[XFS] Quota inode has no parent.
[safe/jmp/linux-2.6] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_clnt.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_alloc.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_quota.h"
31 #include "xfs_mount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_itable.h"
42 #include "xfs_rtalloc.h"
43 #include "xfs_error.h"
44 #include "xfs_bmap.h"
45 #include "xfs_rw.h"
46 #include "xfs_acl.h"
47 #include "xfs_attr.h"
48 #include "xfs_buf_item.h"
49 #include "xfs_trans_space.h"
50 #include "xfs_utils.h"
51 #include "xfs_qm.h"
52
53 /*
54  * The global quota manager. There is only one of these for the entire
55  * system, _not_ one per file system. XQM keeps track of the overall
56  * quota functionality, including maintaining the freelist and hash
57  * tables of dquots.
58  */
59 mutex_t         xfs_Gqm_lock;
60 struct xfs_qm   *xfs_Gqm;
61 uint            ndquot;
62
63 kmem_zone_t     *qm_dqzone;
64 kmem_zone_t     *qm_dqtrxzone;
65 static struct shrinker *xfs_qm_shaker;
66
67 static cred_t   xfs_zerocr;
68
69 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
70 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
71
72 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
73 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
74 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
75 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
76
77 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
78 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
79 STATIC int      xfs_qm_shake(int, gfp_t);
80
81 #ifdef DEBUG
82 extern mutex_t  qcheck_lock;
83 #endif
84
85 #ifdef QUOTADEBUG
86 #define XQM_LIST_PRINT(l, NXT, title) \
87 { \
88         xfs_dquot_t     *dqp; int i = 0; \
89         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
90         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
91                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
92                                   "bcnt = %d, icnt = %d, refs = %d", \
93                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
94                         DQFLAGTO_TYPESTR(dqp),       \
95                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
96                         (int) be64_to_cpu(dqp->q_core.d_icount), \
97                         (int) dqp->q_nrefs);  } \
98 }
99 #else
100 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
101 #endif
102
103 /*
104  * Initialize the XQM structure.
105  * Note that there is not one quota manager per file system.
106  */
107 STATIC struct xfs_qm *
108 xfs_Gqm_init(void)
109 {
110         xfs_dqhash_t    *udqhash, *gdqhash;
111         xfs_qm_t        *xqm;
112         size_t          hsize;
113         uint            i;
114
115         /*
116          * Initialize the dquot hash tables.
117          */
118         udqhash = kmem_zalloc_greedy(&hsize,
119                                      XFS_QM_HASHSIZE_LOW, XFS_QM_HASHSIZE_HIGH,
120                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
121         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
122         hsize /= sizeof(xfs_dqhash_t);
123         ndquot = hsize << 8;
124
125         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
126         xqm->qm_dqhashmask = hsize - 1;
127         xqm->qm_usr_dqhtable = udqhash;
128         xqm->qm_grp_dqhtable = gdqhash;
129         ASSERT(xqm->qm_usr_dqhtable != NULL);
130         ASSERT(xqm->qm_grp_dqhtable != NULL);
131
132         for (i = 0; i < hsize; i++) {
133                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
134                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
135         }
136
137         /*
138          * Freelist of all dquots of all file systems
139          */
140         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
141
142         /*
143          * dquot zone. we register our own low-memory callback.
144          */
145         if (!qm_dqzone) {
146                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
147                                                 "xfs_dquots");
148                 qm_dqzone = xqm->qm_dqzone;
149         } else
150                 xqm->qm_dqzone = qm_dqzone;
151
152         xfs_qm_shaker = set_shrinker(DEFAULT_SEEKS, xfs_qm_shake);
153
154         /*
155          * The t_dqinfo portion of transactions.
156          */
157         if (!qm_dqtrxzone) {
158                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
159                                                    "xfs_dqtrx");
160                 qm_dqtrxzone = xqm->qm_dqtrxzone;
161         } else
162                 xqm->qm_dqtrxzone = qm_dqtrxzone;
163
164         atomic_set(&xqm->qm_totaldquots, 0);
165         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
166         xqm->qm_nrefs = 0;
167 #ifdef DEBUG
168         mutex_init(&qcheck_lock);
169 #endif
170         return xqm;
171 }
172
173 /*
174  * Destroy the global quota manager when its reference count goes to zero.
175  */
176 STATIC void
177 xfs_qm_destroy(
178         struct xfs_qm   *xqm)
179 {
180         int             hsize, i;
181
182         ASSERT(xqm != NULL);
183         ASSERT(xqm->qm_nrefs == 0);
184         remove_shrinker(xfs_qm_shaker);
185         hsize = xqm->qm_dqhashmask + 1;
186         for (i = 0; i < hsize; i++) {
187                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
188                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
189         }
190         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
191         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
192         xqm->qm_usr_dqhtable = NULL;
193         xqm->qm_grp_dqhtable = NULL;
194         xqm->qm_dqhashmask = 0;
195         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
196 #ifdef DEBUG
197         mutex_destroy(&qcheck_lock);
198 #endif
199         kmem_free(xqm, sizeof(xfs_qm_t));
200 }
201
202 /*
203  * Called at mount time to let XQM know that another file system is
204  * starting quotas. This isn't crucial information as the individual mount
205  * structures are pretty independent, but it helps the XQM keep a
206  * global view of what's going on.
207  */
208 /* ARGSUSED */
209 STATIC int
210 xfs_qm_hold_quotafs_ref(
211         struct xfs_mount *mp)
212 {
213         /*
214          * Need to lock the xfs_Gqm structure for things like this. For example,
215          * the structure could disappear between the entry to this routine and
216          * a HOLD operation if not locked.
217          */
218         XFS_QM_LOCK(xfs_Gqm);
219
220         if (xfs_Gqm == NULL)
221                 xfs_Gqm = xfs_Gqm_init();
222         /*
223          * We can keep a list of all filesystems with quotas mounted for
224          * debugging and statistical purposes, but ...
225          * Just take a reference and get out.
226          */
227         XFS_QM_HOLD(xfs_Gqm);
228         XFS_QM_UNLOCK(xfs_Gqm);
229
230         return 0;
231 }
232
233
234 /*
235  * Release the reference that a filesystem took at mount time,
236  * so that we know when we need to destroy the entire quota manager.
237  */
238 /* ARGSUSED */
239 STATIC void
240 xfs_qm_rele_quotafs_ref(
241         struct xfs_mount *mp)
242 {
243         xfs_dquot_t     *dqp, *nextdqp;
244
245         ASSERT(xfs_Gqm);
246         ASSERT(xfs_Gqm->qm_nrefs > 0);
247
248         /*
249          * Go thru the freelist and destroy all inactive dquots.
250          */
251         xfs_qm_freelist_lock(xfs_Gqm);
252
253         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
254              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
255                 xfs_dqlock(dqp);
256                 nextdqp = dqp->dq_flnext;
257                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
258                         ASSERT(dqp->q_mount == NULL);
259                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
260                         ASSERT(dqp->HL_PREVP == NULL);
261                         ASSERT(dqp->MPL_PREVP == NULL);
262                         XQM_FREELIST_REMOVE(dqp);
263                         xfs_dqunlock(dqp);
264                         xfs_qm_dqdestroy(dqp);
265                 } else {
266                         xfs_dqunlock(dqp);
267                 }
268                 dqp = nextdqp;
269         }
270         xfs_qm_freelist_unlock(xfs_Gqm);
271
272         /*
273          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
274          * be restarted.
275          */
276         XFS_QM_LOCK(xfs_Gqm);
277         XFS_QM_RELE(xfs_Gqm);
278         if (xfs_Gqm->qm_nrefs == 0) {
279                 xfs_qm_destroy(xfs_Gqm);
280                 xfs_Gqm = NULL;
281         }
282         XFS_QM_UNLOCK(xfs_Gqm);
283 }
284
285 /*
286  * This is called at mount time from xfs_mountfs to initialize the quotainfo
287  * structure and start the global quota manager (xfs_Gqm) if it hasn't done
288  * so already.  Note that the superblock has not been read in yet.
289  */
290 void
291 xfs_qm_mount_quotainit(
292         xfs_mount_t     *mp,
293         uint            flags)
294 {
295         /*
296          * User, projects or group quotas has to be on.
297          */
298         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_PQUOTA | XFSMNT_GQUOTA));
299
300         /*
301          * Initialize the flags in the mount structure. From this point
302          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
303          * Note that we enforce nothing if accounting is off.
304          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
305          * It isn't necessary to take the quotaoff lock to do this; this is
306          * called from mount.
307          */
308         if (flags & XFSMNT_UQUOTA) {
309                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
310                 if (flags & XFSMNT_UQUOTAENF)
311                         mp->m_qflags |= XFS_UQUOTA_ENFD;
312         }
313         if (flags & XFSMNT_GQUOTA) {
314                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
315                 if (flags & XFSMNT_GQUOTAENF)
316                         mp->m_qflags |= XFS_OQUOTA_ENFD;
317         } else if (flags & XFSMNT_PQUOTA) {
318                 mp->m_qflags |= (XFS_PQUOTA_ACCT | XFS_PQUOTA_ACTIVE);
319                 if (flags & XFSMNT_PQUOTAENF)
320                         mp->m_qflags |= XFS_OQUOTA_ENFD;
321         }
322 }
323
324 /*
325  * Just destroy the quotainfo structure.
326  */
327 void
328 xfs_qm_unmount_quotadestroy(
329         xfs_mount_t     *mp)
330 {
331         if (mp->m_quotainfo)
332                 xfs_qm_destroy_quotainfo(mp);
333 }
334
335
336 /*
337  * This is called from xfs_mountfs to start quotas and initialize all
338  * necessary data structures like quotainfo.  This is also responsible for
339  * running a quotacheck as necessary.  We are guaranteed that the superblock
340  * is consistently read in at this point.
341  */
342 int
343 xfs_qm_mount_quotas(
344         xfs_mount_t     *mp,
345         int             mfsi_flags)
346 {
347         unsigned long   s;
348         int             error = 0;
349         uint            sbf;
350
351
352         /*
353          * If quotas on realtime volumes is not supported, we disable
354          * quotas immediately.
355          */
356         if (mp->m_sb.sb_rextents) {
357                 cmn_err(CE_NOTE,
358                         "Cannot turn on quotas for realtime filesystem %s",
359                         mp->m_fsname);
360                 mp->m_qflags = 0;
361                 goto write_changes;
362         }
363
364         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
365
366         /*
367          * Allocate the quotainfo structure inside the mount struct, and
368          * create quotainode(s), and change/rev superblock if necessary.
369          */
370         if ((error = xfs_qm_init_quotainfo(mp))) {
371                 /*
372                  * We must turn off quotas.
373                  */
374                 ASSERT(mp->m_quotainfo == NULL);
375                 mp->m_qflags = 0;
376                 goto write_changes;
377         }
378         /*
379          * If any of the quotas are not consistent, do a quotacheck.
