Pull bugzilla-7880 into release branch
[linux-2.6] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_bit.h"
21 #include "xfs_log.h"
22 #include "xfs_inum.h"
23 #include "xfs_clnt.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_alloc.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_quota.h"
31 #include "xfs_mount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_btree.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_itable.h"
42 #include "xfs_rtalloc.h"
43 #include "xfs_error.h"
44 #include "xfs_bmap.h"
45 #include "xfs_rw.h"
46 #include "xfs_acl.h"
47 #include "xfs_attr.h"
48 #include "xfs_buf_item.h"
49 #include "xfs_trans_space.h"
50 #include "xfs_utils.h"
51 #include "xfs_qm.h"
52
53 /*
54  * The global quota manager. There is only one of these for the entire
55  * system, _not_ one per file system. XQM keeps track of the overall
56  * quota functionality, including maintaining the freelist and hash
57  * tables of dquots.
58  */
59 mutex_t         xfs_Gqm_lock;
60 struct xfs_qm   *xfs_Gqm;
61 uint            ndquot;
62
63 kmem_zone_t     *qm_dqzone;
64 kmem_zone_t     *qm_dqtrxzone;
65
66 static cred_t   xfs_zerocr;
67
68 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
69 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
70
71 STATIC void     xfs_qm_freelist_init(xfs_frlist_t *);
72 STATIC void     xfs_qm_freelist_destroy(xfs_frlist_t *);
73 STATIC int      xfs_qm_mplist_nowait(xfs_mount_t *);
74 STATIC int      xfs_qm_dqhashlock_nowait(xfs_dquot_t *);
75
76 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
77 STATIC int      xfs_qm_init_quotainfo(xfs_mount_t *);
78 STATIC int      xfs_qm_shake(int, gfp_t);
79
80 static struct shrinker xfs_qm_shaker = {
81         .shrink = xfs_qm_shake,
82         .seeks = DEFAULT_SEEKS,
83 };
84
85 #ifdef DEBUG
86 extern mutex_t  qcheck_lock;
87 #endif
88
89 #ifdef QUOTADEBUG
90 #define XQM_LIST_PRINT(l, NXT, title) \
91 { \
92         xfs_dquot_t     *dqp; int i = 0; \
93         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
94         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
95                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
96                                   "bcnt = %d, icnt = %d, refs = %d", \
97                         ++i, (int) be32_to_cpu(dqp->q_core.d_id), \
98                         DQFLAGTO_TYPESTR(dqp),       \
99                         (int) be64_to_cpu(dqp->q_core.d_bcount), \
100                         (int) be64_to_cpu(dqp->q_core.d_icount), \
101                         (int) dqp->q_nrefs);  } \
102 }
103 #else
104 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
105 #endif
106
107 /*
108  * Initialize the XQM structure.
109  * Note that there is not one quota manager per file system.
110  */
111 STATIC struct xfs_qm *
112 xfs_Gqm_init(void)
113 {
114         xfs_dqhash_t    *udqhash, *gdqhash;
115         xfs_qm_t        *xqm;
116         size_t          hsize;
117         uint            i;
118
119         /*
120          * Initialize the dquot hash tables.
121          */
122         udqhash = kmem_zalloc_greedy(&hsize,
123                                      XFS_QM_HASHSIZE_LOW, XFS_QM_HASHSIZE_HIGH,
124                                      KM_SLEEP | KM_MAYFAIL | KM_LARGE);
125         gdqhash = kmem_zalloc(hsize, KM_SLEEP | KM_LARGE);
126         hsize /= sizeof(xfs_dqhash_t);
127         ndquot = hsize << 8;
128
129         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
130         xqm->qm_dqhashmask = hsize - 1;
131         xqm->qm_usr_dqhtable = udqhash;
132         xqm->qm_grp_dqhtable = gdqhash;
133         ASSERT(xqm->qm_usr_dqhtable != NULL);
134         ASSERT(xqm->qm_grp_dqhtable != NULL);
135
136         for (i = 0; i < hsize; i++) {
137                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
138                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
139         }
140
141         /*
142          * Freelist of all dquots of all file systems
143          */
144         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
145
146         /*
147          * dquot zone. we register our own low-memory callback.
