[XFS] Always use struct xfs_btree_block instead of short / longform
[linux-2.6] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 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_clnt.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_alloc.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_quota.h"
31 #include "xfs_mount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_itable.h"
42 #include "xfs_rtalloc.h"
43 #include "xfs_error.h"
44 #include "xfs_bmap.h"
45 #include "xfs_rw.h"
46 #include "xfs_acl.h"
47 #include "xfs_attr.h"
48 #include "xfs_buf_item.h"
49 #include "xfs_trans_space.h"
50 #include "xfs_utils.h"
51 #include "xfs_qm.h"
52
53 /*
54  * The global quota manager. There is only one of these for the entire
55  * system, _not_ one per file system. XQM keeps track of the overall
56  * quota functionality, including maintaining the freelist and hash
57  * tables of dquots.
58  */
59 mutex_t         xfs_Gqm_lock;
60 struct xfs_qm   *xfs_Gqm;
61 uint            ndquot;
62
63 kmem_zone_t     *qm_dqzone;
64 kmem_zone_t     *qm_dqtrxzone;
65
66 static cred_t   xfs_zerocr;
67
68 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
69 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
70
71 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
72 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
73 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
74 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
75
76 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
77 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
78 STATIC int      xfs_qm_shake(int, gfp_t);
79
80 static struct shrinker xfs_qm_shaker = {
81         .shrink = xfs_qm_shake,
82         .seeks = DEFAULT_SEEKS,
83 };
84
85 #ifdef DEBUG
86 extern mutex_t  qcheck_lock;
87 #endif
88
89 #ifdef QUOTADEBUG
90 #define XQM_LIST_PRINT(l, NXT, title) \
91 { \
92         xfs_dquot_t     *dqp; int i = 0; \
93         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
94         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
95                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
96                                   "bcnt = %d, icnt = %d, refs = %d", \
97                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
98                         DQFLAGTO_TYPESTR(dqp),       \
99                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
100                         (int) be64_to_cpu(dqp->q_core.d_icount), \
101                         (int) dqp->q_nrefs);  } \
102 }
103 #else
104 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
105 #endif
106
107 /*
108  * Initialize the XQM structure.
109  * Note that there is not one quota manager per file system.
110  */
111 STATIC struct xfs_qm *
112 xfs_Gqm_init(void)
113 {
114         xfs_dqhash_t    *udqhash, *gdqhash;
115         xfs_qm_t        *xqm;
116         size_t          hsize;
117         uint            i;
118
119         /*
120          * Initialize the dquot hash tables.
121          */
122         udqhash = kmem_zalloc_greedy(&hsize,
123                                      XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t),
124                                      XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t),
125                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
126         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
127         hsize /= sizeof(xfs_dqhash_t);
128         ndquot = hsize << 8;
129
130         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
131         xqm->qm_dqhashmask = hsize - 1;
132         xqm->qm_usr_dqhtable = udqhash;
133         xqm->qm_grp_dqhtable = gdqhash;
134         ASSERT(xqm->qm_usr_dqhtable != NULL);
135         ASSERT(xqm->qm_grp_dqhtable != NULL);
136
137         for (i = 0; i < hsize; i++) {
138                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
139                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
140         }
141
142         /*
143          * Freelist of all dquots of all file systems
144          */
145         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
146
147         /*
148          * dquot zone. we register our own low-memory callback.
149          */
150         if (!qm_dqzone) {
151                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
152                                                 "xfs_dquots");
153                 qm_dqzone = xqm->qm_dqzone;
154         } else
155                 xqm->qm_dqzone = qm_dqzone;
156
157         register_shrinker(&xfs_qm_shaker);
158
159         /*
160          * The t_dqinfo portion of transactions.
161          */
162         if (!qm_dqtrxzone) {
163                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
164                                                    "xfs_dqtrx");
165                 qm_dqtrxzone = xqm->qm_dqtrxzone;
166         } else
167                 xqm->qm_dqtrxzone = qm_dqtrxzone;
168
169         atomic_set(&xqm->qm_totaldquots, 0);
170         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
171         xqm->qm_nrefs = 0;
172 #ifdef DEBUG
173         mutex_init(&qcheck_lock);
174 #endif
175         return xqm;
176 }
177
178 /*
179  * Destroy the global quota manager when its reference count goes to zero.
180  */
181 STATIC void
182 xfs_qm_destroy(
183         struct xfs_qm   *xqm)
184 {
185         int             hsize, i;
186
187         ASSERT(xqm != NULL);
188         ASSERT(xqm->qm_nrefs == 0);
189         unregister_shrinker(&xfs_qm_shaker);
190         hsize = xqm->qm_dqhashmask + 1;
191         for (i = 0; i < hsize; i++) {
192                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
193                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
194         }
195         kmem_free(xqm->qm_usr_dqhtable);
196         kmem_free(xqm->qm_grp_dqhtable);
197         xqm->qm_usr_dqhtable = NULL;
198         xqm->qm_grp_dqhtable = NULL;
199         xqm->qm_dqhashmask = 0;
200         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
201 #ifdef DEBUG
202         mutex_destroy(&qcheck_lock);
203 #endif
204         kmem_free(xqm);
205 }
206
207 /*
208  * Called at mount time to let XQM know that another file system is
209  * starting quotas. This isn't crucial information as the individual mount
210  * structures are pretty independent, but it helps the XQM keep a
211  * global view of what's going on.
212  */
213 /* ARGSUSED */
214 STATIC int
215 xfs_qm_hold_quotafs_ref(
216         struct xfs_mount *mp)
217 {
218         /*
219          * Need to lock the xfs_Gqm structure for things like this. For example,
220          * the structure could disappear between the entry to this routine and
221          * a HOLD operation if not locked.
222          */
223         XFS_QM_LOCK(xfs_Gqm);
224
225         if (xfs_Gqm == NULL)
226                 xfs_Gqm = xfs_Gqm_init();
227         /*
228          * We can keep a list of all filesystems with quotas mounted for
229          * debugging and statistical purposes, but ...
230          * Just take a reference and get out.
231          */
232         XFS_QM_HOLD(xfs_Gqm);
233         XFS_QM_UNLOCK(xfs_Gqm);
234
235         return 0;
236 }
237
238
239 /*
240  * Release the reference that a filesystem took at mount time,
241  * so that we know when we need to destroy the entire quota manager.
242  */
243 /* ARGSUSED */
244 STATIC void
245 xfs_qm_rele_quotafs_ref(
246         struct xfs_mount *mp)
247 {
248         xfs_dquot_t     *dqp, *nextdqp;
249
250         ASSERT(xfs_Gqm);
251         ASSERT(xfs_Gqm->qm_nrefs > 0);
252
253         /*
254          * Go thru the freelist and destroy all inactive dquots.
255          */
256         xfs_qm_freelist_lock(xfs_Gqm);
257
258         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
259              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
260                 xfs_dqlock(dqp);
261                 nextdqp = dqp->dq_flnext;
262                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
263                         ASSERT(dqp->q_mount == NULL);
264                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
265                         ASSERT(dqp->HL_PREVP == NULL);
266                         ASSERT(dqp->MPL_PREVP == NULL);
267                         XQM_FREELIST_REMOVE(dqp);
268                         xfs_dqunlock(dqp);
269                         xfs_qm_dqdestroy(dqp);
270                 } else {
271                         xfs_dqunlock(dqp);
272                 }
273                 dqp = nextdqp;
274         }
275         xfs_qm_freelist_unlock(xfs_Gqm);
276
277         /*
278          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
279          * be restarted.
280          */
281         XFS_QM_LOCK(xfs_Gqm);
282         XFS_QM_RELE(xfs_Gqm);
283         if (xfs_Gqm->qm_nrefs == 0) {
284                 xfs_qm_destroy(xfs_Gqm);
285                 xfs_Gqm = NULL;
286         }
287         XFS_QM_UNLOCK(xfs_Gqm);
288 }
289
290 /*
291  * Just destroy the quotainfo structure.
292  */
293 void
294 xfs_qm_unmount_quotadestroy(
295         xfs_mount_t     *mp)
296 {
297         if (mp->m_quotainfo)
298                 xfs_qm_destroy_quotainfo(mp);
299 }
300
301
302 /*
303  * This is called from xfs_mountfs to start quotas and initialize all
304  * necessary data structures like quotainfo.  This is also responsible for
305  * running a quotacheck as necessary.  We are guaranteed that the superblock
306  * is consistently read in at this point.
307  *
308  * If we fail here, the mount will continue with quota turned off. We don't
309  * need to inidicate success or failure at all.
310  */
311 void
312 xfs_qm_mount_quotas(
313         xfs_mount_t     *mp)
314 {
315         int             error = 0;
316         uint            sbf;
317
318         /*
319          * If quotas on realtime volumes is not supported, we disable
320          * quotas immediately.
321          */
322         if (mp->m_sb.sb_rextents) {
323                 cmn_err(CE_NOTE,
324                         "Cannot turn on quotas for realtime filesystem %s",
325                         mp->m_fsname);
326                 mp->m_qflags = 0;
327                 goto write_changes;
328         }
329
330         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
331
332         /*
333          * Allocate the quotainfo structure inside the mount struct, and
334          * create quotainode(s), and change/rev superblock if necessary.
335          */
336         error = xfs_qm_init_quotainfo(mp);
337         if (error) {
338                 /*
339                  * We must turn off quotas.
340                  */
341                 ASSERT(mp->m_quotainfo == NULL);
342                 mp->m_qflags = 0;
343                 goto write_changes;
344         }
345         /*
346          * If any of the quotas are not consistent, do a quotacheck.
347          */
348         if (XFS_QM_NEED_QUOTACHECK(mp)) {
349                 error = xfs_qm_quotacheck(mp);
350                 if (error) {
351                         /* Quotacheck failed and disabled quotas. */
352                         return;
353                 }
354         }
355         /* 
356          * If one type of quotas is off, then it will lose its
357          * quotachecked status, since we won't be doing accounting for
358          * that type anymore.
359          */
360         if (!XFS_IS_UQUOTA_ON(mp))
361                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
362         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp)))
363                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
364
365  write_changes:
366         /*
367          * We actually don't have to acquire the m_sb_lock at all.
368          * This can only be called from mount, and that's single threaded. XXX
369          */
370         spin_lock(&mp->m_sb_lock);
371         sbf = mp->m_sb.sb_qflags;
372         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
373         spin_unlock(&mp->m_sb_lock);
374
375         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
376                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
377                         /*
378                          * We could only have been turning quotas off.
379                          * We aren't in very good shape actually because
380                          * the incore structures are convinced that quotas are
381                          * off, but the on disk superblock doesn't know that !