380          */
381         if (XFS_QM_NEED_QUOTACHECK(mp) &&
382                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
383                 if ((error = xfs_qm_quotacheck(mp))) {
384                         /* Quotacheck has failed and quotas have
385                          * been disabled.
386                          */
387                         return XFS_ERROR(error);
388                 }
389         }
390         /* 
391          * If one type of quotas is off, then it will lose its
392          * quotachecked status, since we won't be doing accounting for
393          * that type anymore.
394          */
395         if (!XFS_IS_UQUOTA_ON(mp)) {
396                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
397         }
398         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) {
399                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
400         }
401
402  write_changes:
403         /*
404          * We actually don't have to acquire the SB_LOCK at all.
405          * This can only be called from mount, and that's single threaded. XXX
406          */
407         s = XFS_SB_LOCK(mp);
408         sbf = mp->m_sb.sb_qflags;
409         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
410         XFS_SB_UNLOCK(mp, s);
411
412         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
413                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
414                         /*
415                          * We could only have been turning quotas off.
416                          * We aren't in very good shape actually because
417                          * the incore structures are convinced that quotas are
418                          * off, but the on disk superblock doesn't know that !
419                          */
420                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
421                         xfs_fs_cmn_err(CE_ALERT, mp,
422                                 "XFS mount_quotas: Superblock update failed!");
423                 }
424         }
425
426         if (error) {
427                 xfs_fs_cmn_err(CE_WARN, mp,
428                         "Failed to initialize disk quotas.");
429         }
430         return XFS_ERROR(error);
431 }
432
433 /*
434  * Called from the vfsops layer.
435  */
436 int
437 xfs_qm_unmount_quotas(
438         xfs_mount_t     *mp)
439 {
440         xfs_inode_t     *uqp, *gqp;
441         int             error = 0;
442
443         /*
444          * Release the dquots that root inode, et al might be holding,
445          * before we flush quotas and blow away the quotainfo structure.
446          */
447         ASSERT(mp->m_rootip);
448         xfs_qm_dqdetach(mp->m_rootip);
449         if (mp->m_rbmip)
450                 xfs_qm_dqdetach(mp->m_rbmip);
451         if (mp->m_rsumip)
452                 xfs_qm_dqdetach(mp->m_rsumip);
453
454         /*
455          * Flush out the quota inodes.
456          */
457         uqp = gqp = NULL;
458         if (mp->m_quotainfo) {
459                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
460                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
461                         xfs_iflock(uqp);
462                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
463                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
464                         if (unlikely(error == EFSCORRUPTED)) {
465                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
466                                                  XFS_ERRLEVEL_LOW, mp);
467                                 goto out;
468                         }
469                 }
470                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
471                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
472                         xfs_iflock(gqp);
473                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
474                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
475                         if (unlikely(error == EFSCORRUPTED)) {
476                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
477                                                  XFS_ERRLEVEL_LOW, mp);
478                                 goto out;
479                         }
480                 }
481         }
482         if (uqp) {
483                  XFS_PURGE_INODE(uqp);
484                  mp->m_quotainfo->qi_uquotaip = NULL;
485         }
486         if (gqp) {
487                 XFS_PURGE_INODE(gqp);
488                 mp->m_quotainfo->qi_gquotaip = NULL;
489         }
490 out:
491         return XFS_ERROR(error);
492 }
493
494 /*
495  * Flush all dquots of the given file system to disk. The dquots are
496  * _not_ purged from memory here, just their data written to disk.
497  */
498 STATIC int
499 xfs_qm_dqflush_all(
500         xfs_mount_t     *mp,
501         int             flags)
502 {
503         int             recl;
504         xfs_dquot_t     *dqp;
505         int             niters;
506         int             error;
507
508         if (mp->m_quotainfo == NULL)
509                 return 0;
510         niters = 0;
511 again:
512         xfs_qm_mplist_lock(mp);
513         FOREACH_DQUOT_IN_MP(dqp, mp) {
514                 xfs_dqlock(dqp);
515                 if (! XFS_DQ_IS_DIRTY(dqp)) {
516                         xfs_dqunlock(dqp);
517                         continue;
518                 }
519                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
520                 /* XXX a sentinel would be better */
521                 recl = XFS_QI_MPLRECLAIMS(mp);
522                 if (! xfs_qm_dqflock_nowait(dqp)) {
523                         /*
524                          * If we can't grab the flush lock then check
525                          * to see if the dquot has been flushed delayed
526                          * write.  If so, grab its buffer and send it
527                          * out immediately.  We'll be able to acquire
528                          * the flush lock when the I/O completes.
529                          */
530                         xfs_qm_dqflock_pushbuf_wait(dqp);
531                 }
532                 /*
533                  * Let go of the mplist lock. We don't want to hold it
534                  * across a disk write.
535                  */
536                 xfs_qm_mplist_unlock(mp);
537                 error = xfs_qm_dqflush(dqp, flags);
538                 xfs_dqunlock(dqp);
539                 if (error)
540                         return error;
541
542                 xfs_qm_mplist_lock(mp);
543                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
544                         xfs_qm_mplist_unlock(mp);
545                         /* XXX restart limit */
546                         goto again;
547                 }
548         }
549
550         xfs_qm_mplist_unlock(mp);
551         /* return ! busy */
552         return 0;
553 }
554 /*
555  * Release the group dquot pointers the user dquots may be
556  * carrying around as a hint. mplist is locked on entry and exit.
557  */
558 STATIC void
559 xfs_qm_detach_gdquots(
560         xfs_mount_t     *mp)
561 {
562         xfs_dquot_t     *dqp, *gdqp;
563         int             nrecl;
564
565  again:
566         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
567         dqp = XFS_QI_MPLNEXT(mp);
568         while (dqp) {
569                 xfs_dqlock(dqp);
570                 if ((gdqp = dqp->q_gdquot)) {
571                         xfs_dqlock(gdqp);
572                         dqp->q_gdquot = NULL;
573                 }
574                 xfs_dqunlock(dqp);
575
576                 if (gdqp) {
577                         /*
578                          * Can't hold the mplist lock across a dqput.
579                          * XXXmust convert to marker based iterations here.
580                          */
581                         nrecl = XFS_QI_MPLRECLAIMS(mp);
582                         xfs_qm_mplist_unlock(mp);
583                         xfs_qm_dqput(gdqp);
584
585                         xfs_qm_mplist_lock(mp);
586                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
587                                 goto again;
588                 }
589                 dqp = dqp->MPL_NEXT;
590         }
591 }
592
593 /*
594  * Go through all the incore dquots of this file system and take them
595  * off the mplist and hashlist, if the dquot type matches the dqtype
596  * parameter. This is used when turning off quota accounting for
597  * users and/or groups, as well as when the filesystem is unmounting.
598  */
599 STATIC int
600 xfs_qm_dqpurge_int(
601         xfs_mount_t     *mp,
602         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
603 {
604         xfs_dquot_t     *dqp;
605         uint            dqtype;
606         int             nrecl;
607         xfs_dquot_t     *nextdqp;
608         int             nmisses;
609
610         if (mp->m_quotainfo == NULL)
611                 return 0;
612
613         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
614         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
615         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
616
617         xfs_qm_mplist_lock(mp);
618
619         /*
620          * In the first pass through all incore dquots of this filesystem,
621          * we release the group dquot pointers the user dquots may be
622          * carrying around as a hint. We need to do this irrespective of
623          * what's being turned off.
624          */
625         xfs_qm_detach_gdquots(mp);
626
627       again:
628         nmisses = 0;
629         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
630         /*
631          * Try to get rid of all of the unwanted dquots. The idea is to
632          * get them off mplist and hashlist, but leave them on freelist.
633          */
634         dqp = XFS_QI_MPLNEXT(mp);
635         while (dqp) {
636                 /*
637                  * It's OK to look at the type without taking dqlock here.
638                  * We're holding the mplist lock here, and that's needed for
639                  * a dqreclaim.
640                  */
641                 if ((dqp->dq_flags & dqtype) == 0) {
642                         dqp = dqp->MPL_NEXT;
643                         continue;
644                 }
645
646                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
647                         nrecl = XFS_QI_MPLRECLAIMS(mp);
648                         xfs_qm_mplist_unlock(mp);
649                         XFS_DQ_HASH_LOCK(dqp->q_hash);
650                         xfs_qm_mplist_lock(mp);
651
652                         /*
653                          * XXXTheoretically, we can get into a very long
654                          * ping pong game here.
655                          * No one can be adding dquots to the mplist at
656                          * this point, but somebody might be taking things off.
657                          */
658                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
659                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
660                                 goto again;
661                         }
662                 }
663
664                 /*
665                  * Take the dquot off the mplist and hashlist. It may remain on
666                  * freelist in INACTIVE state.
667                  */
668                 nextdqp = dqp->MPL_NEXT;
669                 nmisses += xfs_qm_dqpurge(dqp, flags);
670                 dqp = nextdqp;
671         }
672         xfs_qm_mplist_unlock(mp);
673         return nmisses;
674 }
675
676 int
677 xfs_qm_dqpurge_all(
678         xfs_mount_t     *mp,
679         uint            flags)
680 {
681         int             ndquots;
682
683         /*
684          * Purge the dquot cache.
685          * None of the dquots should really be busy at this point.
686          */
687         if (mp->m_quotainfo) {
688                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
689                         delay(ndquots * 10);
690                 }
691         }
692         return 0;
693 }
694
695 STATIC int
696 xfs_qm_dqattach_one(
697         xfs_inode_t     *ip,
698         xfs_dqid_t      id,
699         uint            type,
700         uint            doalloc,
701         uint            dolock,
702         xfs_dquot_t     *udqhint, /* hint */
703         xfs_dquot_t     **IO_idqpp)
704 {
705         xfs_dquot_t     *dqp;
706         int             error;
707
708         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
709         error = 0;
710         /*
711          * See if we already have it in the inode itself. IO_idqpp is
712          * &i_udquot or &i_gdquot. This made the code look weird, but
713          * made the logic a lot simpler.
714          */
715         if ((dqp = *IO_idqpp)) {
716                 if (dolock)
717                         xfs_dqlock(dqp);
718                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
719                 goto done;
720         }
721
722         /*
723          * udqhint is the i_udquot field in inode, and is non-NULL only
724          * when the type arg is group/project. Its purpose is to save a
725          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
726          * the user dquot.
727          */
728         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
729         if (udqhint && !dolock)
730                 xfs_dqlock(udqhint);
731
732         /*
733          * No need to take dqlock to look at the id.
734          * The ID can't change until it gets reclaimed, and it won't
735          * be reclaimed as long as we have a ref from inode and we hold
736          * the ilock.
