]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - fs/xfs/quota/xfs_qm.c
[XFS] Update license/copyright notices to match the prefered SGI
[linux-2.6-omap-h63xx.git] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_clnt.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir.h"
28 #include "xfs_dir2.h"
29 #include "xfs_alloc.h"
30 #include "xfs_dmapi.h"
31 #include "xfs_quota.h"
32 #include "xfs_mount.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_dir_sf.h"
37 #include "xfs_dir2_sf.h"
38 #include "xfs_attr_sf.h"
39 #include "xfs_dinode.h"
40 #include "xfs_inode.h"
41 #include "xfs_btree.h"
42 #include "xfs_ialloc.h"
43 #include "xfs_itable.h"
44 #include "xfs_rtalloc.h"
45 #include "xfs_error.h"
46 #include "xfs_bmap.h"
47 #include "xfs_rw.h"
48 #include "xfs_acl.h"
49 #include "xfs_cap.h"
50 #include "xfs_mac.h"
51 #include "xfs_attr.h"
52 #include "xfs_buf_item.h"
53 #include "xfs_trans_space.h"
54 #include "xfs_utils.h"
55 #include "xfs_qm.h"
56
57 /*
58  * The global quota manager. There is only one of these for the entire
59  * system, _not_ one per file system. XQM keeps track of the overall
60  * quota functionality, including maintaining the freelist and hash
61  * tables of dquots.
62  */
63 mutex_t xfs_Gqm_lock;
64 struct xfs_qm   *xfs_Gqm;
65
66 kmem_zone_t     *qm_dqzone;
67 kmem_zone_t     *qm_dqtrxzone;
68 STATIC kmem_shaker_t    xfs_qm_shaker;
69
70 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
71 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
72
73 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
74 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
75 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
76 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
77
78 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
79 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
80 STATIC int      xfs_qm_shake(int, unsigned int);
81
82 #ifdef DEBUG
83 extern mutex_t  qcheck_lock;
84 #endif
85
86 #ifdef QUOTADEBUG
87 #define XQM_LIST_PRINT(l, NXT, title) \
88 { \
89         xfs_dquot_t     *dqp; int i = 0; \
90         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
91         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
92                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
93                                   "bcnt = %d, icnt = %d, refs = %d", \
94                         ++i, (int) INT_GET(dqp->q_core.d_id, ARCH_CONVERT), \
95                         DQFLAGTO_TYPESTR(dqp),       \
96                         (int) INT_GET(dqp->q_core.d_bcount, ARCH_CONVERT), \
97                         (int) INT_GET(dqp->q_core.d_icount, ARCH_CONVERT), \
98                         (int) dqp->q_nrefs);  } \
99 }
100 #else
101 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
102 #endif
103
104 /*
105  * Initialize the XQM structure.
106  * Note that there is not one quota manager per file system.
107  */
108 STATIC struct xfs_qm *
109 xfs_Gqm_init(void)
110 {
111         xfs_qm_t                *xqm;
112         int                     hsize, i;
113
114         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
115         ASSERT(xqm);
116
117         /*
118          * Initialize the dquot hash tables.
119          */
120         hsize = (DQUOT_HASH_HEURISTIC < XFS_QM_NCSIZE_THRESHOLD) ?
121                 XFS_QM_HASHSIZE_LOW : XFS_QM_HASHSIZE_HIGH;
122         xqm->qm_dqhashmask = hsize - 1;
123
124         xqm->qm_usr_dqhtable = (xfs_dqhash_t *)kmem_zalloc(hsize *
125                                                       sizeof(xfs_dqhash_t),
126                                                       KM_SLEEP);
127         xqm->qm_grp_dqhtable = (xfs_dqhash_t *)kmem_zalloc(hsize *
128                                                       sizeof(xfs_dqhash_t),
129                                                       KM_SLEEP);
130         ASSERT(xqm->qm_usr_dqhtable != NULL);
131         ASSERT(xqm->qm_grp_dqhtable != NULL);
132
133         for (i = 0; i < hsize; i++) {
134                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
135                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
136         }
137
138         /*
139          * Freelist of all dquots of all file systems
140          */
141         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
142
143         /*
144          * dquot zone. we register our own low-memory callback.
145          */
146         if (!qm_dqzone) {
147                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
148                                                 "xfs_dquots");
149                 qm_dqzone = xqm->qm_dqzone;
150         } else
151                 xqm->qm_dqzone = qm_dqzone;
152
153         xfs_qm_shaker = kmem_shake_register(xfs_qm_shake);
154
155         /*
156          * The t_dqinfo portion of transactions.
157          */
158         if (!qm_dqtrxzone) {
159                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
160                                                    "xfs_dqtrx");
161                 qm_dqtrxzone = xqm->qm_dqtrxzone;
162         } else
163                 xqm->qm_dqtrxzone = qm_dqtrxzone;
164
165         atomic_set(&xqm->qm_totaldquots, 0);
166         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
167         xqm->qm_nrefs = 0;
168 #ifdef DEBUG
169         mutex_init(&qcheck_lock, MUTEX_DEFAULT, "qchk");
170 #endif
171         return xqm;
172 }
173
174 /*
175  * Destroy the global quota manager when its reference count goes to zero.
176  */
177 STATIC void
178 xfs_qm_destroy(
179         struct xfs_qm   *xqm)
180 {
181         int             hsize, i;
182
183         ASSERT(xqm != NULL);
184         ASSERT(xqm->qm_nrefs == 0);
185         kmem_shake_deregister(xfs_qm_shaker);
186         hsize = xqm->qm_dqhashmask + 1;
187         for (i = 0; i < hsize; i++) {
188                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
189                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
190         }
191         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
192         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
193         xqm->qm_usr_dqhtable = NULL;
194         xqm->qm_grp_dqhtable = NULL;
195         xqm->qm_dqhashmask = 0;
196         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
197 #ifdef DEBUG
198         mutex_destroy(&qcheck_lock);
199 #endif
200         kmem_free(xqm, sizeof(xfs_qm_t));
201 }
202
203 /*
204  * Called at mount time to let XQM know that another file system is
205  * starting quotas. This isn't crucial information as the individual mount
206  * structures are pretty independent, but it helps the XQM keep a
207  * global view of what's going on.
208  */
209 /* ARGSUSED */
210 STATIC int
211 xfs_qm_hold_quotafs_ref(
212         struct xfs_mount *mp)
213 {
214         /*
215          * Need to lock the xfs_Gqm structure for things like this. For example,
216          * the structure could disappear between the entry to this routine and
217          * a HOLD operation if not locked.
218          */
219         XFS_QM_LOCK(xfs_Gqm);
220
221         if (xfs_Gqm == NULL)
222                 xfs_Gqm = xfs_Gqm_init();
223         /*
224          * We can keep a list of all filesystems with quotas mounted for
225          * debugging and statistical purposes, but ...
226          * Just take a reference and get out.
227          */
228         XFS_QM_HOLD(xfs_Gqm);
229         XFS_QM_UNLOCK(xfs_Gqm);
230
231         return 0;
232 }
233
234
235 /*
236  * Release the reference that a filesystem took at mount time,
237  * so that we know when we need to destroy the entire quota manager.
238  */
239 /* ARGSUSED */
240 STATIC void
241 xfs_qm_rele_quotafs_ref(
242         struct xfs_mount *mp)
243 {
244         xfs_dquot_t     *dqp, *nextdqp;
245
246         ASSERT(xfs_Gqm);
247         ASSERT(xfs_Gqm->qm_nrefs > 0);
248
249         /*
250          * Go thru the freelist and destroy all inactive dquots.
251          */
252         xfs_qm_freelist_lock(xfs_Gqm);
253
254         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
255              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
256                 xfs_dqlock(dqp);
257                 nextdqp = dqp->dq_flnext;
258                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
259                         ASSERT(dqp->q_mount == NULL);
260                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
261                         ASSERT(dqp->HL_PREVP == NULL);
262                         ASSERT(dqp->MPL_PREVP == NULL);
263                         XQM_FREELIST_REMOVE(dqp);
264                         xfs_dqunlock(dqp);
265                         xfs_qm_dqdestroy(dqp);
266                 } else {
267                         xfs_dqunlock(dqp);
268                 }
269                 dqp = nextdqp;
270         }
271         xfs_qm_freelist_unlock(xfs_Gqm);
272
273         /*
274          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
275          * be restarted.
276          */
277         XFS_QM_LOCK(xfs_Gqm);
278         XFS_QM_RELE(xfs_Gqm);
279         if (xfs_Gqm->qm_nrefs == 0) {
280                 xfs_qm_destroy(xfs_Gqm);
281                 xfs_Gqm = NULL;
282         }
283         XFS_QM_UNLOCK(xfs_Gqm);
284 }
285
286 /*
287  * This is called at mount time from xfs_mountfs to initialize the quotainfo
288  * structure and start the global quotamanager (xfs_Gqm) if it hasn't done
289  * so already.  Note that the superblock has not been read in yet.
290  */
291 void
292 xfs_qm_mount_quotainit(
293         xfs_mount_t     *mp,
294         uint            flags)
295 {
296         /*
297          * User, projects or group quotas has to be on.
298          */
299         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_PQUOTA | XFSMNT_GQUOTA));
300
301         /*
302          * Initialize the flags in the mount structure. From this point
303          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
304          * Note that we enforce nothing if accounting is off.
305          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
306          * It isn't necessary to take the quotaoff lock to do this; this is
307          * called from mount.
308          */
309         if (flags & XFSMNT_UQUOTA) {
310                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
311                 if (flags & XFSMNT_UQUOTAENF)
312                         mp->m_qflags |= XFS_UQUOTA_ENFD;
313         }
314         if (flags & XFSMNT_GQUOTA) {
315                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
316                 if (flags & XFSMNT_GQUOTAENF)
317                         mp->m_qflags |= XFS_OQUOTA_ENFD;
318         } else if (flags & XFSMNT_PQUOTA) {
319                 mp->m_qflags |= (XFS_PQUOTA_ACCT | XFS_PQUOTA_ACTIVE);
320                 if (flags & XFSMNT_PQUOTAENF)
321                         mp->m_qflags |= XFS_OQUOTA_ENFD;
322         }
323 }
324
325 /*
326  * Just destroy the quotainfo structure.
327  */
328 void
329 xfs_qm_unmount_quotadestroy(
330         xfs_mount_t     *mp)
331 {
332         if (mp->m_quotainfo)
333                 xfs_qm_destroy_quotainfo(mp);
334 }
335
336
337 /*
338  * This is called from xfs_mountfs to start quotas and initialize all
339  * necessary data structures like quotainfo.  This is also responsible for
340  * running a quotacheck as necessary.  We are guaranteed that the superblock
341  * is consistently read in at this point.
342  */
343 int
344 xfs_qm_mount_quotas(
345         xfs_mount_t     *mp,
346         int             mfsi_flags)
347 {
348         unsigned long   s;
349         int             error = 0;
350         uint            sbf;
351
352
353         /*
354          * If quotas on realtime volumes is not supported, we disable
355          * quotas immediately.
356          */
357         if (mp->m_sb.sb_rextents) {
358                 cmn_err(CE_NOTE,
359                         "Cannot turn on quotas for realtime filesystem %s",
360                         mp->m_fsname);
361                 mp->m_qflags = 0;
362                 goto write_changes;
363         }
364
365         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
366
367         /*
368          * Allocate the quotainfo structure inside the mount struct, and
369          * create quotainode(s), and change/rev superblock if necessary.
