Merge git://git.kernel.org/pub/scm/linux/kernel/git/bunk/trivial
[linux-2.6.git] / 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_dir.h"
28 #include "xfs_dir2.h"
29 #include "xfs_alloc.h"
30 #include "xfs_dmapi.h"
31 #include "xfs_quota.h"
32 #include "xfs_mount.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_dir_sf.h"
37 #include "xfs_dir2_sf.h"
38 #include "xfs_attr_sf.h"
39 #include "xfs_dinode.h"
40 #include "xfs_inode.h"
41 #include "xfs_btree.h"
42 #include "xfs_ialloc.h"
43 #include "xfs_itable.h"
44 #include "xfs_rtalloc.h"
45 #include "xfs_error.h"
46 #include "xfs_bmap.h"
47 #include "xfs_rw.h"
48 #include "xfs_acl.h"
49 #include "xfs_cap.h"
50 #include "xfs_mac.h"
51 #include "xfs_attr.h"
52 #include "xfs_buf_item.h"
53 #include "xfs_trans_space.h"
54 #include "xfs_utils.h"
55 #include "xfs_qm.h"
56
57 /*
58  * The global quota manager. There is only one of these for the entire
59  * system, _not_ one per file system. XQM keeps track of the overall
60  * quota functionality, including maintaining the freelist and hash
61  * tables of dquots.
62  */
63 mutex_t         xfs_Gqm_lock;
64 struct xfs_qm   *xfs_Gqm;
65 uint            ndquot;
66
67 kmem_zone_t     *qm_dqzone;
68 kmem_zone_t     *qm_dqtrxzone;
69 STATIC kmem_shaker_t    xfs_qm_shaker;
70
71 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
72 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
73
74 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
75 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
76 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
77 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
78
79 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
80 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
81 STATIC int      xfs_qm_shake(int, gfp_t);
82
83 #ifdef DEBUG
84 extern mutex_t  qcheck_lock;
85 #endif
86
87 #ifdef QUOTADEBUG
88 #define XQM_LIST_PRINT(l, NXT, title) \
89 { \
90         xfs_dquot_t     *dqp; int i = 0; \
91         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
92         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
93                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
94                                   "bcnt = %d, icnt = %d, refs = %d", \
95                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
96                         DQFLAGTO_TYPESTR(dqp),       \
97                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
98                         (int) be64_to_cpu(dqp->q_core.d_icount), \
99                         (int) dqp->q_nrefs);  } \
100 }
101 #else
102 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
103 #endif
104
105 /*
106  * Initialize the XQM structure.
107  * Note that there is not one quota manager per file system.
108  */
109 STATIC struct xfs_qm *
110 xfs_Gqm_init(void)
111 {
112         xfs_dqhash_t    *udqhash, *gdqhash;
113         xfs_qm_t        *xqm;
114         uint            i, hsize, flags = KM_SLEEP | KM_MAYFAIL;
115
116         /*
117          * Initialize the dquot hash tables.
118          */
119         hsize = XFS_QM_HASHSIZE_HIGH;
120         while (!(udqhash = kmem_zalloc(hsize * sizeof(xfs_dqhash_t), flags))) {
121                 if ((hsize >>= 1) <= XFS_QM_HASHSIZE_LOW)
122                         flags = KM_SLEEP;
123         }
124         gdqhash = kmem_zalloc(hsize * sizeof(xfs_dqhash_t), KM_SLEEP);
125         ndquot = hsize << 8;
126
127         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
128         xqm->qm_dqhashmask = hsize - 1;
129         xqm->qm_usr_dqhtable = udqhash;
130         xqm->qm_grp_dqhtable = gdqhash;
131         ASSERT(xqm->qm_usr_dqhtable != NULL);
132         ASSERT(xqm->qm_grp_dqhtable != NULL);
133
134         for (i = 0; i < hsize; i++) {
135                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
136                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
137         }
138
139         /*
140          * Freelist of all dquots of all file systems
141          */
142         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
143
144         /*
145          * dquot zone. we register our own low-memory callback.
146          */
147         if (!qm_dqzone) {
148                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
149                                                 "xfs_dquots");
150                 qm_dqzone = xqm->qm_dqzone;
151         } else
152                 xqm->qm_dqzone = qm_dqzone;
153
154         xfs_qm_shaker = kmem_shake_register(xfs_qm_shake);
155
156         /*
157          * The t_dqinfo portion of transactions.
158          */
159         if (!qm_dqtrxzone) {
160                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
161                                                    "xfs_dqtrx");
162                 qm_dqtrxzone = xqm->qm_dqtrxzone;
163         } else
164                 xqm->qm_dqtrxzone = qm_dqtrxzone;
165
166         atomic_set(&xqm->qm_totaldquots, 0);
167         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
168         xqm->qm_nrefs = 0;
169 #ifdef DEBUG
170         xfs_mutex_init(&qcheck_lock, MUTEX_DEFAULT, "qchk");
171 #endif
172         return xqm;
173 }
174
175 /*
176  * Destroy the global quota manager when its reference count goes to zero.
177  */
178 STATIC void
179 xfs_qm_destroy(
180         struct xfs_qm   *xqm)
181 {
182         int             hsize, i;
183
184         ASSERT(xqm != NULL);
185         ASSERT(xqm->qm_nrefs == 0);
186         kmem_shake_deregister(xfs_qm_shaker);
187         hsize = xqm->qm_dqhashmask + 1;
188         for (i = 0; i < hsize; i++) {
189                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
190                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
191         }
192         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
193         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
194         xqm->qm_usr_dqhtable = NULL;
195         xqm->qm_grp_dqhtable = NULL;
196         xqm->qm_dqhashmask = 0;
197         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
198 #ifdef DEBUG
199         mutex_destroy(&qcheck_lock);
200 #endif
201         kmem_free(xqm, sizeof(xfs_qm_t));
202 }
203
204 /*
205  * Called at mount time to let XQM know that another file system is
206  * starting quotas. This isn't crucial information as the individual mount
207  * structures are pretty independent, but it helps the XQM keep a
208  * global view of what's going on.
209  */
210 /* ARGSUSED */
211 STATIC int
212 xfs_qm_hold_quotafs_ref(
213         struct xfs_mount *mp)
214 {
215         /*
216          * Need to lock the xfs_Gqm structure for things like this. For example,
217          * the structure could disappear between the entry to this routine and
218          * a HOLD operation if not locked.
219          */
220         XFS_QM_LOCK(xfs_Gqm);
221
222         if (xfs_Gqm == NULL)
223                 xfs_Gqm = xfs_Gqm_init();
224         /*
225          * We can keep a list of all filesystems with quotas mounted for
226          * debugging and statistical purposes, but ...
227          * Just take a reference and get out.
228          */
229         XFS_QM_HOLD(xfs_Gqm);
230         XFS_QM_UNLOCK(xfs_Gqm);
231
232         return 0;
233 }
234
235
236 /*
237  * Release the reference that a filesystem took at mount time,
238  * so that we know when we need to destroy the entire quota manager.
239  */
240 /* ARGSUSED */
241 STATIC void
242 xfs_qm_rele_quotafs_ref(
243         struct xfs_mount *mp)
244 {
245         xfs_dquot_t     *dqp, *nextdqp;
246
247         ASSERT(xfs_Gqm);
248         ASSERT(xfs_Gqm->qm_nrefs > 0);
249
250         /*
251          * Go thru the freelist and destroy all inactive dquots.
252          */
253         xfs_qm_freelist_lock(xfs_Gqm);
254
255         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
256              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
257                 xfs_dqlock(dqp);
258                 nextdqp = dqp->dq_flnext;
259                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
260                         ASSERT(dqp->q_mount == NULL);
261                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
262                         ASSERT(dqp->HL_PREVP == NULL);
263                         ASSERT(dqp->MPL_PREVP == NULL);
264                         XQM_FREELIST_REMOVE(dqp);
265                         xfs_dqunlock(dqp);
266                         xfs_qm_dqdestroy(dqp);
267                 } else {
268                         xfs_dqunlock(dqp);
269                 }
270                 dqp = nextdqp;
271         }
272         xfs_qm_freelist_unlock(xfs_Gqm);
273
274         /*
275          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
276          * be restarted.
277          */
278         XFS_QM_LOCK(xfs_Gqm);
279         XFS_QM_RELE(xfs_Gqm);
280         if (xfs_Gqm->qm_nrefs == 0) {
281                 xfs_qm_destroy(xfs_Gqm);
282                 xfs_Gqm = NULL;
283         }
284         XFS_QM_UNLOCK(xfs_Gqm);
285 }
286
287 /*
288  * This is called at mount time from xfs_mountfs to initialize the quotainfo
289  * structure and start the global quotamanager (xfs_Gqm) if it hasn't done
290  * so already.  Note that the superblock has not been read in yet.
291  */
292 void
293 xfs_qm_mount_quotainit(
294         xfs_mount_t     *mp,
295         uint            flags)
296 {
297         /*
298          * User, projects or group quotas has to be on.
299          */
300         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_PQUOTA | XFSMNT_GQUOTA));
301
302         /*
303          * Initialize the flags in the mount structure. From this point
304          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
305          * Note that we enforce nothing if accounting is off.
306          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
307          * It isn't necessary to take the quotaoff lock to do this; this is
308          * called from mount.
309          */
310         if (flags & XFSMNT_UQUOTA) {
311                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
312                 if (flags & XFSMNT_UQUOTAENF)
313                         mp->m_qflags |= XFS_UQUOTA_ENFD;
314         }
315         if (flags & XFSMNT_GQUOTA) {
316                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
317                 if (flags & XFSMNT_GQUOTAENF)
318                         mp->m_qflags |= XFS_OQUOTA_ENFD;
319         } else if (flags & XFSMNT_PQUOTA) {
320                 mp->m_qflags |= (XFS_PQUOTA_ACCT | XFS_PQUOTA_ACTIVE);
321                 if (flags & XFSMNT_PQUOTAENF)
322                         mp->m_qflags |= XFS_OQUOTA_ENFD;
323         }
324 }
325
326 /*
327  * Just destroy the quotainfo structure.
328  */
329 void
330 xfs_qm_unmount_quotadestroy(
331         xfs_mount_t     *mp)
332 {
333         if (mp->m_quotainfo)
334                 xfs_qm_destroy_quotainfo(mp);
335 }
336
337
338 /*
339  * This is called from xfs_mountfs to start quotas and initialize all
340  * necessary data structures like quotainfo.  This is also responsible for
341  * running a quotacheck as necessary.  We are guaranteed that the superblock
342  * is consistently read in at this point.
343  */
344 int
345 xfs_qm_mount_quotas(
346         xfs_mount_t     *mp,
347         int             mfsi_flags)
348 {
349         unsigned long   s;
350         int             error = 0;
351         uint            sbf;
352
353
354         /*
355          * If quotas on realtime volumes is not supported, we disable
356          * quotas immediately.