737          */
738         if (udqhint &&
739             (dqp = udqhint->q_gdquot) &&
740             (be32_to_cpu(dqp->q_core.d_id) == id)) {
741                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
742                 xfs_dqlock(dqp);
743                 XFS_DQHOLD(dqp);
744                 ASSERT(*IO_idqpp == NULL);
745                 *IO_idqpp = dqp;
746                 if (!dolock) {
747                         xfs_dqunlock(dqp);
748                         xfs_dqunlock(udqhint);
749                 }
750                 goto done;
751         }
752         /*
753          * We can't hold a dquot lock when we call the dqget code.
754          * We'll deadlock in no time, because of (not conforming to)
755          * lock ordering - the inodelock comes before any dquot lock,
756          * and we may drop and reacquire the ilock in xfs_qm_dqget().
757          */
758         if (udqhint)
759                 xfs_dqunlock(udqhint);
760         /*
761          * Find the dquot from somewhere. This bumps the
762          * reference count of dquot and returns it locked.
763          * This can return ENOENT if dquot didn't exist on
764          * disk and we didn't ask it to allocate;
765          * ESRCH if quotas got turned off suddenly.
766          */
767         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
768                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
769                 if (udqhint && dolock)
770                         xfs_dqlock(udqhint);
771                 goto done;
772         }
773
774         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
775         /*
776          * dqget may have dropped and re-acquired the ilock, but it guarantees
777          * that the dquot returned is the one that should go in the inode.
778          */
779         *IO_idqpp = dqp;
780         ASSERT(dqp);
781         ASSERT(XFS_DQ_IS_LOCKED(dqp));
782         if (! dolock) {
783                 xfs_dqunlock(dqp);
784                 goto done;
785         }
786         if (! udqhint)
787                 goto done;
788
789         ASSERT(udqhint);
790         ASSERT(dolock);
791         ASSERT(XFS_DQ_IS_LOCKED(dqp));
792         if (! xfs_qm_dqlock_nowait(udqhint)) {
793                 xfs_dqunlock(dqp);
794                 xfs_dqlock(udqhint);
795                 xfs_dqlock(dqp);
796         }
797       done:
798 #ifdef QUOTADEBUG
799         if (udqhint) {
800                 if (dolock)
801                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
802         }
803         if (! error) {
804                 if (dolock)
805                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
806         }
807 #endif
808         return error;
809 }
810
811
812 /*
813  * Given a udquot and gdquot, attach a ptr to the group dquot in the
814  * udquot as a hint for future lookups. The idea sounds simple, but the
815  * execution isn't, because the udquot might have a group dquot attached
816  * already and getting rid of that gets us into lock ordering constraints.
817  * The process is complicated more by the fact that the dquots may or may not
818  * be locked on entry.
819  */
820 STATIC void
821 xfs_qm_dqattach_grouphint(
822         xfs_dquot_t     *udq,
823         xfs_dquot_t     *gdq,
824         uint            locked)
825 {
826         xfs_dquot_t     *tmp;
827
828 #ifdef QUOTADEBUG
829         if (locked) {
830                 ASSERT(XFS_DQ_IS_LOCKED(udq));
831                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
832         }
833 #endif
834         if (! locked)
835                 xfs_dqlock(udq);
836
837         if ((tmp = udq->q_gdquot)) {
838                 if (tmp == gdq) {
839                         if (! locked)
840                                 xfs_dqunlock(udq);
841                         return;
842                 }
843
844                 udq->q_gdquot = NULL;
845                 /*
846                  * We can't keep any dqlocks when calling dqrele,
847                  * because the freelist lock comes before dqlocks.
848                  */
849                 xfs_dqunlock(udq);
850                 if (locked)
851                         xfs_dqunlock(gdq);
852                 /*
853                  * we took a hard reference once upon a time in dqget,
854                  * so give it back when the udquot no longer points at it
855                  * dqput() does the unlocking of the dquot.
856                  */
857                 xfs_qm_dqrele(tmp);
858
859                 xfs_dqlock(udq);
860                 xfs_dqlock(gdq);
861
862         } else {
863                 ASSERT(XFS_DQ_IS_LOCKED(udq));
864                 if (! locked) {
865                         xfs_dqlock(gdq);
866                 }
867         }
868
869         ASSERT(XFS_DQ_IS_LOCKED(udq));
870         ASSERT(XFS_DQ_IS_LOCKED(gdq));
871         /*
872          * Somebody could have attached a gdquot here,
873          * when we dropped the uqlock. If so, just do nothing.
874          */
875         if (udq->q_gdquot == NULL) {
876                 XFS_DQHOLD(gdq);
877                 udq->q_gdquot = gdq;
878         }
879         if (! locked) {
880                 xfs_dqunlock(gdq);
881                 xfs_dqunlock(udq);
882         }
883 }
884
885
886 /*
887  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
888  * into account.
889  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
890  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
891  * much made this code a complete mess, but it has been pretty useful.
892  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
893  * Inode may get unlocked and relocked in here, and the caller must deal with
894  * the consequences.
895  */
896 int
897 xfs_qm_dqattach(
898         xfs_inode_t     *ip,
899         uint            flags)
900 {
901         xfs_mount_t     *mp = ip->i_mount;
902         uint            nquotas = 0;
903         int             error = 0;
904
905         if ((! XFS_IS_QUOTA_ON(mp)) ||
906             (! XFS_NOT_DQATTACHED(mp, ip)) ||
907             (ip->i_ino == mp->m_sb.sb_uquotino) ||
908             (ip->i_ino == mp->m_sb.sb_gquotino))
909                 return 0;
910
911         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
912                XFS_ISLOCKED_INODE_EXCL(ip));
913
914         if (! (flags & XFS_QMOPT_ILOCKED))
915                 xfs_ilock(ip, XFS_ILOCK_EXCL);
916
917         if (XFS_IS_UQUOTA_ON(mp)) {
918                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
919                                                 flags & XFS_QMOPT_DQALLOC,
920                                                 flags & XFS_QMOPT_DQLOCK,
921                                                 NULL, &ip->i_udquot);
922                 if (error)
923                         goto done;
924                 nquotas++;
925         }
926         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
927         if (XFS_IS_OQUOTA_ON(mp)) {
928                 error = XFS_IS_GQUOTA_ON(mp) ?
929                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
930                                                 flags & XFS_QMOPT_DQALLOC,
931                                                 flags & XFS_QMOPT_DQLOCK,
932                                                 ip->i_udquot, &ip->i_gdquot) :
933                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
934                                                 flags & XFS_QMOPT_DQALLOC,
935                                                 flags & XFS_QMOPT_DQLOCK,
936                                                 ip->i_udquot, &ip->i_gdquot);
937                 /*
938                  * Don't worry about the udquot that we may have
939                  * attached above. It'll get detached, if not already.
940                  */
941                 if (error)
942                         goto done;
943                 nquotas++;
944         }
945
946         /*
947          * Attach this group quota to the user quota as a hint.
948          * This WON'T, in general, result in a thrash.
949          */
950         if (nquotas == 2) {
951                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
952                 ASSERT(ip->i_udquot);
953                 ASSERT(ip->i_gdquot);
954
955                 /*
956                  * We may or may not have the i_udquot locked at this point,
957                  * but this check is OK since we don't depend on the i_gdquot to
958                  * be accurate 100% all the time. It is just a hint, and this
959                  * will succeed in general.
960                  */
961                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
962                         goto done;
963                 /*
964                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
965                  */
966                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
967                                          flags & XFS_QMOPT_DQLOCK);
968         }
969
970       done:
971
972 #ifdef QUOTADEBUG
973         if (! error) {
974                 if (ip->i_udquot) {
975                         if (flags & XFS_QMOPT_DQLOCK)
976                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
977                 }
978                 if (ip->i_gdquot) {
979                         if (flags & XFS_QMOPT_DQLOCK)
980                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
981                 }
982                 if (XFS_IS_UQUOTA_ON(mp))
983                         ASSERT(ip->i_udquot);
984                 if (XFS_IS_OQUOTA_ON(mp))
985                         ASSERT(ip->i_gdquot);
986         }
987 #endif
988
989         if (! (flags & XFS_QMOPT_ILOCKED))
990                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
991
992 #ifdef QUOTADEBUG
993         else
994                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
995 #endif
996         return error;
997 }
998
999 /*
1000  * Release dquots (and their references) if any.
1001  * The inode should be locked EXCL except when this's called by
1002  * xfs_ireclaim.
1003  */
1004 void
1005 xfs_qm_dqdetach(
1006         xfs_inode_t     *ip)
1007 {
1008         if (!(ip->i_udquot || ip->i_gdquot))
1009                 return;
1010
1011         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1012         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1013         if (ip->i_udquot) {
1014                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1015                 xfs_qm_dqrele(ip->i_udquot);
1016                 ip->i_udquot = NULL;
1017         }
1018         if (ip->i_gdquot) {
1019                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
1020                 xfs_qm_dqrele(ip->i_gdquot);
1021                 ip->i_gdquot = NULL;
1022         }
1023 }
1024
1025 /*
1026  * This is called by VFS_SYNC and flags arg determines the caller,
1027  * and its motives, as done in xfs_sync.
1028  *
1029  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1030  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1031  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1032  */
1033
1034 int
1035 xfs_qm_sync(
1036         xfs_mount_t     *mp,
1037         short           flags)
1038 {
1039         int             recl, restarts;
1040         xfs_dquot_t     *dqp;
1041         uint            flush_flags;
1042         boolean_t       nowait;
1043         int             error;
1044
1045         restarts = 0;
1046         /*
1047          * We won't block unless we are asked to.
1048          */
1049         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1050
1051   again:
1052         xfs_qm_mplist_lock(mp);
1053         /*
1054          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1055          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1056          * when we have the mplist lock, we know that dquots will be consistent
1057          * as long as we have it locked.
1058          */
1059         if (! XFS_IS_QUOTA_ON(mp)) {
1060                 xfs_qm_mplist_unlock(mp);
1061                 return 0;
1062         }
1063         FOREACH_DQUOT_IN_MP(dqp, mp) {
1064                 /*
1065                  * If this is vfs_sync calling, then skip the dquots that
1066                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1067                  * This is very similar to what xfs_sync does with inodes.
1068                  */
1069                 if (flags & SYNC_BDFLUSH) {
1070                         if (! XFS_DQ_IS_DIRTY(dqp))
1071                                 continue;
1072                 }
1073
1074                 if (nowait) {
1075                         /*
1076                          * Try to acquire the dquot lock. We are NOT out of
1077                          * lock order, but we just don't want to wait for this
1078                          * lock, unless somebody wanted us to.
1079                          */
1080                         if (! xfs_qm_dqlock_nowait(dqp))
1081                                 continue;
1082                 } else {
1083                         xfs_dqlock(dqp);
1084                 }
1085
1086                 /*
1087                  * Now, find out for sure if this dquot is dirty or not.