370          */
371         if ((error = xfs_qm_init_quotainfo(mp))) {
372                 /*
373                  * We must turn off quotas.
374                  */
375                 ASSERT(mp->m_quotainfo == NULL);
376                 mp->m_qflags = 0;
377                 goto write_changes;
378         }
379         /*
380          * If any of the quotas are not consistent, do a quotacheck.
381          */
382         if (XFS_QM_NEED_QUOTACHECK(mp) &&
383                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
384                 if ((error = xfs_qm_quotacheck(mp))) {
385                         /* Quotacheck has failed and quotas have
386                          * been disabled.
387                          */
388                         return XFS_ERROR(error);
389                 }
390         }
391
392  write_changes:
393         /*
394          * We actually don't have to acquire the SB_LOCK at all.
395          * This can only be called from mount, and that's single threaded. XXX
396          */
397         s = XFS_SB_LOCK(mp);
398         sbf = mp->m_sb.sb_qflags;
399         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
400         XFS_SB_UNLOCK(mp, s);
401
402         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
403                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
404                         /*
405                          * We could only have been turning quotas off.
406                          * We aren't in very good shape actually because
407                          * the incore structures are convinced that quotas are
408                          * off, but the on disk superblock doesn't know that !
409                          */
410                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
411                         xfs_fs_cmn_err(CE_ALERT, mp,
412                                 "XFS mount_quotas: Superblock update failed!");
413                 }
414         }
415
416         if (error) {
417                 xfs_fs_cmn_err(CE_WARN, mp,
418                         "Failed to initialize disk quotas.");
419         }
420         return XFS_ERROR(error);
421 }
422
423 /*
424  * Called from the vfsops layer.
425  */
426 int
427 xfs_qm_unmount_quotas(
428         xfs_mount_t     *mp)
429 {
430         xfs_inode_t     *uqp, *gqp;
431         int             error = 0;
432
433         /*
434          * Release the dquots that root inode, et al might be holding,
435          * before we flush quotas and blow away the quotainfo structure.
436          */
437         ASSERT(mp->m_rootip);
438         xfs_qm_dqdetach(mp->m_rootip);
439         if (mp->m_rbmip)
440                 xfs_qm_dqdetach(mp->m_rbmip);
441         if (mp->m_rsumip)
442                 xfs_qm_dqdetach(mp->m_rsumip);
443
444         /*
445          * Flush out the quota inodes.
446          */
447         uqp = gqp = NULL;
448         if (mp->m_quotainfo) {
449                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
450                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
451                         xfs_iflock(uqp);
452                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
453                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
454                         if (unlikely(error == EFSCORRUPTED)) {
455                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
456                                                  XFS_ERRLEVEL_LOW, mp);
457                                 goto out;
458                         }
459                 }
460                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
461                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
462                         xfs_iflock(gqp);
463                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
464                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
465                         if (unlikely(error == EFSCORRUPTED)) {
466                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
467                                                  XFS_ERRLEVEL_LOW, mp);
468                                 goto out;
469                         }
470                 }
471         }
472         if (uqp) {
473                  XFS_PURGE_INODE(uqp);
474                  mp->m_quotainfo->qi_uquotaip = NULL;
475         }
476         if (gqp) {
477                 XFS_PURGE_INODE(gqp);
478                 mp->m_quotainfo->qi_gquotaip = NULL;
479         }
480 out:
481         return XFS_ERROR(error);
482 }
483
484 /*
485  * Flush all dquots of the given file system to disk. The dquots are
486  * _not_ purged from memory here, just their data written to disk.
487  */
488 STATIC int
489 xfs_qm_dqflush_all(
490         xfs_mount_t     *mp,
491         int             flags)
492 {
493         int             recl;
494         xfs_dquot_t     *dqp;
495         int             niters;
496         int             error;
497
498         if (mp->m_quotainfo == NULL)
499                 return (0);
500         niters = 0;
501 again:
502         xfs_qm_mplist_lock(mp);
503         FOREACH_DQUOT_IN_MP(dqp, mp) {
504                 xfs_dqlock(dqp);
505                 if (! XFS_DQ_IS_DIRTY(dqp)) {
506                         xfs_dqunlock(dqp);
507                         continue;
508                 }
509                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
510                 /* XXX a sentinel would be better */
511                 recl = XFS_QI_MPLRECLAIMS(mp);
512                 if (! xfs_qm_dqflock_nowait(dqp)) {
513                         /*
514                          * If we can't grab the flush lock then check
515                          * to see if the dquot has been flushed delayed
516                          * write.  If so, grab its buffer and send it
517                          * out immediately.  We'll be able to acquire
518                          * the flush lock when the I/O completes.
519                          */
520                         xfs_qm_dqflock_pushbuf_wait(dqp);
521                 }
522                 /*
523                  * Let go of the mplist lock. We don't want to hold it
524                  * across a disk write.
525                  */
526                 xfs_qm_mplist_unlock(mp);
527                 error = xfs_qm_dqflush(dqp, flags);
528                 xfs_dqunlock(dqp);
529                 if (error)
530                         return (error);
531
532                 xfs_qm_mplist_lock(mp);
533                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
534                         xfs_qm_mplist_unlock(mp);
535                         /* XXX restart limit */
536                         goto again;
537                 }
538         }
539
540         xfs_qm_mplist_unlock(mp);
541         /* return ! busy */
542         return (0);
543 }
544 /*
545  * Release the group dquot pointers the user dquots may be
546  * carrying around as a hint. mplist is locked on entry and exit.
547  */
548 STATIC void
549 xfs_qm_detach_gdquots(
550         xfs_mount_t     *mp)
551 {
552         xfs_dquot_t     *dqp, *gdqp;
553         int             nrecl;
554
555  again:
556         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
557         dqp = XFS_QI_MPLNEXT(mp);
558         while (dqp) {
559                 xfs_dqlock(dqp);
560                 if ((gdqp = dqp->q_gdquot)) {
561                         xfs_dqlock(gdqp);
562                         dqp->q_gdquot = NULL;
563                 }
564                 xfs_dqunlock(dqp);
565
566                 if (gdqp) {
567                         /*
568                          * Can't hold the mplist lock across a dqput.
569                          * XXXmust convert to marker based iterations here.
570                          */
571                         nrecl = XFS_QI_MPLRECLAIMS(mp);
572                         xfs_qm_mplist_unlock(mp);
573                         xfs_qm_dqput(gdqp);
574
575                         xfs_qm_mplist_lock(mp);
576                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
577                                 goto again;
578                 }
579                 dqp = dqp->MPL_NEXT;
580         }
581 }
582
583 /*
584  * Go through all the incore dquots of this file system and take them
585  * off the mplist and hashlist, if the dquot type matches the dqtype
586  * parameter. This is used when turning off quota accounting for
587  * users and/or groups, as well as when the filesystem is unmounting.
588  */
589 STATIC int
590 xfs_qm_dqpurge_int(
591         xfs_mount_t     *mp,
592         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
593 {
594         xfs_dquot_t     *dqp;
595         uint            dqtype;
596         int             nrecl;
597         xfs_dquot_t     *nextdqp;
598         int             nmisses;
599
600         if (mp->m_quotainfo == NULL)
601                 return (0);
602
603         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
604         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
605         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
606
607         xfs_qm_mplist_lock(mp);
608
609         /*
610          * In the first pass through all incore dquots of this filesystem,
611          * we release the group dquot pointers the user dquots may be
612          * carrying around as a hint. We need to do this irrespective of
613          * what's being turned off.
614          */
615         xfs_qm_detach_gdquots(mp);
616
617       again:
618         nmisses = 0;
619         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
620         /*
621          * Try to get rid of all of the unwanted dquots. The idea is to
622          * get them off mplist and hashlist, but leave them on freelist.
623          */
624         dqp = XFS_QI_MPLNEXT(mp);
625         while (dqp) {
626                 /*
627                  * It's OK to look at the type without taking dqlock here.
628                  * We're holding the mplist lock here, and that's needed for
629                  * a dqreclaim.
630                  */
631                 if ((dqp->dq_flags & dqtype) == 0) {
632                         dqp = dqp->MPL_NEXT;
633                         continue;
634                 }
635
636                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
637                         nrecl = XFS_QI_MPLRECLAIMS(mp);
638                         xfs_qm_mplist_unlock(mp);
639                         XFS_DQ_HASH_LOCK(dqp->q_hash);
640                         xfs_qm_mplist_lock(mp);
641
642                         /*
643                          * XXXTheoretically, we can get into a very long
644                          * ping pong game here.
645                          * No one can be adding dquots to the mplist at
646                          * this point, but somebody might be taking things off.
647                          */
648                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
649                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
650                                 goto again;
651                         }
652                 }
653
654                 /*
655                  * Take the dquot off the mplist and hashlist. It may remain on
656                  * freelist in INACTIVE state.
657                  */
658                 nextdqp = dqp->MPL_NEXT;
659                 nmisses += xfs_qm_dqpurge(dqp, flags);
660                 dqp = nextdqp;
661         }
662         xfs_qm_mplist_unlock(mp);
663         return nmisses;
664 }
665
666 int
667 xfs_qm_dqpurge_all(
668         xfs_mount_t     *mp,
669         uint            flags)
670 {
671         int             ndquots;
672
673         /*
674          * Purge the dquot cache.
675          * None of the dquots should really be busy at this point.
676          */
677         if (mp->m_quotainfo) {
678                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
679                         delay(ndquots * 10);
680                 }
681         }
682         return 0;
683 }
684
685 STATIC int
686 xfs_qm_dqattach_one(
687         xfs_inode_t     *ip,
688         xfs_dqid_t      id,
689         uint            type,
690         uint            doalloc,
691         uint            dolock,
692         xfs_dquot_t     *udqhint, /* hint */
693         xfs_dquot_t     **IO_idqpp)
694 {
695         xfs_dquot_t     *dqp;
696         int             error;
697
698         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
699         error = 0;
700         /*
701          * See if we already have it in the inode itself. IO_idqpp is
702          * &i_udquot or &i_gdquot. This made the code look weird, but
703          * made the logic a lot simpler.
704          */
705         if ((dqp = *IO_idqpp)) {
706                 if (dolock)
707                         xfs_dqlock(dqp);
708                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
709                 goto done;
710         }
711
712         /*
713          * udqhint is the i_udquot field in inode, and is non-NULL only
714          * when the type arg is group/project. Its purpose is to save a
715          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
716          * the user dquot.
717          */
718         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
719         if (udqhint && !dolock)
720                 xfs_dqlock(udqhint);
721
722         /*
723          * No need to take dqlock to look at the id.
724          * The ID can't change until it gets reclaimed, and it won't
725          * be reclaimed as long as we have a ref from inode and we hold
726          * the ilock.
727          */
728         if (udqhint &&
729             (dqp = udqhint->q_gdquot) &&
730             (INT_GET(dqp->q_core.d_id, ARCH_CONVERT) == id)) {
731                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
732                 xfs_dqlock(dqp);
733                 XFS_DQHOLD(dqp);
734                 ASSERT(*IO_idqpp == NULL);
735                 *IO_idqpp = dqp;
736                 if (!dolock) {
737                         xfs_dqunlock(dqp);
738                         xfs_dqunlock(udqhint);
739                 }
740                 goto done;
741         }
742         /*
743          * We can't hold a dquot lock when we call the dqget code.
744          * We'll deadlock in no time, because of (not conforming to)
745          * lock ordering - the inodelock comes before any dquot lock,
746          * and we may drop and reacquire the ilock in xfs_qm_dqget().