357          */
358         if (mp->m_sb.sb_rextents) {
359                 cmn_err(CE_NOTE,
360                         "Cannot turn on quotas for realtime filesystem %s",
361                         mp->m_fsname);
362                 mp->m_qflags = 0;
363                 goto write_changes;
364         }
365
366         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
367
368         /*
369          * Allocate the quotainfo structure inside the mount struct, and
370          * create quotainode(s), and change/rev superblock if necessary.
371          */
372         if ((error = xfs_qm_init_quotainfo(mp))) {
373                 /*
374                  * We must turn off quotas.
375                  */
376                 ASSERT(mp->m_quotainfo == NULL);
377                 mp->m_qflags = 0;
378                 goto write_changes;
379         }
380         /*
381          * If any of the quotas are not consistent, do a quotacheck.
382          */
383         if (XFS_QM_NEED_QUOTACHECK(mp) &&
384                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
385                 if ((error = xfs_qm_quotacheck(mp))) {
386                         /* Quotacheck has failed and quotas have
387                          * been disabled.
388                          */
389                         return XFS_ERROR(error);
390                 }
391         }
392
393  write_changes:
394         /*
395          * We actually don't have to acquire the SB_LOCK at all.
396          * This can only be called from mount, and that's single threaded. XXX
397          */
398         s = XFS_SB_LOCK(mp);
399         sbf = mp->m_sb.sb_qflags;
400         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
401         XFS_SB_UNLOCK(mp, s);
402
403         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
404                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
405                         /*
406                          * We could only have been turning quotas off.
407                          * We aren't in very good shape actually because
408                          * the incore structures are convinced that quotas are
409                          * off, but the on disk superblock doesn't know that !
410                          */
411                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
412                         xfs_fs_cmn_err(CE_ALERT, mp,
413                                 "XFS mount_quotas: Superblock update failed!");
414                 }
415         }
416
417         if (error) {
418                 xfs_fs_cmn_err(CE_WARN, mp,
419                         "Failed to initialize disk quotas.");
420         }
421         return XFS_ERROR(error);
422 }
423
424 /*
425  * Called from the vfsops layer.
426  */
427 int
428 xfs_qm_unmount_quotas(
429         xfs_mount_t     *mp)
430 {
431         xfs_inode_t     *uqp, *gqp;
432         int             error = 0;
433
434         /*
435          * Release the dquots that root inode, et al might be holding,
436          * before we flush quotas and blow away the quotainfo structure.
437          */
438         ASSERT(mp->m_rootip);
439         xfs_qm_dqdetach(mp->m_rootip);
440         if (mp->m_rbmip)
441                 xfs_qm_dqdetach(mp->m_rbmip);
442         if (mp->m_rsumip)
443                 xfs_qm_dqdetach(mp->m_rsumip);
444
445         /*
446          * Flush out the quota inodes.
447          */
448         uqp = gqp = NULL;
449         if (mp->m_quotainfo) {
450                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
451                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
452                         xfs_iflock(uqp);
453                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
454                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
455                         if (unlikely(error == EFSCORRUPTED)) {
456                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
457                                                  XFS_ERRLEVEL_LOW, mp);
458                                 goto out;
459                         }
460                 }
461                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
462                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
463                         xfs_iflock(gqp);
464                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
465                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
466                         if (unlikely(error == EFSCORRUPTED)) {
467                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
468                                                  XFS_ERRLEVEL_LOW, mp);
469                                 goto out;
470                         }
471                 }
472         }
473         if (uqp) {
474                  XFS_PURGE_INODE(uqp);
475                  mp->m_quotainfo->qi_uquotaip = NULL;
476         }
477         if (gqp) {
478                 XFS_PURGE_INODE(gqp);
479                 mp->m_quotainfo->qi_gquotaip = NULL;
480         }
481 out:
482         return XFS_ERROR(error);
483 }
484
485 /*
486  * Flush all dquots of the given file system to disk. The dquots are
487  * _not_ purged from memory here, just their data written to disk.
488  */
489 STATIC int
490 xfs_qm_dqflush_all(
491         xfs_mount_t     *mp,
492         int             flags)
493 {
494         int             recl;
495         xfs_dquot_t     *dqp;
496         int             niters;
497         int             error;
498
499         if (mp->m_quotainfo == NULL)
500                 return 0;
501         niters = 0;
502 again:
503         xfs_qm_mplist_lock(mp);
504         FOREACH_DQUOT_IN_MP(dqp, mp) {
505                 xfs_dqlock(dqp);
506                 if (! XFS_DQ_IS_DIRTY(dqp)) {
507                         xfs_dqunlock(dqp);
508                         continue;
509                 }
510                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
511                 /* XXX a sentinel would be better */
512                 recl = XFS_QI_MPLRECLAIMS(mp);
513                 if (! xfs_qm_dqflock_nowait(dqp)) {
514                         /*
515                          * If we can't grab the flush lock then check
516                          * to see if the dquot has been flushed delayed
517                          * write.  If so, grab its buffer and send it
518                          * out immediately.  We'll be able to acquire
519                          * the flush lock when the I/O completes.
520                          */
521                         xfs_qm_dqflock_pushbuf_wait(dqp);
522                 }
523                 /*
524                  * Let go of the mplist lock. We don't want to hold it
525                  * across a disk write.
526                  */
527                 xfs_qm_mplist_unlock(mp);
528                 error = xfs_qm_dqflush(dqp, flags);
529                 xfs_dqunlock(dqp);
530                 if (error)
531                         return error;
532
533                 xfs_qm_mplist_lock(mp);
534                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
535                         xfs_qm_mplist_unlock(mp);
536                         /* XXX restart limit */
537                         goto again;
538                 }
539         }
540
541         xfs_qm_mplist_unlock(mp);
542         /* return ! busy */
543         return 0;
544 }
545 /*
546  * Release the group dquot pointers the user dquots may be
547  * carrying around as a hint. mplist is locked on entry and exit.
548  */
549 STATIC void
550 xfs_qm_detach_gdquots(
551         xfs_mount_t     *mp)
552 {
553         xfs_dquot_t     *dqp, *gdqp;
554         int             nrecl;
555
556  again:
557         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
558         dqp = XFS_QI_MPLNEXT(mp);
559         while (dqp) {
560                 xfs_dqlock(dqp);
561                 if ((gdqp = dqp->q_gdquot)) {
562                         xfs_dqlock(gdqp);
563                         dqp->q_gdquot = NULL;
564                 }
565                 xfs_dqunlock(dqp);
566
567                 if (gdqp) {
568                         /*
569                          * Can't hold the mplist lock across a dqput.
570                          * XXXmust convert to marker based iterations here.
571                          */
572                         nrecl = XFS_QI_MPLRECLAIMS(mp);
573                         xfs_qm_mplist_unlock(mp);
574                         xfs_qm_dqput(gdqp);
575
576                         xfs_qm_mplist_lock(mp);
577                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
578                                 goto again;
579                 }
580                 dqp = dqp->MPL_NEXT;
581         }
582 }
583
584 /*
585  * Go through all the incore dquots of this file system and take them
586  * off the mplist and hashlist, if the dquot type matches the dqtype
587  * parameter. This is used when turning off quota accounting for
588  * users and/or groups, as well as when the filesystem is unmounting.
589  */
590 STATIC int
591 xfs_qm_dqpurge_int(
592         xfs_mount_t     *mp,
593         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
594 {
595         xfs_dquot_t     *dqp;
596         uint            dqtype;
597         int             nrecl;
598         xfs_dquot_t     *nextdqp;
599         int             nmisses;
600
601         if (mp->m_quotainfo == NULL)
602                 return 0;
603
604         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
605         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
606         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
607
608         xfs_qm_mplist_lock(mp);
609
610         /*
611          * In the first pass through all incore dquots of this filesystem,
612          * we release the group dquot pointers the user dquots may be
613          * carrying around as a hint. We need to do this irrespective of
614          * what's being turned off.
615          */
616         xfs_qm_detach_gdquots(mp);
617
618       again:
619         nmisses = 0;
620         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
621         /*
622          * Try to get rid of all of the unwanted dquots. The idea is to
623          * get them off mplist and hashlist, but leave them on freelist.
624          */
625         dqp = XFS_QI_MPLNEXT(mp);
626         while (dqp) {
627                 /*
628                  * It's OK to look at the type without taking dqlock here.
629                  * We're holding the mplist lock here, and that's needed for
630                  * a dqreclaim.
631                  */
632                 if ((dqp->dq_flags & dqtype) == 0) {
633                         dqp = dqp->MPL_NEXT;
634                         continue;
635                 }
636
637                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
638                         nrecl = XFS_QI_MPLRECLAIMS(mp);
639                         xfs_qm_mplist_unlock(mp);
640                         XFS_DQ_HASH_LOCK(dqp->q_hash);
641                         xfs_qm_mplist_lock(mp);
642
643                         /*
644                          * XXXTheoretically, we can get into a very long
645                          * ping pong game here.
646                          * No one can be adding dquots to the mplist at
647                          * this point, but somebody might be taking things off.
648                          */
649                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
650                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
651                                 goto again;
652                         }
653                 }
654
655                 /*
656                  * Take the dquot off the mplist and hashlist. It may remain on
657                  * freelist in INACTIVE state.
658                  */
659                 nextdqp = dqp->MPL_NEXT;
660                 nmisses += xfs_qm_dqpurge(dqp, flags);
661                 dqp = nextdqp;
662         }
663         xfs_qm_mplist_unlock(mp);
664         return nmisses;
665 }
666
667 int
668 xfs_qm_dqpurge_all(
669         xfs_mount_t     *mp,
670         uint            flags)
671 {
672         int             ndquots;
673
674         /*
675          * Purge the dquot cache.
676          * None of the dquots should really be busy at this point.
677          */
678         if (mp->m_quotainfo) {
679                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
680                         delay(ndquots * 10);
681                 }
682         }
683         return 0;
684 }
685
686 STATIC int
687 xfs_qm_dqattach_one(
688         xfs_inode_t     *ip,
689         xfs_dqid_t      id,
690         uint            type,
691         uint            doalloc,
692         uint            dolock,
693         xfs_dquot_t     *udqhint, /* hint */
694         xfs_dquot_t     **IO_idqpp)
695 {
696         xfs_dquot_t     *dqp;
697         int             error;
698
699         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
700         error = 0;
701         /*
702          * See if we already have it in the inode itself. IO_idqpp is
703          * &i_udquot or &i_gdquot. This made the code look weird, but
704          * made the logic a lot simpler.
705          */
706         if ((dqp = *IO_idqpp)) {
707                 if (dolock)
708                         xfs_dqlock(dqp);
709                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
710                 goto done;
711         }
712
713         /*
714          * udqhint is the i_udquot field in inode, and is non-NULL only
715          * when the type arg is group/project. Its purpose is to save a
716          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
717          * the user dquot.
718          */
719         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
720         if (udqhint && !dolock)
721                 xfs_dqlock(udqhint);
722
723         /*
724          * No need to take dqlock to look at the id.
725          * The ID can't change until it gets reclaimed, and it won't
726          * be reclaimed as long as we have a ref from inode and we hold
727          * the ilock.