1088                  */
1089                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1090                         xfs_dqunlock(dqp);
1091                         continue;
1092                 }
1093
1094                 /* XXX a sentinel would be better */
1095                 recl = XFS_QI_MPLRECLAIMS(mp);
1096                 if (! xfs_qm_dqflock_nowait(dqp)) {
1097                         if (nowait) {
1098                                 xfs_dqunlock(dqp);
1099                                 continue;
1100                         }
1101                         /*
1102                          * If we can't grab the flush lock then if the caller
1103                          * really wanted us to give this our best shot, so
1104                          * see if we can give a push to the buffer before we wait
1105                          * on the flush lock. At this point, we know that
1106                          * even though the dquot is being flushed,
1107                          * it has (new) dirty data.
1108                          */
1109                         xfs_qm_dqflock_pushbuf_wait(dqp);
1110                 }
1111                 /*
1112                  * Let go of the mplist lock. We don't want to hold it
1113                  * across a disk write
1114                  */
1115                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1116                 xfs_qm_mplist_unlock(mp);
1117                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1118                 error = xfs_qm_dqflush(dqp, flush_flags);
1119                 xfs_dqunlock(dqp);
1120                 if (error && XFS_FORCED_SHUTDOWN(mp))
1121                         return 0;       /* Need to prevent umount failure */
1122                 else if (error)
1123                         return error;
1124
1125                 xfs_qm_mplist_lock(mp);
1126                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1127                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1128                                 break;
1129
1130                         xfs_qm_mplist_unlock(mp);
1131                         goto again;
1132                 }
1133         }
1134
1135         xfs_qm_mplist_unlock(mp);
1136         return 0;
1137 }
1138
1139
1140 /*
1141  * This initializes all the quota information that's kept in the
1142  * mount structure
1143  */
1144 STATIC int
1145 xfs_qm_init_quotainfo(
1146         xfs_mount_t     *mp)
1147 {
1148         xfs_quotainfo_t *qinf;
1149         int             error;
1150         xfs_dquot_t     *dqp;
1151
1152         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1153
1154         /*
1155          * Tell XQM that we exist as soon as possible.
1156          */
1157         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1158                 return error;
1159         }
1160
1161         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1162
1163         /*
1164          * See if quotainodes are setup, and if not, allocate them,
1165          * and change the superblock accordingly.
1166          */
1167         if ((error = xfs_qm_init_quotainos(mp))) {
1168                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1169                 mp->m_quotainfo = NULL;
1170                 return error;
1171         }
1172
1173         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1174         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1175         qinf->qi_dqreclaims = 0;
1176
1177         /* mutex used to serialize quotaoffs */
1178         mutex_init(&qinf->qi_quotaofflock);
1179
1180         /* Precalc some constants */
1181         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1182         ASSERT(qinf->qi_dqchunklen);
1183         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1184         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1185
1186         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1187
1188         /*
1189          * We try to get the limits from the superuser's limits fields.
1190          * This is quite hacky, but it is standard quota practice.
1191          * We look at the USR dquot with id == 0 first, but if user quotas
1192          * are not enabled we goto the GRP dquot with id == 0.
1193          * We don't really care to keep separate default limits for user
1194          * and group quotas, at least not at this point.
1195          */
1196         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1197                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1198                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1199                                 XFS_DQ_PROJ),
1200                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1201                              &dqp);
1202         if (! error) {
1203                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1204
1205                 /*
1206                  * The warnings and timers set the grace period given to
1207                  * a user or group before he or she can not perform any
1208                  * more writing. If it is zero, a default is used.
1209                  */
1210                 qinf->qi_btimelimit = ddqp->d_btimer ?
1211                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1212                 qinf->qi_itimelimit = ddqp->d_itimer ?
1213                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1214                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1215                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1216                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1217                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1218                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1219                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1220                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1221                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1222                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1223                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1224                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1225                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1226                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1227                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1228  
1229                 /*
1230                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1231                  * we don't want this dquot cached. We haven't done a
1232                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1233                  */
1234                 xfs_qm_dqdestroy(dqp);
1235         } else {
1236                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1237                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1238                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1239                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1240                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1241                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1242         }
1243
1244         return 0;
1245 }
1246
1247
1248 /*
1249  * Gets called when unmounting a filesystem or when all quotas get
1250  * turned off.
1251  * This purges the quota inodes, destroys locks and frees itself.
1252  */
1253 void
1254 xfs_qm_destroy_quotainfo(
1255         xfs_mount_t     *mp)
1256 {
1257         xfs_quotainfo_t *qi;
1258
1259         qi = mp->m_quotainfo;
1260         ASSERT(qi != NULL);
1261         ASSERT(xfs_Gqm != NULL);
1262
1263         /*
1264          * Release the reference that XQM kept, so that we know
1265          * when the XQM structure should be freed. We cannot assume
1266          * that xfs_Gqm is non-null after this point.
1267          */
1268         xfs_qm_rele_quotafs_ref(mp);
1269
1270         spinlock_destroy(&qi->qi_pinlock);
1271         xfs_qm_list_destroy(&qi->qi_dqlist);
1272
1273         if (qi->qi_uquotaip) {
1274                 XFS_PURGE_INODE(qi->qi_uquotaip);
1275                 qi->qi_uquotaip = NULL; /* paranoia */
1276         }
1277         if (qi->qi_gquotaip) {
1278                 XFS_PURGE_INODE(qi->qi_gquotaip);
1279                 qi->qi_gquotaip = NULL;
1280         }
1281         mutex_destroy(&qi->qi_quotaofflock);
1282         kmem_free(qi, sizeof(xfs_quotainfo_t));
1283         mp->m_quotainfo = NULL;
1284 }
1285
1286
1287
1288 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1289
1290 /* ARGSUSED */
1291 STATIC void
1292 xfs_qm_list_init(
1293         xfs_dqlist_t    *list,
1294         char            *str,
1295         int             n)
1296 {
1297         mutex_init(&list->qh_lock);
1298         list->qh_next = NULL;
1299         list->qh_version = 0;
1300         list->qh_nelems = 0;
1301 }
1302
1303 STATIC void
1304 xfs_qm_list_destroy(
1305         xfs_dqlist_t    *list)
1306 {
1307         mutex_destroy(&(list->qh_lock));
1308 }
1309
1310
1311 /*
1312  * Stripped down version of dqattach. This doesn't attach, or even look at the
1313  * dquots attached to the inode. The rationale is that there won't be any
1314  * attached at the time this is called from quotacheck.
1315  */
1316 STATIC int
1317 xfs_qm_dqget_noattach(
1318         xfs_inode_t     *ip,
1319         xfs_dquot_t     **O_udqpp,
1320         xfs_dquot_t     **O_gdqpp)
1321 {
1322         int             error;
1323         xfs_mount_t     *mp;
1324         xfs_dquot_t     *udqp, *gdqp;
1325
1326         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1327         mp = ip->i_mount;
1328         udqp = NULL;
1329         gdqp = NULL;
1330
1331         if (XFS_IS_UQUOTA_ON(mp)) {
1332                 ASSERT(ip->i_udquot == NULL);
1333                 /*
1334                  * We want the dquot allocated if it doesn't exist.
1335                  */
1336                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1337                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1338                                          &udqp))) {
1339                         /*
1340                          * Shouldn't be able to turn off quotas here.
1341                          */
1342                         ASSERT(error != ESRCH);
1343                         ASSERT(error != ENOENT);
1344                         return error;
1345                 }
1346                 ASSERT(udqp);
1347         }
1348
1349         if (XFS_IS_OQUOTA_ON(mp)) {
1350                 ASSERT(ip->i_gdquot == NULL);
1351                 if (udqp)
1352                         xfs_dqunlock(udqp);
1353                 error = XFS_IS_GQUOTA_ON(mp) ?
1354                                 xfs_qm_dqget(mp, ip,
1355                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1356                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1357                                              &gdqp) :
1358                                 xfs_qm_dqget(mp, ip,
1359                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1360                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1361                                              &gdqp);
1362                 if (error) {
1363                         if (udqp)
1364                                 xfs_qm_dqrele(udqp);
1365                         ASSERT(error != ESRCH);
1366                         ASSERT(error != ENOENT);
1367                         return error;
1368                 }
1369                 ASSERT(gdqp);
1370
1371                 /* Reacquire the locks in the right order */
1372                 if (udqp) {
1373                         if (! xfs_qm_dqlock_nowait(udqp)) {
1374                                 xfs_dqunlock(gdqp);
1375                                 xfs_dqlock(udqp);
1376                                 xfs_dqlock(gdqp);
1377                         }
1378                 }
1379         }
1380
1381         *O_udqpp = udqp;
1382         *O_gdqpp = gdqp;
1383
1384 #ifdef QUOTADEBUG
1385         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1386         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1387 #endif
1388         return 0;
1389 }
1390
1391 /*
1392  * Create an inode and return with a reference already taken, but unlocked
1393  * This is how we create quota inodes
1394  */
1395 STATIC int
1396 xfs_qm_qino_alloc(
1397         xfs_mount_t     *mp,
1398         xfs_inode_t     **ip,
1399         __int64_t       sbfields,
1400         uint            flags)
1401 {
1402         xfs_trans_t     *tp;
1403         int             error;
1404         unsigned long   s;
1405         int             committed;
1406
1407         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1408         if ((error = xfs_trans_reserve(tp,
1409                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1410                                       XFS_CREATE_LOG_RES(mp), 0,
1411                                       XFS_TRANS_PERM_LOG_RES,
1412                                       XFS_CREATE_LOG_COUNT))) {
1413                 xfs_trans_cancel(tp, 0);
1414                 return error;
1415         }
1416
1417         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1418                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1419                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1420                                  XFS_TRANS_ABORT);
1421                 return error;
1422         }
1423
1424         /*
1425          * Keep an extra reference to this quota inode. This inode is
1426          * locked exclusively and joined to the transaction already.
1427          */
1428         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1429         VN_HOLD(XFS_ITOV((*ip)));
1430
1431         /*
1432          * Make the changes in the superblock, and log those too.
1433          * sbfields arg may contain fields other than *QUOTINO;
1434          * VERSIONNUM for example.