747          */
748         if (udqhint)
749                 xfs_dqunlock(udqhint);
750         /*
751          * Find the dquot from somewhere. This bumps the
752          * reference count of dquot and returns it locked.
753          * This can return ENOENT if dquot didn't exist on
754          * disk and we didn't ask it to allocate;
755          * ESRCH if quotas got turned off suddenly.
756          */
757         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
758                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
759                 if (udqhint && dolock)
760                         xfs_dqlock(udqhint);
761                 goto done;
762         }
763
764         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
765         /*
766          * dqget may have dropped and re-acquired the ilock, but it guarantees
767          * that the dquot returned is the one that should go in the inode.
768          */
769         *IO_idqpp = dqp;
770         ASSERT(dqp);
771         ASSERT(XFS_DQ_IS_LOCKED(dqp));
772         if (! dolock) {
773                 xfs_dqunlock(dqp);
774                 goto done;
775         }
776         if (! udqhint)
777                 goto done;
778
779         ASSERT(udqhint);
780         ASSERT(dolock);
781         ASSERT(XFS_DQ_IS_LOCKED(dqp));
782         if (! xfs_qm_dqlock_nowait(udqhint)) {
783                 xfs_dqunlock(dqp);
784                 xfs_dqlock(udqhint);
785                 xfs_dqlock(dqp);
786         }
787       done:
788 #ifdef QUOTADEBUG
789         if (udqhint) {
790                 if (dolock)
791                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
792         }
793         if (! error) {
794                 if (dolock)
795                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
796         }
797 #endif
798         return (error);
799 }
800
801
802 /*
803  * Given a udquot and gdquot, attach a ptr to the group dquot in the
804  * udquot as a hint for future lookups. The idea sounds simple, but the
805  * execution isn't, because the udquot might have a group dquot attached
806  * already and getting rid of that gets us into lock ordering contraints.
807  * The process is complicated more by the fact that the dquots may or may not
808  * be locked on entry.
809  */
810 STATIC void
811 xfs_qm_dqattach_grouphint(
812         xfs_dquot_t     *udq,
813         xfs_dquot_t     *gdq,
814         uint            locked)
815 {
816         xfs_dquot_t     *tmp;
817
818 #ifdef QUOTADEBUG
819         if (locked) {
820                 ASSERT(XFS_DQ_IS_LOCKED(udq));
821                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
822         }
823 #endif
824         if (! locked)
825                 xfs_dqlock(udq);
826
827         if ((tmp = udq->q_gdquot)) {
828                 if (tmp == gdq) {
829                         if (! locked)
830                                 xfs_dqunlock(udq);
831                         return;
832                 }
833
834                 udq->q_gdquot = NULL;
835                 /*
836                  * We can't keep any dqlocks when calling dqrele,
837                  * because the freelist lock comes before dqlocks.
838                  */
839                 xfs_dqunlock(udq);
840                 if (locked)
841                         xfs_dqunlock(gdq);
842                 /*
843                  * we took a hard reference once upon a time in dqget,
844                  * so give it back when the udquot no longer points at it
845                  * dqput() does the unlocking of the dquot.
846                  */
847                 xfs_qm_dqrele(tmp);
848
849                 xfs_dqlock(udq);
850                 xfs_dqlock(gdq);
851
852         } else {
853                 ASSERT(XFS_DQ_IS_LOCKED(udq));
854                 if (! locked) {
855                         xfs_dqlock(gdq);
856                 }
857         }
858
859         ASSERT(XFS_DQ_IS_LOCKED(udq));
860         ASSERT(XFS_DQ_IS_LOCKED(gdq));
861         /*
862          * Somebody could have attached a gdquot here,
863          * when we dropped the uqlock. If so, just do nothing.
864          */
865         if (udq->q_gdquot == NULL) {
866                 XFS_DQHOLD(gdq);
867                 udq->q_gdquot = gdq;
868         }
869         if (! locked) {
870                 xfs_dqunlock(gdq);
871                 xfs_dqunlock(udq);
872         }
873 }
874
875
876 /*
877  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
878  * into account.
879  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
880  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
881  * much made this code a complete mess, but it has been pretty useful.
882  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
883  * Inode may get unlocked and relocked in here, and the caller must deal with
884  * the consequences.
885  */
886 int
887 xfs_qm_dqattach(
888         xfs_inode_t     *ip,
889         uint            flags)
890 {
891         xfs_mount_t     *mp = ip->i_mount;
892         uint            nquotas = 0;
893         int             error = 0;
894
895         if ((! XFS_IS_QUOTA_ON(mp)) ||
896             (! XFS_NOT_DQATTACHED(mp, ip)) ||
897             (ip->i_ino == mp->m_sb.sb_uquotino) ||
898             (ip->i_ino == mp->m_sb.sb_gquotino))
899                 return (0);
900
901         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
902                XFS_ISLOCKED_INODE_EXCL(ip));
903
904         if (! (flags & XFS_QMOPT_ILOCKED))
905                 xfs_ilock(ip, XFS_ILOCK_EXCL);
906
907         if (XFS_IS_UQUOTA_ON(mp)) {
908                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
909                                                 flags & XFS_QMOPT_DQALLOC,
910                                                 flags & XFS_QMOPT_DQLOCK,
911                                                 NULL, &ip->i_udquot);
912                 if (error)
913                         goto done;
914                 nquotas++;
915         }
916         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
917         if (XFS_IS_OQUOTA_ON(mp)) {
918                 error = XFS_IS_GQUOTA_ON(mp) ?
919                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
920                                                 flags & XFS_QMOPT_DQALLOC,
921                                                 flags & XFS_QMOPT_DQLOCK,
922                                                 ip->i_udquot, &ip->i_gdquot) :
923                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
924                                                 flags & XFS_QMOPT_DQALLOC,
925                                                 flags & XFS_QMOPT_DQLOCK,
926                                                 ip->i_udquot, &ip->i_gdquot);
927                 /*
928                  * Don't worry about the udquot that we may have
929                  * attached above. It'll get detached, if not already.
930                  */
931                 if (error)
932                         goto done;
933                 nquotas++;
934         }
935
936         /*
937          * Attach this group quota to the user quota as a hint.
938          * This WON'T, in general, result in a thrash.
939          */
940         if (nquotas == 2) {
941                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
942                 ASSERT(ip->i_udquot);
943                 ASSERT(ip->i_gdquot);
944
945                 /*
946                  * We may or may not have the i_udquot locked at this point,
947                  * but this check is OK since we don't depend on the i_gdquot to
948                  * be accurate 100% all the time. It is just a hint, and this
949                  * will succeed in general.
950                  */
951                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
952                         goto done;
953                 /*
954                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
955                  */
956                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
957                                          flags & XFS_QMOPT_DQLOCK);
958         }
959
960       done:
961
962 #ifdef QUOTADEBUG
963         if (! error) {
964                 if (ip->i_udquot) {
965                         if (flags & XFS_QMOPT_DQLOCK)
966                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
967                 }
968                 if (ip->i_gdquot) {
969                         if (flags & XFS_QMOPT_DQLOCK)
970                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
971                 }
972                 if (XFS_IS_UQUOTA_ON(mp))
973                         ASSERT(ip->i_udquot);
974                 if (XFS_IS_OQUOTA_ON(mp))
975                         ASSERT(ip->i_gdquot);
976         }
977 #endif
978
979         if (! (flags & XFS_QMOPT_ILOCKED))
980                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
981
982 #ifdef QUOTADEBUG
983         else
984                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
985 #endif
986         return (error);
987 }
988
989 /*
990  * Release dquots (and their references) if any.
991  * The inode should be locked EXCL except when this's called by
992  * xfs_ireclaim.
993  */
994 void
995 xfs_qm_dqdetach(
996         xfs_inode_t     *ip)
997 {
998         if (!(ip->i_udquot || ip->i_gdquot))
999                 return;
1000
1001         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1002         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1003         if (ip->i_udquot) {
1004                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1005                 xfs_qm_dqrele(ip->i_udquot);
1006                 ip->i_udquot = NULL;
1007         }
1008         if (ip->i_gdquot) {
1009                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
1010                 xfs_qm_dqrele(ip->i_gdquot);
1011                 ip->i_gdquot = NULL;
1012         }
1013 }
1014
1015 /*
1016  * This is called by VFS_SYNC and flags arg determines the caller,
1017  * and its motives, as done in xfs_sync.
1018  *
1019  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1020  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1021  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1022  */
1023
1024 int
1025 xfs_qm_sync(
1026         xfs_mount_t     *mp,
1027         short           flags)
1028 {
1029         int             recl, restarts;
1030         xfs_dquot_t     *dqp;
1031         uint            flush_flags;
1032         boolean_t       nowait;
1033         int             error;
1034
1035         restarts = 0;
1036         /*
1037          * We won't block unless we are asked to.
1038          */
1039         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1040
1041   again:
1042         xfs_qm_mplist_lock(mp);
1043         /*
1044          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1045          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1046          * when we have the mplist lock, we know that dquots will be consistent
1047          * as long as we have it locked.
1048          */
1049         if (! XFS_IS_QUOTA_ON(mp)) {
1050                 xfs_qm_mplist_unlock(mp);
1051                 return (0);
1052         }
1053         FOREACH_DQUOT_IN_MP(dqp, mp) {
1054                 /*
1055                  * If this is vfs_sync calling, then skip the dquots that
1056                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1057                  * This is very similar to what xfs_sync does with inodes.
1058                  */
1059                 if (flags & SYNC_BDFLUSH) {
1060                         if (! XFS_DQ_IS_DIRTY(dqp))
1061                                 continue;
1062                 }
1063
1064                 if (nowait) {
1065                         /*
1066                          * Try to acquire the dquot lock. We are NOT out of
1067                          * lock order, but we just don't want to wait for this
1068                          * lock, unless somebody wanted us to.
1069                          */
1070                         if (! xfs_qm_dqlock_nowait(dqp))
1071                                 continue;
1072                 } else {
1073                         xfs_dqlock(dqp);
1074                 }
1075
1076                 /*
1077                  * Now, find out for sure if this dquot is dirty or not.
1078                  */
1079                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1080                         xfs_dqunlock(dqp);
1081                         continue;
1082                 }
1083
1084                 /* XXX a sentinel would be better */
1085                 recl = XFS_QI_MPLRECLAIMS(mp);
1086                 if (! xfs_qm_dqflock_nowait(dqp)) {
1087                         if (nowait) {
1088                                 xfs_dqunlock(dqp);
1089                                 continue;
1090                         }
1091                         /*
1092                          * If we can't grab the flush lock then if the caller
1093                          * really wanted us to give this our best shot,
1094                          * see if we can give a push to the buffer before we wait
1095                          * on the flush lock. At this point, we know that
1096                          * eventhough the dquot is being flushed,
1097                          * it has (new) dirty data.
1098                          */
1099                         xfs_qm_dqflock_pushbuf_wait(dqp);
1100                 }
1101                 /*
1102                  * Let go of the mplist lock. We don't want to hold it
1103                  * across a disk write
1104                  */
1105                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1106                 xfs_qm_mplist_unlock(mp);
1107                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1108                 error = xfs_qm_dqflush(dqp, flush_flags);
1109                 xfs_dqunlock(dqp);
1110                 if (error && XFS_FORCED_SHUTDOWN(mp))
1111                         return(0);      /* Need to prevent umount failure */
1112                 else if (error)
1113                         return (error);
1114
1115                 xfs_qm_mplist_lock(mp);
1116                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1117                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1118                                 break;
1119
1120                         xfs_qm_mplist_unlock(mp);
1121                         goto again;
1122                 }
1123         }
1124
1125         xfs_qm_mplist_unlock(mp);
1126         return (0);
1127 }
1128
1129
1130 /*
1131  * This initializes all the quota information that's kept in the
1132  * mount structure
1133  */
1134 STATIC int
1135 xfs_qm_init_quotainfo(
1136         xfs_mount_t     *mp)
1137 {
1138         xfs_quotainfo_t *qinf;
1139         int             error;
1140         xfs_dquot_t     *dqp;
1141
1142         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1143
1144         /*
1145          * Tell XQM that we exist as soon as possible.
