        *dqp)
936 {
937         struct xfs_quotainfo    *qi = dqp->q_mount->m_quotainfo;
938         struct xfs_dquot        *gdqp;
939
940         trace_xfs_dqput_free(dqp);
941
942         mutex_lock(&qi->qi_lru_lock);
943         if (list_empty(&dqp->q_lru)) {
944                 list_add_tail(&dqp->q_lru, &qi->qi_lru_list);
945                 qi->qi_lru_count++;
946                 XFS_STATS_INC(xs_qm_dquot_unused);
947         }
948         mutex_unlock(&qi->qi_lru_lock);
949
950         /*
951          * If we just added a udquot to the freelist, then we want to release
952          * the gdquot reference that it (probably) has. Otherwise it'll keep
953          * the gdquot from getting reclaimed.
954          */
955         gdqp = dqp->q_gdquot;
956         if (gdqp) {
957                 xfs_dqlock(gdqp);
958                 dqp->q_gdquot = NULL;
959         }
960         xfs_dqunlock(dqp);
961
962         /*
963          * If we had a group quota hint, release it now.
964          */
965         if (gdqp)
966                 xfs_qm_dqput(gdqp);
967 }
968
969 /*
970  * Release a reference to the dquot (decrement ref-count) and unlock it.
971  *
972  * If there is a group quota attached to this dquot, carefully release that
973  * too without tripping over deadlocks'n'stuff.
974  */
975 void
976 xfs_qm_dqput(
977         struct xfs_dquot        *dqp)
978 {
979         ASSERT(dqp->q_nrefs > 0);
980         ASSERT(XFS_DQ_IS_LOCKED(dqp));
981
982         trace_xfs_dqput(dqp);
983
984         if (--dqp->q_nrefs > 0)
985                 xfs_dqunlock(dqp);
986         else
987                 xfs_qm_dqput_final(dqp);
988 }
989
990 /*
991  * Release a dquot. Flush it if dirty, then dqput() it.
992  * dquot must not be locked.
993  */
994 void
995 xfs_qm_dqrele(
996         xfs_dquot_t     *dqp)
997 {
998         if (!dqp)
999                 return;
1000
1001         trace_xfs_dqrele(dqp);
1002
1003         xfs_dqlock(dqp);
1004         /*
1005          * We don't care to flush it if the dquot is dirty here.
1006          * That will create stutters that we want to avoid.
1007          * Instead we do a delayed write when we try to reclaim
1008          * a dirty dquot. Also xfs_sync will take part of the burden...
1009          */
1010         xfs_qm_dqput(dqp);
1011 }
1012
1013 /*
1014  * This is the dquot flushing I/O completion routine.  It is called
1015  * from interrupt level when the buffer containing the dquot is
1016  * flushed to disk.  It is responsible for removing the dquot logitem
1017  * from the AIL if it has not been re-logged, and unlocking the dquot's
1018  * flush lock. This behavior is very similar to that of inodes..
1019  */
1020 STATIC void
1021 xfs_qm_dqflush_done(
1022         struct xfs_buf          *bp,
1023         struct xfs_log_item     *lip)
1024 {
1025         xfs_dq_logitem_t        *qip = (struct xfs_dq_logitem *)lip;
1026         xfs_dquot_t             *dqp = qip->qli_dquot;
1027         struct xfs_ail          *ailp = lip->li_ailp;
1028
1029         /*
1030          * We only want to pull the item from the AIL if its
1031          * location in the log has not changed since we started the flush.
1032          * Thus, we only bother if the dquot's lsn has
1033          * not changed. First we check the lsn outside the lock
1034          * since it's cheaper, and then we recheck while
1035          * holding the lock before removing the dquot from the AIL.