1435          */
1436         s = XFS_SB_LOCK(mp);
1437         if (flags & XFS_QMOPT_SBVERSION) {
1438 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1439                 unsigned oldv = mp->m_sb.sb_versionnum;
1440 #endif
1441                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1442                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1443                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1444                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1445                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1446
1447                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1448                 mp->m_sb.sb_uquotino = NULLFSINO;
1449                 mp->m_sb.sb_gquotino = NULLFSINO;
1450
1451                 /* qflags will get updated _after_ quotacheck */
1452                 mp->m_sb.sb_qflags = 0;
1453 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1454                 cmn_err(CE_NOTE,
1455                         "Old superblock version %x, converting to %x.",
1456                         oldv, mp->m_sb.sb_versionnum);
1457 #endif
1458         }
1459         if (flags & XFS_QMOPT_UQUOTA)
1460                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1461         else
1462                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1463         XFS_SB_UNLOCK(mp, s);
1464         xfs_mod_sb(tp, sbfields);
1465
1466         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1467                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1468                 return error;
1469         }
1470         return 0;
1471 }
1472
1473
1474 STATIC int
1475 xfs_qm_reset_dqcounts(
1476         xfs_mount_t     *mp,
1477         xfs_buf_t       *bp,
1478         xfs_dqid_t      id,
1479         uint            type)
1480 {
1481         xfs_disk_dquot_t        *ddq;
1482         int                     j;
1483
1484         xfs_buftrace("RESET DQUOTS", bp);
1485         /*
1486          * Reset all counters and timers. They'll be
1487          * started afresh by xfs_qm_quotacheck.
1488          */
1489 #ifdef DEBUG
1490         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1491         do_div(j, sizeof(xfs_dqblk_t));
1492         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1493 #endif
1494         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1495         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1496                 /*
1497                  * Do a sanity check, and if needed, repair the dqblk. Don't
1498                  * output any warnings because it's perfectly possible to
1499                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1500                  */
1501                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1502                                       "xfs_quotacheck");
1503                 ddq->d_bcount = 0;
1504                 ddq->d_icount = 0;
1505                 ddq->d_rtbcount = 0;
1506                 ddq->d_btimer = 0;
1507                 ddq->d_itimer = 0;
1508                 ddq->d_rtbtimer = 0;
1509                 ddq->d_bwarns = 0;
1510                 ddq->d_iwarns = 0;
1511                 ddq->d_rtbwarns = 0;
1512                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1513         }
1514
1515         return 0;
1516 }
1517
1518 STATIC int
1519 xfs_qm_dqiter_bufs(
1520         xfs_mount_t     *mp,
1521         xfs_dqid_t      firstid,
1522         xfs_fsblock_t   bno,
1523         xfs_filblks_t   blkcnt,
1524         uint            flags)
1525 {
1526         xfs_buf_t       *bp;
1527         int             error;
1528         int             notcommitted;
1529         int             incr;
1530         int             type;
1531
1532         ASSERT(blkcnt > 0);
1533         notcommitted = 0;
1534         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1535                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1536         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1537                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1538         error = 0;
1539
1540         /*
1541          * Blkcnt arg can be a very big number, and might even be
1542          * larger than the log itself. So, we have to break it up into
1543          * manageable-sized transactions.
1544          * Note that we don't start a permanent transaction here; we might
1545          * not be able to get a log reservation for the whole thing up front,
1546          * and we don't really care to either, because we just discard
1547          * everything if we were to crash in the middle of this loop.
1548          */
1549         while (blkcnt--) {
1550                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1551                               XFS_FSB_TO_DADDR(mp, bno),
1552                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1553                 if (error)
1554                         break;
1555
1556                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1557                 xfs_bdwrite(mp, bp);
1558                 /*
1559                  * goto the next block.
1560                  */
1561                 bno++;
1562                 firstid += XFS_QM_DQPERBLK(mp);
1563         }
1564         return error;
1565 }
1566
1567 /*
1568  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1569  * caller supplied function for every chunk of dquots that we find.
1570  */
1571 STATIC int
1572 xfs_qm_dqiterate(
1573         xfs_mount_t     *mp,
1574         xfs_inode_t     *qip,
1575         uint            flags)
1576 {
1577         xfs_bmbt_irec_t         *map;
1578         int                     i, nmaps;       /* number of map entries */
1579         int                     error;          /* return value */
1580         xfs_fileoff_t           lblkno;
1581         xfs_filblks_t           maxlblkcnt;
1582         xfs_dqid_t              firstid;
1583         xfs_fsblock_t           rablkno;
1584         xfs_filblks_t           rablkcnt;
1585
1586         error = 0;
1587         /*
1588          * This looks racy, but we can't keep an inode lock across a
1589          * trans_reserve. But, this gets called during quotacheck, and that
1590          * happens only at mount time which is single threaded.
1591          */
1592         if (qip->i_d.di_nblocks == 0)
1593                 return 0;
1594
1595         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1596
1597         lblkno = 0;
1598         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1599         do {
1600                 nmaps = XFS_DQITER_MAP_SIZE;
1601                 /*
1602                  * We aren't changing the inode itself. Just changing
1603                  * some of its data. No new blocks are added here, and
1604                  * the inode is never added to the transaction.
1605                  */
1606                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1607                 error = xfs_bmapi(NULL, qip, lblkno,
1608                                   maxlblkcnt - lblkno,
1609                                   XFS_BMAPI_METADATA,
1610                                   NULL,
1611                                   0, map, &nmaps, NULL, NULL);
1612                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1613                 if (error)
1614                         break;
1615
1616                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1617                 for (i = 0; i < nmaps; i++) {
1618                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1619                         ASSERT(map[i].br_blockcount);
1620
1621
1622                         lblkno += map[i].br_blockcount;
1623
1624                         if (map[i].br_startblock == HOLESTARTBLOCK)
1625                                 continue;
1626
1627                         firstid = (xfs_dqid_t) map[i].br_startoff *
1628                                 XFS_QM_DQPERBLK(mp);
1629                         /*
1630                          * Do a read-ahead on the next extent.
1631                          */
1632                         if ((i+1 < nmaps) &&
1633                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1634                                 rablkcnt =  map[i+1].br_blockcount;
1635                                 rablkno = map[i+1].br_startblock;
1636                                 while (rablkcnt--) {
1637                                         xfs_baread(mp->m_ddev_targp,
1638                                                XFS_FSB_TO_DADDR(mp, rablkno),
1639                                                (int)XFS_QI_DQCHUNKLEN(mp));
1640                                         rablkno++;
1641                                 }
1642                         }
1643                         /*
1644                          * Iterate thru all the blks in the extent and
1645                          * reset the counters of all the dquots inside them.
1646                          */
1647                         if ((error = xfs_qm_dqiter_bufs(mp,
1648                                                        firstid,
1649                                                        map[i].br_startblock,
1650                                                        map[i].br_blockcount,
1651                                                        flags))) {
1652                                 break;
1653                         }
1654                 }
1655
1656                 if (error)
1657                         break;
1658         } while (nmaps > 0);
1659
1660         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1661
1662         return error;
1663 }
1664
1665 /*
1666  * Called by dqusage_adjust in doing a quotacheck.
1667  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1668  * this updates its incore copy as well as the buffer copy. This is
1669  * so that once the quotacheck is done, we can just log all the buffers,
1670  * as opposed to logging numerous updates to individual dquots.
1671  */
1672 STATIC void
1673 xfs_qm_quotacheck_dqadjust(
1674         xfs_dquot_t             *dqp,
1675         xfs_qcnt_t              nblks,
1676         xfs_qcnt_t              rtblks)
1677 {
1678         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1679         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1680         /*
1681          * Adjust the inode count and the block count to reflect this inode's
1682          * resource usage.
1683          */
1684         be64_add(&dqp->q_core.d_icount, 1);
1685         dqp->q_res_icount++;
1686         if (nblks) {
1687                 be64_add(&dqp->q_core.d_bcount, nblks);
1688                 dqp->q_res_bcount += nblks;
1689         }
1690         if (rtblks) {
1691                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1692                 dqp->q_res_rtbcount += rtblks;
1693         }
1694
1695         /*
1696          * Set default limits, adjust timers (since we changed usages)
1697          */
1698         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1699                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1700                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1701         }
1702
1703         dqp->dq_flags |= XFS_DQ_DIRTY;
1704 }
1705
1706 STATIC int
1707 xfs_qm_get_rtblks(
1708         xfs_inode_t     *ip,
1709         xfs_qcnt_t      *O_rtblks)
1710 {
1711         xfs_filblks_t   rtblks;                 /* total rt blks */
1712         xfs_extnum_t    idx;                    /* extent record index */
1713         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1714         xfs_extnum_t    nextents;               /* number of extent entries */
1715         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1716         int             error;
1717
1718         ASSERT(XFS_IS_REALTIME_INODE(ip));
1719         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1720         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1721                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1722                         return error;
1723         }
1724         rtblks = 0;
1725         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1726         for (idx = 0; idx < nextents; idx++) {
1727                 ep = xfs_iext_get_ext(ifp, idx);
1728                 rtblks += xfs_bmbt_get_blockcount(ep);
1729         }
1730         *O_rtblks = (xfs_qcnt_t)rtblks;
1731         return 0;
1732 }
1733
1734 /*
1735  * callback routine supplied to bulkstat(). Given an inumber, find its
1736  * dquots and update them to account for resources taken by that inode.
1737  */
1738 /* ARGSUSED */
1739 STATIC int
1740 xfs_qm_dqusage_adjust(
1741         xfs_mount_t     *mp,            /* mount point for filesystem */
1742         xfs_ino_t       ino,            /* inode number to get data for */
1743         void            __user *buffer, /* not used */
1744         int             ubsize,         /* not used */
1745         void            *private_data,  /* not used */
1746         xfs_daddr_t     bno,            /* starting block of inode cluster */
1747         int             *ubused,        /* not used */
1748         void            *dip,           /* on-disk inode pointer (not used) */
1749         int             *res)           /* result code value */
1750 {
1751         xfs_inode_t     *ip;
1752         xfs_dquot_t     *udqp, *gdqp;
1753         xfs_qcnt_t      nblks, rtblks;
1754         int             error;
1755
1756         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1757
1758         /*
1759          * rootino must have its resources accounted for, not so with the quota
1760          * inodes.
1761          */
1762         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1763                 *res = BULKSTAT_RV_NOTHING;
1764                 return XFS_ERROR(EINVAL);
1765         }
1766
1767         /*
1768          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1769          * interface expects the inode to be exclusively locked because that's
1770          * the case in all other instances. It's OK that we do this because
1771          * quotacheck is done only at mount time.
1772          */
1773         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1774                 *res = BULKSTAT_RV_NOTHING;
1775                 return error;
1776         }
1777
1778         if (ip->i_d.di_mode == 0) {
1779                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1780                 *res = BULKSTAT_RV_NOTHING;
1781                 return XFS_ERROR(ENOENT);
1782         }
1783
1784         /*
1785          * Obtain the locked dquots. In case of an error (eg. allocation
1786          * fails for ENOSPC), we return the negative of the error number
1787          * to bulkstat, so that it can get propagated to quotacheck() and
1788          * making us disable quotas for the file system.
1789          */
1790         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1791                 xfs_iput(ip, XFS_ILOCK_EXCL);
1792                 *res = BULKSTAT_RV_GIVEUP;
1793                 return error;
1794         }
1795
1796         rtblks = 0;
1797         if (! XFS_IS_REALTIME_INODE(ip)) {
1798                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1799         } else {
1800                 /*
1801                  * Walk thru the extent list and count the realtime blocks.