1146          */
1147         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1148                 return (error);
1149         }
1150
1151         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1152
1153         /*
1154          * See if quotainodes are setup, and if not, allocate them,
1155          * and change the superblock accordingly.
1156          */
1157         if ((error = xfs_qm_init_quotainos(mp))) {
1158                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1159                 mp->m_quotainfo = NULL;
1160                 return (error);
1161         }
1162
1163         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1164         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1165         qinf->qi_dqreclaims = 0;
1166
1167         /* mutex used to serialize quotaoffs */
1168         mutex_init(&qinf->qi_quotaofflock, MUTEX_DEFAULT, "qoff");
1169
1170         /* Precalc some constants */
1171         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1172         ASSERT(qinf->qi_dqchunklen);
1173         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1174         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1175
1176         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1177
1178         /*
1179          * We try to get the limits from the superuser's limits fields.
1180          * This is quite hacky, but it is standard quota practice.
1181          * We look at the USR dquot with id == 0 first, but if user quotas
1182          * are not enabled we goto the GRP dquot with id == 0.
1183          * We don't really care to keep separate default limits for user
1184          * and group quotas, at least not at this point.
1185          */
1186         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1187                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1188                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1189                                 XFS_DQ_PROJ),
1190                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1191                              &dqp);
1192         if (! error) {
1193                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1194
1195                 /*
1196                  * The warnings and timers set the grace period given to
1197                  * a user or group before he or she can not perform any
1198                  * more writing. If it is zero, a default is used.
1199                  */
1200                 qinf->qi_btimelimit =
1201                                 INT_GET(ddqp->d_btimer, ARCH_CONVERT) ?
1202                                 INT_GET(ddqp->d_btimer, ARCH_CONVERT) :
1203                                 XFS_QM_BTIMELIMIT;
1204                 qinf->qi_itimelimit =
1205                                 INT_GET(ddqp->d_itimer, ARCH_CONVERT) ?
1206                                 INT_GET(ddqp->d_itimer, ARCH_CONVERT) :
1207                                 XFS_QM_ITIMELIMIT;
1208                 qinf->qi_rtbtimelimit =
1209                                 INT_GET(ddqp->d_rtbtimer, ARCH_CONVERT) ?
1210                                 INT_GET(ddqp->d_rtbtimer, ARCH_CONVERT) :
1211                                 XFS_QM_RTBTIMELIMIT;
1212                 qinf->qi_bwarnlimit =
1213                                 INT_GET(ddqp->d_bwarns, ARCH_CONVERT) ?
1214                                 INT_GET(ddqp->d_bwarns, ARCH_CONVERT) :
1215                                 XFS_QM_BWARNLIMIT;
1216                 qinf->qi_iwarnlimit =
1217                                 INT_GET(ddqp->d_iwarns, ARCH_CONVERT) ?
1218                                 INT_GET(ddqp->d_iwarns, ARCH_CONVERT) :
1219                                 XFS_QM_IWARNLIMIT;
1220                 qinf->qi_rtbwarnlimit =
1221                                 INT_GET(ddqp->d_rtbwarns, ARCH_CONVERT) ?
1222                                 INT_GET(ddqp->d_rtbwarns, ARCH_CONVERT) :
1223                                 XFS_QM_RTBWARNLIMIT;
1224                 qinf->qi_bhardlimit =
1225                                 INT_GET(ddqp->d_blk_hardlimit, ARCH_CONVERT);
1226                 qinf->qi_bsoftlimit =
1227                                 INT_GET(ddqp->d_blk_softlimit, ARCH_CONVERT);
1228                 qinf->qi_ihardlimit =
1229                                 INT_GET(ddqp->d_ino_hardlimit, ARCH_CONVERT);
1230                 qinf->qi_isoftlimit =
1231                                 INT_GET(ddqp->d_ino_softlimit, ARCH_CONVERT);
1232                 qinf->qi_rtbhardlimit =
1233                                 INT_GET(ddqp->d_rtb_hardlimit, ARCH_CONVERT);
1234                 qinf->qi_rtbsoftlimit =
1235                                 INT_GET(ddqp->d_rtb_softlimit, ARCH_CONVERT);
1236  
1237                 /*
1238                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1239                  * we don't want this dquot cached. We haven't done a
1240                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1241                  */
1242                 xfs_qm_dqdestroy(dqp);
1243         } else {
1244                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1245                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1246                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1247                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1248                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1249                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1250         }
1251
1252         return (0);
1253 }
1254
1255
1256 /*
1257  * Gets called when unmounting a filesystem or when all quotas get
1258  * turned off.
1259  * This purges the quota inodes, destroys locks and frees itself.
1260  */
1261 void
1262 xfs_qm_destroy_quotainfo(
1263         xfs_mount_t     *mp)
1264 {
1265         xfs_quotainfo_t *qi;
1266
1267         qi = mp->m_quotainfo;
1268         ASSERT(qi != NULL);
1269         ASSERT(xfs_Gqm != NULL);
1270
1271         /*
1272          * Release the reference that XQM kept, so that we know
1273          * when the XQM structure should be freed. We cannot assume
1274          * that xfs_Gqm is non-null after this point.
1275          */
1276         xfs_qm_rele_quotafs_ref(mp);
1277
1278         spinlock_destroy(&qi->qi_pinlock);
1279         xfs_qm_list_destroy(&qi->qi_dqlist);
1280
1281         if (qi->qi_uquotaip) {
1282                 XFS_PURGE_INODE(qi->qi_uquotaip);
1283                 qi->qi_uquotaip = NULL; /* paranoia */
1284         }
1285         if (qi->qi_gquotaip) {
1286                 XFS_PURGE_INODE(qi->qi_gquotaip);
1287                 qi->qi_gquotaip = NULL;
1288         }
1289         mutex_destroy(&qi->qi_quotaofflock);
1290         kmem_free(qi, sizeof(xfs_quotainfo_t));
1291         mp->m_quotainfo = NULL;
1292 }
1293
1294
1295
1296 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1297
1298 /* ARGSUSED */
1299 STATIC void
1300 xfs_qm_list_init(
1301         xfs_dqlist_t    *list,
1302         char            *str,
1303         int             n)
1304 {
1305         mutex_init(&list->qh_lock, MUTEX_DEFAULT, str);
1306         list->qh_next = NULL;
1307         list->qh_version = 0;
1308         list->qh_nelems = 0;
1309 }
1310
1311 STATIC void
1312 xfs_qm_list_destroy(
1313         xfs_dqlist_t    *list)
1314 {
1315         mutex_destroy(&(list->qh_lock));
1316 }
1317
1318
1319 /*
1320  * Stripped down version of dqattach. This doesn't attach, or even look at the
1321  * dquots attached to the inode. The rationale is that there won't be any
1322  * attached at the time this is called from quotacheck.
1323  */
1324 STATIC int
1325 xfs_qm_dqget_noattach(
1326         xfs_inode_t     *ip,
1327         xfs_dquot_t     **O_udqpp,
1328         xfs_dquot_t     **O_gdqpp)
1329 {
1330         int             error;
1331         xfs_mount_t     *mp;
1332         xfs_dquot_t     *udqp, *gdqp;
1333
1334         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1335         mp = ip->i_mount;
1336         udqp = NULL;
1337         gdqp = NULL;
1338
1339         if (XFS_IS_UQUOTA_ON(mp)) {
1340                 ASSERT(ip->i_udquot == NULL);
1341                 /*
1342                  * We want the dquot allocated if it doesn't exist.
1343                  */
1344                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1345                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1346                                          &udqp))) {
1347                         /*
1348                          * Shouldn't be able to turn off quotas here.
1349                          */
1350                         ASSERT(error != ESRCH);
1351                         ASSERT(error != ENOENT);
1352                         return (error);
1353                 }
1354                 ASSERT(udqp);
1355         }
1356
1357         if (XFS_IS_OQUOTA_ON(mp)) {
1358                 ASSERT(ip->i_gdquot == NULL);
1359                 if (udqp)
1360                         xfs_dqunlock(udqp);
1361                 error = XFS_IS_GQUOTA_ON(mp) ?
1362                                 xfs_qm_dqget(mp, ip,
1363                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1364                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1365                                              &gdqp) :
1366                                 xfs_qm_dqget(mp, ip,
1367                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1368                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1369                                              &gdqp);
1370                 if (error) {
1371                         if (udqp)
1372                                 xfs_qm_dqrele(udqp);
1373                         ASSERT(error != ESRCH);
1374                         ASSERT(error != ENOENT);
1375                         return (error);
1376                 }
1377                 ASSERT(gdqp);
1378
1379                 /* Reacquire the locks in the right order */
1380                 if (udqp) {
1381                         if (! xfs_qm_dqlock_nowait(udqp)) {
1382                                 xfs_dqunlock(gdqp);
1383                                 xfs_dqlock(udqp);
1384                                 xfs_dqlock(gdqp);
1385                         }
1386                 }
1387         }
1388
1389         *O_udqpp = udqp;
1390         *O_gdqpp = gdqp;
1391
1392 #ifdef QUOTADEBUG
1393         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1394         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1395 #endif
1396         return (0);
1397 }
1398
1399 /*
1400  * Create an inode and return with a reference already taken, but unlocked
1401  * This is how we create quota inodes
1402  */
1403 STATIC int
1404 xfs_qm_qino_alloc(
1405         xfs_mount_t     *mp,
1406         xfs_inode_t     **ip,
1407         __int64_t       sbfields,
1408         uint            flags)
1409 {
1410         xfs_trans_t     *tp;
1411         int             error;
1412         unsigned long s;
1413         cred_t          zerocr;
1414         int             committed;
1415
1416         tp = xfs_trans_alloc(mp,XFS_TRANS_QM_QINOCREATE);
1417         if ((error = xfs_trans_reserve(tp,
1418                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1419                                       XFS_CREATE_LOG_RES(mp), 0,
1420                                       XFS_TRANS_PERM_LOG_RES,
1421                                       XFS_CREATE_LOG_COUNT))) {
1422                 xfs_trans_cancel(tp, 0);
1423                 return (error);
1424         }
1425         memset(&zerocr, 0, sizeof(zerocr));
1426
1427         if ((error = xfs_dir_ialloc(&tp, mp->m_rootip, S_IFREG, 1, 0,
1428                                    &zerocr, 0, 1, ip, &committed))) {
1429                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1430                                  XFS_TRANS_ABORT);
1431                 return (error);
1432         }
1433
1434         /*
1435          * Keep an extra reference to this quota inode. This inode is
1436          * locked exclusively and joined to the transaction already.
1437          */
1438         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1439         VN_HOLD(XFS_ITOV((*ip)));
1440
1441         /*
1442          * Make the changes in the superblock, and log those too.
1443          * sbfields arg may contain fields other than *QUOTINO;
1444          * VERSIONNUM for example.