728          */
729         if (udqhint &&
730             (dqp = udqhint->q_gdquot) &&
731             (be32_to_cpu(dqp->q_core.d_id) == id)) {
732                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
733                 xfs_dqlock(dqp);
734                 XFS_DQHOLD(dqp);
735                 ASSERT(*IO_idqpp == NULL);
736                 *IO_idqpp = dqp;
737                 if (!dolock) {
738                         xfs_dqunlock(dqp);
739                         xfs_dqunlock(udqhint);
740                 }
741                 goto done;
742         }
743         /*
744          * We can't hold a dquot lock when we call the dqget code.
745          * We'll deadlock in no time, because of (not conforming to)
746          * lock ordering - the inodelock comes before any dquot lock,
747          * and we may drop and reacquire the ilock in xfs_qm_dqget().
748          */
749         if (udqhint)
750                 xfs_dqunlock(udqhint);
751         /*
752          * Find the dquot from somewhere. This bumps the
753          * reference count of dquot and returns it locked.
754          * This can return ENOENT if dquot didn't exist on
755          * disk and we didn't ask it to allocate;
756          * ESRCH if quotas got turned off suddenly.
757          */
758         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
759                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
760                 if (udqhint && dolock)
761                         xfs_dqlock(udqhint);
762                 goto done;
763         }
764
765         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
766         /*
767          * dqget may have dropped and re-acquired the ilock, but it guarantees
768          * that the dquot returned is the one that should go in the inode.
769          */
770         *IO_idqpp = dqp;
771         ASSERT(dqp);
772         ASSERT(XFS_DQ_IS_LOCKED(dqp));
773         if (! dolock) {
774                 xfs_dqunlock(dqp);
775                 goto done;
776         }
777         if (! udqhint)
778                 goto done;
779
780         ASSERT(udqhint);
781         ASSERT(dolock);
782         ASSERT(XFS_DQ_IS_LOCKED(dqp));
783         if (! xfs_qm_dqlock_nowait(udqhint)) {
784                 xfs_dqunlock(dqp);
785                 xfs_dqlock(udqhint);
786                 xfs_dqlock(dqp);
787         }
788       done:
789 #ifdef QUOTADEBUG
790         if (udqhint) {
791                 if (dolock)
792                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
793         }
794         if (! error) {
795                 if (dolock)
796                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
797         }
798 #endif
799         return error;
800 }
801
802
803 /*
804  * Given a udquot and gdquot, attach a ptr to the group dquot in the
805  * udquot as a hint for future lookups. The idea sounds simple, but the
806  * execution isn't, because the udquot might have a group dquot attached
807  * already and getting rid of that gets us into lock ordering contraints.
808  * The process is complicated more by the fact that the dquots may or may not
809  * be locked on entry.
810  */
811 STATIC void
812 xfs_qm_dqattach_grouphint(
813         xfs_dquot_t     *udq,
814         xfs_dquot_t     *gdq,
815         uint            locked)
816 {
817         xfs_dquot_t     *tmp;
818
819 #ifdef QUOTADEBUG
820         if (locked) {
821                 ASSERT(XFS_DQ_IS_LOCKED(udq));
822                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
823         }
824 #endif
825         if (! locked)
826                 xfs_dqlock(udq);
827
828         if ((tmp = udq->q_gdquot)) {
829                 if (tmp == gdq) {
830                         if (! locked)
831                                 xfs_dqunlock(udq);
832                         return;
833                 }
834
835                 udq->q_gdquot = NULL;
836                 /*
837                  * We can't keep any dqlocks when calling dqrele,
838                  * because the freelist lock comes before dqlocks.
839                  */
840                 xfs_dqunlock(udq);
841                 if (locked)
842                         xfs_dqunlock(gdq);
843                 /*
844                  * we took a hard reference once upon a time in dqget,
845                  * so give it back when the udquot no longer points at it
846                  * dqput() does the unlocking of the dquot.
847                  */
848                 xfs_qm_dqrele(tmp);
849
850                 xfs_dqlock(udq);
851                 xfs_dqlock(gdq);
852
853         } else {
854                 ASSERT(XFS_DQ_IS_LOCKED(udq));
855                 if (! locked) {
856                         xfs_dqlock(gdq);
857                 }
858         }
859
860         ASSERT(XFS_DQ_IS_LOCKED(udq));
861         ASSERT(XFS_DQ_IS_LOCKED(gdq));
862         /*
863          * Somebody could have attached a gdquot here,
864          * when we dropped the uqlock. If so, just do nothing.
865          */
866         if (udq->q_gdquot == NULL) {
867                 XFS_DQHOLD(gdq);
868                 udq->q_gdquot = gdq;
869         }
870         if (! locked) {
871                 xfs_dqunlock(gdq);
872                 xfs_dqunlock(udq);
873         }
874 }
875
876
877 /*
878  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
879  * into account.
880  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
881  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
882  * much made this code a complete mess, but it has been pretty useful.
883  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
884  * Inode may get unlocked and relocked in here, and the caller must deal with
885  * the consequences.
886  */
887 int
888 xfs_qm_dqattach(
889         xfs_inode_t     *ip,
890         uint            flags)
891 {
892         xfs_mount_t     *mp = ip->i_mount;
893         uint            nquotas = 0;
894         int             error = 0;
895
896         if ((! XFS_IS_QUOTA_ON(mp)) ||
897             (! XFS_NOT_DQATTACHED(mp, ip)) ||
898             (ip->i_ino == mp->m_sb.sb_uquotino) ||
899             (ip->i_ino == mp->m_sb.sb_gquotino))
900                 return 0;
901
902         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
903                XFS_ISLOCKED_INODE_EXCL(ip));
904
905         if (! (flags & XFS_QMOPT_ILOCKED))
906                 xfs_ilock(ip, XFS_ILOCK_EXCL);
907
908         if (XFS_IS_UQUOTA_ON(mp)) {
909                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
910                                                 flags & XFS_QMOPT_DQALLOC,
911                                                 flags & XFS_QMOPT_DQLOCK,
912                                                 NULL, &ip->i_udquot);
913                 if (error)
914                         goto done;
915                 nquotas++;
916         }
917         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
918         if (XFS_IS_OQUOTA_ON(mp)) {
919                 error = XFS_IS_GQUOTA_ON(mp) ?
920                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
921                                                 flags & XFS_QMOPT_DQALLOC,
922                                                 flags & XFS_QMOPT_DQLOCK,
923                                                 ip->i_udquot, &ip->i_gdquot) :
924                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
925                                                 flags & XFS_QMOPT_DQALLOC,
926                                                 flags & XFS_QMOPT_DQLOCK,
927                                                 ip->i_udquot, &ip->i_gdquot);
928                 /*
929                  * Don't worry about the udquot that we may have
930                  * attached above. It'll get detached, if not already.
931                  */
932                 if (error)
933                         goto done;
934                 nquotas++;
935         }
936
937         /*
938          * Attach this group quota to the user quota as a hint.
939          * This WON'T, in general, result in a thrash.
940          */
941         if (nquotas == 2) {
942                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
943                 ASSERT(ip->i_udquot);
944                 ASSERT(ip->i_gdquot);
945
946                 /*
947                  * We may or may not have the i_udquot locked at this point,
948                  * but this check is OK since we don't depend on the i_gdquot to
949                  * be accurate 100% all the time. It is just a hint, and this
950                  * will succeed in general.
951                  */
952                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
953                         goto done;
954                 /*
955                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
956                  */
957                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
958                                          flags & XFS_QMOPT_DQLOCK);
959         }
960
961       done:
962
963 #ifdef QUOTADEBUG
964         if (! error) {
965                 if (ip->i_udquot) {
966                         if (flags & XFS_QMOPT_DQLOCK)
967                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
968                 }
969                 if (ip->i_gdquot) {
970                         if (flags & XFS_QMOPT_DQLOCK)
971                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
972                 }
973                 if (XFS_IS_UQUOTA_ON(mp))
974                         ASSERT(ip->i_udquot);
975                 if (XFS_IS_OQUOTA_ON(mp))
976                         ASSERT(ip->i_gdquot);
977         }
978 #endif
979
980         if (! (flags & XFS_QMOPT_ILOCKED))
981                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
982
983 #ifdef QUOTADEBUG
984         else
985                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
986 #endif
987         return error;
988 }
989
990 /*
991  * Release dquots (and their references) if any.
992  * The inode should be locked EXCL except when this's called by
993  * xfs_ireclaim.
994  */
995 void
996 xfs_qm_dqdetach(
997         xfs_inode_t     *ip)
998 {
999         if (!(ip->i_udquot || ip->i_gdquot))
1000                 return;
1001
1002         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1003         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1004         if (ip->i_udquot) {
1005                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1006                 xfs_qm_dqrele(ip->i_udquot);
1007                 ip->i_udquot = NULL;
1008         }
1009         if (ip->i_gdquot) {
1010                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
1011                 xfs_qm_dqrele(ip->i_gdquot);
1012                 ip->i_gdquot = NULL;
1013         }
1014 }
1015
1016 /*
1017  * This is called by VFS_SYNC and flags arg determines the caller,
1018  * and its motives, as done in xfs_sync.
1019  *
1020  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1021  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1022  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1023  */
1024
1025 int
1026 xfs_qm_sync(
1027         xfs_mount_t     *mp,
1028         short           flags)
1029 {
1030         int             recl, restarts;
1031         xfs_dquot_t     *dqp;
1032         uint            flush_flags;
1033         boolean_t       nowait;
1034         int             error;
1035
1036         restarts = 0;
1037         /*
1038          * We won't block unless we are asked to.
1039          */
1040         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1041
1042   again:
1043         xfs_qm_mplist_lock(mp);
1044         /*
1045          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1046          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1047          * when we have the mplist lock, we know that dquots will be consistent
1048          * as long as we have it locked.
1049          */
1050         if (! XFS_IS_QUOTA_ON(mp)) {
1051                 xfs_qm_mplist_unlock(mp);
1052                 return 0;
1053         }
1054         FOREACH_DQUOT_IN_MP(dqp, mp) {
1055                 /*
1056                  * If this is vfs_sync calling, then skip the dquots that
1057                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1058                  * This is very similar to what xfs_sync does with inodes.
1059                  */
1060                 if (flags & SYNC_BDFLUSH) {
1061                         if (! XFS_DQ_IS_DIRTY(dqp))
1062                                 continue;
1063                 }
1064
1065                 if (nowait) {
1066                         /*
1067                          * Try to acquire the dquot lock. We are NOT out of
1068                          * lock order, but we just don't want to wait for this
1069                          * lock, unless somebody wanted us to.
1070                          */
1071                         if (! xfs_qm_dqlock_nowait(dqp))
1072                                 continue;
1073                 } else {
1074                         xfs_dqlock(dqp);
1075                 }
1076
1077                 /*
1078                  * Now, find out for sure if this dquot is dirty or not.