382                          */
383                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
384                         xfs_fs_cmn_err(CE_ALERT, mp,
385                                 "XFS mount_quotas: Superblock update failed!");
386                 }
387         }
388
389         if (error) {
390                 xfs_fs_cmn_err(CE_WARN, mp,
391                         "Failed to initialize disk quotas.");
392         }
393         return;
394 }
395
396 /*
397  * Called from the vfsops layer.
398  */
399 int
400 xfs_qm_unmount_quotas(
401         xfs_mount_t     *mp)
402 {
403         xfs_inode_t     *uqp, *gqp;
404         int             error = 0;
405
406         /*
407          * Release the dquots that root inode, et al might be holding,
408          * before we flush quotas and blow away the quotainfo structure.
409          */
410         ASSERT(mp->m_rootip);
411         xfs_qm_dqdetach(mp->m_rootip);
412         if (mp->m_rbmip)
413                 xfs_qm_dqdetach(mp->m_rbmip);
414         if (mp->m_rsumip)
415                 xfs_qm_dqdetach(mp->m_rsumip);
416
417         /*
418          * Flush out the quota inodes.
419          */
420         uqp = gqp = NULL;
421         if (mp->m_quotainfo) {
422                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
423                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
424                         xfs_iflock(uqp);
425                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
426                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
427                         if (unlikely(error == EFSCORRUPTED)) {
428                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
429                                                  XFS_ERRLEVEL_LOW, mp);
430                                 goto out;
431                         }
432                 }
433                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
434                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
435                         xfs_iflock(gqp);
436                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
437                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
438                         if (unlikely(error == EFSCORRUPTED)) {
439                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
440                                                  XFS_ERRLEVEL_LOW, mp);
441                                 goto out;
442                         }
443                 }
444         }
445         if (uqp) {
446                  IRELE(uqp);
447                  mp->m_quotainfo->qi_uquotaip = NULL;
448         }
449         if (gqp) {
450                 IRELE(gqp);
451                 mp->m_quotainfo->qi_gquotaip = NULL;
452         }
453 out:
454         return XFS_ERROR(error);
455 }
456
457 /*
458  * Flush all dquots of the given file system to disk. The dquots are
459  * _not_ purged from memory here, just their data written to disk.
460  */
461 STATIC int
462 xfs_qm_dqflush_all(
463         xfs_mount_t     *mp,
464         int             flags)
465 {
466         int             recl;
467         xfs_dquot_t     *dqp;
468         int             niters;
469         int             error;
470
471         if (mp->m_quotainfo == NULL)
472                 return 0;
473         niters = 0;
474 again:
475         xfs_qm_mplist_lock(mp);
476         FOREACH_DQUOT_IN_MP(dqp, mp) {
477                 xfs_dqlock(dqp);
478                 if (! XFS_DQ_IS_DIRTY(dqp)) {
479                         xfs_dqunlock(dqp);
480                         continue;
481                 }
482                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
483                 /* XXX a sentinel would be better */
484                 recl = XFS_QI_MPLRECLAIMS(mp);
485                 if (!xfs_dqflock_nowait(dqp)) {
486                         /*
487                          * If we can't grab the flush lock then check
488                          * to see if the dquot has been flushed delayed
489                          * write.  If so, grab its buffer and send it
490                          * out immediately.  We'll be able to acquire
491                          * the flush lock when the I/O completes.
492                          */
493                         xfs_qm_dqflock_pushbuf_wait(dqp);
494                 }
495                 /*
496                  * Let go of the mplist lock. We don't want to hold it
497                  * across a disk write.
498                  */
499                 xfs_qm_mplist_unlock(mp);
500                 error = xfs_qm_dqflush(dqp, flags);
501                 xfs_dqunlock(dqp);
502                 if (error)
503                         return error;
504
505                 xfs_qm_mplist_lock(mp);
506                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
507                         xfs_qm_mplist_unlock(mp);
508                         /* XXX restart limit */
509                         goto again;
510                 }
511         }
512
513         xfs_qm_mplist_unlock(mp);
514         /* return ! busy */
515         return 0;
516 }
517 /*
518  * Release the group dquot pointers the user dquots may be
519  * carrying around as a hint. mplist is locked on entry and exit.
520  */
521 STATIC void
522 xfs_qm_detach_gdquots(
523         xfs_mount_t     *mp)
524 {
525         xfs_dquot_t     *dqp, *gdqp;
526         int             nrecl;
527
528  again:
529         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
530         dqp = XFS_QI_MPLNEXT(mp);
531         while (dqp) {
532                 xfs_dqlock(dqp);
533                 if ((gdqp = dqp->q_gdquot)) {
534                         xfs_dqlock(gdqp);
535                         dqp->q_gdquot = NULL;
536                 }
537                 xfs_dqunlock(dqp);
538
539                 if (gdqp) {
540                         /*
541                          * Can't hold the mplist lock across a dqput.
542                          * XXXmust convert to marker based iterations here.
543                          */
544                         nrecl = XFS_QI_MPLRECLAIMS(mp);
545                         xfs_qm_mplist_unlock(mp);
546                         xfs_qm_dqput(gdqp);
547
548                         xfs_qm_mplist_lock(mp);
549                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
550                                 goto again;
551                 }
552                 dqp = dqp->MPL_NEXT;
553         }
554 }
555
556 /*
557  * Go through all the incore dquots of this file system and take them
558  * off the mplist and hashlist, if the dquot type matches the dqtype
559  * parameter. This is used when turning off quota accounting for
560  * users and/or groups, as well as when the filesystem is unmounting.
561  */
562 STATIC int
563 xfs_qm_dqpurge_int(
564         xfs_mount_t     *mp,
565         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
566 {
567         xfs_dquot_t     *dqp;
568         uint            dqtype;
569         int             nrecl;
570         xfs_dquot_t     *nextdqp;
571         int             nmisses;
572
573         if (mp->m_quotainfo == NULL)
574                 return 0;
575
576         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
577         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
578         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
579
580         xfs_qm_mplist_lock(mp);
581
582         /*
583          * In the first pass through all incore dquots of this filesystem,
584          * we release the group dquot pointers the user dquots may be
585          * carrying around as a hint. We need to do this irrespective of
586          * what's being turned off.
587          */
588         xfs_qm_detach_gdquots(mp);
589
590       again:
591         nmisses = 0;
592         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
593         /*
594          * Try to get rid of all of the unwanted dquots. The idea is to
595          * get them off mplist and hashlist, but leave them on freelist.
596          */
597         dqp = XFS_QI_MPLNEXT(mp);
598         while (dqp) {
599                 /*
600                  * It's OK to look at the type without taking dqlock here.
601                  * We're holding the mplist lock here, and that's needed for
602                  * a dqreclaim.
603                  */
604                 if ((dqp->dq_flags & dqtype) == 0) {
605                         dqp = dqp->MPL_NEXT;
606                         continue;
607                 }
608
609                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
610                         nrecl = XFS_QI_MPLRECLAIMS(mp);
611                         xfs_qm_mplist_unlock(mp);
612                         XFS_DQ_HASH_LOCK(dqp->q_hash);
613                         xfs_qm_mplist_lock(mp);
614
615                         /*
616                          * XXXTheoretically, we can get into a very long
617                          * ping pong game here.
618                          * No one can be adding dquots to the mplist at
619                          * this point, but somebody might be taking things off.
620                          */
621                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
622                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
623                                 goto again;
624                         }
625                 }
626
627                 /*
628                  * Take the dquot off the mplist and hashlist. It may remain on
629                  * freelist in INACTIVE state.
630                  */
631                 nextdqp = dqp->MPL_NEXT;
632                 nmisses += xfs_qm_dqpurge(dqp);
633                 dqp = nextdqp;
634         }
635         xfs_qm_mplist_unlock(mp);
636         return nmisses;
637 }
638
639 int
640 xfs_qm_dqpurge_all(
641         xfs_mount_t     *mp,
642         uint            flags)
643 {
644         int             ndquots;
645
646         /*
647          * Purge the dquot cache.
648          * None of the dquots should really be busy at this point.
649          */
650         if (mp->m_quotainfo) {
651                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
652                         delay(ndquots * 10);
653                 }
654         }
655         return 0;
656 }
657
658 STATIC int
659 xfs_qm_dqattach_one(
660         xfs_inode_t     *ip,
661         xfs_dqid_t      id,
662         uint            type,
663         uint            doalloc,
664         uint            dolock,
665         xfs_dquot_t     *udqhint, /* hint */
666         xfs_dquot_t     **IO_idqpp)
667 {
668         xfs_dquot_t     *dqp;
669         int             error;
670
671         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
672         error = 0;
673         /*
674          * See if we already have it in the inode itself. IO_idqpp is
675          * &i_udquot or &i_gdquot. This made the code look weird, but
676          * made the logic a lot simpler.
677          */
678         if ((dqp = *IO_idqpp)) {
679                 if (dolock)
680                         xfs_dqlock(dqp);
681                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
682                 goto done;
683         }
684
685         /*
686          * udqhint is the i_udquot field in inode, and is non-NULL only
687          * when the type arg is group/project. Its purpose is to save a
688          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
689          * the user dquot.
690          */
691         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
692         if (udqhint && !dolock)
693                 xfs_dqlock(udqhint);
694
695         /*
696          * No need to take dqlock to look at the id.
697          * The ID can't change until it gets reclaimed, and it won't
698          * be reclaimed as long as we have a ref from inode and we hold
699          * the ilock.
700          */
701         if (udqhint &&
702             (dqp = udqhint->q_gdquot) &&
703             (be32_to_cpu(dqp->q_core.d_id) == id)) {
704                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
705                 xfs_dqlock(dqp);
706                 XFS_DQHOLD(dqp);
707                 ASSERT(*IO_idqpp == NULL);
708                 *IO_idqpp = dqp;
709                 if (!dolock) {
710                         xfs_dqunlock(dqp);
711                         xfs_dqunlock(udqhint);
712                 }
713                 goto done;
714         }
715         /*
716          * We can't hold a dquot lock when we call the dqget code.
717          * We'll deadlock in no time, because of (not conforming to)
718          * lock ordering - the inodelock comes before any dquot lock,
719          * and we may drop and reacquire the ilock in xfs_qm_dqget().
720          */
721         if (udqhint)
722                 xfs_dqunlock(udqhint);
723         /*
724          * Find the dquot from somewhere. This bumps the
725          * reference count of dquot and returns it locked.
726          * This can return ENOENT if dquot didn't exist on
727          * disk and we didn't ask it to allocate;
728          * ESRCH if quotas got turned off suddenly.
729          */
730         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
731                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
732                 if (udqhint && dolock)
733                         xfs_dqlock(udqhint);
734                 goto done;
735         }
736
737         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
738         /*
739          * dqget may have dropped and re-acquired the ilock, but it guarantees
740          * that the dquot returned is the one that should go in the inode.