1802                  */
1803                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1804                         xfs_iput(ip, XFS_ILOCK_EXCL);
1805                         if (udqp)
1806                                 xfs_qm_dqput(udqp);
1807                         if (gdqp)
1808                                 xfs_qm_dqput(gdqp);
1809                         *res = BULKSTAT_RV_GIVEUP;
1810                         return error;
1811                 }
1812                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1813         }
1814         ASSERT(ip->i_delayed_blks == 0);
1815
1816         /*
1817          * We can't release the inode while holding its dquot locks.
1818          * The inode can go into inactive and might try to acquire the dquotlocks.
1819          * So, just unlock here and do a vn_rele at the end.
1820          */
1821         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1822
1823         /*
1824          * Add the (disk blocks and inode) resources occupied by this
1825          * inode to its dquots. We do this adjustment in the incore dquot,
1826          * and also copy the changes to its buffer.
1827          * We don't care about putting these changes in a transaction
1828          * envelope because if we crash in the middle of a 'quotacheck'
1829          * we have to start from the beginning anyway.
1830          * Once we're done, we'll log all the dquot bufs.
1831          *
1832          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1833          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1834          */
1835         if (XFS_IS_UQUOTA_ON(mp)) {
1836                 ASSERT(udqp);
1837                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1838                 xfs_qm_dqput(udqp);
1839         }
1840         if (XFS_IS_OQUOTA_ON(mp)) {
1841                 ASSERT(gdqp);
1842                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1843                 xfs_qm_dqput(gdqp);
1844         }
1845         /*
1846          * Now release the inode. This will send it to 'inactive', and
1847          * possibly even free blocks.
1848          */
1849         VN_RELE(XFS_ITOV(ip));
1850
1851         /*
1852          * Goto next inode.
1853          */
1854         *res = BULKSTAT_RV_DIDONE;
1855         return 0;
1856 }
1857
1858 /*
1859  * Walk thru all the filesystem inodes and construct a consistent view
1860  * of the disk quota world. If the quotacheck fails, disable quotas.
1861  */
1862 int
1863 xfs_qm_quotacheck(
1864         xfs_mount_t     *mp)
1865 {
1866         int             done, count, error;
1867         xfs_ino_t       lastino;
1868         size_t          structsz;
1869         xfs_inode_t     *uip, *gip;
1870         uint            flags;
1871
1872         count = INT_MAX;
1873         structsz = 1;
1874         lastino = 0;
1875         flags = 0;
1876
1877         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1878         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1879
1880         /*
1881          * There should be no cached dquots. The (simplistic) quotacheck
1882          * algorithm doesn't like that.
1883          */
1884         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1885
1886         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1887
1888         /*
1889          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1890          * their counters to zero. We need a clean slate.
1891          * We don't log our changes till later.
1892          */
1893         if ((uip = XFS_QI_UQIP(mp))) {
1894                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1895                         goto error_return;
1896                 flags |= XFS_UQUOTA_CHKD;
1897         }
1898
1899         if ((gip = XFS_QI_GQIP(mp))) {
1900                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1901                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1902                         goto error_return;
1903                 flags |= XFS_OQUOTA_CHKD;
1904         }
1905
1906         do {
1907                 /*
1908                  * Iterate thru all the inodes in the file system,
1909                  * adjusting the corresponding dquot counters in core.
1910                  */
1911                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1912                                      xfs_qm_dqusage_adjust, NULL,
1913                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1914                         break;
1915
1916         } while (! done);
1917
1918         /*
1919          * We can get this error if we couldn't do a dquot allocation inside
1920          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1921          * dirty dquots that might be cached, we just want to get rid of them
1922          * and turn quotaoff. The dquots won't be attached to any of the inodes
1923          * at this point (because we intentionally didn't in dqget_noattach).
1924          */
1925         if (error) {
1926                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1927                 goto error_return;
1928         }
1929         /*
1930          * We've made all the changes that we need to make incore.
1931          * Now flush_them down to disk buffers.
1932          */
1933         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1934
1935         /*
1936          * We didn't log anything, because if we crashed, we'll have to
1937          * start the quotacheck from scratch anyway. However, we must make
1938          * sure that our dquot changes are secure before we put the
1939          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1940          * flush.
1941          */
1942         XFS_bflush(mp->m_ddev_targp);
1943
1944         /*
1945          * If one type of quotas is off, then it will lose its
1946          * quotachecked status, since we won't be doing accounting for
1947          * that type anymore.
1948          */
1949         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1950         mp->m_qflags |= flags;
1951
1952         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1953
1954  error_return:
1955         if (error) {
1956                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1957                         "Disabling quotas.",
1958                         mp->m_fsname, error);
1959                 /*
1960                  * We must turn off quotas.
1961                  */
1962                 ASSERT(mp->m_quotainfo != NULL);
1963                 ASSERT(xfs_Gqm != NULL);
1964                 xfs_qm_destroy_quotainfo(mp);
1965                 (void)xfs_mount_reset_sbqflags(mp);
1966         } else {
1967                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1968         }
1969         return (error);
1970 }
1971
1972 /*
1973  * This is called after the superblock has been read in and we're ready to
1974  * iget the quota inodes.
1975  */
1976 STATIC int
1977 xfs_qm_init_quotainos(
1978         xfs_mount_t     *mp)
1979 {
1980         xfs_inode_t     *uip, *gip;
1981         int             error;
1982         __int64_t       sbflags;
1983         uint            flags;
1984
1985         ASSERT(mp->m_quotainfo);
1986         uip = gip = NULL;
1987         sbflags = 0;
1988         flags = 0;
1989
1990         /*
1991          * Get the uquota and gquota inodes
1992          */
1993         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1994                 if (XFS_IS_UQUOTA_ON(mp) &&
1995                     mp->m_sb.sb_uquotino != NULLFSINO) {
1996                         ASSERT(mp->m_sb.sb_uquotino > 0);
1997                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
1998                                              0, 0, &uip, 0)))
1999                                 return XFS_ERROR(error);
2000                 }
2001                 if (XFS_IS_OQUOTA_ON(mp) &&
2002                     mp->m_sb.sb_gquotino != NULLFSINO) {
2003                         ASSERT(mp->m_sb.sb_gquotino > 0);
2004                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2005                                              0, 0, &gip, 0))) {
2006                                 if (uip)
2007                                         VN_RELE(XFS_ITOV(uip));
2008                                 return XFS_ERROR(error);
2009                         }
2010                 }
2011         } else {
2012                 flags |= XFS_QMOPT_SBVERSION;
2013                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2014                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2015         }
2016
2017         /*
2018          * Create the two inodes, if they don't exist already. The changes
2019          * made above will get added to a transaction and logged in one of
2020          * the qino_alloc calls below.  If the device is readonly,
2021          * temporarily switch to read-write to do this.
2022          */
2023         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2024                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2025                                               sbflags | XFS_SB_UQUOTINO,
2026                                               flags | XFS_QMOPT_UQUOTA)))
2027                         return XFS_ERROR(error);
2028
2029                 flags &= ~XFS_QMOPT_SBVERSION;
2030         }
2031         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
2032                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
2033                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2034                 error = xfs_qm_qino_alloc(mp, &gip,
2035                                           sbflags | XFS_SB_GQUOTINO, flags);
2036                 if (error) {
2037                         if (uip)
2038                                 VN_RELE(XFS_ITOV(uip));
2039
2040                         return XFS_ERROR(error);
2041                 }
2042         }
2043
2044         XFS_QI_UQIP(mp) = uip;
2045         XFS_QI_GQIP(mp) = gip;
2046
2047         return 0;
2048 }
2049
2050
2051 /*
2052  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2053  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2054  * favor the lookup function ...
2055  * XXXsup merge this with qm_reclaim_one().
2056  */
2057 STATIC int
2058 xfs_qm_shake_freelist(
2059         int howmany)
2060 {
2061         int             nreclaimed;
2062         xfs_dqhash_t    *hash;
2063         xfs_dquot_t     *dqp, *nextdqp;
2064         int             restarts;
2065         int             nflushes;
2066
2067         if (howmany <= 0)
2068                 return 0;
2069
2070         nreclaimed = 0;
2071         restarts = 0;
2072         nflushes = 0;
2073
2074 #ifdef QUOTADEBUG
2075         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2076 #endif
2077         /* lock order is : hashchainlock, freelistlock, mplistlock */
2078  tryagain:
2079         xfs_qm_freelist_lock(xfs_Gqm);
2080
2081         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2082              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2083               nreclaimed < howmany); ) {
2084                 xfs_dqlock(dqp);
2085
2086                 /*
2087                  * We are racing with dqlookup here. Naturally we don't
2088                  * want to reclaim a dquot that lookup wants.
2089                  */
2090                 if (dqp->dq_flags & XFS_DQ_WANT) {
2091                         xfs_dqunlock(dqp);
2092                         xfs_qm_freelist_unlock(xfs_Gqm);
2093                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2094                                 return nreclaimed;
2095                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2096                         goto tryagain;
2097                 }
2098
2099                 /*
2100                  * If the dquot is inactive, we are assured that it is
2101                  * not on the mplist or the hashlist, and that makes our
2102                  * life easier.
2103                  */
2104                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2105                         ASSERT(dqp->q_mount == NULL);
2106                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2107                         ASSERT(dqp->HL_PREVP == NULL);
2108                         ASSERT(dqp->MPL_PREVP == NULL);
2109                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2110                         nextdqp = dqp->dq_flnext;
2111                         goto off_freelist;
2112                 }
2113
2114                 ASSERT(dqp->MPL_PREVP);
2115                 /*
2116                  * Try to grab the flush lock. If this dquot is in the process of
2117                  * getting flushed to disk, we don't want to reclaim it.
2118                  */
2119                 if (! xfs_qm_dqflock_nowait(dqp)) {
2120                         xfs_dqunlock(dqp);
2121                         dqp = dqp->dq_flnext;
2122                         continue;
2123                 }
2124
2125                 /*
2126                  * We have the flush lock so we know that this is not in the
2127                  * process of being flushed. So, if this is dirty, flush it
2128                  * DELWRI so that we don't get a freelist infested with
2129                  * dirty dquots.
2130                  */
2131                 if (XFS_DQ_IS_DIRTY(dqp)) {
2132                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2133                         /*
2134                          * We flush it delayed write, so don't bother
2135                          * releasing the mplock.
2136                          */
2137                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2138                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2139                         dqp = dqp->dq_flnext;
2140                         continue;
2141                 }
2142                 /*
2143                  * We're trying to get the hashlock out of order. This races
2144                  * with dqlookup; so, we giveup and goto the next dquot if
2145                  * we couldn't get the hashlock. This way, we won't starve
2146                  * a dqlookup process that holds the hashlock that is
2147                  * waiting for the freelist lock.