1079                  */
1080                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1081                         xfs_dqunlock(dqp);
1082                         continue;
1083                 }
1084
1085                 /* XXX a sentinel would be better */
1086                 recl = XFS_QI_MPLRECLAIMS(mp);
1087                 if (! xfs_qm_dqflock_nowait(dqp)) {
1088                         if (nowait) {
1089                                 xfs_dqunlock(dqp);
1090                                 continue;
1091                         }
1092                         /*
1093                          * If we can't grab the flush lock then if the caller
1094                          * really wanted us to give this our best shot,
1095                          * see if we can give a push to the buffer before we wait
1096                          * on the flush lock. At this point, we know that
1097                          * eventhough the dquot is being flushed,
1098                          * it has (new) dirty data.
1099                          */
1100                         xfs_qm_dqflock_pushbuf_wait(dqp);
1101                 }
1102                 /*
1103                  * Let go of the mplist lock. We don't want to hold it
1104                  * across a disk write
1105                  */
1106                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1107                 xfs_qm_mplist_unlock(mp);
1108                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1109                 error = xfs_qm_dqflush(dqp, flush_flags);
1110                 xfs_dqunlock(dqp);
1111                 if (error && XFS_FORCED_SHUTDOWN(mp))
1112                         return 0;       /* Need to prevent umount failure */
1113                 else if (error)
1114                         return error;
1115
1116                 xfs_qm_mplist_lock(mp);
1117                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1118                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1119                                 break;
1120
1121                         xfs_qm_mplist_unlock(mp);
1122                         goto again;
1123                 }
1124         }
1125
1126         xfs_qm_mplist_unlock(mp);
1127         return 0;
1128 }
1129
1130
1131 /*
1132  * This initializes all the quota information that's kept in the
1133  * mount structure
1134  */
1135 STATIC int
1136 xfs_qm_init_quotainfo(
1137         xfs_mount_t     *mp)
1138 {
1139         xfs_quotainfo_t *qinf;
1140         int             error;
1141         xfs_dquot_t     *dqp;
1142
1143         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1144
1145         /*
1146          * Tell XQM that we exist as soon as possible.
1147          */
1148         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1149                 return error;
1150         }
1151
1152         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1153
1154         /*
1155          * See if quotainodes are setup, and if not, allocate them,
1156          * and change the superblock accordingly.
1157          */
1158         if ((error = xfs_qm_init_quotainos(mp))) {
1159                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1160                 mp->m_quotainfo = NULL;
1161                 return error;
1162         }
1163
1164         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1165         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1166         qinf->qi_dqreclaims = 0;
1167
1168         /* mutex used to serialize quotaoffs */
1169         mutex_init(&qinf->qi_quotaofflock);
1170
1171         /* Precalc some constants */
1172         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1173         ASSERT(qinf->qi_dqchunklen);
1174         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1175         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1176
1177         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1178
1179         /*
1180          * We try to get the limits from the superuser's limits fields.
1181          * This is quite hacky, but it is standard quota practice.
1182          * We look at the USR dquot with id == 0 first, but if user quotas
1183          * are not enabled we goto the GRP dquot with id == 0.
1184          * We don't really care to keep separate default limits for user
1185          * and group quotas, at least not at this point.
1186          */
1187         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1188                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1189                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1190                                 XFS_DQ_PROJ),
1191                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1192                              &dqp);
1193         if (! error) {
1194                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1195
1196                 /*
1197                  * The warnings and timers set the grace period given to
1198                  * a user or group before he or she can not perform any
1199                  * more writing. If it is zero, a default is used.
1200                  */
1201                 qinf->qi_btimelimit = ddqp->d_btimer ?
1202                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1203                 qinf->qi_itimelimit = ddqp->d_itimer ?
1204                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1205                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1206                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1207                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1208                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1209                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1210                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1211                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1212                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1213                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1214                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1215                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1216                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1217                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1218                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1219  
1220                 /*
1221                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1222                  * we don't want this dquot cached. We haven't done a
1223                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1224                  */
1225                 xfs_qm_dqdestroy(dqp);
1226         } else {
1227                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1228                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1229                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1230                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1231                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1232                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1233         }
1234
1235         return 0;
1236 }
1237
1238
1239 /*
1240  * Gets called when unmounting a filesystem or when all quotas get
1241  * turned off.
1242  * This purges the quota inodes, destroys locks and frees itself.
1243  */
1244 void
1245 xfs_qm_destroy_quotainfo(
1246         xfs_mount_t     *mp)
1247 {
1248         xfs_quotainfo_t *qi;
1249
1250         qi = mp->m_quotainfo;
1251         ASSERT(qi != NULL);
1252         ASSERT(xfs_Gqm != NULL);
1253
1254         /*
1255          * Release the reference that XQM kept, so that we know
1256          * when the XQM structure should be freed. We cannot assume
1257          * that xfs_Gqm is non-null after this point.
1258          */
1259         xfs_qm_rele_quotafs_ref(mp);
1260
1261         spinlock_destroy(&qi->qi_pinlock);
1262         xfs_qm_list_destroy(&qi->qi_dqlist);
1263
1264         if (qi->qi_uquotaip) {
1265                 XFS_PURGE_INODE(qi->qi_uquotaip);
1266                 qi->qi_uquotaip = NULL; /* paranoia */
1267         }
1268         if (qi->qi_gquotaip) {
1269                 XFS_PURGE_INODE(qi->qi_gquotaip);
1270                 qi->qi_gquotaip = NULL;
1271         }
1272         mutex_destroy(&qi->qi_quotaofflock);
1273         kmem_free(qi, sizeof(xfs_quotainfo_t));
1274         mp->m_quotainfo = NULL;
1275 }
1276
1277
1278
1279 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1280
1281 /* ARGSUSED */
1282 STATIC void
1283 xfs_qm_list_init(
1284         xfs_dqlist_t    *list,
1285         char            *str,
1286         int             n)
1287 {
1288         mutex_init(&list->qh_lock);
1289         list->qh_next = NULL;
1290         list->qh_version = 0;
1291         list->qh_nelems = 0;
1292 }
1293
1294 STATIC void
1295 xfs_qm_list_destroy(
1296         xfs_dqlist_t    *list)
1297 {
1298         mutex_destroy(&(list->qh_lock));
1299 }
1300
1301
1302 /*
1303  * Stripped down version of dqattach. This doesn't attach, or even look at the
1304  * dquots attached to the inode. The rationale is that there won't be any
1305  * attached at the time this is called from quotacheck.
1306  */
1307 STATIC int
1308 xfs_qm_dqget_noattach(
1309         xfs_inode_t     *ip,
1310         xfs_dquot_t     **O_udqpp,
1311         xfs_dquot_t     **O_gdqpp)
1312 {
1313         int             error;
1314         xfs_mount_t     *mp;
1315         xfs_dquot_t     *udqp, *gdqp;
1316
1317         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1318         mp = ip->i_mount;
1319         udqp = NULL;
1320         gdqp = NULL;
1321
1322         if (XFS_IS_UQUOTA_ON(mp)) {
1323                 ASSERT(ip->i_udquot == NULL);
1324                 /*
1325                  * We want the dquot allocated if it doesn't exist.
1326                  */
1327                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1328                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1329                                          &udqp))) {
1330                         /*
1331                          * Shouldn't be able to turn off quotas here.
1332                          */
1333                         ASSERT(error != ESRCH);
1334                         ASSERT(error != ENOENT);
1335                         return error;
1336                 }
1337                 ASSERT(udqp);
1338         }
1339
1340         if (XFS_IS_OQUOTA_ON(mp)) {
1341                 ASSERT(ip->i_gdquot == NULL);
1342                 if (udqp)
1343                         xfs_dqunlock(udqp);
1344                 error = XFS_IS_GQUOTA_ON(mp) ?
1345                                 xfs_qm_dqget(mp, ip,
1346                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1347                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1348                                              &gdqp) :
1349                                 xfs_qm_dqget(mp, ip,
1350                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1351                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1352                                              &gdqp);
1353                 if (error) {
1354                         if (udqp)
1355                                 xfs_qm_dqrele(udqp);
1356                         ASSERT(error != ESRCH);
1357                         ASSERT(error != ENOENT);
1358                         return error;
1359                 }
1360                 ASSERT(gdqp);
1361
1362                 /* Reacquire the locks in the right order */
1363                 if (udqp) {
1364                         if (! xfs_qm_dqlock_nowait(udqp)) {
1365                                 xfs_dqunlock(gdqp);
1366                                 xfs_dqlock(udqp);
1367                                 xfs_dqlock(gdqp);
1368                         }
1369                 }
1370         }
1371
1372         *O_udqpp = udqp;
1373         *O_gdqpp = gdqp;
1374
1375 #ifdef QUOTADEBUG
1376         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1377         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1378 #endif
1379         return 0;
1380 }
1381
1382 /*
1383  * Create an inode and return with a reference already taken, but unlocked
1384  * This is how we create quota inodes
1385  */
1386 STATIC int
1387 xfs_qm_qino_alloc(
1388         xfs_mount_t     *mp,
1389         xfs_inode_t     **ip,
1390         __int64_t       sbfields,
1391         uint            flags)
1392 {
1393         xfs_trans_t     *tp;
1394         int             error;
1395         unsigned long   s;
1396         cred_t          zerocr;
1397         xfs_inode_t     zeroino;
1398         int             committed;
1399
1400         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1401         if ((error = xfs_trans_reserve(tp,
1402                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1403                                       XFS_CREATE_LOG_RES(mp), 0,
1404                                       XFS_TRANS_PERM_LOG_RES,
1405                                       XFS_CREATE_LOG_COUNT))) {
1406                 xfs_trans_cancel(tp, 0);
1407                 return error;
1408         }
1409         memset(&zerocr, 0, sizeof(zerocr));
1410         memset(&zeroino, 0, sizeof(zeroino));
1411
1412         if ((error = xfs_dir_ialloc(&tp, &zeroino, S_IFREG, 1, 0,
1413                                    &zerocr, 0, 1, ip, &committed))) {
1414                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1415                                  XFS_TRANS_ABORT);
1416                 return error;
1417         }
1418
1419         /*
1420          * Keep an extra reference to this quota inode. This inode is
1421          * locked exclusively and joined to the transaction already.
1422          */
1423         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1424         VN_HOLD(XFS_ITOV((*ip)));
1425
1426         /*
1427          * Make the changes in the superblock, and log those too.
1428          * sbfields arg may contain fields other than *QUOTINO;
1429          * VERSIONNUM for example.