741          */
742         *IO_idqpp = dqp;
743         ASSERT(dqp);
744         ASSERT(XFS_DQ_IS_LOCKED(dqp));
745         if (! dolock) {
746                 xfs_dqunlock(dqp);
747                 goto done;
748         }
749         if (! udqhint)
750                 goto done;
751
752         ASSERT(udqhint);
753         ASSERT(dolock);
754         ASSERT(XFS_DQ_IS_LOCKED(dqp));
755         if (! xfs_qm_dqlock_nowait(udqhint)) {
756                 xfs_dqunlock(dqp);
757                 xfs_dqlock(udqhint);
758                 xfs_dqlock(dqp);
759         }
760       done:
761 #ifdef QUOTADEBUG
762         if (udqhint) {
763                 if (dolock)
764                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
765         }
766         if (! error) {
767                 if (dolock)
768                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
769         }
770 #endif
771         return error;
772 }
773
774
775 /*
776  * Given a udquot and gdquot, attach a ptr to the group dquot in the
777  * udquot as a hint for future lookups. The idea sounds simple, but the
778  * execution isn't, because the udquot might have a group dquot attached
779  * already and getting rid of that gets us into lock ordering constraints.
780  * The process is complicated more by the fact that the dquots may or may not
781  * be locked on entry.
782  */
783 STATIC void
784 xfs_qm_dqattach_grouphint(
785         xfs_dquot_t     *udq,
786         xfs_dquot_t     *gdq,
787         uint            locked)
788 {
789         xfs_dquot_t     *tmp;
790
791 #ifdef QUOTADEBUG
792         if (locked) {
793                 ASSERT(XFS_DQ_IS_LOCKED(udq));
794                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
795         }
796 #endif
797         if (! locked)
798                 xfs_dqlock(udq);
799
800         if ((tmp = udq->q_gdquot)) {
801                 if (tmp == gdq) {
802                         if (! locked)
803                                 xfs_dqunlock(udq);
804                         return;
805                 }
806
807                 udq->q_gdquot = NULL;
808                 /*
809                  * We can't keep any dqlocks when calling dqrele,
810                  * because the freelist lock comes before dqlocks.
811                  */
812                 xfs_dqunlock(udq);
813                 if (locked)
814                         xfs_dqunlock(gdq);
815                 /*
816                  * we took a hard reference once upon a time in dqget,
817                  * so give it back when the udquot no longer points at it
818                  * dqput() does the unlocking of the dquot.
819                  */
820                 xfs_qm_dqrele(tmp);
821
822                 xfs_dqlock(udq);
823                 xfs_dqlock(gdq);
824
825         } else {
826                 ASSERT(XFS_DQ_IS_LOCKED(udq));
827                 if (! locked) {
828                         xfs_dqlock(gdq);
829                 }
830         }
831
832         ASSERT(XFS_DQ_IS_LOCKED(udq));
833         ASSERT(XFS_DQ_IS_LOCKED(gdq));
834         /*
835          * Somebody could have attached a gdquot here,
836          * when we dropped the uqlock. If so, just do nothing.
837          */
838         if (udq->q_gdquot == NULL) {
839                 XFS_DQHOLD(gdq);
840                 udq->q_gdquot = gdq;
841         }
842         if (! locked) {
843                 xfs_dqunlock(gdq);
844                 xfs_dqunlock(udq);
845         }
846 }
847
848
849 /*
850  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
851  * into account.
852  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
853  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
854  * much made this code a complete mess, but it has been pretty useful.
855  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
856  * Inode may get unlocked and relocked in here, and the caller must deal with
857  * the consequences.
858  */
859 int
860 xfs_qm_dqattach(
861         xfs_inode_t     *ip,
862         uint            flags)
863 {
864         xfs_mount_t     *mp = ip->i_mount;
865         uint            nquotas = 0;
866         int             error = 0;
867
868         if ((! XFS_IS_QUOTA_ON(mp)) ||
869             (! XFS_NOT_DQATTACHED(mp, ip)) ||
870             (ip->i_ino == mp->m_sb.sb_uquotino) ||
871             (ip->i_ino == mp->m_sb.sb_gquotino))
872                 return 0;
873
874         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
875                xfs_isilocked(ip, XFS_ILOCK_EXCL));
876
877         if (! (flags & XFS_QMOPT_ILOCKED))
878                 xfs_ilock(ip, XFS_ILOCK_EXCL);
879
880         if (XFS_IS_UQUOTA_ON(mp)) {
881                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
882                                                 flags & XFS_QMOPT_DQALLOC,
883                                                 flags & XFS_QMOPT_DQLOCK,
884                                                 NULL, &ip->i_udquot);
885                 if (error)
886                         goto done;
887                 nquotas++;
888         }
889
890         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
891         if (XFS_IS_OQUOTA_ON(mp)) {
892                 error = XFS_IS_GQUOTA_ON(mp) ?
893                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
894                                                 flags & XFS_QMOPT_DQALLOC,
895                                                 flags & XFS_QMOPT_DQLOCK,
896                                                 ip->i_udquot, &ip->i_gdquot) :
897                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
898                                                 flags & XFS_QMOPT_DQALLOC,
899                                                 flags & XFS_QMOPT_DQLOCK,
900                                                 ip->i_udquot, &ip->i_gdquot);
901                 /*
902                  * Don't worry about the udquot that we may have
903                  * attached above. It'll get detached, if not already.
904                  */
905                 if (error)
906                         goto done;
907                 nquotas++;
908         }
909
910         /*
911          * Attach this group quota to the user quota as a hint.
912          * This WON'T, in general, result in a thrash.
913          */
914         if (nquotas == 2) {
915                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
916                 ASSERT(ip->i_udquot);
917                 ASSERT(ip->i_gdquot);
918
919                 /*
920                  * We may or may not have the i_udquot locked at this point,
921                  * but this check is OK since we don't depend on the i_gdquot to
922                  * be accurate 100% all the time. It is just a hint, and this
923                  * will succeed in general.
924                  */
925                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
926                         goto done;
927                 /*
928                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
929                  */
930                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
931                                          flags & XFS_QMOPT_DQLOCK);
932         }
933
934       done:
935
936 #ifdef QUOTADEBUG
937         if (! error) {
938                 if (ip->i_udquot) {
939                         if (flags & XFS_QMOPT_DQLOCK)
940                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
941                 }
942                 if (ip->i_gdquot) {
943                         if (flags & XFS_QMOPT_DQLOCK)
944                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
945                 }
946                 if (XFS_IS_UQUOTA_ON(mp))
947                         ASSERT(ip->i_udquot);
948                 if (XFS_IS_OQUOTA_ON(mp))
949                         ASSERT(ip->i_gdquot);
950         }
951 #endif
952
953         if (! (flags & XFS_QMOPT_ILOCKED))
954                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
955
956 #ifdef QUOTADEBUG
957         else
958                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
959 #endif
960         return error;
961 }
962
963 /*
964  * Release dquots (and their references) if any.
965  * The inode should be locked EXCL except when this's called by
966  * xfs_ireclaim.
967  */
968 void
969 xfs_qm_dqdetach(
970         xfs_inode_t     *ip)
971 {
972         if (!(ip->i_udquot || ip->i_gdquot))
973                 return;
974
975         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
976         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
977         if (ip->i_udquot) {
978                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
979                 xfs_qm_dqrele(ip->i_udquot);
980                 ip->i_udquot = NULL;
981         }
982         if (ip->i_gdquot) {
983                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
984                 xfs_qm_dqrele(ip->i_gdquot);
985                 ip->i_gdquot = NULL;
986         }
987 }
988
989 /*
990  * This is called by VFS_SYNC and flags arg determines the caller,
991  * and its motives, as done in xfs_sync.
992  *
993  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
994  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
995  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
996  */
997
998 int
999 xfs_qm_sync(
1000         xfs_mount_t     *mp,
1001         int             flags)
1002 {
1003         int             recl, restarts;
1004         xfs_dquot_t     *dqp;
1005         uint            flush_flags;
1006         boolean_t       nowait;
1007         int             error;
1008
1009         if (! XFS_IS_QUOTA_ON(mp))
1010                 return 0;
1011
1012         restarts = 0;
1013         /*
1014          * We won't block unless we are asked to.
1015          */
1016         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1017
1018   again:
1019         xfs_qm_mplist_lock(mp);
1020         /*
1021          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1022          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1023          * when we have the mplist lock, we know that dquots will be consistent
1024          * as long as we have it locked.
1025          */
1026         if (! XFS_IS_QUOTA_ON(mp)) {
1027                 xfs_qm_mplist_unlock(mp);
1028                 return 0;
1029         }
1030         FOREACH_DQUOT_IN_MP(dqp, mp) {
1031                 /*
1032                  * If this is vfs_sync calling, then skip the dquots that
1033                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1034                  * This is very similar to what xfs_sync does with inodes.
1035                  */
1036                 if (flags & SYNC_BDFLUSH) {
1037                         if (! XFS_DQ_IS_DIRTY(dqp))
1038                                 continue;
1039                 }
1040
1041                 if (nowait) {
1042                         /*
1043                          * Try to acquire the dquot lock. We are NOT out of
1044                          * lock order, but we just don't want to wait for this
1045                          * lock, unless somebody wanted us to.
1046                          */
1047                         if (! xfs_qm_dqlock_nowait(dqp))
1048                                 continue;
1049                 } else {
1050                         xfs_dqlock(dqp);
1051                 }
1052
1053                 /*
1054                  * Now, find out for sure if this dquot is dirty or not.
1055                  */
1056                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1057                         xfs_dqunlock(dqp);
1058                         continue;
1059                 }
1060
1061                 /* XXX a sentinel would be better */
1062                 recl = XFS_QI_MPLRECLAIMS(mp);
1063                 if (!xfs_dqflock_nowait(dqp)) {
1064                         if (nowait) {
1065                                 xfs_dqunlock(dqp);
1066                                 continue;
1067                         }
1068                         /*
1069                          * If we can't grab the flush lock then if the caller
1070                          * really wanted us to give this our best shot, so
1071                          * see if we can give a push to the buffer before we wait
1072                          * on the flush lock. At this point, we know that
1073                          * even though the dquot is being flushed,
1074                          * it has (new) dirty data.