2148                  */
2149                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2150                         xfs_dqfunlock(dqp);
2151                         xfs_dqunlock(dqp);
2152                         dqp = dqp->dq_flnext;
2153                         continue;
2154                 }
2155                 /*
2156                  * This races with dquot allocation code as well as dqflush_all
2157                  * and reclaim code. So, if we failed to grab the mplist lock,
2158                  * giveup everything and start over.
2159                  */
2160                 hash = dqp->q_hash;
2161                 ASSERT(hash);
2162                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2163                         /* XXX put a sentinel so that we can come back here */
2164                         xfs_dqfunlock(dqp);
2165                         xfs_dqunlock(dqp);
2166                         XFS_DQ_HASH_UNLOCK(hash);
2167                         xfs_qm_freelist_unlock(xfs_Gqm);
2168                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2169                                 return nreclaimed;
2170                         goto tryagain;
2171                 }
2172                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2173 #ifdef QUOTADEBUG
2174                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2175                         dqp, be32_to_cpu(dqp->q_core.d_id));
2176 #endif
2177                 ASSERT(dqp->q_nrefs == 0);
2178                 nextdqp = dqp->dq_flnext;
2179                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2180                 XQM_HASHLIST_REMOVE(hash, dqp);
2181                 xfs_dqfunlock(dqp);
2182                 xfs_qm_mplist_unlock(dqp->q_mount);
2183                 XFS_DQ_HASH_UNLOCK(hash);
2184
2185  off_freelist:
2186                 XQM_FREELIST_REMOVE(dqp);
2187                 xfs_dqunlock(dqp);
2188                 nreclaimed++;
2189                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2190                 xfs_qm_dqdestroy(dqp);
2191                 dqp = nextdqp;
2192         }
2193         xfs_qm_freelist_unlock(xfs_Gqm);
2194         return nreclaimed;
2195 }
2196
2197
2198 /*
2199  * The kmem_shake interface is invoked when memory is running low.
2200  */
2201 /* ARGSUSED */
2202 STATIC int
2203 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2204 {
2205         int     ndqused, nfree, n;
2206
2207         if (!kmem_shake_allow(gfp_mask))
2208                 return 0;
2209         if (!xfs_Gqm)
2210                 return 0;
2211
2212         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2213         /* incore dquots in all f/s's */
2214         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2215
2216         ASSERT(ndqused >= 0);
2217
2218         if (nfree <= ndqused && nfree < ndquot)
2219                 return 0;
2220
2221         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2222         n = nfree - ndqused - ndquot;           /* # over target */
2223
2224         return xfs_qm_shake_freelist(MAX(nfree, n));
2225 }
2226
2227
2228 /*
2229  * Just pop the least recently used dquot off the freelist and
2230  * recycle it. The returned dquot is locked.
2231  */
2232 STATIC xfs_dquot_t *
2233 xfs_qm_dqreclaim_one(void)
2234 {
2235         xfs_dquot_t     *dqpout;
2236         xfs_dquot_t     *dqp;
2237         int             restarts;
2238         int             nflushes;
2239
2240         restarts = 0;
2241         dqpout = NULL;
2242         nflushes = 0;
2243
2244         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2245  startagain:
2246         xfs_qm_freelist_lock(xfs_Gqm);
2247
2248         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2249                 xfs_dqlock(dqp);
2250
2251                 /*
2252                  * We are racing with dqlookup here. Naturally we don't
2253                  * want to reclaim a dquot that lookup wants. We release the
2254                  * freelist lock and start over, so that lookup will grab
2255                  * both the dquot and the freelistlock.
2256                  */
2257                 if (dqp->dq_flags & XFS_DQ_WANT) {
2258                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2259                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2260                         xfs_dqunlock(dqp);
2261                         xfs_qm_freelist_unlock(xfs_Gqm);
2262                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2263                                 return NULL;
2264                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2265                         goto startagain;
2266                 }
2267
2268                 /*
2269                  * If the dquot is inactive, we are assured that it is
2270                  * not on the mplist or the hashlist, and that makes our
2271                  * life easier.
2272                  */
2273                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2274                         ASSERT(dqp->q_mount == NULL);
2275                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2276                         ASSERT(dqp->HL_PREVP == NULL);
2277                         ASSERT(dqp->MPL_PREVP == NULL);
2278                         XQM_FREELIST_REMOVE(dqp);
2279                         xfs_dqunlock(dqp);
2280                         dqpout = dqp;
2281                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2282                         break;
2283                 }
2284
2285                 ASSERT(dqp->q_hash);
2286                 ASSERT(dqp->MPL_PREVP);
2287
2288                 /*
2289                  * Try to grab the flush lock. If this dquot is in the process of
2290                  * getting flushed to disk, we don't want to reclaim it.
2291                  */
2292                 if (! xfs_qm_dqflock_nowait(dqp)) {
2293                         xfs_dqunlock(dqp);
2294                         continue;
2295                 }
2296
2297                 /*
2298                  * We have the flush lock so we know that this is not in the
2299                  * process of being flushed. So, if this is dirty, flush it
2300                  * DELWRI so that we don't get a freelist infested with
2301                  * dirty dquots.
2302                  */
2303                 if (XFS_DQ_IS_DIRTY(dqp)) {
2304                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2305                         /*
2306                          * We flush it delayed write, so don't bother
2307                          * releasing the freelist lock.
2308                          */
2309                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2310                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2311                         continue;
2312                 }
2313
2314                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2315                         xfs_dqfunlock(dqp);
2316                         xfs_dqunlock(dqp);
2317                         continue;
2318                 }
2319
2320                 if (! xfs_qm_dqhashlock_nowait(dqp))
2321                         goto mplistunlock;
2322
2323                 ASSERT(dqp->q_nrefs == 0);
2324                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2325                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2326                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2327                 XQM_FREELIST_REMOVE(dqp);
2328                 dqpout = dqp;
2329                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2330  mplistunlock:
2331                 xfs_qm_mplist_unlock(dqp->q_mount);
2332                 xfs_dqfunlock(dqp);
2333                 xfs_dqunlock(dqp);
2334                 if (dqpout)
2335                         break;
2336         }
2337
2338         xfs_qm_freelist_unlock(xfs_Gqm);
2339         return dqpout;
2340 }
2341
2342
2343 /*------------------------------------------------------------------*/
2344
2345 /*
2346  * Return a new incore dquot. Depending on the number of
2347  * dquots in the system, we either allocate a new one on the kernel heap,
2348  * or reclaim a free one.
2349  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2350  * to reclaim an existing one from the freelist.
2351  */
2352 boolean_t
2353 xfs_qm_dqalloc_incore(
2354         xfs_dquot_t **O_dqpp)
2355 {
2356         xfs_dquot_t     *dqp;
2357
2358         /*
2359          * Check against high water mark to see if we want to pop
2360          * a nincompoop dquot off the freelist.
2361          */
2362         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2363                 /*
2364                  * Try to recycle a dquot from the freelist.
2365                  */
2366                 if ((dqp = xfs_qm_dqreclaim_one())) {
2367                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2368                         /*
2369                          * Just zero the core here. The rest will get
2370                          * reinitialized by caller. XXX we shouldn't even
2371                          * do this zero ...
2372                          */
2373                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2374                         *O_dqpp = dqp;
2375                         return B_FALSE;
2376                 }
2377                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2378         }
2379
2380         /*
2381          * Allocate a brand new dquot on the kernel heap and return it
2382          * to the caller to initialize.
2383          */
2384         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2385         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2386         atomic_inc(&xfs_Gqm->qm_totaldquots);
2387
2388         return B_TRUE;
2389 }
2390
2391
2392 /*
2393  * Start a transaction and write the incore superblock changes to
2394  * disk. flags parameter indicates which fields have changed.
2395  */
2396 int
2397 xfs_qm_write_sb_changes(
2398         xfs_mount_t     *mp,
2399         __int64_t       flags)
2400 {
2401         xfs_trans_t     *tp;
2402         int             error;
2403
2404 #ifdef QUOTADEBUG
2405         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2406 #endif
2407         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2408         if ((error = xfs_trans_reserve(tp, 0,
2409                                       mp->m_sb.sb_sectsize + 128, 0,
2410                                       0,
2411                                       XFS_DEFAULT_LOG_COUNT))) {
2412                 xfs_trans_cancel(tp, 0);
2413                 return error;
2414         }
2415
2416         xfs_mod_sb(tp, flags);
2417         (void) xfs_trans_commit(tp, 0);
2418
2419         return 0;
2420 }
2421
2422
2423 /* --------------- utility functions for vnodeops ---------------- */
2424
2425
2426 /*
2427  * Given an inode, a uid and gid (from cred_t) make sure that we have
2428  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2429  * quotas by creating this file.
2430  * This also attaches dquot(s) to the given inode after locking it,
2431  * and returns the dquots corresponding to the uid and/or gid.
2432  *
2433  * in   : inode (unlocked)
2434  * out  : udquot, gdquot with references taken and unlocked
2435  */
2436 int
2437 xfs_qm_vop_dqalloc(
2438         xfs_mount_t     *mp,
2439         xfs_inode_t     *ip,
2440         uid_t           uid,
2441         gid_t           gid,
2442         prid_t          prid,
2443         uint            flags,
2444         xfs_dquot_t     **O_udqpp,
2445         xfs_dquot_t     **O_gdqpp)
2446 {
2447         int             error;
2448         xfs_dquot_t     *uq, *gq;
2449         uint            lockflags;
2450
2451         if (!XFS_IS_QUOTA_ON(mp))
2452                 return 0;
2453
2454         lockflags = XFS_ILOCK_EXCL;
2455         xfs_ilock(ip, lockflags);
2456
2457         if ((flags & XFS_QMOPT_INHERIT) &&
2458             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2459                 gid = ip->i_d.di_gid;
2460
2461         /*
2462          * Attach the dquot(s) to this inode, doing a dquot allocation
2463          * if necessary. The dquot(s) will not be locked.
2464          */
2465         if (XFS_NOT_DQATTACHED(mp, ip)) {
2466                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2467                                             XFS_QMOPT_ILOCKED))) {
2468                         xfs_iunlock(ip, lockflags);
2469                         return error;
2470                 }
2471         }
2472
2473         uq = gq = NULL;
2474         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2475                 if (ip->i_d.di_uid != uid) {
2476                         /*
2477                          * What we need is the dquot that has this uid, and
2478                          * if we send the inode to dqget, the uid of the inode
2479                          * takes priority over what's sent in the uid argument.
2480                          * We must unlock inode here before calling dqget if
2481                          * we're not sending the inode, because otherwise
2482                          * we'll deadlock by doing trans_reserve while
2483                          * holding ilock.
2484                          */
2485                         xfs_iunlock(ip, lockflags);
2486                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2487                                                  XFS_DQ_USER,
2488                                                  XFS_QMOPT_DQALLOC |
2489                                                  XFS_QMOPT_DOWARN,
2490                                                  &uq))) {
2491                                 ASSERT(error != ENOENT);
2492                                 return error;
2493                         }
2494                         /*
2495                          * Get the ilock in the right order.