148          */
149         if (!qm_dqzone) {
150                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
151                                                 "xfs_dquots");
152                 qm_dqzone = xqm->qm_dqzone;
153         } else
154                 xqm->qm_dqzone = qm_dqzone;
155
156         register_shrinker(&xfs_qm_shaker);
157
158         /*
159          * The t_dqinfo portion of transactions.
160          */
161         if (!qm_dqtrxzone) {
162                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
163                                                    "xfs_dqtrx");
164                 qm_dqtrxzone = xqm->qm_dqtrxzone;
165         } else
166                 xqm->qm_dqtrxzone = qm_dqtrxzone;
167
168         atomic_set(&xqm->qm_totaldquots, 0);
169         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
170         xqm->qm_nrefs = 0;
171 #ifdef DEBUG
172         mutex_init(&qcheck_lock);
173 #endif
174         return xqm;
175 }
176
177 /*
178  * Destroy the global quota manager when its reference count goes to zero.
179  */
180 STATIC void
181 xfs_qm_destroy(
182         struct xfs_qm   *xqm)
183 {
184         int             hsize, i;
185
186         ASSERT(xqm != NULL);
187         ASSERT(xqm->qm_nrefs == 0);
188         unregister_shrinker(&xfs_qm_shaker);
189         hsize = xqm->qm_dqhashmask + 1;
190         for (i = 0; i < hsize; i++) {
191                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
192                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
193         }
194         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
195         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
196         xqm->qm_usr_dqhtable = NULL;
197         xqm->qm_grp_dqhtable = NULL;
198         xqm->qm_dqhashmask = 0;
199         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
200 #ifdef DEBUG
201         mutex_destroy(&qcheck_lock);
202 #endif
203         kmem_free(xqm, sizeof(xfs_qm_t));
204 }
205
206 /*
207  * Called at mount time to let XQM know that another file system is
208  * starting quotas. This isn't crucial information as the individual mount
209  * structures are pretty independent, but it helps the XQM keep a
210  * global view of what's going on.
211  */
212 /* ARGSUSED */
213 STATIC int
214 xfs_qm_hold_quotafs_ref(
215         struct xfs_mount *mp)
216 {
217         /*
218          * Need to lock the xfs_Gqm structure for things like this. For example,
219          * the structure could disappear between the entry to this routine and
220          * a HOLD operation if not locked.
221          */
222         XFS_QM_LOCK(xfs_Gqm);
223
224         if (xfs_Gqm == NULL)
225                 xfs_Gqm = xfs_Gqm_init();
226         /*
227          * We can keep a list of all filesystems with quotas mounted for
228          * debugging and statistical purposes, but ...
229          * Just take a reference and get out.
230          */
231         XFS_QM_HOLD(xfs_Gqm);
232         XFS_QM_UNLOCK(xfs_Gqm);
233
234         return 0;
235 }
236
237
238 /*
239  * Release the reference that a filesystem took at mount time,
240  * so that we know when we need to destroy the entire quota manager.
241  */
242 /* ARGSUSED */
243 STATIC void
244 xfs_qm_rele_quotafs_ref(
245         struct xfs_mount *mp)
246 {
247         xfs_dquot_t     *dqp, *nextdqp;
248
249         ASSERT(xfs_Gqm);
250         ASSERT(xfs_Gqm->qm_nrefs > 0);
251
252         /*
253          * Go thru the freelist and destroy all inactive dquots.
254          */
255         xfs_qm_freelist_lock(xfs_Gqm);
256
257         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
258              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
259                 xfs_dqlock(dqp);
260                 nextdqp = dqp->dq_flnext;
261                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
262                         ASSERT(dqp->q_mount == NULL);
263                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
264                         ASSERT(dqp->HL_PREVP == NULL);
265                         ASSERT(dqp->MPL_PREVP == NULL);
266                         XQM_FREELIST_REMOVE(dqp);
267                         xfs_dqunlock(dqp);
268                         xfs_qm_dqdestroy(dqp);
269                 } else {
270                         xfs_dqunlock(dqp);
271                 }
272                 dqp = nextdqp;
273         }
274         xfs_qm_freelist_unlock(xfs_Gqm);
275
276         /*
277          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
278          * be restarted.