1075                          */
1076                         xfs_qm_dqflock_pushbuf_wait(dqp);
1077                 }
1078                 /*
1079                  * Let go of the mplist lock. We don't want to hold it
1080                  * across a disk write
1081                  */
1082                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1083                 xfs_qm_mplist_unlock(mp);
1084                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1085                 error = xfs_qm_dqflush(dqp, flush_flags);
1086                 xfs_dqunlock(dqp);
1087                 if (error && XFS_FORCED_SHUTDOWN(mp))
1088                         return 0;       /* Need to prevent umount failure */
1089                 else if (error)
1090                         return error;
1091
1092                 xfs_qm_mplist_lock(mp);
1093                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1094                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1095                                 break;
1096
1097                         xfs_qm_mplist_unlock(mp);
1098                         goto again;
1099                 }
1100         }
1101
1102         xfs_qm_mplist_unlock(mp);
1103         return 0;
1104 }
1105
1106
1107 /*
1108  * This initializes all the quota information that's kept in the
1109  * mount structure
1110  */
1111 STATIC int
1112 xfs_qm_init_quotainfo(
1113         xfs_mount_t     *mp)
1114 {
1115         xfs_quotainfo_t *qinf;
1116         int             error;
1117         xfs_dquot_t     *dqp;
1118
1119         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1120
1121         /*
1122          * Tell XQM that we exist as soon as possible.
1123          */
1124         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1125                 return error;
1126         }
1127
1128         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1129
1130         /*
1131          * See if quotainodes are setup, and if not, allocate them,
1132          * and change the superblock accordingly.
1133          */
1134         if ((error = xfs_qm_init_quotainos(mp))) {
1135                 kmem_free(qinf);
1136                 mp->m_quotainfo = NULL;
1137                 return error;
1138         }
1139
1140         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1141         qinf->qi_dqreclaims = 0;
1142
1143         /* mutex used to serialize quotaoffs */
1144         mutex_init(&qinf->qi_quotaofflock);
1145
1146         /* Precalc some constants */
1147         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1148         ASSERT(qinf->qi_dqchunklen);
1149         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1150         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1151
1152         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1153
1154         /*
1155          * We try to get the limits from the superuser's limits fields.
1156          * This is quite hacky, but it is standard quota practice.
1157          * We look at the USR dquot with id == 0 first, but if user quotas
1158          * are not enabled we goto the GRP dquot with id == 0.
1159          * We don't really care to keep separate default limits for user
1160          * and group quotas, at least not at this point.
1161          */
1162         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1163                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1164                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1165                                 XFS_DQ_PROJ),
1166                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1167                              &dqp);
1168         if (! error) {
1169                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1170
1171                 /*
1172                  * The warnings and timers set the grace period given to
1173                  * a user or group before he or she can not perform any
1174                  * more writing. If it is zero, a default is used.
1175                  */
1176                 qinf->qi_btimelimit = ddqp->d_btimer ?
1177                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1178                 qinf->qi_itimelimit = ddqp->d_itimer ?
1179                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1180                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1181                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1182                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1183                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1184                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1185                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1186                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1187                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1188                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1189                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1190                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1191                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1192                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1193                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1194  
1195                 /*
1196                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1197                  * we don't want this dquot cached. We haven't done a
1198                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1199                  */
1200                 xfs_qm_dqdestroy(dqp);
1201         } else {
1202                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1203                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1204                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1205                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1206                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1207                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1208         }
1209
1210         return 0;
1211 }
1212
1213
1214 /*
1215  * Gets called when unmounting a filesystem or when all quotas get
1216  * turned off.
1217  * This purges the quota inodes, destroys locks and frees itself.
1218  */
1219 void
1220 xfs_qm_destroy_quotainfo(
1221         xfs_mount_t     *mp)
1222 {
1223         xfs_quotainfo_t *qi;
1224
1225         qi = mp->m_quotainfo;
1226         ASSERT(qi != NULL);
1227         ASSERT(xfs_Gqm != NULL);
1228
1229         /*
1230          * Release the reference that XQM kept, so that we know
1231          * when the XQM structure should be freed. We cannot assume
1232          * that xfs_Gqm is non-null after this point.
1233          */
1234         xfs_qm_rele_quotafs_ref(mp);
1235
1236         xfs_qm_list_destroy(&qi->qi_dqlist);
1237
1238         if (qi->qi_uquotaip) {
1239                 IRELE(qi->qi_uquotaip);
1240                 qi->qi_uquotaip = NULL; /* paranoia */
1241         }
1242         if (qi->qi_gquotaip) {
1243                 IRELE(qi->qi_gquotaip);
1244                 qi->qi_gquotaip = NULL;
1245         }
1246         mutex_destroy(&qi->qi_quotaofflock);
1247         kmem_free(qi);
1248         mp->m_quotainfo = NULL;
1249 }
1250
1251
1252
1253 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1254
1255 /* ARGSUSED */
1256 STATIC void
1257 xfs_qm_list_init(
1258         xfs_dqlist_t    *list,
1259         char            *str,
1260         int             n)
1261 {
1262         mutex_init(&list->qh_lock);
1263         list->qh_next = NULL;
1264         list->qh_version = 0;
1265         list->qh_nelems = 0;
1266 }
1267
1268 STATIC void
1269 xfs_qm_list_destroy(
1270         xfs_dqlist_t    *list)
1271 {
1272         mutex_destroy(&(list->qh_lock));
1273 }
1274
1275
1276 /*
1277  * Stripped down version of dqattach. This doesn't attach, or even look at the
1278  * dquots attached to the inode. The rationale is that there won't be any
1279  * attached at the time this is called from quotacheck.
1280  */
1281 STATIC int
1282 xfs_qm_dqget_noattach(
1283         xfs_inode_t     *ip,
1284         xfs_dquot_t     **O_udqpp,
1285         xfs_dquot_t     **O_gdqpp)
1286 {
1287         int             error;
1288         xfs_mount_t     *mp;
1289         xfs_dquot_t     *udqp, *gdqp;
1290
1291         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1292         mp = ip->i_mount;
1293         udqp = NULL;
1294         gdqp = NULL;
1295
1296         if (XFS_IS_UQUOTA_ON(mp)) {
1297                 ASSERT(ip->i_udquot == NULL);
1298                 /*
1299                  * We want the dquot allocated if it doesn't exist.
1300                  */
1301                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1302                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1303                                          &udqp))) {
1304                         /*
1305                          * Shouldn't be able to turn off quotas here.
1306                          */
1307                         ASSERT(error != ESRCH);
1308                         ASSERT(error != ENOENT);
1309                         return error;
1310                 }
1311                 ASSERT(udqp);
1312         }
1313
1314         if (XFS_IS_OQUOTA_ON(mp)) {
1315                 ASSERT(ip->i_gdquot == NULL);
1316                 if (udqp)
1317                         xfs_dqunlock(udqp);
1318                 error = XFS_IS_GQUOTA_ON(mp) ?
1319                                 xfs_qm_dqget(mp, ip,
1320                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1321                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1322                                              &gdqp) :
1323                                 xfs_qm_dqget(mp, ip,
1324                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1325                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1326                                              &gdqp);
1327                 if (error) {
1328                         if (udqp)
1329                                 xfs_qm_dqrele(udqp);
1330                         ASSERT(error != ESRCH);
1331                         ASSERT(error != ENOENT);
1332                         return error;
1333                 }
1334                 ASSERT(gdqp);
1335
1336                 /* Reacquire the locks in the right order */
1337                 if (udqp) {
1338                         if (! xfs_qm_dqlock_nowait(udqp)) {
1339                                 xfs_dqunlock(gdqp);
1340                                 xfs_dqlock(udqp);
1341                                 xfs_dqlock(gdqp);
1342                         }
1343                 }
1344         }
1345
1346         *O_udqpp = udqp;
1347         *O_gdqpp = gdqp;
1348
1349 #ifdef QUOTADEBUG
1350         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1351         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1352 #endif
1353         return 0;
1354 }
1355
1356 /*
1357  * Create an inode and return with a reference already taken, but unlocked
1358  * This is how we create quota inodes
1359  */
1360 STATIC int
1361 xfs_qm_qino_alloc(
1362         xfs_mount_t     *mp,
1363         xfs_inode_t     **ip,
1364         __int64_t       sbfields,
1365         uint            flags)
1366 {
1367         xfs_trans_t     *tp;
1368         int             error;
1369         int             committed;
1370
1371         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1372         if ((error = xfs_trans_reserve(tp,
1373                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1374                                       XFS_CREATE_LOG_RES(mp), 0,
1375                                       XFS_TRANS_PERM_LOG_RES,
1376                                       XFS_CREATE_LOG_COUNT))) {
1377                 xfs_trans_cancel(tp, 0);
1378                 return error;
1379         }
1380
1381         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1382                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1383                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1384                                  XFS_TRANS_ABORT);
1385                 return error;
1386         }
1387
1388         /*
1389          * Keep an extra reference to this quota inode. This inode is
1390          * locked exclusively and joined to the transaction already.
1391          */
1392         ASSERT(xfs_isilocked(*ip, XFS_ILOCK_EXCL));
1393         IHOLD(*ip);
1394
1395         /*
1396          * Make the changes in the superblock, and log those too.
1397          * sbfields arg may contain fields other than *QUOTINO;
1398          * VERSIONNUM for example.
1399          */
1400         spin_lock(&mp->m_sb_lock);
1401         if (flags & XFS_QMOPT_SBVERSION) {
1402 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1403                 unsigned oldv = mp->m_sb.sb_versionnum;
1404 #endif
1405                 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb));
1406                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1407                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1408                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1409                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1410
1411                 xfs_sb_version_addquota(&mp->m_sb);
1412                 mp->m_sb.sb_uquotino = NULLFSINO;
1413                 mp->m_sb.sb_gquotino = NULLFSINO;
1414
1415                 /* qflags will get updated _after_ quotacheck */
1416                 mp->m_sb.sb_qflags = 0;
1417 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1418                 cmn_err(CE_NOTE,
1419                         "Old superblock version %x, converting to %x.",
1420                         oldv, mp->m_sb.sb_versionnum);
1421 #endif
1422         }
1423         if (flags & XFS_QMOPT_UQUOTA)
1424                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1425         else
1426                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1427         spin_unlock(&mp->m_sb_lock);
1428         xfs_mod_sb(tp, sbfields);
1429
1430         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1431                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1432                 return error;
1433         }
1434         return 0;
1435 }
1436
1437
1438 STATIC void
1439 xfs_qm_reset_dqcounts(
1440         xfs_mount_t     *mp,
1441         xfs_buf_t       *bp,
1442         xfs_dqid_t      id,
1443         uint            type)
1444 {
1445         xfs_disk_dquot_t        *ddq;
1446         int                     j;
1447
1448         xfs_buftrace("RESET DQUOTS", bp);
1449         /*
1450          * Reset all counters and timers. They'll be
1451          * started afresh by xfs_qm_quotacheck.