1036          */
1037         if ((lip->li_flags & XFS_LI_IN_AIL) &&
1038             lip->li_lsn == qip->qli_flush_lsn) {
1039
1040                 /* xfs_trans_ail_delete() drops the AIL lock. */
1041                 spin_lock(&ailp->xa_lock);
1042                 if (lip->li_lsn == qip->qli_flush_lsn)
1043                         xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1044                 else
1045                         spin_unlock(&ailp->xa_lock);
1046         }
1047
1048         /*
1049          * Release the dq's flush lock since we're done with it.
1050          */
1051         xfs_dqfunlock(dqp);
1052 }
1053
1054 /*
1055  * Write a modified dquot to disk.
1056  * The dquot must be locked and the flush lock too taken by caller.
1057  * The flush lock will not be unlocked until the dquot reaches the disk,
1058  * but the dquot is free to be unlocked and modified by the caller
1059  * in the interim. Dquot is still locked on return. This behavior is
1060  * identical to that of inodes.
1061  */
1062 int
1063 xfs_qm_dqflush(
1064         struct xfs_dquot        *dqp,
1065         struct xfs_buf          **bpp)
1066 {
1067         struct xfs_mount        *mp = dqp->q_mount;
1068         struct xfs_buf          *bp;
1069         struct xfs_disk_dquot   *ddqp;
1070         int                     error;
1071
1072         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1073         ASSERT(!completion_done(&dqp->q_flush));
1074
1075         trace_xfs_dqflush(dqp);
1076
1077         *bpp = NULL;
1078
1079         xfs_qm_dqunpin_wait(dqp);
1080
1081         /*
1082          * This may have been unpinned because the filesystem is shutting
1083          * down forcibly. If that's the case we must not write this dquot
1084          * to disk, because the log record didn't make it to disk.
1085          *
1086          * We also have to remove the log item from the AIL in this case,
1087          * as we wait for an emptry AIL as part of the unmount process.
1088          */
1089         if (XFS_FORCED_SHUTDOWN(mp)) {
1090                 struct xfs_log_item     *lip = &dqp->q_logitem.qli_item;
1091                 dqp->dq_flags &= ~XFS_DQ_DIRTY;
1092
1093                 spin_lock(&mp->m_ail->xa_lock);
1094                 if (lip->li_flags & XFS_LI_IN_AIL)
1095                         xfs_trans_ail_delete(mp->m_ail, lip,
1096                                              SHUTDOWN_CORRUPT_INCORE);
1097                 else
1098                         spin_unlock(&mp->m_ail->xa_lock);
1099                 error = XFS_ERROR(EIO);
1100                 goto out_unlock;
1101         }
1102
1103         /*
1104          * Get the buffer containing the on-disk dquot
1105          */
1106         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dqp->q_blkno,
1107                                    mp->m_quotainfo->qi_dqchunklen, 0, &bp, NULL);
1108         if (error)
1109                 goto out_unlock;
1110
1111         /*
1112          * Calculate the location of the dquot inside the buffer.
1113          */
1114         ddqp = bp->b_addr + dqp->q_bufoffset;
1115
1116         /*
1117          * A simple sanity check in case we got a corrupted dquot..
1118          */
1119         error = xfs_qm_dqcheck(mp, &dqp->q_core, be32_to_cpu(ddqp->d_id), 0,
1120                            XFS_QMOPT_DOWARN, "dqflush (incore copy)");
1121         if (error) {
1122                 xfs_buf_relse(bp);
1123                 xfs_dqfunlock(dqp);
1124                 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1125                 return XFS_ERROR(EIO);
1126         }
1127
1128         /* This is the only portion of data that needs to persist */
1129         memcpy(ddqp, &dqp->q_core, sizeof(xfs_disk_dquot_t));
1130
1131         /*
1132          * Clear the dirty field and remember the flush lsn for later use.