279          */
280         XFS_QM_LOCK(xfs_Gqm);
281         XFS_QM_RELE(xfs_Gqm);
282         if (xfs_Gqm->qm_nrefs == 0) {
283                 xfs_qm_destroy(xfs_Gqm);
284                 xfs_Gqm = NULL;
285         }
286         XFS_QM_UNLOCK(xfs_Gqm);
287 }
288
289 /*
290  * This is called at mount time from xfs_mountfs to initialize the quotainfo
291  * structure and start the global quota manager (xfs_Gqm) if it hasn't done
292  * so already.  Note that the superblock has not been read in yet.
293  */
294 void
295 xfs_qm_mount_quotainit(
296         xfs_mount_t     *mp,
297         uint            flags)
298 {
299         /*
300          * User, projects or group quotas has to be on.
301          */
302         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_PQUOTA | XFSMNT_GQUOTA));
303
304         /*
305          * Initialize the flags in the mount structure. From this point
306          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
307          * Note that we enforce nothing if accounting is off.
308          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
309          * It isn't necessary to take the quotaoff lock to do this; this is
310          * called from mount.
311          */
312         if (flags & XFSMNT_UQUOTA) {
313                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
314                 if (flags & XFSMNT_UQUOTAENF)
315                         mp->m_qflags |= XFS_UQUOTA_ENFD;
316         }
317         if (flags & XFSMNT_GQUOTA) {
318                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
319                 if (flags & XFSMNT_GQUOTAENF)
320                         mp->m_qflags |= XFS_OQUOTA_ENFD;
321         } else if (flags & XFSMNT_PQUOTA) {
322                 mp->m_qflags |= (XFS_PQUOTA_ACCT | XFS_PQUOTA_ACTIVE);
323                 if (flags & XFSMNT_PQUOTAENF)
324                         mp->m_qflags |= XFS_OQUOTA_ENFD;
325         }
326 }
327
328 /*
329  * Just destroy the quotainfo structure.
330  */
331 void
332 xfs_qm_unmount_quotadestroy(
333         xfs_mount_t     *mp)
334 {
335         if (mp->m_quotainfo)
336                 xfs_qm_destroy_quotainfo(mp);
337 }
338
339
340 /*
341  * This is called from xfs_mountfs to start quotas and initialize all
342  * necessary data structures like quotainfo.  This is also responsible for
343  * running a quotacheck as necessary.  We are guaranteed that the superblock
344  * is consistently read in at this point.
345  */
346 int
347 xfs_qm_mount_quotas(
348         xfs_mount_t     *mp,
349         int             mfsi_flags)
350 {
351         unsigned long   s;
352         int             error = 0;
353         uint            sbf;
354
355
356         /*
357          * If quotas on realtime volumes is not supported, we disable
358          * quotas immediately.
359          */
360         if (mp->m_sb.sb_rextents) {
361                 cmn_err(CE_NOTE,
362                         "Cannot turn on quotas for realtime filesystem %s",
363                         mp->m_fsname);
364                 mp->m_qflags = 0;
365                 goto write_changes;
366         }
367
368         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
369
370         /*
371          * Allocate the quotainfo structure inside the mount struct, and
372          * create quotainode(s), and change/rev superblock if necessary.
373          */
374         if ((error = xfs_qm_init_quotainfo(mp))) {
375                 /*
376                  * We must turn off quotas.
377                  */
378                 ASSERT(mp->m_quotainfo == NULL);
379                 mp->m_qflags = 0;
380                 goto write_changes;
381         }
382         /*
383          * If any of the quotas are not consistent, do a quotacheck.
384          */
385         if (XFS_QM_NEED_QUOTACHECK(mp) &&
386                 !(mfsi_flags & XFS_MFSI_NO_QUOTACHECK)) {
387                 if ((error = xfs_qm_quotacheck(mp))) {
388                         /* Quotacheck has failed and quotas have
389                          * been disabled.
390                          */
391                         return XFS_ERROR(error);
392                 }
393         }
394         /* 
395          * If one type of quotas is off, then it will lose its
396          * quotachecked status, since we won't be doing accounting for
397          * that type anymore.
398          */
399         if (!XFS_IS_UQUOTA_ON(mp)) {
400                 mp->m_qflags &= ~XFS_UQUOTA_CHKD;
401         }
402         if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) {
403                 mp->m_qflags &= ~XFS_OQUOTA_CHKD;
404         }
405
406  write_changes:
407         /*
408          * We actually don't have to acquire the SB_LOCK at all.