2496                          */
2497                         xfs_dqunlock(uq);
2498                         lockflags = XFS_ILOCK_SHARED;
2499                         xfs_ilock(ip, lockflags);
2500                 } else {
2501                         /*
2502                          * Take an extra reference, because we'll return
2503                          * this to caller
2504                          */
2505                         ASSERT(ip->i_udquot);
2506                         uq = ip->i_udquot;
2507                         xfs_dqlock(uq);
2508                         XFS_DQHOLD(uq);
2509                         xfs_dqunlock(uq);
2510                 }
2511         }
2512         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2513                 if (ip->i_d.di_gid != gid) {
2514                         xfs_iunlock(ip, lockflags);
2515                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2516                                                  XFS_DQ_GROUP,
2517                                                  XFS_QMOPT_DQALLOC |
2518                                                  XFS_QMOPT_DOWARN,
2519                                                  &gq))) {
2520                                 if (uq)
2521                                         xfs_qm_dqrele(uq);
2522                                 ASSERT(error != ENOENT);
2523                                 return error;
2524                         }
2525                         xfs_dqunlock(gq);
2526                         lockflags = XFS_ILOCK_SHARED;
2527                         xfs_ilock(ip, lockflags);
2528                 } else {
2529                         ASSERT(ip->i_gdquot);
2530                         gq = ip->i_gdquot;
2531                         xfs_dqlock(gq);
2532                         XFS_DQHOLD(gq);
2533                         xfs_dqunlock(gq);
2534                 }
2535         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2536                 if (ip->i_d.di_projid != prid) {
2537                         xfs_iunlock(ip, lockflags);
2538                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2539                                                  XFS_DQ_PROJ,
2540                                                  XFS_QMOPT_DQALLOC |
2541                                                  XFS_QMOPT_DOWARN,
2542                                                  &gq))) {
2543                                 if (uq)
2544                                         xfs_qm_dqrele(uq);
2545                                 ASSERT(error != ENOENT);
2546                                 return (error);
2547                         }
2548                         xfs_dqunlock(gq);
2549                         lockflags = XFS_ILOCK_SHARED;
2550                         xfs_ilock(ip, lockflags);
2551                 } else {
2552                         ASSERT(ip->i_gdquot);
2553                         gq = ip->i_gdquot;
2554                         xfs_dqlock(gq);
2555                         XFS_DQHOLD(gq);
2556                         xfs_dqunlock(gq);
2557                 }
2558         }
2559         if (uq)
2560                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2561
2562         xfs_iunlock(ip, lockflags);
2563         if (O_udqpp)
2564                 *O_udqpp = uq;
2565         else if (uq)
2566                 xfs_qm_dqrele(uq);
2567         if (O_gdqpp)
2568                 *O_gdqpp = gq;
2569         else if (gq)
2570                 xfs_qm_dqrele(gq);
2571         return 0;
2572 }
2573
2574 /*
2575  * Actually transfer ownership, and do dquot modifications.
2576  * These were already reserved.
2577  */
2578 xfs_dquot_t *
2579 xfs_qm_vop_chown(
2580         xfs_trans_t     *tp,
2581         xfs_inode_t     *ip,
2582         xfs_dquot_t     **IO_olddq,
2583         xfs_dquot_t     *newdq)
2584 {
2585         xfs_dquot_t     *prevdq;
2586         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2587                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2588
2589         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2590         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2591
2592         /* old dquot */
2593         prevdq = *IO_olddq;
2594         ASSERT(prevdq);
2595         ASSERT(prevdq != newdq);
2596
2597         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2598         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2599
2600         /* the sparkling new dquot */
2601         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2602         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2603
2604         /*
2605          * Take an extra reference, because the inode
2606          * is going to keep this dquot pointer even
2607          * after the trans_commit.
2608          */
2609         xfs_dqlock(newdq);
2610         XFS_DQHOLD(newdq);
2611         xfs_dqunlock(newdq);
2612         *IO_olddq = newdq;
2613
2614         return prevdq;
2615 }
2616
2617 /*
2618  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2619  */
2620 int
2621 xfs_qm_vop_chown_reserve(
2622         xfs_trans_t     *tp,
2623         xfs_inode_t     *ip,
2624         xfs_dquot_t     *udqp,
2625         xfs_dquot_t     *gdqp,
2626         uint            flags)
2627 {
2628         int             error;
2629         xfs_mount_t     *mp;
2630         uint            delblks, blkflags, prjflags = 0;
2631         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2632
2633         ASSERT(XFS_ISLOCKED_INODE(ip));
2634         mp = ip->i_mount;
2635         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2636
2637         delblks = ip->i_delayed_blks;
2638         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2639         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2640                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2641
2642         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2643             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2644                 delblksudq = udqp;
2645                 /*
2646                  * If there are delayed allocation blocks, then we have to
2647                  * unreserve those from the old dquot, and add them to the
2648                  * new dquot.
2649                  */
2650                 if (delblks) {
2651                         ASSERT(ip->i_udquot);
2652                         unresudq = ip->i_udquot;
2653                 }
2654         }
2655         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2656                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2657                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2658                         prjflags = XFS_QMOPT_ENOSPC;
2659
2660                 if (prjflags ||
2661                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2662                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2663                         delblksgdq = gdqp;
2664                         if (delblks) {
2665                                 ASSERT(ip->i_gdquot);
2666                                 unresgdq = ip->i_gdquot;
2667                         }
2668                 }
2669         }
2670
2671         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2672                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2673                                 flags | blkflags | prjflags)))
2674                 return (error);
2675
2676         /*
2677          * Do the delayed blks reservations/unreservations now. Since, these
2678          * are done without the help of a transaction, if a reservation fails
2679          * its previous reservations won't be automatically undone by trans
2680          * code. So, we have to do it manually here.
2681          */
2682         if (delblks) {
2683                 /*
2684                  * Do the reservations first. Unreservation can't fail.
2685                  */
2686                 ASSERT(delblksudq || delblksgdq);
2687                 ASSERT(unresudq || unresgdq);
2688                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2689                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2690                                 flags | blkflags | prjflags)))
2691                         return (error);
2692                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2693                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2694                                 blkflags);
2695         }
2696
2697         return (0);
2698 }
2699
2700 int
2701 xfs_qm_vop_rename_dqattach(
2702         xfs_inode_t     **i_tab)
2703 {
2704         xfs_inode_t     *ip;
2705         int             i;
2706         int             error;
2707
2708         ip = i_tab[0];
2709
2710         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2711                 return 0;
2712
2713         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2714                 error = xfs_qm_dqattach(ip, 0);
2715                 if (error)
2716                         return error;
2717         }
2718         for (i = 1; (i < 4 && i_tab[i]); i++) {
2719                 /*
2720                  * Watch out for duplicate entries in the table.
2721                  */
2722                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2723                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2724                                 error = xfs_qm_dqattach(ip, 0);
2725                                 if (error)
2726                                         return error;
2727                         }
2728                 }
2729         }
2730         return 0;
2731 }
2732
2733 void
2734 xfs_qm_vop_dqattach_and_dqmod_newinode(
2735         xfs_trans_t     *tp,
2736         xfs_inode_t     *ip,
2737         xfs_dquot_t     *udqp,
2738         xfs_dquot_t     *gdqp)
2739 {
2740         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2741                 return;
2742
2743         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2744         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2745
2746         if (udqp) {
2747                 xfs_dqlock(udqp);
2748                 XFS_DQHOLD(udqp);
2749                 xfs_dqunlock(udqp);
2750                 ASSERT(ip->i_udquot == NULL);
2751                 ip->i_udquot = udqp;
2752                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2753                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2754                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2755         }
2756         if (gdqp) {
2757                 xfs_dqlock(gdqp);
2758                 XFS_DQHOLD(gdqp);
2759                 xfs_dqunlock(gdqp);
2760                 ASSERT(ip->i_gdquot == NULL);
2761                 ip->i_gdquot = gdqp;
2762                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2763                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2764                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2765                                 be32_to_cpu(gdqp->q_core.d_id));
2766                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2767         }
2768 }
2769
2770 /* ------------- list stuff -----------------*/
2771 STATIC void
2772 xfs_qm_freelist_init(xfs_frlist_t *ql)
2773 {
2774         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2775         mutex_init(&ql->qh_lock);
2776         ql->qh_version = 0;
2777         ql->qh_nelems = 0;
2778 }
2779
2780 STATIC void
2781 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2782 {
2783         xfs_dquot_t     *dqp, *nextdqp;
2784
2785         mutex_lock(&ql->qh_lock);
2786         for (dqp = ql->qh_next;
2787              dqp != (xfs_dquot_t *)ql; ) {
2788                 xfs_dqlock(dqp);
2789                 nextdqp = dqp->dq_flnext;
2790 #ifdef QUOTADEBUG
2791                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2792 #endif
2793                 XQM_FREELIST_REMOVE(dqp);
2794                 xfs_dqunlock(dqp);
2795                 xfs_qm_dqdestroy(dqp);
2796                 dqp = nextdqp;
2797         }
2798         mutex_unlock(&ql->qh_lock);
2799         mutex_destroy(&ql->qh_lock);
2800
2801         ASSERT(ql->qh_nelems == 0);
2802 }
2803
2804 STATIC void
2805 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2806 {
2807         dq->dq_flnext = ql->qh_next;
2808         dq->dq_flprev = (xfs_dquot_t *)ql;
2809         ql->qh_next = dq;
2810         dq->dq_flnext->dq_flprev = dq;
2811         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2812         xfs_Gqm->qm_dqfreelist.qh_version++;
2813 }
2814
2815 void
2816 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2817 {
2818         xfs_dquot_t *next = dq->dq_flnext;
2819         xfs_dquot_t *prev = dq->dq_flprev;
2820
2821         next->dq_flprev = prev;
2822         prev->dq_flnext = next;
2823         dq->dq_flnext = dq->dq_flprev = dq;
2824         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2825         xfs_Gqm->qm_dqfreelist.qh_version++;
2826 }
2827
2828 void
2829 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2830 {
2831         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2832 }
2833
2834 STATIC int
2835 xfs_qm_dqhashlock_nowait(
2836         xfs_dquot_t *dqp)
2837 {
2838         int locked;
2839
2840         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2841         return locked;
2842 }
2843
2844 int
2845 xfs_qm_freelist_lock_nowait(
2846         xfs_qm_t *xqm)
2847 {
2848         int locked;
2849
2850         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2851         return locked;
2852 }
2853
2854 STATIC int
2855 xfs_qm_mplist_nowait(
2856         xfs_mount_t     *mp)
2857 {
2858         int locked;
2859
2860         ASSERT(mp->m_quotainfo);
2861         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2862         return locked;
2863 }