1430          */
1431         s = XFS_SB_LOCK(mp);
1432         if (flags & XFS_QMOPT_SBVERSION) {
1433 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1434                 unsigned oldv = mp->m_sb.sb_versionnum;
1435 #endif
1436                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1437                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1438                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1439                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1440                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1441
1442                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1443                 mp->m_sb.sb_uquotino = NULLFSINO;
1444                 mp->m_sb.sb_gquotino = NULLFSINO;
1445
1446                 /* qflags will get updated _after_ quotacheck */
1447                 mp->m_sb.sb_qflags = 0;
1448 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1449                 cmn_err(CE_NOTE,
1450                         "Old superblock version %x, converting to %x.",
1451                         oldv, mp->m_sb.sb_versionnum);
1452 #endif
1453         }
1454         if (flags & XFS_QMOPT_UQUOTA)
1455                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1456         else
1457                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1458         XFS_SB_UNLOCK(mp, s);
1459         xfs_mod_sb(tp, sbfields);
1460
1461         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES,
1462                                      NULL))) {
1463                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1464                 return error;
1465         }
1466         return 0;
1467 }
1468
1469
1470 STATIC int
1471 xfs_qm_reset_dqcounts(
1472         xfs_mount_t     *mp,
1473         xfs_buf_t       *bp,
1474         xfs_dqid_t      id,
1475         uint            type)
1476 {
1477         xfs_disk_dquot_t        *ddq;
1478         int                     j;
1479
1480         xfs_buftrace("RESET DQUOTS", bp);
1481         /*
1482          * Reset all counters and timers. They'll be
1483          * started afresh by xfs_qm_quotacheck.
1484          */
1485 #ifdef DEBUG
1486         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1487         do_div(j, sizeof(xfs_dqblk_t));
1488         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1489 #endif
1490         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1491         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1492                 /*
1493                  * Do a sanity check, and if needed, repair the dqblk. Don't
1494                  * output any warnings because it's perfectly possible to
1495                  * find unitialized dquot blks. See comment in xfs_qm_dqcheck.
1496                  */
1497                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1498                                       "xfs_quotacheck");
1499                 ddq->d_bcount = 0;
1500                 ddq->d_icount = 0;
1501                 ddq->d_rtbcount = 0;
1502                 ddq->d_btimer = 0;
1503                 ddq->d_itimer = 0;
1504                 ddq->d_rtbtimer = 0;
1505                 ddq->d_bwarns = 0;
1506                 ddq->d_iwarns = 0;
1507                 ddq->d_rtbwarns = 0;
1508                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1509         }
1510
1511         return 0;
1512 }
1513
1514 STATIC int
1515 xfs_qm_dqiter_bufs(
1516         xfs_mount_t     *mp,
1517         xfs_dqid_t      firstid,
1518         xfs_fsblock_t   bno,
1519         xfs_filblks_t   blkcnt,
1520         uint            flags)
1521 {
1522         xfs_buf_t       *bp;
1523         int             error;
1524         int             notcommitted;
1525         int             incr;
1526         int             type;
1527
1528         ASSERT(blkcnt > 0);
1529         notcommitted = 0;
1530         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1531                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1532         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1533                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1534         error = 0;
1535
1536         /*
1537          * Blkcnt arg can be a very big number, and might even be
1538          * larger than the log itself. So, we have to break it up into
1539          * manageable-sized transactions.
1540          * Note that we don't start a permanent transaction here; we might
1541          * not be able to get a log reservation for the whole thing up front,
1542          * and we don't really care to either, because we just discard
1543          * everything if we were to crash in the middle of this loop.
1544          */
1545         while (blkcnt--) {
1546                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1547                               XFS_FSB_TO_DADDR(mp, bno),
1548                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1549                 if (error)
1550                         break;
1551
1552                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1553                 xfs_bdwrite(mp, bp);
1554                 /*
1555                  * goto the next block.
1556                  */
1557                 bno++;
1558                 firstid += XFS_QM_DQPERBLK(mp);
1559         }
1560         return error;
1561 }
1562
1563 /*
1564  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1565  * caller supplied function for every chunk of dquots that we find.
1566  */
1567 STATIC int
1568 xfs_qm_dqiterate(
1569         xfs_mount_t     *mp,
1570         xfs_inode_t     *qip,
1571         uint            flags)
1572 {
1573         xfs_bmbt_irec_t         *map;
1574         int                     i, nmaps;       /* number of map entries */
1575         int                     error;          /* return value */
1576         xfs_fileoff_t           lblkno;
1577         xfs_filblks_t           maxlblkcnt;
1578         xfs_dqid_t              firstid;
1579         xfs_fsblock_t           rablkno;
1580         xfs_filblks_t           rablkcnt;
1581
1582         error = 0;
1583         /*
1584          * This looks racey, but we can't keep an inode lock across a
1585          * trans_reserve. But, this gets called during quotacheck, and that
1586          * happens only at mount time which is single threaded.
1587          */
1588         if (qip->i_d.di_nblocks == 0)
1589                 return 0;
1590
1591         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1592
1593         lblkno = 0;
1594         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1595         do {
1596                 nmaps = XFS_DQITER_MAP_SIZE;
1597                 /*
1598                  * We aren't changing the inode itself. Just changing
1599                  * some of its data. No new blocks are added here, and
1600                  * the inode is never added to the transaction.
1601                  */
1602                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1603                 error = xfs_bmapi(NULL, qip, lblkno,
1604                                   maxlblkcnt - lblkno,
1605                                   XFS_BMAPI_METADATA,
1606                                   NULL,
1607                                   0, map, &nmaps, NULL);
1608                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1609                 if (error)
1610                         break;
1611
1612                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1613                 for (i = 0; i < nmaps; i++) {
1614                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1615                         ASSERT(map[i].br_blockcount);
1616
1617
1618                         lblkno += map[i].br_blockcount;
1619
1620                         if (map[i].br_startblock == HOLESTARTBLOCK)
1621                                 continue;
1622
1623                         firstid = (xfs_dqid_t) map[i].br_startoff *
1624                                 XFS_QM_DQPERBLK(mp);
1625                         /*
1626                          * Do a read-ahead on the next extent.
1627                          */
1628                         if ((i+1 < nmaps) &&
1629                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1630                                 rablkcnt =  map[i+1].br_blockcount;
1631                                 rablkno = map[i+1].br_startblock;
1632                                 while (rablkcnt--) {
1633                                         xfs_baread(mp->m_ddev_targp,
1634                                                XFS_FSB_TO_DADDR(mp, rablkno),
1635                                                (int)XFS_QI_DQCHUNKLEN(mp));
1636                                         rablkno++;
1637                                 }
1638                         }
1639                         /*
1640                          * Iterate thru all the blks in the extent and
1641                          * reset the counters of all the dquots inside them.
1642                          */
1643                         if ((error = xfs_qm_dqiter_bufs(mp,
1644                                                        firstid,
1645                                                        map[i].br_startblock,
1646                                                        map[i].br_blockcount,
1647                                                        flags))) {
1648                                 break;
1649                         }
1650                 }
1651
1652                 if (error)
1653                         break;
1654         } while (nmaps > 0);
1655
1656         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1657
1658         return error;
1659 }
1660
1661 /*
1662  * Called by dqusage_adjust in doing a quotacheck.
1663  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1664  * this updates its incore copy as well as the buffer copy. This is
1665  * so that once the quotacheck is done, we can just log all the buffers,
1666  * as opposed to logging numerous updates to individual dquots.
1667  */
1668 STATIC void
1669 xfs_qm_quotacheck_dqadjust(
1670         xfs_dquot_t             *dqp,
1671         xfs_qcnt_t              nblks,
1672         xfs_qcnt_t              rtblks)
1673 {
1674         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1675         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1676         /*
1677          * Adjust the inode count and the block count to reflect this inode's
1678          * resource usage.
1679          */
1680         be64_add(&dqp->q_core.d_icount, 1);
1681         dqp->q_res_icount++;
1682         if (nblks) {
1683                 be64_add(&dqp->q_core.d_bcount, nblks);
1684                 dqp->q_res_bcount += nblks;
1685         }
1686         if (rtblks) {
1687                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1688                 dqp->q_res_rtbcount += rtblks;
1689         }
1690
1691         /*
1692          * Set default limits, adjust timers (since we changed usages)
1693          */
1694         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1695                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1696                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1697         }
1698
1699         dqp->dq_flags |= XFS_DQ_DIRTY;
1700 }
1701
1702 STATIC int
1703 xfs_qm_get_rtblks(
1704         xfs_inode_t     *ip,
1705         xfs_qcnt_t      *O_rtblks)
1706 {
1707         xfs_filblks_t   rtblks;                 /* total rt blks */
1708         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1709         xfs_extnum_t    nextents;               /* number of extent entries */
1710         xfs_bmbt_rec_t  *base;                  /* base of extent array */
1711         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1712         int             error;
1713
1714         ASSERT(XFS_IS_REALTIME_INODE(ip));
1715         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1716         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1717                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1718                         return error;
1719         }
1720         rtblks = 0;
1721         nextents = ifp->if_bytes / sizeof(xfs_bmbt_rec_t);
1722         base = &ifp->if_u1.if_extents[0];
1723         for (ep = base; ep < &base[nextents]; ep++)
1724                 rtblks += xfs_bmbt_get_blockcount(ep);
1725         *O_rtblks = (xfs_qcnt_t)rtblks;
1726         return 0;
1727 }
1728
1729 /*
1730  * callback routine supplied to bulkstat(). Given an inumber, find its
1731  * dquots and update them to account for resources taken by that inode.
1732  */
1733 /* ARGSUSED */
1734 STATIC int
1735 xfs_qm_dqusage_adjust(
1736         xfs_mount_t     *mp,            /* mount point for filesystem */
1737         xfs_ino_t       ino,            /* inode number to get data for */
1738         void            __user *buffer, /* not used */
1739         int             ubsize,         /* not used */
1740         void            *private_data,  /* not used */
1741         xfs_daddr_t     bno,            /* starting block of inode cluster */
1742         int             *ubused,        /* not used */
1743         void            *dip,           /* on-disk inode pointer (not used) */
1744         int             *res)           /* result code value */
1745 {
1746         xfs_inode_t     *ip;
1747         xfs_dquot_t     *udqp, *gdqp;
1748         xfs_qcnt_t      nblks, rtblks;
1749         int             error;
1750
1751         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1752
1753         /*
1754          * rootino must have its resources accounted for, not so with the quota
1755          * inodes.
1756          */
1757         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1758                 *res = BULKSTAT_RV_NOTHING;
1759                 return XFS_ERROR(EINVAL);
1760         }
1761
1762         /*
1763          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1764          * interface expects the inode to be exclusively locked because that's
1765          * the case in all other instances. It's OK that we do this because
1766          * quotacheck is done only at mount time.
1767          */
1768         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1769                 *res = BULKSTAT_RV_NOTHING;
1770                 return error;
1771         }
1772
1773         if (ip->i_d.di_mode == 0) {
1774                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1775                 *res = BULKSTAT_RV_NOTHING;
1776                 return XFS_ERROR(ENOENT);
1777         }
1778
1779         /*
1780          * Obtain the locked dquots. In case of an error (eg. allocation
1781          * fails for ENOSPC), we return the negative of the error number
1782          * to bulkstat, so that it can get propagated to quotacheck() and
1783          * making us disable quotas for the file system.