409          * This can only be called from mount, and that's single threaded. XXX
410          */
411         s = XFS_SB_LOCK(mp);
412         sbf = mp->m_sb.sb_qflags;
413         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
414         XFS_SB_UNLOCK(mp, s);
415
416         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
417                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
418                         /*
419                          * We could only have been turning quotas off.
420                          * We aren't in very good shape actually because
421                          * the incore structures are convinced that quotas are
422                          * off, but the on disk superblock doesn't know that !
423                          */
424                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
425                         xfs_fs_cmn_err(CE_ALERT, mp,
426                                 "XFS mount_quotas: Superblock update failed!");
427                 }
428         }
429
430         if (error) {
431                 xfs_fs_cmn_err(CE_WARN, mp,
432                         "Failed to initialize disk quotas.");
433         }
434         return XFS_ERROR(error);
435 }
436
437 /*
438  * Called from the vfsops layer.
439  */
440 int
441 xfs_qm_unmount_quotas(
442         xfs_mount_t     *mp)
443 {
444         xfs_inode_t     *uqp, *gqp;
445         int             error = 0;
446
447         /*
448          * Release the dquots that root inode, et al might be holding,
449          * before we flush quotas and blow away the quotainfo structure.
450          */
451         ASSERT(mp->m_rootip);
452         xfs_qm_dqdetach(mp->m_rootip);
453         if (mp->m_rbmip)
454                 xfs_qm_dqdetach(mp->m_rbmip);
455         if (mp->m_rsumip)
456                 xfs_qm_dqdetach(mp->m_rsumip);
457
458         /*
459          * Flush out the quota inodes.
460          */
461         uqp = gqp = NULL;
462         if (mp->m_quotainfo) {
463                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
464                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
465                         xfs_iflock(uqp);
466                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
467                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
468                         if (unlikely(error == EFSCORRUPTED)) {
469                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
470                                                  XFS_ERRLEVEL_LOW, mp);
471                                 goto out;
472                         }
473                 }
474                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
475                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
476                         xfs_iflock(gqp);
477                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
478                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
479                         if (unlikely(error == EFSCORRUPTED)) {
480                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
481                                                  XFS_ERRLEVEL_LOW, mp);
482                                 goto out;
483                         }
484                 }
485         }
486         if (uqp) {
487                  XFS_PURGE_INODE(uqp);
488                  mp->m_quotainfo->qi_uquotaip = NULL;
489         }
490         if (gqp) {
491                 XFS_PURGE_INODE(gqp);
492                 mp->m_quotainfo->qi_gquotaip = NULL;
493         }
494 out:
495         return XFS_ERROR(error);
496 }
497
498 /*
499  * Flush all dquots of the given file system to disk. The dquots are
500  * _not_ purged from memory here, just their data written to disk.
501  */
502 STATIC int
503 xfs_qm_dqflush_all(
504         xfs_mount_t     *mp,
505         int             flags)
506 {
507         int             recl;
508         xfs_dquot_t     *dqp;
509         int             niters;
510         int             error;
511
512         if (mp->m_quotainfo == NULL)
513                 return 0;
514         niters = 0;
515 again:
516         xfs_qm_mplist_lock(mp);
517         FOREACH_DQUOT_IN_MP(dqp, mp) {
518                 xfs_dqlock(dqp);
519                 if (! XFS_DQ_IS_DIRTY(dqp)) {
520                         xfs_dqunlock(dqp);
521                         continue;
522                 }
523                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
524                 /* XXX a sentinel would be better */
525                 recl = XFS_QI_MPLRECLAIMS(mp);
526                 if (! xfs_qm_dqflock_nowait(dqp)) {
527                         /*
528                          * If we can't grab the flush lock then check
529                          * to see if the dquot has been flushed delayed
530                          * write.  If so, grab its buffer and send it
531                          * out immediately.  We'll be able to acquire
532                          * the flush lock when the I/O completes.
533                          */
534                         xfs_qm_dqflock_pushbuf_wait(dqp);
535                 }
536                 /*
537                  * Let go of the mplist lock. We don't want to hold it
538                  * across a disk write.
539                  */
540                 xfs_qm_mplist_unlock(mp);
541                 error = xfs_qm_dqflush(dqp, flags);
542                 xfs_dqunlock(dqp);
543                 if (error)
544                         return error;
545
546                 xfs_qm_mplist_lock(mp);
547                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
548                         xfs_qm_mplist_unlock(mp);
549                         /* XXX restart limit */
550                         goto again;
551                 }
552         }
553
554         xfs_qm_mplist_unlock(mp);
555         /* return ! busy */
556         return 0;
557 }
558 /*
559  * Release the group dquot pointers the user dquots may be
560  * carrying around as a hint. mplist is locked on entry and exit.