1452          */
1453 #ifdef DEBUG
1454         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1455         do_div(j, sizeof(xfs_dqblk_t));
1456         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1457 #endif
1458         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1459         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1460                 /*
1461                  * Do a sanity check, and if needed, repair the dqblk. Don't
1462                  * output any warnings because it's perfectly possible to
1463                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1464                  */
1465                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1466                                       "xfs_quotacheck");
1467                 ddq->d_bcount = 0;
1468                 ddq->d_icount = 0;
1469                 ddq->d_rtbcount = 0;
1470                 ddq->d_btimer = 0;
1471                 ddq->d_itimer = 0;
1472                 ddq->d_rtbtimer = 0;
1473                 ddq->d_bwarns = 0;
1474                 ddq->d_iwarns = 0;
1475                 ddq->d_rtbwarns = 0;
1476                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1477         }
1478 }
1479
1480 STATIC int
1481 xfs_qm_dqiter_bufs(
1482         xfs_mount_t     *mp,
1483         xfs_dqid_t      firstid,
1484         xfs_fsblock_t   bno,
1485         xfs_filblks_t   blkcnt,
1486         uint            flags)
1487 {
1488         xfs_buf_t       *bp;
1489         int             error;
1490         int             notcommitted;
1491         int             incr;
1492         int             type;
1493
1494         ASSERT(blkcnt > 0);
1495         notcommitted = 0;
1496         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1497                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1498         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1499                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1500         error = 0;
1501
1502         /*
1503          * Blkcnt arg can be a very big number, and might even be
1504          * larger than the log itself. So, we have to break it up into
1505          * manageable-sized transactions.
1506          * Note that we don't start a permanent transaction here; we might
1507          * not be able to get a log reservation for the whole thing up front,
1508          * and we don't really care to either, because we just discard
1509          * everything if we were to crash in the middle of this loop.
1510          */
1511         while (blkcnt--) {
1512                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1513                               XFS_FSB_TO_DADDR(mp, bno),
1514                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1515                 if (error)
1516                         break;
1517
1518                 xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1519                 xfs_bdwrite(mp, bp);
1520                 /*
1521                  * goto the next block.
1522                  */
1523                 bno++;
1524                 firstid += XFS_QM_DQPERBLK(mp);
1525         }
1526         return error;
1527 }
1528
1529 /*
1530  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1531  * caller supplied function for every chunk of dquots that we find.
1532  */
1533 STATIC int
1534 xfs_qm_dqiterate(
1535         xfs_mount_t     *mp,
1536         xfs_inode_t     *qip,
1537         uint            flags)
1538 {
1539         xfs_bmbt_irec_t         *map;
1540         int                     i, nmaps;       /* number of map entries */
1541         int                     error;          /* return value */
1542         xfs_fileoff_t           lblkno;
1543         xfs_filblks_t           maxlblkcnt;
1544         xfs_dqid_t              firstid;
1545         xfs_fsblock_t           rablkno;
1546         xfs_filblks_t           rablkcnt;
1547
1548         error = 0;
1549         /*
1550          * This looks racy, but we can't keep an inode lock across a
1551          * trans_reserve. But, this gets called during quotacheck, and that
1552          * happens only at mount time which is single threaded.
1553          */
1554         if (qip->i_d.di_nblocks == 0)
1555                 return 0;
1556
1557         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1558
1559         lblkno = 0;
1560         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1561         do {
1562                 nmaps = XFS_DQITER_MAP_SIZE;
1563                 /*
1564                  * We aren't changing the inode itself. Just changing
1565                  * some of its data. No new blocks are added here, and
1566                  * the inode is never added to the transaction.
1567                  */
1568                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1569                 error = xfs_bmapi(NULL, qip, lblkno,
1570                                   maxlblkcnt - lblkno,
1571                                   XFS_BMAPI_METADATA,
1572                                   NULL,
1573                                   0, map, &nmaps, NULL, NULL);
1574                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1575                 if (error)
1576                         break;
1577
1578                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1579                 for (i = 0; i < nmaps; i++) {
1580                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1581                         ASSERT(map[i].br_blockcount);
1582
1583
1584                         lblkno += map[i].br_blockcount;
1585
1586                         if (map[i].br_startblock == HOLESTARTBLOCK)
1587                                 continue;
1588
1589                         firstid = (xfs_dqid_t) map[i].br_startoff *
1590                                 XFS_QM_DQPERBLK(mp);
1591                         /*
1592                          * Do a read-ahead on the next extent.
1593                          */
1594                         if ((i+1 < nmaps) &&
1595                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1596                                 rablkcnt =  map[i+1].br_blockcount;
1597                                 rablkno = map[i+1].br_startblock;
1598                                 while (rablkcnt--) {
1599                                         xfs_baread(mp->m_ddev_targp,
1600                                                XFS_FSB_TO_DADDR(mp, rablkno),
1601                                                (int)XFS_QI_DQCHUNKLEN(mp));
1602                                         rablkno++;
1603                                 }
1604                         }
1605                         /*
1606                          * Iterate thru all the blks in the extent and
1607                          * reset the counters of all the dquots inside them.
1608                          */
1609                         if ((error = xfs_qm_dqiter_bufs(mp,
1610                                                        firstid,
1611                                                        map[i].br_startblock,
1612                                                        map[i].br_blockcount,
1613                                                        flags))) {
1614                                 break;
1615                         }
1616                 }
1617
1618                 if (error)
1619                         break;
1620         } while (nmaps > 0);
1621
1622         kmem_free(map);
1623
1624         return error;
1625 }
1626
1627 /*
1628  * Called by dqusage_adjust in doing a quotacheck.
1629  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1630  * this updates its incore copy as well as the buffer copy. This is
1631  * so that once the quotacheck is done, we can just log all the buffers,
1632  * as opposed to logging numerous updates to individual dquots.
1633  */
1634 STATIC void
1635 xfs_qm_quotacheck_dqadjust(
1636         xfs_dquot_t             *dqp,
1637         xfs_qcnt_t              nblks,
1638         xfs_qcnt_t              rtblks)
1639 {
1640         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1641         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1642         /*
1643          * Adjust the inode count and the block count to reflect this inode's
1644          * resource usage.
1645          */
1646         be64_add_cpu(&dqp->q_core.d_icount, 1);
1647         dqp->q_res_icount++;
1648         if (nblks) {
1649                 be64_add_cpu(&dqp->q_core.d_bcount, nblks);
1650                 dqp->q_res_bcount += nblks;
1651         }
1652         if (rtblks) {
1653                 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks);
1654                 dqp->q_res_rtbcount += rtblks;
1655         }
1656
1657         /*
1658          * Set default limits, adjust timers (since we changed usages)
1659          */
1660         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1661                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1662                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1663         }
1664
1665         dqp->dq_flags |= XFS_DQ_DIRTY;
1666 }
1667
1668 STATIC int
1669 xfs_qm_get_rtblks(
1670         xfs_inode_t     *ip,
1671         xfs_qcnt_t      *O_rtblks)
1672 {
1673         xfs_filblks_t   rtblks;                 /* total rt blks */
1674         xfs_extnum_t    idx;                    /* extent record index */
1675         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1676         xfs_extnum_t    nextents;               /* number of extent entries */
1677         int             error;
1678
1679         ASSERT(XFS_IS_REALTIME_INODE(ip));
1680         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1681         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1682                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1683                         return error;
1684         }
1685         rtblks = 0;
1686         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1687         for (idx = 0; idx < nextents; idx++)
1688                 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx));
1689         *O_rtblks = (xfs_qcnt_t)rtblks;
1690         return 0;
1691 }
1692
1693 /*
1694  * callback routine supplied to bulkstat(). Given an inumber, find its
1695  * dquots and update them to account for resources taken by that inode.
1696  */
1697 /* ARGSUSED */
1698 STATIC int
1699 xfs_qm_dqusage_adjust(
1700         xfs_mount_t     *mp,            /* mount point for filesystem */
1701         xfs_ino_t       ino,            /* inode number to get data for */
1702         void            __user *buffer, /* not used */
1703         int             ubsize,         /* not used */
1704         void            *private_data,  /* not used */
1705         xfs_daddr_t     bno,            /* starting block of inode cluster */
1706         int             *ubused,        /* not used */
1707         void            *dip,           /* on-disk inode pointer (not used) */
1708         int             *res)           /* result code value */
1709 {
1710         xfs_inode_t     *ip;
1711         xfs_dquot_t     *udqp, *gdqp;
1712         xfs_qcnt_t      nblks, rtblks;
1713         int             error;
1714
1715         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1716
1717         /*
1718          * rootino must have its resources accounted for, not so with the quota
1719          * inodes.
1720          */
1721         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1722                 *res = BULKSTAT_RV_NOTHING;
1723                 return XFS_ERROR(EINVAL);
1724         }
1725
1726         /*
1727          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1728          * interface expects the inode to be exclusively locked because that's
1729          * the case in all other instances. It's OK that we do this because
1730          * quotacheck is done only at mount time.
1731          */
1732         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1733                 *res = BULKSTAT_RV_NOTHING;
1734                 return error;
1735         }
1736
1737         /*
1738          * Obtain the locked dquots. In case of an error (eg. allocation
1739          * fails for ENOSPC), we return the negative of the error number
1740          * to bulkstat, so that it can get propagated to quotacheck() and
1741          * making us disable quotas for the file system.
1742          */
1743         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1744                 xfs_iput(ip, XFS_ILOCK_EXCL);
1745                 *res = BULKSTAT_RV_GIVEUP;
1746                 return error;
1747         }
1748
1749         rtblks = 0;
1750         if (! XFS_IS_REALTIME_INODE(ip)) {
1751                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1752         } else {
1753                 /*
1754                  * Walk thru the extent list and count the realtime blocks.
1755                  */
1756                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1757                         xfs_iput(ip, XFS_ILOCK_EXCL);
1758                         if (udqp)
1759                                 xfs_qm_dqput(udqp);
1760                         if (gdqp)
1761                                 xfs_qm_dqput(gdqp);
1762                         *res = BULKSTAT_RV_GIVEUP;
1763                         return error;
1764                 }
1765                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1766         }
1767         ASSERT(ip->i_delayed_blks == 0);
1768
1769         /*
1770          * We can't release the inode while holding its dquot locks.
1771          * The inode can go into inactive and might try to acquire the dquotlocks.
1772          * So, just unlock here and do a vn_rele at the end.
1773          */
1774         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1775
1776         /*
1777          * Add the (disk blocks and inode) resources occupied by this
1778          * inode to its dquots. We do this adjustment in the incore dquot,
1779          * and also copy the changes to its buffer.
1780          * We don't care about putting these changes in a transaction
1781          * envelope because if we crash in the middle of a 'quotacheck'
1782          * we have to start from the beginning anyway.