1445          */
1446         s = XFS_SB_LOCK(mp);
1447         if (flags & XFS_QMOPT_SBVERSION) {
1448 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1449                 unsigned oldv = mp->m_sb.sb_versionnum;
1450 #endif
1451                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1452                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1453                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1454                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1455                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1456
1457                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1458                 mp->m_sb.sb_uquotino = NULLFSINO;
1459                 mp->m_sb.sb_gquotino = NULLFSINO;
1460
1461                 /* qflags will get updated _after_ quotacheck */
1462                 mp->m_sb.sb_qflags = 0;
1463 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1464                 cmn_err(CE_NOTE,
1465                         "Old superblock version %x, converting to %x.",
1466                         oldv, mp->m_sb.sb_versionnum);
1467 #endif
1468         }
1469         if (flags & XFS_QMOPT_UQUOTA)
1470                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1471         else
1472                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1473         XFS_SB_UNLOCK(mp, s);
1474         xfs_mod_sb(tp, sbfields);
1475
1476         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES,
1477                                      NULL))) {
1478                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1479                 return (error);
1480         }
1481         return (0);
1482 }
1483
1484
1485 STATIC int
1486 xfs_qm_reset_dqcounts(
1487         xfs_mount_t     *mp,
1488         xfs_buf_t       *bp,
1489         xfs_dqid_t      id,
1490         uint            type)
1491 {
1492         xfs_disk_dquot_t        *ddq;
1493         int                     j;
1494
1495         xfs_buftrace("RESET DQUOTS", bp);
1496         /*
1497          * Reset all counters and timers. They'll be
1498          * started afresh by xfs_qm_quotacheck.
1499          */
1500 #ifdef DEBUG
1501         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1502         do_div(j, sizeof(xfs_dqblk_t));
1503         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1504 #endif
1505         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1506         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1507                 /*
1508                  * Do a sanity check, and if needed, repair the dqblk. Don't
1509                  * output any warnings because it's perfectly possible to
1510                  * find unitialized dquot blks. See comment in xfs_qm_dqcheck.
1511                  */
1512                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1513                                       "xfs_quotacheck");
1514                 INT_SET(ddq->d_bcount, ARCH_CONVERT, 0ULL);
1515                 INT_SET(ddq->d_icount, ARCH_CONVERT, 0ULL);
1516                 INT_SET(ddq->d_rtbcount, ARCH_CONVERT, 0ULL);
1517                 INT_SET(ddq->d_btimer, ARCH_CONVERT, (time_t)0);
1518                 INT_SET(ddq->d_itimer, ARCH_CONVERT, (time_t)0);
1519                 INT_SET(ddq->d_rtbtimer, ARCH_CONVERT, (time_t)0);
1520                 INT_SET(ddq->d_bwarns, ARCH_CONVERT, 0UL);
1521                 INT_SET(ddq->d_iwarns, ARCH_CONVERT, 0UL);
1522                 INT_SET(ddq->d_rtbwarns, ARCH_CONVERT, 0UL);
1523                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1524         }
1525
1526         return (0);
1527 }
1528
1529 STATIC int
1530 xfs_qm_dqiter_bufs(
1531         xfs_mount_t     *mp,
1532         xfs_dqid_t      firstid,
1533         xfs_fsblock_t   bno,
1534         xfs_filblks_t   blkcnt,
1535         uint            flags)
1536 {
1537         xfs_buf_t       *bp;
1538         int             error;
1539         int             notcommitted;
1540         int             incr;
1541         int             type;
1542
1543         ASSERT(blkcnt > 0);
1544         notcommitted = 0;
1545         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1546                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1547         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1548                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1549         error = 0;
1550
1551         /*
1552          * Blkcnt arg can be a very big number, and might even be
1553          * larger than the log itself. So, we have to break it up into
1554          * manageable-sized transactions.
1555          * Note that we don't start a permanent transaction here; we might
1556          * not be able to get a log reservation for the whole thing up front,
1557          * and we don't really care to either, because we just discard
1558          * everything if we were to crash in the middle of this loop.
1559          */
1560         while (blkcnt--) {
1561                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1562                               XFS_FSB_TO_DADDR(mp, bno),
1563                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1564                 if (error)
1565                         break;
1566
1567                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1568                 xfs_bdwrite(mp, bp);
1569                 /*
1570                  * goto the next block.
1571                  */
1572                 bno++;
1573                 firstid += XFS_QM_DQPERBLK(mp);
1574         }
1575         return (error);
1576 }
1577
1578 /*
1579  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1580  * caller supplied function for every chunk of dquots that we find.
1581  */
1582 STATIC int
1583 xfs_qm_dqiterate(
1584         xfs_mount_t     *mp,
1585         xfs_inode_t     *qip,
1586         uint            flags)
1587 {
1588         xfs_bmbt_irec_t         *map;
1589         int                     i, nmaps;       /* number of map entries */
1590         int                     error;          /* return value */
1591         xfs_fileoff_t           lblkno;
1592         xfs_filblks_t           maxlblkcnt;
1593         xfs_dqid_t              firstid;
1594         xfs_fsblock_t           rablkno;
1595         xfs_filblks_t           rablkcnt;
1596
1597         error = 0;
1598         /*
1599          * This looks racey, but we can't keep an inode lock across a
1600          * trans_reserve. But, this gets called during quotacheck, and that
1601          * happens only at mount time which is single threaded.
1602          */
1603         if (qip->i_d.di_nblocks == 0)
1604                 return (0);
1605
1606         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1607
1608         lblkno = 0;
1609         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1610         do {
1611                 nmaps = XFS_DQITER_MAP_SIZE;
1612                 /*
1613                  * We aren't changing the inode itself. Just changing
1614                  * some of its data. No new blocks are added here, and
1615                  * the inode is never added to the transaction.
1616                  */
1617                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1618                 error = xfs_bmapi(NULL, qip, lblkno,
1619                                   maxlblkcnt - lblkno,
1620                                   XFS_BMAPI_METADATA,
1621                                   NULL,
1622                                   0, map, &nmaps, NULL);
1623                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1624                 if (error)
1625                         break;
1626
1627                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1628                 for (i = 0; i < nmaps; i++) {
1629                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1630                         ASSERT(map[i].br_blockcount);
1631
1632
1633                         lblkno += map[i].br_blockcount;
1634
1635                         if (map[i].br_startblock == HOLESTARTBLOCK)
1636                                 continue;
1637
1638                         firstid = (xfs_dqid_t) map[i].br_startoff *
1639                                 XFS_QM_DQPERBLK(mp);
1640                         /*
1641                          * Do a read-ahead on the next extent.
1642                          */
1643                         if ((i+1 < nmaps) &&
1644                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1645                                 rablkcnt =  map[i+1].br_blockcount;
1646                                 rablkno = map[i+1].br_startblock;
1647                                 while (rablkcnt--) {
1648                                         xfs_baread(mp->m_ddev_targp,
1649                                                XFS_FSB_TO_DADDR(mp, rablkno),
1650                                                (int)XFS_QI_DQCHUNKLEN(mp));
1651                                         rablkno++;
1652                                 }
1653                         }
1654                         /*
1655                          * Iterate thru all the blks in the extent and
1656                          * reset the counters of all the dquots inside them.
1657                          */
1658                         if ((error = xfs_qm_dqiter_bufs(mp,
1659                                                        firstid,
1660                                                        map[i].br_startblock,
1661                                                        map[i].br_blockcount,
1662                                                        flags))) {
1663                                 break;
1664                         }
1665                 }
1666
1667                 if (error)
1668                         break;
1669         } while (nmaps > 0);
1670
1671         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1672
1673         return (error);
1674 }
1675
1676 /*
1677  * Called by dqusage_adjust in doing a quotacheck.
1678  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1679  * this updates its incore copy as well as the buffer copy. This is
1680  * so that once the quotacheck is done, we can just log all the buffers,
1681  * as opposed to logging numerous updates to individual dquots.
1682  */
1683 STATIC void
1684 xfs_qm_quotacheck_dqadjust(
1685         xfs_dquot_t             *dqp,
1686         xfs_qcnt_t              nblks,
1687         xfs_qcnt_t              rtblks)
1688 {
1689         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1690         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1691         /*
1692          * Adjust the inode count and the block count to reflect this inode's
1693          * resource usage.
1694          */
1695         INT_MOD(dqp->q_core.d_icount, ARCH_CONVERT, +1);
1696         dqp->q_res_icount++;
1697         if (nblks) {
1698                 INT_MOD(dqp->q_core.d_bcount, ARCH_CONVERT, nblks);
1699                 dqp->q_res_bcount += nblks;
1700         }
1701         if (rtblks) {
1702                 INT_MOD(dqp->q_core.d_rtbcount, ARCH_CONVERT, rtblks);
1703                 dqp->q_res_rtbcount += rtblks;
1704         }
1705
1706         /*
1707          * Set default limits, adjust timers (since we changed usages)
1708          */
1709         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1710                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1711                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1712         }
1713
1714         dqp->dq_flags |= XFS_DQ_DIRTY;
1715 }
1716
1717 STATIC int
1718 xfs_qm_get_rtblks(
1719         xfs_inode_t     *ip,
1720         xfs_qcnt_t      *O_rtblks)
1721 {
1722         xfs_filblks_t   rtblks;                 /* total rt blks */
1723         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1724         xfs_extnum_t    nextents;               /* number of extent entries */
1725         xfs_bmbt_rec_t  *base;                  /* base of extent array */
1726         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1727         int             error;
1728
1729         ASSERT(XFS_IS_REALTIME_INODE(ip));
1730         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1731         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1732                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1733                         return (error);
1734         }
1735         rtblks = 0;
1736         nextents = ifp->if_bytes / sizeof(xfs_bmbt_rec_t);
1737         base = &ifp->if_u1.if_extents[0];
1738         for (ep = base; ep < &base[nextents]; ep++)
1739                 rtblks += xfs_bmbt_get_blockcount(ep);
1740         *O_rtblks = (xfs_qcnt_t)rtblks;
1741         return (0);
1742 }
1743
1744 /*
1745  * callback routine supplied to bulkstat(). Given an inumber, find its
1746  * dquots and update them to account for resources taken by that inode.
1747  */
1748 /* ARGSUSED */
1749 STATIC int
1750 xfs_qm_dqusage_adjust(
1751         xfs_mount_t     *mp,            /* mount point for filesystem */
1752         xfs_ino_t       ino,            /* inode number to get data for */
1753         void            __user *buffer, /* not used */
1754         int             ubsize,         /* not used */
1755         void            *private_data,  /* not used */
1756         xfs_daddr_t     bno,            /* starting block of inode cluster */
1757         int             *ubused,        /* not used */
1758         void            *dip,           /* on-disk inode pointer (not used) */
1759         int             *res)           /* result code value */
1760 {
1761         xfs_inode_t     *ip;
1762         xfs_dquot_t     *udqp, *gdqp;
1763         xfs_qcnt_t      nblks, rtblks;
1764         int             error;
1765
1766         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1767
1768         /*
1769          * rootino must have its resources accounted for, not so with the quota
1770          * inodes.
1771          */
1772         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1773                 *res = BULKSTAT_RV_NOTHING;
1774                 return XFS_ERROR(EINVAL);
1775         }
1776
1777         /*
1778          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1779          * interface expects the inode to be exclusively locked because that's
1780          * the case in all other instances. It's OK that we do this because
1781          * quotacheck is done only at mount time.