561  */
562 STATIC void
563 xfs_qm_detach_gdquots(
564         xfs_mount_t     *mp)
565 {
566         xfs_dquot_t     *dqp, *gdqp;
567         int             nrecl;
568
569  again:
570         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
571         dqp = XFS_QI_MPLNEXT(mp);
572         while (dqp) {
573                 xfs_dqlock(dqp);
574                 if ((gdqp = dqp->q_gdquot)) {
575                         xfs_dqlock(gdqp);
576                         dqp->q_gdquot = NULL;
577                 }
578                 xfs_dqunlock(dqp);
579
580                 if (gdqp) {
581                         /*
582                          * Can't hold the mplist lock across a dqput.
583                          * XXXmust convert to marker based iterations here.
584                          */
585                         nrecl = XFS_QI_MPLRECLAIMS(mp);
586                         xfs_qm_mplist_unlock(mp);
587                         xfs_qm_dqput(gdqp);
588
589                         xfs_qm_mplist_lock(mp);
590                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
591                                 goto again;
592                 }
593                 dqp = dqp->MPL_NEXT;
594         }
595 }
596
597 /*
598  * Go through all the incore dquots of this file system and take them
599  * off the mplist and hashlist, if the dquot type matches the dqtype
600  * parameter. This is used when turning off quota accounting for
601  * users and/or groups, as well as when the filesystem is unmounting.
602  */
603 STATIC int
604 xfs_qm_dqpurge_int(
605         xfs_mount_t     *mp,
606         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
607 {
608         xfs_dquot_t     *dqp;
609         uint            dqtype;
610         int             nrecl;
611         xfs_dquot_t     *nextdqp;
612         int             nmisses;
613
614         if (mp->m_quotainfo == NULL)
615                 return 0;
616
617         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
618         dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
619         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
620
621         xfs_qm_mplist_lock(mp);
622
623         /*
624          * In the first pass through all incore dquots of this filesystem,
625          * we release the group dquot pointers the user dquots may be
626          * carrying around as a hint. We need to do this irrespective of
627          * what's being turned off.
628          */
629         xfs_qm_detach_gdquots(mp);
630
631       again:
632         nmisses = 0;
633         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
634         /*
635          * Try to get rid of all of the unwanted dquots. The idea is to
636          * get them off mplist and hashlist, but leave them on freelist.
637          */
638         dqp = XFS_QI_MPLNEXT(mp);
639         while (dqp) {
640                 /*
641                  * It's OK to look at the type without taking dqlock here.
642                  * We're holding the mplist lock here, and that's needed for
643                  * a dqreclaim.
644                  */
645                 if ((dqp->dq_flags & dqtype) == 0) {
646                         dqp = dqp->MPL_NEXT;
647                         continue;
648                 }
649
650                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
651                         nrecl = XFS_QI_MPLRECLAIMS(mp);
652                         xfs_qm_mplist_unlock(mp);
653                         XFS_DQ_HASH_LOCK(dqp->q_hash);
654                         xfs_qm_mplist_lock(mp);
655
656                         /*
657                          * XXXTheoretically, we can get into a very long
658                          * ping pong game here.
659                          * No one can be adding dquots to the mplist at
660                          * this point, but somebody might be taking things off.
661                          */
662                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
663                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
664                                 goto again;
665                         }
666                 }
667
668                 /*
669                  * Take the dquot off the mplist and hashlist. It may remain on
670                  * freelist in INACTIVE state.
671                  */
672                 nextdqp = dqp->MPL_NEXT;
673                 nmisses += xfs_qm_dqpurge(dqp, flags);
674                 dqp = nextdqp;
675         }
676         xfs_qm_mplist_unlock(mp);
677         return nmisses;
678 }
679
680 int
681 xfs_qm_dqpurge_all(
682         xfs_mount_t     *mp,
683         uint            flags)
684 {
685         int             ndquots;
686
687         /*
688          * Purge the dquot cache.
689          * None of the dquots should really be busy at this point.