1783          * Once we're done, we'll log all the dquot bufs.
1784          *
1785          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1786          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1787          */
1788         if (XFS_IS_UQUOTA_ON(mp)) {
1789                 ASSERT(udqp);
1790                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1791                 xfs_qm_dqput(udqp);
1792         }
1793         if (XFS_IS_OQUOTA_ON(mp)) {
1794                 ASSERT(gdqp);
1795                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1796                 xfs_qm_dqput(gdqp);
1797         }
1798         /*
1799          * Now release the inode. This will send it to 'inactive', and
1800          * possibly even free blocks.
1801          */
1802         IRELE(ip);
1803
1804         /*
1805          * Goto next inode.
1806          */
1807         *res = BULKSTAT_RV_DIDONE;
1808         return 0;
1809 }
1810
1811 /*
1812  * Walk thru all the filesystem inodes and construct a consistent view
1813  * of the disk quota world. If the quotacheck fails, disable quotas.
1814  */
1815 int
1816 xfs_qm_quotacheck(
1817         xfs_mount_t     *mp)
1818 {
1819         int             done, count, error;
1820         xfs_ino_t       lastino;
1821         size_t          structsz;
1822         xfs_inode_t     *uip, *gip;
1823         uint            flags;
1824
1825         count = INT_MAX;
1826         structsz = 1;
1827         lastino = 0;
1828         flags = 0;
1829
1830         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1831         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1832
1833         /*
1834          * There should be no cached dquots. The (simplistic) quotacheck
1835          * algorithm doesn't like that.
1836          */
1837         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1838
1839         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1840
1841         /*
1842          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1843          * their counters to zero. We need a clean slate.
1844          * We don't log our changes till later.
1845          */
1846         if ((uip = XFS_QI_UQIP(mp))) {
1847                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1848                         goto error_return;
1849                 flags |= XFS_UQUOTA_CHKD;
1850         }
1851
1852         if ((gip = XFS_QI_GQIP(mp))) {
1853                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1854                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1855                         goto error_return;
1856                 flags |= XFS_OQUOTA_CHKD;
1857         }
1858
1859         do {
1860                 /*
1861                  * Iterate thru all the inodes in the file system,
1862                  * adjusting the corresponding dquot counters in core.
1863                  */
1864                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1865                                      xfs_qm_dqusage_adjust, NULL,
1866                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1867                         break;
1868
1869         } while (! done);
1870
1871         /*
1872          * We've made all the changes that we need to make incore.
1873          * Flush them down to disk buffers if everything was updated
1874          * successfully.
1875          */
1876         if (!error)
1877                 error = xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1878
1879         /*
1880          * We can get this error if we couldn't do a dquot allocation inside
1881          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1882          * dirty dquots that might be cached, we just want to get rid of them
1883          * and turn quotaoff. The dquots won't be attached to any of the inodes
1884          * at this point (because we intentionally didn't in dqget_noattach).
1885          */
1886         if (error) {
1887                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1888                 goto error_return;
1889         }
1890
1891         /*
1892          * We didn't log anything, because if we crashed, we'll have to
1893          * start the quotacheck from scratch anyway. However, we must make
1894          * sure that our dquot changes are secure before we put the
1895          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1896          * flush.
1897          */
1898         XFS_bflush(mp->m_ddev_targp);
1899
1900         /*
1901          * If one type of quotas is off, then it will lose its
1902          * quotachecked status, since we won't be doing accounting for
1903          * that type anymore.
1904          */
1905         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1906         mp->m_qflags |= flags;
1907
1908         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1909
1910  error_return:
1911         if (error) {
1912                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1913                         "Disabling quotas.",
1914                         mp->m_fsname, error);
1915                 /*
1916                  * We must turn off quotas.
1917                  */
1918                 ASSERT(mp->m_quotainfo != NULL);
1919                 ASSERT(xfs_Gqm != NULL);
1920                 xfs_qm_destroy_quotainfo(mp);
1921                 if (xfs_mount_reset_sbqflags(mp)) {
1922                         cmn_err(CE_WARN, "XFS quotacheck %s: "
1923                                 "Failed to reset quota flags.", mp->m_fsname);
1924                 }
1925         } else {
1926                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1927         }
1928         return (error);
1929 }
1930
1931 /*
1932  * This is called after the superblock has been read in and we're ready to
1933  * iget the quota inodes.
1934  */
1935 STATIC int
1936 xfs_qm_init_quotainos(
1937         xfs_mount_t     *mp)
1938 {
1939         xfs_inode_t     *uip, *gip;
1940         int             error;
1941         __int64_t       sbflags;
1942         uint            flags;
1943
1944         ASSERT(mp->m_quotainfo);
1945         uip = gip = NULL;
1946         sbflags = 0;
1947         flags = 0;
1948
1949         /*
1950          * Get the uquota and gquota inodes
1951          */
1952         if (xfs_sb_version_hasquota(&mp->m_sb)) {
1953                 if (XFS_IS_UQUOTA_ON(mp) &&
1954                     mp->m_sb.sb_uquotino != NULLFSINO) {
1955                         ASSERT(mp->m_sb.sb_uquotino > 0);
1956                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1957                                              0, 0, &uip, 0)))
1958                                 return XFS_ERROR(error);
1959                 }
1960                 if (XFS_IS_OQUOTA_ON(mp) &&
1961                     mp->m_sb.sb_gquotino != NULLFSINO) {
1962                         ASSERT(mp->m_sb.sb_gquotino > 0);
1963                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
1964                                              0, 0, &gip, 0))) {
1965                                 if (uip)
1966                                         IRELE(uip);
1967                                 return XFS_ERROR(error);
1968                         }
1969                 }
1970         } else {
1971                 flags |= XFS_QMOPT_SBVERSION;
1972                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1973                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
1974         }
1975
1976         /*
1977          * Create the two inodes, if they don't exist already. The changes
1978          * made above will get added to a transaction and logged in one of
1979          * the qino_alloc calls below.  If the device is readonly,
1980          * temporarily switch to read-write to do this.
1981          */
1982         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
1983                 if ((error = xfs_qm_qino_alloc(mp, &uip,
1984                                               sbflags | XFS_SB_UQUOTINO,
1985                                               flags | XFS_QMOPT_UQUOTA)))
1986                         return XFS_ERROR(error);
1987
1988                 flags &= ~XFS_QMOPT_SBVERSION;
1989         }
1990         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
1991                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
1992                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
1993                 error = xfs_qm_qino_alloc(mp, &gip,
1994                                           sbflags | XFS_SB_GQUOTINO, flags);
1995                 if (error) {
1996                         if (uip)
1997                                 IRELE(uip);
1998
1999                         return XFS_ERROR(error);
2000                 }
2001         }
2002
2003         XFS_QI_UQIP(mp) = uip;
2004         XFS_QI_GQIP(mp) = gip;
2005
2006         return 0;
2007 }
2008
2009
2010 /*
2011  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2012  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2013  * favor the lookup function ...
2014  * XXXsup merge this with qm_reclaim_one().
2015  */
2016 STATIC int
2017 xfs_qm_shake_freelist(
2018         int howmany)
2019 {
2020         int             nreclaimed;
2021         xfs_dqhash_t    *hash;
2022         xfs_dquot_t     *dqp, *nextdqp;
2023         int             restarts;
2024         int             nflushes;
2025
2026         if (howmany <= 0)
2027                 return 0;
2028
2029         nreclaimed = 0;
2030         restarts = 0;
2031         nflushes = 0;
2032
2033 #ifdef QUOTADEBUG
2034         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2035 #endif
2036         /* lock order is : hashchainlock, freelistlock, mplistlock */
2037  tryagain:
2038         xfs_qm_freelist_lock(xfs_Gqm);
2039
2040         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2041              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2042               nreclaimed < howmany); ) {
2043                 xfs_dqlock(dqp);
2044
2045                 /*
2046                  * We are racing with dqlookup here. Naturally we don't
2047                  * want to reclaim a dquot that lookup wants.
2048                  */
2049                 if (dqp->dq_flags & XFS_DQ_WANT) {
2050                         xfs_dqunlock(dqp);
2051                         xfs_qm_freelist_unlock(xfs_Gqm);
2052                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2053                                 return nreclaimed;
2054                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2055                         goto tryagain;
2056                 }
2057
2058                 /*
2059                  * If the dquot is inactive, we are assured that it is
2060                  * not on the mplist or the hashlist, and that makes our
2061                  * life easier.
2062                  */
2063                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2064                         ASSERT(dqp->q_mount == NULL);
2065                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2066                         ASSERT(dqp->HL_PREVP == NULL);
2067                         ASSERT(dqp->MPL_PREVP == NULL);
2068                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2069                         nextdqp = dqp->dq_flnext;
2070                         goto off_freelist;
2071                 }
2072
2073                 ASSERT(dqp->MPL_PREVP);
2074                 /*
2075                  * Try to grab the flush lock. If this dquot is in the process of
2076                  * getting flushed to disk, we don't want to reclaim it.
2077                  */
2078                 if (!xfs_dqflock_nowait(dqp)) {
2079                         xfs_dqunlock(dqp);
2080                         dqp = dqp->dq_flnext;
2081                         continue;
2082                 }
2083
2084                 /*
2085                  * We have the flush lock so we know that this is not in the
2086                  * process of being flushed. So, if this is dirty, flush it
2087                  * DELWRI so that we don't get a freelist infested with
2088                  * dirty dquots.
2089                  */
2090                 if (XFS_DQ_IS_DIRTY(dqp)) {
2091                         int     error;
2092                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2093                         /*
2094                          * We flush it delayed write, so don't bother
2095                          * releasing the mplock.
2096                          */
2097                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2098                         if (error) {
2099                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2100                         "xfs_qm_dqflush_all: dquot %p flush failed", dqp);
2101                         }
2102                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2103                         dqp = dqp->dq_flnext;
2104                         continue;
2105                 }
2106                 /*
2107                  * We're trying to get the hashlock out of order. This races
2108                  * with dqlookup; so, we giveup and goto the next dquot if
2109                  * we couldn't get the hashlock. This way, we won't starve
2110                  * a dqlookup process that holds the hashlock that is
2111                  * waiting for the freelist lock.
2112                  */
2113                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2114                         xfs_dqfunlock(dqp);
2115                         xfs_dqunlock(dqp);
2116                         dqp = dqp->dq_flnext;
2117                         continue;
2118                 }
2119                 /*
2120                  * This races with dquot allocation code as well as dqflush_all
2121                  * and reclaim code. So, if we failed to grab the mplist lock,
2122                  * giveup everything and start over.