1782          */
1783         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1784                 *res = BULKSTAT_RV_NOTHING;
1785                 return (error);
1786         }
1787
1788         if (ip->i_d.di_mode == 0) {
1789                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1790                 *res = BULKSTAT_RV_NOTHING;
1791                 return XFS_ERROR(ENOENT);
1792         }
1793
1794         /*
1795          * Obtain the locked dquots. In case of an error (eg. allocation
1796          * fails for ENOSPC), we return the negative of the error number
1797          * to bulkstat, so that it can get propagated to quotacheck() and
1798          * making us disable quotas for the file system.
1799          */
1800         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1801                 xfs_iput(ip, XFS_ILOCK_EXCL);
1802                 *res = BULKSTAT_RV_GIVEUP;
1803                 return (error);
1804         }
1805
1806         rtblks = 0;
1807         if (! XFS_IS_REALTIME_INODE(ip)) {
1808                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1809         } else {
1810                 /*
1811                  * Walk thru the extent list and count the realtime blocks.
1812                  */
1813                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1814                         xfs_iput(ip, XFS_ILOCK_EXCL);
1815                         if (udqp)
1816                                 xfs_qm_dqput(udqp);
1817                         if (gdqp)
1818                                 xfs_qm_dqput(gdqp);
1819                         *res = BULKSTAT_RV_GIVEUP;
1820                         return (error);
1821                 }
1822                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1823         }
1824         ASSERT(ip->i_delayed_blks == 0);
1825
1826         /*
1827          * We can't release the inode while holding its dquot locks.
1828          * The inode can go into inactive and might try to acquire the dquotlocks.
1829          * So, just unlock here and do a vn_rele at the end.
1830          */
1831         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1832
1833         /*
1834          * Add the (disk blocks and inode) resources occupied by this
1835          * inode to its dquots. We do this adjustment in the incore dquot,
1836          * and also copy the changes to its buffer.
1837          * We don't care about putting these changes in a transaction
1838          * envelope because if we crash in the middle of a 'quotacheck'
1839          * we have to start from the beginning anyway.
1840          * Once we're done, we'll log all the dquot bufs.
1841          *
1842          * The *QUOTA_ON checks below may look pretty racey, but quotachecks
1843          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1844          */
1845         if (XFS_IS_UQUOTA_ON(mp)) {
1846                 ASSERT(udqp);
1847                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1848                 xfs_qm_dqput(udqp);
1849         }
1850         if (XFS_IS_OQUOTA_ON(mp)) {
1851                 ASSERT(gdqp);
1852                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1853                 xfs_qm_dqput(gdqp);
1854         }
1855         /*
1856          * Now release the inode. This will send it to 'inactive', and
1857          * possibly even free blocks.
1858          */
1859         VN_RELE(XFS_ITOV(ip));
1860
1861         /*
1862          * Goto next inode.
1863          */
1864         *res = BULKSTAT_RV_DIDONE;
1865         return (0);
1866 }
1867
1868 /*
1869  * Walk thru all the filesystem inodes and construct a consistent view
1870  * of the disk quota world. If the quotacheck fails, disable quotas.
1871  */
1872 int
1873 xfs_qm_quotacheck(
1874         xfs_mount_t     *mp)
1875 {
1876         int             done, count, error;
1877         xfs_ino_t       lastino;
1878         size_t          structsz;
1879         xfs_inode_t     *uip, *gip;
1880         uint            flags;
1881
1882         count = INT_MAX;
1883         structsz = 1;
1884         lastino = 0;
1885         flags = 0;
1886
1887         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1888         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1889
1890         /*
1891          * There should be no cached dquots. The (simplistic) quotacheck
1892          * algorithm doesn't like that.
1893          */
1894         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1895
1896         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1897
1898         /*
1899          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1900          * their counters to zero. We need a clean slate.
1901          * We don't log our changes till later.
1902          */
1903         if ((uip = XFS_QI_UQIP(mp))) {
1904                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1905                         goto error_return;
1906                 flags |= XFS_UQUOTA_CHKD;
1907         }
1908
1909         if ((gip = XFS_QI_GQIP(mp))) {
1910                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1911                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1912                         goto error_return;
1913                 flags |= XFS_OQUOTA_CHKD;
1914         }
1915
1916         do {
1917                 /*
1918                  * Iterate thru all the inodes in the file system,
1919                  * adjusting the corresponding dquot counters in core.
1920                  */
1921                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1922                                      xfs_qm_dqusage_adjust, NULL,
1923                                      structsz, NULL,
1924                                      BULKSTAT_FG_IGET|BULKSTAT_FG_VFSLOCKED,
1925                                      &done)))
1926                         break;
1927
1928         } while (! done);
1929
1930         /*
1931          * We can get this error if we couldn't do a dquot allocation inside
1932          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1933          * dirty dquots that might be cached, we just want to get rid of them
1934          * and turn quotaoff. The dquots won't be attached to any of the inodes
1935          * at this point (because we intentionally didn't in dqget_noattach).
1936          */
1937         if (error) {
1938                 xfs_qm_dqpurge_all(mp,
1939                                    XFS_QMOPT_UQUOTA|XFS_QMOPT_GQUOTA|
1940                                    XFS_QMOPT_PQUOTA|XFS_QMOPT_QUOTAOFF);
1941                 goto error_return;
1942         }
1943         /*
1944          * We've made all the changes that we need to make incore.
1945          * Now flush_them down to disk buffers.
1946          */
1947         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1948
1949         /*
1950          * We didn't log anything, because if we crashed, we'll have to
1951          * start the quotacheck from scratch anyway. However, we must make
1952          * sure that our dquot changes are secure before we put the
1953          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1954          * flush.
1955          */
1956         XFS_bflush(mp->m_ddev_targp);
1957
1958         /*
1959          * If one type of quotas is off, then it will lose its
1960          * quotachecked status, since we won't be doing accounting for
1961          * that type anymore.
1962          */
1963         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1964         mp->m_qflags |= flags;
1965
1966         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1967
1968  error_return:
1969         if (error) {
1970                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1971                         "Disabling quotas.",
1972                         mp->m_fsname, error);
1973                 /*
1974                  * We must turn off quotas.
1975                  */
1976                 ASSERT(mp->m_quotainfo != NULL);
1977                 ASSERT(xfs_Gqm != NULL);
1978                 xfs_qm_destroy_quotainfo(mp);
1979                 (void)xfs_mount_reset_sbqflags(mp);
1980         } else {
1981                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1982         }
1983         return (error);
1984 }
1985
1986 /*
1987  * This is called after the superblock has been read in and we're ready to
1988  * iget the quota inodes.
1989  */
1990 STATIC int
1991 xfs_qm_init_quotainos(
1992         xfs_mount_t     *mp)
1993 {
1994         xfs_inode_t     *uip, *gip;
1995         int             error;
1996         __int64_t       sbflags;
1997         uint            flags;
1998
1999         ASSERT(mp->m_quotainfo);
2000         uip = gip = NULL;
2001         sbflags = 0;
2002         flags = 0;
2003
2004         /*
2005          * Get the uquota and gquota inodes
2006          */
2007         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
2008                 if (XFS_IS_UQUOTA_ON(mp) &&
2009                     mp->m_sb.sb_uquotino != NULLFSINO) {
2010                         ASSERT(mp->m_sb.sb_uquotino > 0);
2011                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
2012                                              0, 0, &uip, 0)))
2013                                 return XFS_ERROR(error);
2014                 }
2015                 if (XFS_IS_OQUOTA_ON(mp) &&
2016                     mp->m_sb.sb_gquotino != NULLFSINO) {
2017                         ASSERT(mp->m_sb.sb_gquotino > 0);
2018                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2019                                              0, 0, &gip, 0))) {
2020                                 if (uip)
2021                                         VN_RELE(XFS_ITOV(uip));
2022                                 return XFS_ERROR(error);
2023                         }
2024                 }
2025         } else {
2026                 flags |= XFS_QMOPT_SBVERSION;
2027                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2028                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2029         }
2030
2031         /*
2032          * Create the two inodes, if they don't exist already. The changes
2033          * made above will get added to a transaction and logged in one of
2034          * the qino_alloc calls below.  If the device is readonly,
2035          * temporarily switch to read-write to do this.
2036          */
2037         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2038                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2039                                               sbflags | XFS_SB_UQUOTINO,
2040                                               flags | XFS_QMOPT_UQUOTA)))
2041                         return XFS_ERROR(error);
2042
2043                 flags &= ~XFS_QMOPT_SBVERSION;
2044         }
2045         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
2046                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
2047                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2048                 error = xfs_qm_qino_alloc(mp, &gip,
2049                                           sbflags | XFS_SB_GQUOTINO, flags);
2050                 if (error) {
2051                         if (uip)
2052                                 VN_RELE(XFS_ITOV(uip));
2053
2054                         return XFS_ERROR(error);
2055                 }
2056         }
2057
2058         XFS_QI_UQIP(mp) = uip;
2059         XFS_QI_GQIP(mp) = gip;
2060
2061         return (0);
2062 }
2063
2064
2065 /*
2066  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2067  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2068  * favor the lookup function ...
2069  * XXXsup merge this with qm_reclaim_one().
2070  */
2071 STATIC int
2072 xfs_qm_shake_freelist(
2073         int howmany)
2074 {
2075         int             nreclaimed;
2076         xfs_dqhash_t    *hash;
2077         xfs_dquot_t     *dqp, *nextdqp;
2078         int             restarts;
2079         int             nflushes;
2080
2081         if (howmany <= 0)
2082                 return (0);
2083
2084         nreclaimed = 0;
2085         restarts = 0;
2086         nflushes = 0;
2087
2088 #ifdef QUOTADEBUG
2089         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2090 #endif
2091         /* lock order is : hashchainlock, freelistlock, mplistlock */
2092  tryagain:
2093         xfs_qm_freelist_lock(xfs_Gqm);
2094
2095         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2096              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2097               nreclaimed < howmany); ) {
2098                 xfs_dqlock(dqp);
2099
2100                 /*
2101                  * We are racing with dqlookup here. Naturally we don't
2102                  * want to reclaim a dquot that lookup wants.
2103                  */
2104                 if (dqp->dq_flags & XFS_DQ_WANT) {
2105                         xfs_dqunlock(dqp);
2106                         xfs_qm_freelist_unlock(xfs_Gqm);
2107                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2108                                 return (nreclaimed);
2109                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2110                         goto tryagain;
2111                 }
2112
2113                 /*
2114                  * If the dquot is inactive, we are assured that it is
2115                  * not on the mplist or the hashlist, and that makes our
2116                  * life easier.
2117                  */
2118                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2119                         ASSERT(dqp->q_mount == NULL);
2120                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2121                         ASSERT(dqp->HL_PREVP == NULL);
2122                         ASSERT(dqp->MPL_PREVP == NULL);
2123                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2124                         nextdqp = dqp->dq_flnext;
2125                         goto off_freelist;
2126                 }
2127
2128                 ASSERT(dqp->MPL_PREVP);
2129                 /*
2130                  * Try to grab the flush lock. If this dquot is in the process of
2131                  * getting flushed to disk, we don't want to reclaim it.
2132                  */
2133                 if (! xfs_qm_dqflock_nowait(dqp)) {
2134                         xfs_dqunlock(dqp);
2135                         dqp = dqp->dq_flnext;
2136                         continue;
2137                 }
2138
2139                 /*
2140                  * We have the flush lock so we know that this is not in the
2141                  * process of being flushed. So, if this is dirty, flush it
2142                  * DELWRI so that we don't get a freelist infested with
2143                  * dirty dquots.