1784          */
1785         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1786                 xfs_iput(ip, XFS_ILOCK_EXCL);
1787                 *res = BULKSTAT_RV_GIVEUP;
1788                 return error;
1789         }
1790
1791         rtblks = 0;
1792         if (! XFS_IS_REALTIME_INODE(ip)) {
1793                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1794         } else {
1795                 /*
1796                  * Walk thru the extent list and count the realtime blocks.
1797                  */
1798                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1799                         xfs_iput(ip, XFS_ILOCK_EXCL);
1800                         if (udqp)
1801                                 xfs_qm_dqput(udqp);
1802                         if (gdqp)
1803                                 xfs_qm_dqput(gdqp);
1804                         *res = BULKSTAT_RV_GIVEUP;
1805                         return error;
1806                 }
1807                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1808         }
1809         ASSERT(ip->i_delayed_blks == 0);
1810
1811         /*
1812          * We can't release the inode while holding its dquot locks.
1813          * The inode can go into inactive and might try to acquire the dquotlocks.
1814          * So, just unlock here and do a vn_rele at the end.
1815          */
1816         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1817
1818         /*
1819          * Add the (disk blocks and inode) resources occupied by this
1820          * inode to its dquots. We do this adjustment in the incore dquot,
1821          * and also copy the changes to its buffer.
1822          * We don't care about putting these changes in a transaction
1823          * envelope because if we crash in the middle of a 'quotacheck'
1824          * we have to start from the beginning anyway.
1825          * Once we're done, we'll log all the dquot bufs.
1826          *
1827          * The *QUOTA_ON checks below may look pretty racey, but quotachecks
1828          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1829          */
1830         if (XFS_IS_UQUOTA_ON(mp)) {
1831                 ASSERT(udqp);
1832                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1833                 xfs_qm_dqput(udqp);
1834         }
1835         if (XFS_IS_OQUOTA_ON(mp)) {
1836                 ASSERT(gdqp);
1837                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1838                 xfs_qm_dqput(gdqp);
1839         }
1840         /*
1841          * Now release the inode. This will send it to 'inactive', and
1842          * possibly even free blocks.
1843          */
1844         VN_RELE(XFS_ITOV(ip));
1845
1846         /*
1847          * Goto next inode.
1848          */
1849         *res = BULKSTAT_RV_DIDONE;
1850         return 0;
1851 }
1852
1853 /*
1854  * Walk thru all the filesystem inodes and construct a consistent view
1855  * of the disk quota world. If the quotacheck fails, disable quotas.
1856  */
1857 int
1858 xfs_qm_quotacheck(
1859         xfs_mount_t     *mp)
1860 {
1861         int             done, count, error;
1862         xfs_ino_t       lastino;
1863         size_t          structsz;
1864         xfs_inode_t     *uip, *gip;
1865         uint            flags;
1866
1867         count = INT_MAX;
1868         structsz = 1;
1869         lastino = 0;
1870         flags = 0;
1871
1872         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1873         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1874
1875         /*
1876          * There should be no cached dquots. The (simplistic) quotacheck
1877          * algorithm doesn't like that.
1878          */
1879         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1880
1881         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1882
1883         /*
1884          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1885          * their counters to zero. We need a clean slate.
1886          * We don't log our changes till later.
1887          */
1888         if ((uip = XFS_QI_UQIP(mp))) {
1889                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1890                         goto error_return;
1891                 flags |= XFS_UQUOTA_CHKD;
1892         }
1893
1894         if ((gip = XFS_QI_GQIP(mp))) {
1895                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1896                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1897                         goto error_return;
1898                 flags |= XFS_OQUOTA_CHKD;
1899         }
1900
1901         do {
1902                 /*
1903                  * Iterate thru all the inodes in the file system,
1904                  * adjusting the corresponding dquot counters in core.
1905                  */
1906                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1907                                      xfs_qm_dqusage_adjust, NULL,
1908                                      structsz, NULL,
1909                                      BULKSTAT_FG_IGET|BULKSTAT_FG_VFSLOCKED,
1910                                      &done)))
1911                         break;
1912
1913         } while (! done);
1914
1915         /*
1916          * We can get this error if we couldn't do a dquot allocation inside
1917          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1918          * dirty dquots that might be cached, we just want to get rid of them
1919          * and turn quotaoff. The dquots won't be attached to any of the inodes
1920          * at this point (because we intentionally didn't in dqget_noattach).
1921          */
1922         if (error) {
1923                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1924                 goto error_return;
1925         }
1926         /*
1927          * We've made all the changes that we need to make incore.
1928          * Now flush_them down to disk buffers.
1929          */
1930         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1931
1932         /*
1933          * We didn't log anything, because if we crashed, we'll have to
1934          * start the quotacheck from scratch anyway. However, we must make
1935          * sure that our dquot changes are secure before we put the
1936          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1937          * flush.
1938          */
1939         XFS_bflush(mp->m_ddev_targp);
1940
1941         /*
1942          * If one type of quotas is off, then it will lose its
1943          * quotachecked status, since we won't be doing accounting for
1944          * that type anymore.
1945          */
1946         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1947         mp->m_qflags |= flags;
1948
1949         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1950
1951  error_return:
1952         if (error) {
1953                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1954                         "Disabling quotas.",
1955                         mp->m_fsname, error);
1956                 /*
1957                  * We must turn off quotas.
1958                  */
1959                 ASSERT(mp->m_quotainfo != NULL);
1960                 ASSERT(xfs_Gqm != NULL);
1961                 xfs_qm_destroy_quotainfo(mp);
1962                 (void)xfs_mount_reset_sbqflags(mp);
1963         } else {
1964                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1965         }
1966         return (error);
1967 }
1968
1969 /*
1970  * This is called after the superblock has been read in and we're ready to
1971  * iget the quota inodes.
1972  */
1973 STATIC int
1974 xfs_qm_init_quotainos(
1975         xfs_mount_t     *mp)
1976 {
1977         xfs_inode_t     *uip, *gip;
1978         int             error;
1979         __int64_t       sbflags;
1980         uint            flags;
1981
1982         ASSERT(mp->m_quotainfo);
1983         uip = gip = NULL;
1984         sbflags = 0;
1985         flags = 0;
1986
1987         /*
1988          * Get the uquota and gquota inodes
1989          */
1990         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1991                 if (XFS_IS_UQUOTA_ON(mp) &&
1992                     mp->m_sb.sb_uquotino != NULLFSINO) {
1993                         ASSERT(mp->m_sb.sb_uquotino > 0);
1994                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1995                                              0, 0, &uip, 0)))
1996                                 return XFS_ERROR(error);
1997                 }
1998                 if (XFS_IS_OQUOTA_ON(mp) &&
1999                     mp->m_sb.sb_gquotino != NULLFSINO) {
2000                         ASSERT(mp->m_sb.sb_gquotino > 0);
2001                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2002                                              0, 0, &gip, 0))) {
2003                                 if (uip)
2004                                         VN_RELE(XFS_ITOV(uip));
2005                                 return XFS_ERROR(error);
2006                         }
2007                 }
2008         } else {
2009                 flags |= XFS_QMOPT_SBVERSION;
2010                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2011                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2012         }
2013
2014         /*
2015          * Create the two inodes, if they don't exist already. The changes
2016          * made above will get added to a transaction and logged in one of
2017          * the qino_alloc calls below.  If the device is readonly,
2018          * temporarily switch to read-write to do this.
2019          */
2020         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2021                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2022                                               sbflags | XFS_SB_UQUOTINO,
2023                                               flags | XFS_QMOPT_UQUOTA)))
2024                         return XFS_ERROR(error);
2025
2026                 flags &= ~XFS_QMOPT_SBVERSION;
2027         }
2028         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
2029                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
2030                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2031                 error = xfs_qm_qino_alloc(mp, &gip,
2032                                           sbflags | XFS_SB_GQUOTINO, flags);
2033                 if (error) {
2034                         if (uip)
2035                                 VN_RELE(XFS_ITOV(uip));
2036
2037                         return XFS_ERROR(error);
2038                 }
2039         }
2040
2041         XFS_QI_UQIP(mp) = uip;
2042         XFS_QI_GQIP(mp) = gip;
2043
2044         return 0;
2045 }
2046
2047
2048 /*
2049  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2050  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2051  * favor the lookup function ...
2052  * XXXsup merge this with qm_reclaim_one().
2053  */
2054 STATIC int
2055 xfs_qm_shake_freelist(
2056         int howmany)
2057 {
2058         int             nreclaimed;
2059         xfs_dqhash_t    *hash;
2060         xfs_dquot_t     *dqp, *nextdqp;
2061         int             restarts;
2062         int             nflushes;
2063
2064         if (howmany <= 0)
2065                 return 0;
2066
2067         nreclaimed = 0;
2068         restarts = 0;
2069         nflushes = 0;
2070
2071 #ifdef QUOTADEBUG
2072         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2073 #endif
2074         /* lock order is : hashchainlock, freelistlock, mplistlock */
2075  tryagain:
2076         xfs_qm_freelist_lock(xfs_Gqm);
2077
2078         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2079              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2080               nreclaimed < howmany); ) {
2081                 xfs_dqlock(dqp);
2082
2083                 /*
2084                  * We are racing with dqlookup here. Naturally we don't
2085                  * want to reclaim a dquot that lookup wants.
2086                  */
2087                 if (dqp->dq_flags & XFS_DQ_WANT) {
2088                         xfs_dqunlock(dqp);
2089                         xfs_qm_freelist_unlock(xfs_Gqm);
2090                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2091                                 return nreclaimed;
2092                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2093                         goto tryagain;
2094                 }
2095
2096                 /*
2097                  * If the dquot is inactive, we are assured that it is
2098                  * not on the mplist or the hashlist, and that makes our
2099                  * life easier.
2100                  */
2101                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2102                         ASSERT(dqp->q_mount == NULL);
2103                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2104                         ASSERT(dqp->HL_PREVP == NULL);
2105                         ASSERT(dqp->MPL_PREVP == NULL);
2106                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2107                         nextdqp = dqp->dq_flnext;
2108                         goto off_freelist;
2109                 }
2110
2111                 ASSERT(dqp->MPL_PREVP);
2112                 /*
2113                  * Try to grab the flush lock. If this dquot is in the process of
2114                  * getting flushed to disk, we don't want to reclaim it.
2115                  */
2116                 if (! xfs_qm_dqflock_nowait(dqp)) {
2117                         xfs_dqunlock(dqp);
2118                         dqp = dqp->dq_flnext;
2119                         continue;
2120                 }
2121
2122                 /*
2123                  * We have the flush lock so we know that this is not in the
2124                  * process of being flushed. So, if this is dirty, flush it
2125                  * DELWRI so that we don't get a freelist infested with
2126                  * dirty dquots.
2127                  */
2128                 if (XFS_DQ_IS_DIRTY(dqp)) {
2129                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2130                         /*
2131                          * We flush it delayed write, so don't bother
2132                          * releasing the mplock.