2123                  */
2124                 hash = dqp->q_hash;
2125                 ASSERT(hash);
2126                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2127                         /* XXX put a sentinel so that we can come back here */
2128                         xfs_dqfunlock(dqp);
2129                         xfs_dqunlock(dqp);
2130                         XFS_DQ_HASH_UNLOCK(hash);
2131                         xfs_qm_freelist_unlock(xfs_Gqm);
2132                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2133                                 return nreclaimed;
2134                         goto tryagain;
2135                 }
2136                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2137 #ifdef QUOTADEBUG
2138                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2139                         dqp, be32_to_cpu(dqp->q_core.d_id));
2140 #endif
2141                 ASSERT(dqp->q_nrefs == 0);
2142                 nextdqp = dqp->dq_flnext;
2143                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2144                 XQM_HASHLIST_REMOVE(hash, dqp);
2145                 xfs_dqfunlock(dqp);
2146                 xfs_qm_mplist_unlock(dqp->q_mount);
2147                 XFS_DQ_HASH_UNLOCK(hash);
2148
2149  off_freelist:
2150                 XQM_FREELIST_REMOVE(dqp);
2151                 xfs_dqunlock(dqp);
2152                 nreclaimed++;
2153                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2154                 xfs_qm_dqdestroy(dqp);
2155                 dqp = nextdqp;
2156         }
2157         xfs_qm_freelist_unlock(xfs_Gqm);
2158         return nreclaimed;
2159 }
2160
2161
2162 /*
2163  * The kmem_shake interface is invoked when memory is running low.
2164  */
2165 /* ARGSUSED */
2166 STATIC int
2167 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2168 {
2169         int     ndqused, nfree, n;
2170
2171         if (!kmem_shake_allow(gfp_mask))
2172                 return 0;
2173         if (!xfs_Gqm)
2174                 return 0;
2175
2176         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2177         /* incore dquots in all f/s's */
2178         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2179
2180         ASSERT(ndqused >= 0);
2181
2182         if (nfree <= ndqused && nfree < ndquot)
2183                 return 0;
2184
2185         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2186         n = nfree - ndqused - ndquot;           /* # over target */
2187
2188         return xfs_qm_shake_freelist(MAX(nfree, n));
2189 }
2190
2191
2192 /*
2193  * Just pop the least recently used dquot off the freelist and
2194  * recycle it. The returned dquot is locked.
2195  */
2196 STATIC xfs_dquot_t *
2197 xfs_qm_dqreclaim_one(void)
2198 {
2199         xfs_dquot_t     *dqpout;
2200         xfs_dquot_t     *dqp;
2201         int             restarts;
2202         int             nflushes;
2203
2204         restarts = 0;
2205         dqpout = NULL;
2206         nflushes = 0;
2207
2208         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2209  startagain:
2210         xfs_qm_freelist_lock(xfs_Gqm);
2211
2212         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2213                 xfs_dqlock(dqp);
2214
2215                 /*
2216                  * We are racing with dqlookup here. Naturally we don't
2217                  * want to reclaim a dquot that lookup wants. We release the
2218                  * freelist lock and start over, so that lookup will grab
2219                  * both the dquot and the freelistlock.
2220                  */
2221                 if (dqp->dq_flags & XFS_DQ_WANT) {
2222                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2223                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2224                         xfs_dqunlock(dqp);
2225                         xfs_qm_freelist_unlock(xfs_Gqm);
2226                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2227                                 return NULL;
2228                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2229                         goto startagain;
2230                 }
2231
2232                 /*
2233                  * If the dquot is inactive, we are assured that it is
2234                  * not on the mplist or the hashlist, and that makes our
2235                  * life easier.
2236                  */
2237                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2238                         ASSERT(dqp->q_mount == NULL);
2239                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2240                         ASSERT(dqp->HL_PREVP == NULL);
2241                         ASSERT(dqp->MPL_PREVP == NULL);
2242                         XQM_FREELIST_REMOVE(dqp);
2243                         xfs_dqunlock(dqp);
2244                         dqpout = dqp;
2245                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2246                         break;
2247                 }
2248
2249                 ASSERT(dqp->q_hash);
2250                 ASSERT(dqp->MPL_PREVP);
2251
2252                 /*
2253                  * Try to grab the flush lock. If this dquot is in the process of
2254                  * getting flushed to disk, we don't want to reclaim it.
2255                  */
2256                 if (!xfs_dqflock_nowait(dqp)) {
2257                         xfs_dqunlock(dqp);
2258                         continue;
2259                 }
2260
2261                 /*
2262                  * We have the flush lock so we know that this is not in the
2263                  * process of being flushed. So, if this is dirty, flush it
2264                  * DELWRI so that we don't get a freelist infested with
2265                  * dirty dquots.
2266                  */
2267                 if (XFS_DQ_IS_DIRTY(dqp)) {
2268                         int     error;
2269                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2270                         /*
2271                          * We flush it delayed write, so don't bother
2272                          * releasing the freelist lock.
2273                          */
2274                         error = xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2275                         if (error) {
2276                                 xfs_fs_cmn_err(CE_WARN, dqp->q_mount,
2277                         "xfs_qm_dqreclaim: dquot %p flush failed", dqp);
2278                         }
2279                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2280                         continue;
2281                 }
2282
2283                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2284                         xfs_dqfunlock(dqp);
2285                         xfs_dqunlock(dqp);
2286                         continue;
2287                 }
2288
2289                 if (! xfs_qm_dqhashlock_nowait(dqp))
2290                         goto mplistunlock;
2291
2292                 ASSERT(dqp->q_nrefs == 0);
2293                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2294                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2295                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2296                 XQM_FREELIST_REMOVE(dqp);
2297                 dqpout = dqp;
2298                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2299  mplistunlock:
2300                 xfs_qm_mplist_unlock(dqp->q_mount);
2301                 xfs_dqfunlock(dqp);
2302                 xfs_dqunlock(dqp);
2303                 if (dqpout)
2304                         break;
2305         }
2306
2307         xfs_qm_freelist_unlock(xfs_Gqm);
2308         return dqpout;
2309 }
2310
2311
2312 /*------------------------------------------------------------------*/
2313
2314 /*
2315  * Return a new incore dquot. Depending on the number of
2316  * dquots in the system, we either allocate a new one on the kernel heap,
2317  * or reclaim a free one.
2318  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2319  * to reclaim an existing one from the freelist.
2320  */
2321 boolean_t
2322 xfs_qm_dqalloc_incore(
2323         xfs_dquot_t **O_dqpp)
2324 {
2325         xfs_dquot_t     *dqp;
2326
2327         /*
2328          * Check against high water mark to see if we want to pop
2329          * a nincompoop dquot off the freelist.
2330          */
2331         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2332                 /*
2333                  * Try to recycle a dquot from the freelist.
2334                  */
2335                 if ((dqp = xfs_qm_dqreclaim_one())) {
2336                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2337                         /*
2338                          * Just zero the core here. The rest will get
2339                          * reinitialized by caller. XXX we shouldn't even
2340                          * do this zero ...
2341                          */
2342                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2343                         *O_dqpp = dqp;
2344                         return B_FALSE;
2345                 }
2346                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2347         }
2348
2349         /*
2350          * Allocate a brand new dquot on the kernel heap and return it
2351          * to the caller to initialize.
2352          */
2353         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2354         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2355         atomic_inc(&xfs_Gqm->qm_totaldquots);
2356
2357         return B_TRUE;
2358 }
2359
2360
2361 /*
2362  * Start a transaction and write the incore superblock changes to
2363  * disk. flags parameter indicates which fields have changed.
2364  */
2365 int
2366 xfs_qm_write_sb_changes(
2367         xfs_mount_t     *mp,
2368         __int64_t       flags)
2369 {
2370         xfs_trans_t     *tp;
2371         int             error;
2372
2373 #ifdef QUOTADEBUG
2374         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2375 #endif
2376         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2377         if ((error = xfs_trans_reserve(tp, 0,
2378                                       mp->m_sb.sb_sectsize + 128, 0,
2379                                       0,
2380                                       XFS_DEFAULT_LOG_COUNT))) {
2381                 xfs_trans_cancel(tp, 0);
2382                 return error;
2383         }
2384
2385         xfs_mod_sb(tp, flags);
2386         error = xfs_trans_commit(tp, 0);
2387
2388         return error;
2389 }
2390
2391
2392 /* --------------- utility functions for vnodeops ---------------- */
2393
2394
2395 /*
2396  * Given an inode, a uid and gid (from cred_t) make sure that we have
2397  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2398  * quotas by creating this file.
2399  * This also attaches dquot(s) to the given inode after locking it,
2400  * and returns the dquots corresponding to the uid and/or gid.
2401  *
2402  * in   : inode (unlocked)
2403  * out  : udquot, gdquot with references taken and unlocked
2404  */
2405 int
2406 xfs_qm_vop_dqalloc(
2407         xfs_mount_t     *mp,
2408         xfs_inode_t     *ip,
2409         uid_t           uid,
2410         gid_t           gid,
2411         prid_t          prid,
2412         uint            flags,
2413         xfs_dquot_t     **O_udqpp,
2414         xfs_dquot_t     **O_gdqpp)
2415 {
2416         int             error;
2417         xfs_dquot_t     *uq, *gq;
2418         uint            lockflags;
2419
2420         if (!XFS_IS_QUOTA_ON(mp))
2421                 return 0;
2422
2423         lockflags = XFS_ILOCK_EXCL;
2424         xfs_ilock(ip, lockflags);
2425
2426         if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip))
2427                 gid = ip->i_d.di_gid;
2428
2429         /*
2430          * Attach the dquot(s) to this inode, doing a dquot allocation
2431          * if necessary. The dquot(s) will not be locked.
2432          */
2433         if (XFS_NOT_DQATTACHED(mp, ip)) {
2434                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2435                                             XFS_QMOPT_ILOCKED))) {
2436                         xfs_iunlock(ip, lockflags);
2437                         return error;
2438                 }
2439         }
2440
2441         uq = gq = NULL;
2442         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2443                 if (ip->i_d.di_uid != uid) {
2444                         /*
2445                          * What we need is the dquot that has this uid, and
2446                          * if we send the inode to dqget, the uid of the inode
2447                          * takes priority over what's sent in the uid argument.
2448                          * We must unlock inode here before calling dqget if
2449                          * we're not sending the inode, because otherwise
2450                          * we'll deadlock by doing trans_reserve while
2451                          * holding ilock.
2452                          */
2453                         xfs_iunlock(ip, lockflags);
2454                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2455                                                  XFS_DQ_USER,
2456                                                  XFS_QMOPT_DQALLOC |
2457                                                  XFS_QMOPT_DOWARN,
2458                                                  &uq))) {
2459                                 ASSERT(error != ENOENT);
2460                                 return error;
2461                         }
2462                         /*
2463                          * Get the ilock in the right order.