2144                  */
2145                 if (XFS_DQ_IS_DIRTY(dqp)) {
2146                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2147                         /*
2148                          * We flush it delayed write, so don't bother
2149                          * releasing the mplock.
2150                          */
2151                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2152                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2153                         dqp = dqp->dq_flnext;
2154                         continue;
2155                 }
2156                 /*
2157                  * We're trying to get the hashlock out of order. This races
2158                  * with dqlookup; so, we giveup and goto the next dquot if
2159                  * we couldn't get the hashlock. This way, we won't starve
2160                  * a dqlookup process that holds the hashlock that is
2161                  * waiting for the freelist lock.
2162                  */
2163                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2164                         xfs_dqfunlock(dqp);
2165                         xfs_dqunlock(dqp);
2166                         dqp = dqp->dq_flnext;
2167                         continue;
2168                 }
2169                 /*
2170                  * This races with dquot allocation code as well as dqflush_all
2171                  * and reclaim code. So, if we failed to grab the mplist lock,
2172                  * giveup everything and start over.
2173                  */
2174                 hash = dqp->q_hash;
2175                 ASSERT(hash);
2176                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2177                         /* XXX put a sentinel so that we can come back here */
2178                         xfs_dqfunlock(dqp);
2179                         xfs_dqunlock(dqp);
2180                         XFS_DQ_HASH_UNLOCK(hash);
2181                         xfs_qm_freelist_unlock(xfs_Gqm);
2182                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2183                                 return (nreclaimed);
2184                         goto tryagain;
2185                 }
2186                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2187 #ifdef QUOTADEBUG
2188                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2189                         dqp, INT_GET(dqp->q_core.d_id, ARCH_CONVERT));
2190 #endif
2191                 ASSERT(dqp->q_nrefs == 0);
2192                 nextdqp = dqp->dq_flnext;
2193                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2194                 XQM_HASHLIST_REMOVE(hash, dqp);
2195                 xfs_dqfunlock(dqp);
2196                 xfs_qm_mplist_unlock(dqp->q_mount);
2197                 XFS_DQ_HASH_UNLOCK(hash);
2198
2199  off_freelist:
2200                 XQM_FREELIST_REMOVE(dqp);
2201                 xfs_dqunlock(dqp);
2202                 nreclaimed++;
2203                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2204                 xfs_qm_dqdestroy(dqp);
2205                 dqp = nextdqp;
2206         }
2207         xfs_qm_freelist_unlock(xfs_Gqm);
2208         return (nreclaimed);
2209 }
2210
2211
2212 /*
2213  * The kmem_shake interface is invoked when memory is running low.
2214  */
2215 /* ARGSUSED */
2216 STATIC int
2217 xfs_qm_shake(int nr_to_scan, unsigned int gfp_mask)
2218 {
2219         int     ndqused, nfree, n;
2220
2221         if (!kmem_shake_allow(gfp_mask))
2222                 return (0);
2223         if (!xfs_Gqm)
2224                 return (0);
2225
2226         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2227         /* incore dquots in all f/s's */
2228         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2229
2230         ASSERT(ndqused >= 0);
2231
2232         if (nfree <= ndqused && nfree < ndquot)
2233                 return (0);
2234
2235         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2236         n = nfree - ndqused - ndquot;           /* # over target */
2237
2238         return xfs_qm_shake_freelist(MAX(nfree, n));
2239 }
2240
2241
2242 /*
2243  * Just pop the least recently used dquot off the freelist and
2244  * recycle it. The returned dquot is locked.
2245  */
2246 STATIC xfs_dquot_t *
2247 xfs_qm_dqreclaim_one(void)
2248 {
2249         xfs_dquot_t     *dqpout;
2250         xfs_dquot_t     *dqp;
2251         int             restarts;
2252         int             nflushes;
2253
2254         restarts = 0;
2255         dqpout = NULL;
2256         nflushes = 0;
2257
2258         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2259  startagain:
2260         xfs_qm_freelist_lock(xfs_Gqm);
2261
2262         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2263                 xfs_dqlock(dqp);
2264
2265                 /*
2266                  * We are racing with dqlookup here. Naturally we don't
2267                  * want to reclaim a dquot that lookup wants. We release the
2268                  * freelist lock and start over, so that lookup will grab
2269                  * both the dquot and the freelistlock.
2270                  */
2271                 if (dqp->dq_flags & XFS_DQ_WANT) {
2272                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2273                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2274                         xfs_dqunlock(dqp);
2275                         xfs_qm_freelist_unlock(xfs_Gqm);
2276                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2277                                 return (NULL);
2278                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2279                         goto startagain;
2280                 }
2281
2282                 /*
2283                  * If the dquot is inactive, we are assured that it is
2284                  * not on the mplist or the hashlist, and that makes our
2285                  * life easier.
2286                  */
2287                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2288                         ASSERT(dqp->q_mount == NULL);
2289                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2290                         ASSERT(dqp->HL_PREVP == NULL);
2291                         ASSERT(dqp->MPL_PREVP == NULL);
2292                         XQM_FREELIST_REMOVE(dqp);
2293                         xfs_dqunlock(dqp);
2294                         dqpout = dqp;
2295                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2296                         break;
2297                 }
2298
2299                 ASSERT(dqp->q_hash);
2300                 ASSERT(dqp->MPL_PREVP);
2301
2302                 /*
2303                  * Try to grab the flush lock. If this dquot is in the process of
2304                  * getting flushed to disk, we don't want to reclaim it.
2305                  */
2306                 if (! xfs_qm_dqflock_nowait(dqp)) {
2307                         xfs_dqunlock(dqp);
2308                         continue;
2309                 }
2310
2311                 /*
2312                  * We have the flush lock so we know that this is not in the
2313                  * process of being flushed. So, if this is dirty, flush it
2314                  * DELWRI so that we don't get a freelist infested with
2315                  * dirty dquots.
2316                  */
2317                 if (XFS_DQ_IS_DIRTY(dqp)) {
2318                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2319                         /*
2320                          * We flush it delayed write, so don't bother
2321                          * releasing the freelist lock.
2322                          */
2323                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2324                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2325                         continue;
2326                 }
2327
2328                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2329                         xfs_dqfunlock(dqp);
2330                         xfs_dqunlock(dqp);
2331                         continue;
2332                 }
2333
2334                 if (! xfs_qm_dqhashlock_nowait(dqp))
2335                         goto mplistunlock;
2336
2337                 ASSERT(dqp->q_nrefs == 0);
2338                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2339                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2340                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2341                 XQM_FREELIST_REMOVE(dqp);
2342                 dqpout = dqp;
2343                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2344  mplistunlock:
2345                 xfs_qm_mplist_unlock(dqp->q_mount);
2346                 xfs_dqfunlock(dqp);
2347                 xfs_dqunlock(dqp);
2348                 if (dqpout)
2349                         break;
2350         }
2351
2352         xfs_qm_freelist_unlock(xfs_Gqm);
2353         return (dqpout);
2354 }
2355
2356
2357 /*------------------------------------------------------------------*/
2358
2359 /*
2360  * Return a new incore dquot. Depending on the number of
2361  * dquots in the system, we either allocate a new one on the kernel heap,
2362  * or reclaim a free one.
2363  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2364  * to reclaim an existing one from the freelist.
2365  */
2366 boolean_t
2367 xfs_qm_dqalloc_incore(
2368         xfs_dquot_t **O_dqpp)
2369 {
2370         xfs_dquot_t     *dqp;
2371
2372         /*
2373          * Check against high water mark to see if we want to pop
2374          * a nincompoop dquot off the freelist.
2375          */
2376         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2377                 /*
2378                  * Try to recycle a dquot from the freelist.
2379                  */
2380                 if ((dqp = xfs_qm_dqreclaim_one())) {
2381                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2382                         /*
2383                          * Just zero the core here. The rest will get
2384                          * reinitialized by caller. XXX we shouldn't even
2385                          * do this zero ...
2386                          */
2387                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2388                         *O_dqpp = dqp;
2389                         return (B_FALSE);
2390                 }
2391                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2392         }
2393
2394         /*
2395          * Allocate a brand new dquot on the kernel heap and return it
2396          * to the caller to initialize.
2397          */
2398         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2399         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2400         atomic_inc(&xfs_Gqm->qm_totaldquots);
2401
2402         return (B_TRUE);
2403 }
2404
2405
2406 /*
2407  * Start a transaction and write the incore superblock changes to
2408  * disk. flags parameter indicates which fields have changed.
2409  */
2410 int
2411 xfs_qm_write_sb_changes(
2412         xfs_mount_t     *mp,
2413         __int64_t       flags)
2414 {
2415         xfs_trans_t     *tp;
2416         int             error;
2417
2418 #ifdef QUOTADEBUG
2419         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2420 #endif
2421         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2422         if ((error = xfs_trans_reserve(tp, 0,
2423                                       mp->m_sb.sb_sectsize + 128, 0,
2424                                       0,
2425                                       XFS_DEFAULT_LOG_COUNT))) {
2426                 xfs_trans_cancel(tp, 0);
2427                 return (error);
2428         }
2429
2430         xfs_mod_sb(tp, flags);
2431         (void) xfs_trans_commit(tp, 0, NULL);
2432
2433         return (0);
2434 }
2435
2436
2437 /* --------------- utility functions for vnodeops ---------------- */
2438
2439
2440 /*
2441  * Given an inode, a uid and gid (from cred_t) make sure that we have
2442  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2443  * quotas by creating this file.
2444  * This also attaches dquot(s) to the given inode after locking it,
2445  * and returns the dquots corresponding to the uid and/or gid.
2446  *
2447  * in   : inode (unlocked)
2448  * out  : udquot, gdquot with references taken and unlocked
2449  */
2450 int
2451 xfs_qm_vop_dqalloc(
2452         xfs_mount_t     *mp,
2453         xfs_inode_t     *ip,
2454         uid_t           uid,
2455         gid_t           gid,
2456         prid_t          prid,
2457         uint            flags,
2458         xfs_dquot_t     **O_udqpp,
2459         xfs_dquot_t     **O_gdqpp)
2460 {
2461         int             error;
2462         xfs_dquot_t     *uq, *gq;
2463         uint            lockflags;
2464
2465         if (!XFS_IS_QUOTA_ON(mp))
2466                 return 0;
2467
2468         lockflags = XFS_ILOCK_EXCL;
2469         xfs_ilock(ip, lockflags);
2470
2471         if ((flags & XFS_QMOPT_INHERIT) &&
2472             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2473                 gid = ip->i_d.di_gid;
2474
2475         /*
2476          * Attach the dquot(s) to this inode, doing a dquot allocation
2477          * if necessary. The dquot(s) will not be locked.
2478          */
2479         if (XFS_NOT_DQATTACHED(mp, ip)) {
2480                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2481                                             XFS_QMOPT_ILOCKED))) {
2482                         xfs_iunlock(ip, lockflags);
2483                         return (error);
2484                 }
2485         }
2486
2487         uq = gq = NULL;
2488         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2489                 if (ip->i_d.di_uid != uid) {
2490                         /*
2491                          * What we need is the dquot that has this uid, and
2492                          * if we send the inode to dqget, the uid of the inode
2493                          * takes priority over what's sent in the uid argument.
2494                          * We must unlock inode here before calling dqget if
2495                          * we're not sending the inode, because otherwise
2496                          * we'll deadlock by doing trans_reserve while
2497                          * holding ilock.