1133          */
1134         dqp->dq_flags &= ~XFS_DQ_DIRTY;
1135
1136         xfs_trans_ail_copy_lsn(mp->m_ail, &dqp->q_logitem.qli_flush_lsn,
1137                                         &dqp->q_logitem.qli_item.li_lsn);
1138
1139         /*
1140          * copy the lsn into the on-disk dquot now while we have the in memory
1141          * dquot here. This can't be done later in the write verifier as we
1142          * can't get access to the log item at that point in time.
1143          *
1144          * We also calculate the CRC here so that the on-disk dquot in the
1145          * buffer always has a valid CRC. This ensures there is no possibility
1146          * of a dquot without an up-to-date CRC getting to disk.
1147          */
1148         if (xfs_sb_version_hascrc(&mp->m_sb)) {
1149                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddqp;
1150
1151                 dqb->dd_lsn = cpu_to_be64(dqp->q_logitem.qli_item.li_lsn);
1152                 xfs_update_cksum((char *)dqb, sizeof(struct xfs_dqblk),
1153                                  XFS_DQUOT_CRC_OFF);
1154         }
1155
1156         /*
1157          * Attach an iodone routine so that we can remove this dquot from the
1158          * AIL and release the flush lock once the dquot is synced to disk.
1159          */
1160         xfs_buf_attach_iodone(bp, xfs_qm_dqflush_done,
1161                                   &dqp->q_logitem.qli_item);
1162
1163         /*
1164          * If the buffer is pinned then push on the log so we won't
1165          * get stuck waiting in the write for too long.
1166          */
1167         if (xfs_buf_ispinned(bp)) {
1168                 trace_xfs_dqflush_force(dqp);
1169                 xfs_log_force(mp, 0);
1170         }
1171
1172         trace_xfs_dqflush_done(dqp);
1173         *bpp = bp;
1174         return 0;
1175
1176 out_unlock:
1177         xfs_dqfunlock(dqp);
1178         return XFS_ERROR(EIO);
1179 }
1180
1181 /*
1182  * Lock two xfs_dquot structures.
1183  *
1184  * To avoid deadlocks we always lock the quota structure with
1185  * the lowerd id first.
1186  */
1187 void
1188 xfs_dqlock2(
1189         xfs_dquot_t     *d1,
1190         xfs_dquot_t     *d2)
1191 {
1192         if (d1 && d2) {
1193                 ASSERT(d1 != d2);
1194                 if (be32_to_cpu(d1->q_core.d_id) >
1195                     be32_to_cpu(d2->q_core.d_id)) {
1196                         mutex_lock(&d2->q_qlock);
1197                         mutex_lock_nested(&d1->q_qlock, XFS_QLOCK_NESTED);
1198                 } else {
1199                         mutex_lock(&d1->q_qlock);
1200                         mutex_lock_nested(&d2->q_qlock, XFS_QLOCK_NESTED);
1201                 }
1202         } else if (d1) {
1203                 mutex_lock(&d1->q_qlock);
1204         } else if (d2) {
1205                 mutex_lock(&d2->q_qlock);
1206         }
1207 }
1208
1209 int __init
1210 xfs_qm_init(void)
1211 {
1212         xfs_qm_dqzone =
1213                 kmem_zone_init(sizeof(struct xfs_dquot), "xfs_dquot");
1214         if (!xfs_qm_dqzone)
1215                 goto out;
1216
1217         xfs_qm_dqtrxzone =
1218                 kmem_zone_init(sizeof(struct xfs_dquot_acct), "xfs_dqtrx");
1219         if (!xfs_qm_dqtrxzone)
1220                 goto out_free_dqzone;
1221
1222         return 0;
1223
1224 out_free_dqzone:
1225         kmem_zone_destroy(xfs_qm_dqzone);
1226 out:
1227         return -ENOMEM;
1228 }
1229
1230 void
1231 xfs_qm_exit(void)
1232 {
1233         kmem_zone_destroy(xfs_qm_dqtrxzone);
1234         kmem_zone_destroy(xfs_qm_dqzone);
1235 }