2464                          */
2465                         xfs_dqunlock(uq);
2466                         lockflags = XFS_ILOCK_SHARED;
2467                         xfs_ilock(ip, lockflags);
2468                 } else {
2469                         /*
2470                          * Take an extra reference, because we'll return
2471                          * this to caller
2472                          */
2473                         ASSERT(ip->i_udquot);
2474                         uq = ip->i_udquot;
2475                         xfs_dqlock(uq);
2476                         XFS_DQHOLD(uq);
2477                         xfs_dqunlock(uq);
2478                 }
2479         }
2480         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2481                 if (ip->i_d.di_gid != gid) {
2482                         xfs_iunlock(ip, lockflags);
2483                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2484                                                  XFS_DQ_GROUP,
2485                                                  XFS_QMOPT_DQALLOC |
2486                                                  XFS_QMOPT_DOWARN,
2487                                                  &gq))) {
2488                                 if (uq)
2489                                         xfs_qm_dqrele(uq);
2490                                 ASSERT(error != ENOENT);
2491                                 return error;
2492                         }
2493                         xfs_dqunlock(gq);
2494                         lockflags = XFS_ILOCK_SHARED;
2495                         xfs_ilock(ip, lockflags);
2496                 } else {
2497                         ASSERT(ip->i_gdquot);
2498                         gq = ip->i_gdquot;
2499                         xfs_dqlock(gq);
2500                         XFS_DQHOLD(gq);
2501                         xfs_dqunlock(gq);
2502                 }
2503         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2504                 if (ip->i_d.di_projid != prid) {
2505                         xfs_iunlock(ip, lockflags);
2506                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2507                                                  XFS_DQ_PROJ,
2508                                                  XFS_QMOPT_DQALLOC |
2509                                                  XFS_QMOPT_DOWARN,
2510                                                  &gq))) {
2511                                 if (uq)
2512                                         xfs_qm_dqrele(uq);
2513                                 ASSERT(error != ENOENT);
2514                                 return (error);
2515                         }
2516                         xfs_dqunlock(gq);
2517                         lockflags = XFS_ILOCK_SHARED;
2518                         xfs_ilock(ip, lockflags);
2519                 } else {
2520                         ASSERT(ip->i_gdquot);
2521                         gq = ip->i_gdquot;
2522                         xfs_dqlock(gq);
2523                         XFS_DQHOLD(gq);
2524                         xfs_dqunlock(gq);
2525                 }
2526         }
2527         if (uq)
2528                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2529
2530         xfs_iunlock(ip, lockflags);
2531         if (O_udqpp)
2532                 *O_udqpp = uq;
2533         else if (uq)
2534                 xfs_qm_dqrele(uq);
2535         if (O_gdqpp)
2536                 *O_gdqpp = gq;
2537         else if (gq)
2538                 xfs_qm_dqrele(gq);
2539         return 0;
2540 }
2541
2542 /*
2543  * Actually transfer ownership, and do dquot modifications.
2544  * These were already reserved.
2545  */
2546 xfs_dquot_t *
2547 xfs_qm_vop_chown(
2548         xfs_trans_t     *tp,
2549         xfs_inode_t     *ip,
2550         xfs_dquot_t     **IO_olddq,
2551         xfs_dquot_t     *newdq)
2552 {
2553         xfs_dquot_t     *prevdq;
2554         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2555                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2556
2557         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2558         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2559
2560         /* old dquot */
2561         prevdq = *IO_olddq;
2562         ASSERT(prevdq);
2563         ASSERT(prevdq != newdq);
2564
2565         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2566         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2567
2568         /* the sparkling new dquot */
2569         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2570         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2571
2572         /*
2573          * Take an extra reference, because the inode
2574          * is going to keep this dquot pointer even
2575          * after the trans_commit.
2576          */
2577         xfs_dqlock(newdq);
2578         XFS_DQHOLD(newdq);
2579         xfs_dqunlock(newdq);
2580         *IO_olddq = newdq;
2581
2582         return prevdq;
2583 }
2584
2585 /*
2586  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2587  */
2588 int
2589 xfs_qm_vop_chown_reserve(
2590         xfs_trans_t     *tp,
2591         xfs_inode_t     *ip,
2592         xfs_dquot_t     *udqp,
2593         xfs_dquot_t     *gdqp,
2594         uint            flags)
2595 {
2596         int             error;
2597         xfs_mount_t     *mp;
2598         uint            delblks, blkflags, prjflags = 0;
2599         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2600
2601         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2602         mp = ip->i_mount;
2603         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2604
2605         delblks = ip->i_delayed_blks;
2606         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2607         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2608                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2609
2610         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2611             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2612                 delblksudq = udqp;
2613                 /*
2614                  * If there are delayed allocation blocks, then we have to
2615                  * unreserve those from the old dquot, and add them to the
2616                  * new dquot.
2617                  */
2618                 if (delblks) {
2619                         ASSERT(ip->i_udquot);
2620                         unresudq = ip->i_udquot;
2621                 }
2622         }
2623         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2624                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2625                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2626                         prjflags = XFS_QMOPT_ENOSPC;
2627
2628                 if (prjflags ||
2629                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2630                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2631                         delblksgdq = gdqp;
2632                         if (delblks) {
2633                                 ASSERT(ip->i_gdquot);
2634                                 unresgdq = ip->i_gdquot;
2635                         }
2636                 }
2637         }
2638
2639         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2640                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2641                                 flags | blkflags | prjflags)))
2642                 return (error);
2643
2644         /*
2645          * Do the delayed blks reservations/unreservations now. Since, these
2646          * are done without the help of a transaction, if a reservation fails
2647          * its previous reservations won't be automatically undone by trans
2648          * code. So, we have to do it manually here.
2649          */
2650         if (delblks) {
2651                 /*
2652                  * Do the reservations first. Unreservation can't fail.
2653                  */
2654                 ASSERT(delblksudq || delblksgdq);
2655                 ASSERT(unresudq || unresgdq);
2656                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2657                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2658                                 flags | blkflags | prjflags)))
2659                         return (error);
2660                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2661                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2662                                 blkflags);
2663         }
2664
2665         return (0);
2666 }
2667
2668 int
2669 xfs_qm_vop_rename_dqattach(
2670         xfs_inode_t     **i_tab)
2671 {
2672         xfs_inode_t     *ip;
2673         int             i;
2674         int             error;
2675
2676         ip = i_tab[0];
2677
2678         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2679                 return 0;
2680
2681         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2682                 error = xfs_qm_dqattach(ip, 0);
2683                 if (error)
2684                         return error;
2685         }
2686         for (i = 1; (i < 4 && i_tab[i]); i++) {
2687                 /*
2688                  * Watch out for duplicate entries in the table.
2689                  */
2690                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2691                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2692                                 error = xfs_qm_dqattach(ip, 0);
2693                                 if (error)
2694                                         return error;
2695                         }
2696                 }
2697         }
2698         return 0;
2699 }
2700
2701 void
2702 xfs_qm_vop_dqattach_and_dqmod_newinode(
2703         xfs_trans_t     *tp,
2704         xfs_inode_t     *ip,
2705         xfs_dquot_t     *udqp,
2706         xfs_dquot_t     *gdqp)
2707 {
2708         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2709                 return;
2710
2711         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2712         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2713
2714         if (udqp) {
2715                 xfs_dqlock(udqp);
2716                 XFS_DQHOLD(udqp);
2717                 xfs_dqunlock(udqp);
2718                 ASSERT(ip->i_udquot == NULL);
2719                 ip->i_udquot = udqp;
2720                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2721                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2722                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2723         }
2724         if (gdqp) {
2725                 xfs_dqlock(gdqp);
2726                 XFS_DQHOLD(gdqp);
2727                 xfs_dqunlock(gdqp);
2728                 ASSERT(ip->i_gdquot == NULL);
2729                 ip->i_gdquot = gdqp;
2730                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2731                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2732                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2733                                 be32_to_cpu(gdqp->q_core.d_id));
2734                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2735         }
2736 }
2737
2738 /* ------------- list stuff -----------------*/
2739 STATIC void
2740 xfs_qm_freelist_init(xfs_frlist_t *ql)
2741 {
2742         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2743         mutex_init(&ql->qh_lock);
2744         ql->qh_version = 0;
2745         ql->qh_nelems = 0;
2746 }
2747
2748 STATIC void
2749 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2750 {
2751         xfs_dquot_t     *dqp, *nextdqp;
2752
2753         mutex_lock(&ql->qh_lock);
2754         for (dqp = ql->qh_next;
2755              dqp != (xfs_dquot_t *)ql; ) {
2756                 xfs_dqlock(dqp);
2757                 nextdqp = dqp->dq_flnext;
2758 #ifdef QUOTADEBUG
2759                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2760 #endif
2761                 XQM_FREELIST_REMOVE(dqp);
2762                 xfs_dqunlock(dqp);
2763                 xfs_qm_dqdestroy(dqp);
2764                 dqp = nextdqp;
2765         }
2766         mutex_unlock(&ql->qh_lock);
2767         mutex_destroy(&ql->qh_lock);
2768
2769         ASSERT(ql->qh_nelems == 0);
2770 }
2771
2772 STATIC void
2773 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2774 {
2775         dq->dq_flnext = ql->qh_next;
2776         dq->dq_flprev = (xfs_dquot_t *)ql;
2777         ql->qh_next = dq;
2778         dq->dq_flnext->dq_flprev = dq;
2779         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2780         xfs_Gqm->qm_dqfreelist.qh_version++;
2781 }
2782
2783 void
2784 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2785 {
2786         xfs_dquot_t *next = dq->dq_flnext;
2787         xfs_dquot_t *prev = dq->dq_flprev;
2788
2789         next->dq_flprev = prev;
2790         prev->dq_flnext = next;
2791         dq->dq_flnext = dq->dq_flprev = dq;
2792         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2793         xfs_Gqm->qm_dqfreelist.qh_version++;
2794 }
2795
2796 void
2797 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2798 {
2799         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2800 }
2801
2802 STATIC int
2803 xfs_qm_dqhashlock_nowait(
2804         xfs_dquot_t *dqp)
2805 {
2806         int locked;
2807
2808         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2809         return locked;
2810 }
2811
2812 int
2813 xfs_qm_freelist_lock_nowait(
2814         xfs_qm_t *xqm)
2815 {
2816         int locked;
2817
2818         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2819         return locked;
2820 }
2821
2822 STATIC int
2823 xfs_qm_mplist_nowait(
2824         xfs_mount_t     *mp)
2825 {
2826         int locked;
2827
2828         ASSERT(mp->m_quotainfo);
2829         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2830         return locked;
2831 }