2498                          */
2499                         xfs_iunlock(ip, lockflags);
2500                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2501                                                  XFS_DQ_USER,
2502                                                  XFS_QMOPT_DQALLOC |
2503                                                  XFS_QMOPT_DOWARN,
2504                                                  &uq))) {
2505                                 ASSERT(error != ENOENT);
2506                                 return (error);
2507                         }
2508                         /*
2509                          * Get the ilock in the right order.
2510                          */
2511                         xfs_dqunlock(uq);
2512                         lockflags = XFS_ILOCK_SHARED;
2513                         xfs_ilock(ip, lockflags);
2514                 } else {
2515                         /*
2516                          * Take an extra reference, because we'll return
2517                          * this to caller
2518                          */
2519                         ASSERT(ip->i_udquot);
2520                         uq = ip->i_udquot;
2521                         xfs_dqlock(uq);
2522                         XFS_DQHOLD(uq);
2523                         xfs_dqunlock(uq);
2524                 }
2525         }
2526         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2527                 if (ip->i_d.di_gid != gid) {
2528                         xfs_iunlock(ip, lockflags);
2529                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2530                                                  XFS_DQ_GROUP,
2531                                                  XFS_QMOPT_DQALLOC |
2532                                                  XFS_QMOPT_DOWARN,
2533                                                  &gq))) {
2534                                 if (uq)
2535                                         xfs_qm_dqrele(uq);
2536                                 ASSERT(error != ENOENT);
2537                                 return (error);
2538                         }
2539                         xfs_dqunlock(gq);
2540                         lockflags = XFS_ILOCK_SHARED;
2541                         xfs_ilock(ip, lockflags);
2542                 } else {
2543                         ASSERT(ip->i_gdquot);
2544                         gq = ip->i_gdquot;
2545                         xfs_dqlock(gq);
2546                         XFS_DQHOLD(gq);
2547                         xfs_dqunlock(gq);
2548                 }
2549         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2550                 if (ip->i_d.di_projid != prid) {
2551                         xfs_iunlock(ip, lockflags);
2552                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2553                                                  XFS_DQ_PROJ,
2554                                                  XFS_QMOPT_DQALLOC |
2555                                                  XFS_QMOPT_DOWARN,
2556                                                  &gq))) {
2557                                 if (uq)
2558                                         xfs_qm_dqrele(uq);
2559                                 ASSERT(error != ENOENT);
2560                                 return (error);
2561                         }
2562                         xfs_dqunlock(gq);
2563                         lockflags = XFS_ILOCK_SHARED;
2564                         xfs_ilock(ip, lockflags);
2565                 } else {
2566                         ASSERT(ip->i_gdquot);
2567                         gq = ip->i_gdquot;
2568                         xfs_dqlock(gq);
2569                         XFS_DQHOLD(gq);
2570                         xfs_dqunlock(gq);
2571                 }
2572         }
2573         if (uq)
2574                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2575
2576         xfs_iunlock(ip, lockflags);
2577         if (O_udqpp)
2578                 *O_udqpp = uq;
2579         else if (uq)
2580                 xfs_qm_dqrele(uq);
2581         if (O_gdqpp)
2582                 *O_gdqpp = gq;
2583         else if (gq)
2584                 xfs_qm_dqrele(gq);
2585         return (0);
2586 }
2587
2588 /*
2589  * Actually transfer ownership, and do dquot modifications.
2590  * These were already reserved.
2591  */
2592 xfs_dquot_t *
2593 xfs_qm_vop_chown(
2594         xfs_trans_t     *tp,
2595         xfs_inode_t     *ip,
2596         xfs_dquot_t     **IO_olddq,
2597         xfs_dquot_t     *newdq)
2598 {
2599         xfs_dquot_t     *prevdq;
2600         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2601                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2602
2603         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2604         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2605
2606         /* old dquot */
2607         prevdq = *IO_olddq;
2608         ASSERT(prevdq);
2609         ASSERT(prevdq != newdq);
2610
2611         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2612         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2613
2614         /* the sparkling new dquot */
2615         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2616         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2617
2618         /*
2619          * Take an extra reference, because the inode
2620          * is going to keep this dquot pointer even
2621          * after the trans_commit.
2622          */
2623         xfs_dqlock(newdq);
2624         XFS_DQHOLD(newdq);
2625         xfs_dqunlock(newdq);
2626         *IO_olddq = newdq;
2627
2628         return (prevdq);
2629 }
2630
2631 /*
2632  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2633  */
2634 int
2635 xfs_qm_vop_chown_reserve(
2636         xfs_trans_t     *tp,
2637         xfs_inode_t     *ip,
2638         xfs_dquot_t     *udqp,
2639         xfs_dquot_t     *gdqp,
2640         uint            flags)
2641 {
2642         int             error;
2643         xfs_mount_t     *mp;
2644         uint            delblks, blkflags;
2645         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2646
2647         ASSERT(XFS_ISLOCKED_INODE(ip));
2648         mp = ip->i_mount;
2649         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2650
2651         delblks = ip->i_delayed_blks;
2652         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2653         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2654                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2655
2656         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2657             ip->i_d.di_uid != (uid_t)INT_GET(udqp->q_core.d_id, ARCH_CONVERT)) {
2658                 delblksudq = udqp;
2659                 /*
2660                  * If there are delayed allocation blocks, then we have to
2661                  * unreserve those from the old dquot, and add them to the
2662                  * new dquot.
2663                  */
2664                 if (delblks) {
2665                         ASSERT(ip->i_udquot);
2666                         unresudq = ip->i_udquot;
2667                 }
2668         }
2669         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2670                 if ((XFS_IS_GQUOTA_ON(ip->i_mount) && ip->i_d.di_gid !=
2671                                 INT_GET(gdqp->q_core.d_id, ARCH_CONVERT)) ||
2672                     (XFS_IS_PQUOTA_ON(ip->i_mount) && ip->i_d.di_projid !=
2673                                 INT_GET(gdqp->q_core.d_id, ARCH_CONVERT))) {
2674                         delblksgdq = gdqp;
2675                         if (delblks) {
2676                                 ASSERT(ip->i_gdquot);
2677                                 unresgdq = ip->i_gdquot;
2678                         }
2679                 }
2680         }
2681
2682         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2683                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2684                                 flags | blkflags)))
2685                 return (error);
2686
2687         /*
2688          * Do the delayed blks reservations/unreservations now. Since, these
2689          * are done without the help of a transaction, if a reservation fails
2690          * its previous reservations won't be automatically undone by trans
2691          * code. So, we have to do it manually here.
2692          */
2693         if (delblks) {
2694                 /*
2695                  * Do the reservations first. Unreservation can't fail.
2696                  */
2697                 ASSERT(delblksudq || delblksgdq);
2698                 ASSERT(unresudq || unresgdq);
2699                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2700                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2701                                 flags | blkflags)))
2702                         return (error);
2703                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2704                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2705                                 blkflags);
2706         }
2707
2708         return (0);
2709 }
2710
2711 int
2712 xfs_qm_vop_rename_dqattach(
2713         xfs_inode_t     **i_tab)
2714 {
2715         xfs_inode_t     *ip;
2716         int             i;
2717         int             error;
2718
2719         ip = i_tab[0];
2720
2721         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2722                 return (0);
2723
2724         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2725                 error = xfs_qm_dqattach(ip, 0);
2726                 if (error)
2727                         return (error);
2728         }
2729         for (i = 1; (i < 4 && i_tab[i]); i++) {
2730                 /*
2731                  * Watch out for duplicate entries in the table.
2732                  */
2733                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2734                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2735                                 error = xfs_qm_dqattach(ip, 0);
2736                                 if (error)
2737                                         return (error);
2738                         }
2739                 }
2740         }
2741         return (0);
2742 }
2743
2744 void
2745 xfs_qm_vop_dqattach_and_dqmod_newinode(
2746         xfs_trans_t     *tp,
2747         xfs_inode_t     *ip,
2748         xfs_dquot_t     *udqp,
2749         xfs_dquot_t     *gdqp)
2750 {
2751         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2752                 return;
2753
2754         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2755         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2756
2757         if (udqp) {
2758                 xfs_dqlock(udqp);
2759                 XFS_DQHOLD(udqp);
2760                 xfs_dqunlock(udqp);
2761                 ASSERT(ip->i_udquot == NULL);
2762                 ip->i_udquot = udqp;
2763                 ASSERT(ip->i_d.di_uid == INT_GET(udqp->q_core.d_id, ARCH_CONVERT));
2764                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2765         }
2766         if (gdqp) {
2767                 xfs_dqlock(gdqp);
2768                 XFS_DQHOLD(gdqp);
2769                 xfs_dqunlock(gdqp);
2770                 ASSERT(ip->i_gdquot == NULL);
2771                 ip->i_gdquot = gdqp;
2772                 ASSERT(ip->i_d.di_gid == INT_GET(gdqp->q_core.d_id, ARCH_CONVERT));
2773                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2774         }
2775 }
2776
2777 /* ------------- list stuff -----------------*/
2778 STATIC void
2779 xfs_qm_freelist_init(xfs_frlist_t *ql)
2780 {
2781         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2782         mutex_init(&ql->qh_lock, MUTEX_DEFAULT, "dqf");
2783         ql->qh_version = 0;
2784         ql->qh_nelems = 0;
2785 }
2786
2787 STATIC void
2788 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2789 {
2790         xfs_dquot_t     *dqp, *nextdqp;
2791
2792         mutex_lock(&ql->qh_lock, PINOD);
2793         for (dqp = ql->qh_next;
2794              dqp != (xfs_dquot_t *)ql; ) {
2795                 xfs_dqlock(dqp);
2796                 nextdqp = dqp->dq_flnext;
2797 #ifdef QUOTADEBUG
2798                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2799 #endif
2800                 XQM_FREELIST_REMOVE(dqp);
2801                 xfs_dqunlock(dqp);
2802                 xfs_qm_dqdestroy(dqp);
2803                 dqp = nextdqp;
2804         }
2805         /*
2806          * Don't bother about unlocking.
2807          */
2808         mutex_destroy(&ql->qh_lock);
2809
2810         ASSERT(ql->qh_nelems == 0);
2811 }
2812
2813 STATIC void
2814 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2815 {
2816         dq->dq_flnext = ql->qh_next;
2817         dq->dq_flprev = (xfs_dquot_t *)ql;
2818         ql->qh_next = dq;
2819         dq->dq_flnext->dq_flprev = dq;
2820         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2821         xfs_Gqm->qm_dqfreelist.qh_version++;
2822 }
2823
2824 void
2825 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2826 {
2827         xfs_dquot_t *next = dq->dq_flnext;
2828         xfs_dquot_t *prev = dq->dq_flprev;
2829
2830         next->dq_flprev = prev;
2831         prev->dq_flnext = next;
2832         dq->dq_flnext = dq->dq_flprev = dq;
2833         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2834         xfs_Gqm->qm_dqfreelist.qh_version++;
2835 }
2836
2837 void
2838 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2839 {
2840         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2841 }
2842
2843 STATIC int
2844 xfs_qm_dqhashlock_nowait(
2845         xfs_dquot_t *dqp)
2846 {
2847         int locked;
2848
2849         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2850         return (locked);
2851 }
2852
2853 int
2854 xfs_qm_freelist_lock_nowait(
2855         xfs_qm_t *xqm)
2856 {
2857         int locked;
2858
2859         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2860         return (locked);
2861 }
2862
2863 STATIC int
2864 xfs_qm_mplist_nowait(
2865         xfs_mount_t     *mp)
2866 {
2867         int locked;
2868
2869         ASSERT(mp->m_quotainfo);
2870         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2871         return (locked);
2872 }