2133                          */
2134                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2135                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2136                         dqp = dqp->dq_flnext;
2137                         continue;
2138                 }
2139                 /*
2140                  * We're trying to get the hashlock out of order. This races
2141                  * with dqlookup; so, we giveup and goto the next dquot if
2142                  * we couldn't get the hashlock. This way, we won't starve
2143                  * a dqlookup process that holds the hashlock that is
2144                  * waiting for the freelist lock.
2145                  */
2146                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2147                         xfs_dqfunlock(dqp);
2148                         xfs_dqunlock(dqp);
2149                         dqp = dqp->dq_flnext;
2150                         continue;
2151                 }
2152                 /*
2153                  * This races with dquot allocation code as well as dqflush_all
2154                  * and reclaim code. So, if we failed to grab the mplist lock,
2155                  * giveup everything and start over.
2156                  */
2157                 hash = dqp->q_hash;
2158                 ASSERT(hash);
2159                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2160                         /* XXX put a sentinel so that we can come back here */
2161                         xfs_dqfunlock(dqp);
2162                         xfs_dqunlock(dqp);
2163                         XFS_DQ_HASH_UNLOCK(hash);
2164                         xfs_qm_freelist_unlock(xfs_Gqm);
2165                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2166                                 return nreclaimed;
2167                         goto tryagain;
2168                 }
2169                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2170 #ifdef QUOTADEBUG
2171                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2172                         dqp, be32_to_cpu(dqp->q_core.d_id));
2173 #endif
2174                 ASSERT(dqp->q_nrefs == 0);
2175                 nextdqp = dqp->dq_flnext;
2176                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2177                 XQM_HASHLIST_REMOVE(hash, dqp);
2178                 xfs_dqfunlock(dqp);
2179                 xfs_qm_mplist_unlock(dqp->q_mount);
2180                 XFS_DQ_HASH_UNLOCK(hash);
2181
2182  off_freelist:
2183                 XQM_FREELIST_REMOVE(dqp);
2184                 xfs_dqunlock(dqp);
2185                 nreclaimed++;
2186                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2187                 xfs_qm_dqdestroy(dqp);
2188                 dqp = nextdqp;
2189         }
2190         xfs_qm_freelist_unlock(xfs_Gqm);
2191         return nreclaimed;
2192 }
2193
2194
2195 /*
2196  * The kmem_shake interface is invoked when memory is running low.
2197  */
2198 /* ARGSUSED */
2199 STATIC int
2200 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2201 {
2202         int     ndqused, nfree, n;
2203
2204         if (!kmem_shake_allow(gfp_mask))
2205                 return 0;
2206         if (!xfs_Gqm)
2207                 return 0;
2208
2209         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2210         /* incore dquots in all f/s's */
2211         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2212
2213         ASSERT(ndqused >= 0);
2214
2215         if (nfree <= ndqused && nfree < ndquot)
2216                 return 0;
2217
2218         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2219         n = nfree - ndqused - ndquot;           /* # over target */
2220
2221         return xfs_qm_shake_freelist(MAX(nfree, n));
2222 }
2223
2224
2225 /*
2226  * Just pop the least recently used dquot off the freelist and
2227  * recycle it. The returned dquot is locked.
2228  */
2229 STATIC xfs_dquot_t *
2230 xfs_qm_dqreclaim_one(void)
2231 {
2232         xfs_dquot_t     *dqpout;
2233         xfs_dquot_t     *dqp;
2234         int             restarts;
2235         int             nflushes;
2236
2237         restarts = 0;
2238         dqpout = NULL;
2239         nflushes = 0;
2240
2241         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2242  startagain:
2243         xfs_qm_freelist_lock(xfs_Gqm);
2244
2245         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2246                 xfs_dqlock(dqp);
2247
2248                 /*
2249                  * We are racing with dqlookup here. Naturally we don't
2250                  * want to reclaim a dquot that lookup wants. We release the
2251                  * freelist lock and start over, so that lookup will grab
2252                  * both the dquot and the freelistlock.
2253                  */
2254                 if (dqp->dq_flags & XFS_DQ_WANT) {
2255                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2256                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2257                         xfs_dqunlock(dqp);
2258                         xfs_qm_freelist_unlock(xfs_Gqm);
2259                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2260                                 return NULL;
2261                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2262                         goto startagain;
2263                 }
2264
2265                 /*
2266                  * If the dquot is inactive, we are assured that it is
2267                  * not on the mplist or the hashlist, and that makes our
2268                  * life easier.
2269                  */
2270                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2271                         ASSERT(dqp->q_mount == NULL);
2272                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2273                         ASSERT(dqp->HL_PREVP == NULL);
2274                         ASSERT(dqp->MPL_PREVP == NULL);
2275                         XQM_FREELIST_REMOVE(dqp);
2276                         xfs_dqunlock(dqp);
2277                         dqpout = dqp;
2278                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2279                         break;
2280                 }
2281
2282                 ASSERT(dqp->q_hash);
2283                 ASSERT(dqp->MPL_PREVP);
2284
2285                 /*
2286                  * Try to grab the flush lock. If this dquot is in the process of
2287                  * getting flushed to disk, we don't want to reclaim it.
2288                  */
2289                 if (! xfs_qm_dqflock_nowait(dqp)) {
2290                         xfs_dqunlock(dqp);
2291                         continue;
2292                 }
2293
2294                 /*
2295                  * We have the flush lock so we know that this is not in the
2296                  * process of being flushed. So, if this is dirty, flush it
2297                  * DELWRI so that we don't get a freelist infested with
2298                  * dirty dquots.
2299                  */
2300                 if (XFS_DQ_IS_DIRTY(dqp)) {
2301                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2302                         /*
2303                          * We flush it delayed write, so don't bother
2304                          * releasing the freelist lock.
2305                          */
2306                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2307                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2308                         continue;
2309                 }
2310
2311                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2312                         xfs_dqfunlock(dqp);
2313                         xfs_dqunlock(dqp);
2314                         continue;
2315                 }
2316
2317                 if (! xfs_qm_dqhashlock_nowait(dqp))
2318                         goto mplistunlock;
2319
2320                 ASSERT(dqp->q_nrefs == 0);
2321                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2322                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2323                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2324                 XQM_FREELIST_REMOVE(dqp);
2325                 dqpout = dqp;
2326                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2327  mplistunlock:
2328                 xfs_qm_mplist_unlock(dqp->q_mount);
2329                 xfs_dqfunlock(dqp);
2330                 xfs_dqunlock(dqp);
2331                 if (dqpout)
2332                         break;
2333         }
2334
2335         xfs_qm_freelist_unlock(xfs_Gqm);
2336         return dqpout;
2337 }
2338
2339
2340 /*------------------------------------------------------------------*/
2341
2342 /*
2343  * Return a new incore dquot. Depending on the number of
2344  * dquots in the system, we either allocate a new one on the kernel heap,
2345  * or reclaim a free one.
2346  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2347  * to reclaim an existing one from the freelist.
2348  */
2349 boolean_t
2350 xfs_qm_dqalloc_incore(
2351         xfs_dquot_t **O_dqpp)
2352 {
2353         xfs_dquot_t     *dqp;
2354
2355         /*
2356          * Check against high water mark to see if we want to pop
2357          * a nincompoop dquot off the freelist.
2358          */
2359         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2360                 /*
2361                  * Try to recycle a dquot from the freelist.
2362                  */
2363                 if ((dqp = xfs_qm_dqreclaim_one())) {
2364                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2365                         /*
2366                          * Just zero the core here. The rest will get
2367                          * reinitialized by caller. XXX we shouldn't even
2368                          * do this zero ...
2369                          */
2370                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2371                         *O_dqpp = dqp;
2372                         return B_FALSE;
2373                 }
2374                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2375         }
2376
2377         /*
2378          * Allocate a brand new dquot on the kernel heap and return it
2379          * to the caller to initialize.
2380          */
2381         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2382         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2383         atomic_inc(&xfs_Gqm->qm_totaldquots);
2384
2385         return B_TRUE;
2386 }
2387
2388
2389 /*
2390  * Start a transaction and write the incore superblock changes to
2391  * disk. flags parameter indicates which fields have changed.
2392  */
2393 int
2394 xfs_qm_write_sb_changes(
2395         xfs_mount_t     *mp,
2396         __int64_t       flags)
2397 {
2398         xfs_trans_t     *tp;
2399         int             error;
2400
2401 #ifdef QUOTADEBUG
2402         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2403 #endif
2404         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2405         if ((error = xfs_trans_reserve(tp, 0,
2406                                       mp->m_sb.sb_sectsize + 128, 0,
2407                                       0,
2408                                       XFS_DEFAULT_LOG_COUNT))) {
2409                 xfs_trans_cancel(tp, 0);
2410                 return error;
2411         }
2412
2413         xfs_mod_sb(tp, flags);
2414         (void) xfs_trans_commit(tp, 0, NULL);
2415
2416         return 0;
2417 }
2418
2419
2420 /* --------------- utility functions for vnodeops ---------------- */
2421
2422
2423 /*
2424  * Given an inode, a uid and gid (from cred_t) make sure that we have
2425  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2426  * quotas by creating this file.
2427  * This also attaches dquot(s) to the given inode after locking it,
2428  * and returns the dquots corresponding to the uid and/or gid.
2429  *
2430  * in   : inode (unlocked)
2431  * out  : udquot, gdquot with references taken and unlocked
2432  */
2433 int
2434 xfs_qm_vop_dqalloc(
2435         xfs_mount_t     *mp,
2436         xfs_inode_t     *ip,
2437         uid_t           uid,
2438         gid_t           gid,
2439         prid_t          prid,
2440         uint            flags,
2441         xfs_dquot_t     **O_udqpp,
2442         xfs_dquot_t     **O_gdqpp)
2443 {
2444         int             error;
2445         xfs_dquot_t     *uq, *gq;
2446         uint            lockflags;
2447
2448         if (!XFS_IS_QUOTA_ON(mp))
2449                 return 0;
2450
2451         lockflags = XFS_ILOCK_EXCL;
2452         xfs_ilock(ip, lockflags);
2453
2454         if ((flags & XFS_QMOPT_INHERIT) &&
2455             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2456                 gid = ip->i_d.di_gid;
2457
2458         /*
2459          * Attach the dquot(s) to this inode, doing a dquot allocation
2460          * if necessary. The dquot(s) will not be locked.
2461          */
2462         if (XFS_NOT_DQATTACHED(mp, ip)) {
2463                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2464                                             XFS_QMOPT_ILOCKED))) {
2465                         xfs_iunlock(ip, lockflags);
2466                         return error;
2467                 }
2468         }
2469
2470         uq = gq = NULL;
2471         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2472                 if (ip->i_d.di_uid != uid) {
2473                         /*
2474                          * What we need is the dquot that has this uid, and
2475                          * if we send the inode to dqget, the uid of the inode
2476                          * takes priority over what's sent in the uid argument.
2477                          * We must unlock inode here before calling dqget if
2478                          * we're not sending the inode, because otherwise
2479                          * we'll deadlock by doing trans_reserve while
2480                          * holding ilock.
2481                          */
2482                         xfs_iunlock(ip, lockflags);
2483                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2484                                                  XFS_DQ_USER,
2485                                                  XFS_QMOPT_DQALLOC |
2486                                                  XFS_QMOPT_DOWARN,
2487                                                  &uq))) {
2488                                 ASSERT(error != ENOENT);
2489                                 return error;
2490                         }
2491                         /*
2492                          * Get the ilock in the right order.
2493                          */
2494                         xfs_dqunlock(uq);
2495                         lockflags = XFS_ILOCK_SHARED;
2496                         xfs_ilock(ip, lockflags);
2497                 } else {
2498                         /*
2499                          * Take an extra reference, because we'll return
2500                          * this to caller
2501                          */
2502                         ASSERT(ip->i_udquot);
2503                         uq = ip->i_udquot;
2504                         xfs_dqlock(uq);
2505                         XFS_DQHOLD(uq);
2506                         xfs_dqunlock(uq);
2507                 }
2508         }
2509         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2510                 if (ip->i_d.di_gid != gid) {
2511                         xfs_iunlock(ip, lockflags);
2512                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2513                                                  XFS_DQ_GROUP,
2514                                                  XFS_QMOPT_DQALLOC |
2515                                                  XFS_QMOPT_DOWARN,
2516                                                  &gq))) {
2517                                 if (uq)
2518                                         xfs_qm_dqrele(uq);
2519                                 ASSERT(error != ENOENT);
2520                                 return error;
2521                         }
2522                         xfs_dqunlock(gq);
2523                         lockflags = XFS_ILOCK_SHARED;
2524                         xfs_ilock(ip, lockflags);
2525                 } else {
2526                         ASSERT(ip->i_gdquot);
2527                         gq = ip->i_gdquot;
2528                         xfs_dqlock(gq);
2529                         XFS_DQHOLD(gq);
2530                         xfs_dqunlock(gq);
2531                 }
2532         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2533                 if (ip->i_d.di_projid != prid) {
2534                         xfs_iunlock(ip, lockflags);
2535                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2536                                                  XFS_DQ_PROJ,
2537                                                  XFS_QMOPT_DQALLOC |
2538                                                  XFS_QMOPT_DOWARN,
2539                                                  &gq))) {
2540                                 if (uq)
2541                                         xfs_qm_dqrele(uq);
2542                                 ASSERT(error != ENOENT);
2543                                 return (error);
2544                         }
2545                         xfs_dqunlock(gq);
2546                         lockflags = XFS_ILOCK_SHARED;
2547                         xfs_ilock(ip, lockflags);
2548                 } else {
2549                         ASSERT(ip->i_gdquot);
2550                         gq = ip->i_gdquot;
2551                         xfs_dqlock(gq);
2552                         XFS_DQHOLD(gq);
2553                         xfs_dqunlock(gq);
2554                 }
2555         }
2556         if (uq)
2557                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2558
2559         xfs_iunlock(ip, lockflags);
2560         if (O_udqpp)
2561                 *O_udqpp = uq;
2562         else if (uq)
2563                 xfs_qm_dqrele(uq);
2564         if (O_gdqpp)
2565                 *O_gdqpp = gq;
2566         else if (gq)
2567                 xfs_qm_dqrele(gq);
2568         return 0;
2569 }
2570
2571 /*
2572  * Actually transfer ownership, and do dquot modifications.
2573  * These were already reserved.
2574  */
2575 xfs_dquot_t *
2576 xfs_qm_vop_chown(
2577         xfs_trans_t     *tp,
2578         xfs_inode_t     *ip,
2579         xfs_dquot_t     **IO_olddq,
2580         xfs_dquot_t     *newdq)
2581 {
2582         xfs_dquot_t     *prevdq;
2583         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2584                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2585
2586         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2587         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2588
2589         /* old dquot */
2590         prevdq = *IO_olddq;
2591         ASSERT(prevdq);
2592         ASSERT(prevdq != newdq);
2593
2594         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2595         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2596
2597         /* the sparkling new dquot */
2598         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2599         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2600
2601         /*
2602          * Take an extra reference, because the inode
2603          * is going to keep this dquot pointer even
2604          * after the trans_commit.
2605          */
2606         xfs_dqlock(newdq);
2607         XFS_DQHOLD(newdq);
2608         xfs_dqunlock(newdq);
2609         *IO_olddq = newdq;
2610
2611         return prevdq;
2612 }
2613
2614 /*
2615  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2616  */
2617 int
2618 xfs_qm_vop_chown_reserve(
2619         xfs_trans_t     *tp,
2620         xfs_inode_t     *ip,
2621         xfs_dquot_t     *udqp,
2622         xfs_dquot_t     *gdqp,
2623         uint            flags)
2624 {
2625         int             error;
2626         xfs_mount_t     *mp;
2627         uint            delblks, blkflags;
2628         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2629
2630         ASSERT(XFS_ISLOCKED_INODE(ip));
2631         mp = ip->i_mount;
2632         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2633
2634         delblks = ip->i_delayed_blks;
2635         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2636         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2637                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2638
2639         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2640             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2641                 delblksudq = udqp;
2642                 /*
2643                  * If there are delayed allocation blocks, then we have to
2644                  * unreserve those from the old dquot, and add them to the
2645                  * new dquot.
2646                  */
2647                 if (delblks) {
2648                         ASSERT(ip->i_udquot);
2649                         unresudq = ip->i_udquot;
2650                 }
2651         }
2652         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2653                 if ((XFS_IS_GQUOTA_ON(ip->i_mount) &&
2654                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id)) ||
2655                     (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2656                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))) {
2657                         delblksgdq = gdqp;
2658                         if (delblks) {
2659                                 ASSERT(ip->i_gdquot);
2660                                 unresgdq = ip->i_gdquot;
2661                         }
2662                 }
2663         }
2664
2665         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2666                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2667                                 flags | blkflags)))
2668                 return (error);
2669
2670         /*
2671          * Do the delayed blks reservations/unreservations now. Since, these
2672          * are done without the help of a transaction, if a reservation fails
2673          * its previous reservations won't be automatically undone by trans
2674          * code. So, we have to do it manually here.
2675          */
2676         if (delblks) {
2677                 /*
2678                  * Do the reservations first. Unreservation can't fail.
2679                  */
2680                 ASSERT(delblksudq || delblksgdq);
2681                 ASSERT(unresudq || unresgdq);
2682                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2683                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2684                                 flags | blkflags)))
2685                         return (error);
2686                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2687                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2688                                 blkflags);
2689         }
2690
2691         return (0);
2692 }
2693
2694 int
2695 xfs_qm_vop_rename_dqattach(
2696         xfs_inode_t     **i_tab)
2697 {
2698         xfs_inode_t     *ip;
2699         int             i;
2700         int             error;
2701
2702         ip = i_tab[0];
2703
2704         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2705                 return 0;
2706
2707         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2708                 error = xfs_qm_dqattach(ip, 0);
2709                 if (error)
2710                         return error;
2711         }
2712         for (i = 1; (i < 4 && i_tab[i]); i++) {
2713                 /*
2714                  * Watch out for duplicate entries in the table.
2715                  */
2716                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2717                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2718                                 error = xfs_qm_dqattach(ip, 0);
2719                                 if (error)
2720                                         return error;
2721                         }
2722                 }
2723         }
2724         return 0;
2725 }
2726
2727 void
2728 xfs_qm_vop_dqattach_and_dqmod_newinode(
2729         xfs_trans_t     *tp,
2730         xfs_inode_t     *ip,
2731         xfs_dquot_t     *udqp,
2732         xfs_dquot_t     *gdqp)
2733 {
2734         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2735                 return;
2736
2737         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2738         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2739
2740         if (udqp) {
2741                 xfs_dqlock(udqp);
2742                 XFS_DQHOLD(udqp);
2743                 xfs_dqunlock(udqp);
2744                 ASSERT(ip->i_udquot == NULL);
2745                 ip->i_udquot = udqp;
2746                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2747                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2748                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2749         }
2750         if (gdqp) {
2751                 xfs_dqlock(gdqp);
2752                 XFS_DQHOLD(gdqp);
2753                 xfs_dqunlock(gdqp);
2754                 ASSERT(ip->i_gdquot == NULL);
2755                 ip->i_gdquot = gdqp;
2756                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2757                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2758                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2759                                 be32_to_cpu(gdqp->q_core.d_id));
2760                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2761         }
2762 }
2763
2764 /* ------------- list stuff -----------------*/
2765 STATIC void
2766 xfs_qm_freelist_init(xfs_frlist_t *ql)
2767 {
2768         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2769         mutex_init(&ql->qh_lock);
2770         ql->qh_version = 0;
2771         ql->qh_nelems = 0;
2772 }
2773
2774 STATIC void
2775 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2776 {
2777         xfs_dquot_t     *dqp, *nextdqp;
2778
2779         mutex_lock(&ql->qh_lock);
2780         for (dqp = ql->qh_next;
2781              dqp != (xfs_dquot_t *)ql; ) {
2782                 xfs_dqlock(dqp);
2783                 nextdqp = dqp->dq_flnext;
2784 #ifdef QUOTADEBUG
2785                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2786 #endif
2787                 XQM_FREELIST_REMOVE(dqp);
2788                 xfs_dqunlock(dqp);
2789                 xfs_qm_dqdestroy(dqp);
2790                 dqp = nextdqp;
2791         }
2792         /*
2793          * Don't bother about unlocking.
2794          */
2795         mutex_destroy(&ql->qh_lock);
2796
2797         ASSERT(ql->qh_nelems == 0);
2798 }
2799
2800 STATIC void
2801 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2802 {
2803         dq->dq_flnext = ql->qh_next;
2804         dq->dq_flprev = (xfs_dquot_t *)ql;
2805         ql->qh_next = dq;
2806         dq->dq_flnext->dq_flprev = dq;
2807         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2808         xfs_Gqm->qm_dqfreelist.qh_version++;
2809 }
2810
2811 void
2812 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2813 {
2814         xfs_dquot_t *next = dq->dq_flnext;
2815         xfs_dquot_t *prev = dq->dq_flprev;
2816
2817         next->dq_flprev = prev;
2818         prev->dq_flnext = next;
2819         dq->dq_flnext = dq->dq_flprev = dq;
2820         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2821         xfs_Gqm->qm_dqfreelist.qh_version++;
2822 }
2823
2824 void
2825 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2826 {
2827         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2828 }
2829
2830 STATIC int
2831 xfs_qm_dqhashlock_nowait(
2832         xfs_dquot_t *dqp)
2833 {
2834         int locked;
2835
2836         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2837         return locked;
2838 }
2839
2840 int
2841 xfs_qm_freelist_lock_nowait(
2842         xfs_qm_t *xqm)
2843 {
2844         int locked;
2845
2846         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2847         return locked;
2848 }
2849
2850 STATIC int
2851 xfs_qm_mplist_nowait(
2852         xfs_mount_t     *mp)
2853 {
2854         int locked;
2855
2856         ASSERT(mp->m_quotainfo);
2857         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2858         return locked;
2859 }