690          */
691         if (mp->m_quotainfo) {
692                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
693                         delay(ndquots * 10);
694                 }
695         }
696         return 0;
697 }
698
699 STATIC int
700 xfs_qm_dqattach_one(
701         xfs_inode_t     *ip,
702         xfs_dqid_t      id,
703         uint            type,
704         uint            doalloc,
705         uint            dolock,
706         xfs_dquot_t     *udqhint, /* hint */
707         xfs_dquot_t     **IO_idqpp)
708 {
709         xfs_dquot_t     *dqp;
710         int             error;
711
712         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
713         error = 0;
714         /*
715          * See if we already have it in the inode itself. IO_idqpp is
716          * &i_udquot or &i_gdquot. This made the code look weird, but
717          * made the logic a lot simpler.
718          */
719         if ((dqp = *IO_idqpp)) {
720                 if (dolock)
721                         xfs_dqlock(dqp);
722                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
723                 goto done;
724         }
725
726         /*
727          * udqhint is the i_udquot field in inode, and is non-NULL only
728          * when the type arg is group/project. Its purpose is to save a
729          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
730          * the user dquot.
731          */
732         ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
733         if (udqhint && !dolock)
734                 xfs_dqlock(udqhint);
735
736         /*
737          * No need to take dqlock to look at the id.
738          * The ID can't change until it gets reclaimed, and it won't
739          * be reclaimed as long as we have a ref from inode and we hold
740          * the ilock.
741          */
742         if (udqhint &&
743             (dqp = udqhint->q_gdquot) &&
744             (be32_to_cpu(dqp->q_core.d_id) == id)) {
745                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
746                 xfs_dqlock(dqp);
747                 XFS_DQHOLD(dqp);
748                 ASSERT(*IO_idqpp == NULL);
749                 *IO_idqpp = dqp;
750                 if (!dolock) {
751                         xfs_dqunlock(dqp);
752                         xfs_dqunlock(udqhint);
753                 }
754                 goto done;
755         }
756         /*
757          * We can't hold a dquot lock when we call the dqget code.
758          * We'll deadlock in no time, because of (not conforming to)
759          * lock ordering - the inodelock comes before any dquot lock,
760          * and we may drop and reacquire the ilock in xfs_qm_dqget().
761          */
762         if (udqhint)
763                 xfs_dqunlock(udqhint);
764         /*
765          * Find the dquot from somewhere. This bumps the
766          * reference count of dquot and returns it locked.
767          * This can return ENOENT if dquot didn't exist on
768          * disk and we didn't ask it to allocate;
769          * ESRCH if quotas got turned off suddenly.
770          */
771         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
772                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
773                 if (udqhint && dolock)
774                         xfs_dqlock(udqhint);
775                 goto done;
776         }
777
778         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
779         /*
780          * dqget may have dropped and re-acquired the ilock, but it guarantees
781          * that the dquot returned is the one that should go in the inode.
782          */
783         *IO_idqpp = dqp;
784         ASSERT(dqp);
785         ASSERT(XFS_DQ_IS_LOCKED(dqp));
786         if (! dolock) {
787                 xfs_dqunlock(dqp);
788                 goto done;
789         }
790         if (! udqhint)
791                 goto done;
792
793         ASSERT(udqhint);
794         ASSERT(dolock);
795         ASSERT(XFS_DQ_IS_LOCKED(dqp));
796         if (! xfs_qm_dqlock_nowait(udqhint)) {
797                 xfs_dqunlock(dqp);
798                 xfs_dqlock(udqhint);
799                 xfs_dqlock(dqp);
800         }
801       done:
802 #ifdef QUOTADEBUG
803         if (udqhint) {
804                 if (dolock)
805                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
806         }
807         if (! error) {
808                 if (dolock)
809                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
810         }
811 #endif
812         return error;
813 }
814
815
816 /*
817  * Given a udquot and gdquot, attach a ptr to the group dquot in the
818  * udquot as a hint for future lookups. The idea sounds simple, but the
819  * execution isn't, because the udquot might have a group dquot attached
820  * already and getting rid of that gets us into lock ordering constraints.
821  * The process is complicated more by the fact that the dquots may or may not
822  * be locked on entry.
823  */
824 STATIC void
825 xfs_qm_dqattach_grouphint(
826         xfs_dquot_t     *udq,
827         xfs_dquot_t     *gdq,
828         uint            locked)
829 {
830         xfs_dquot_t     *tmp;
831
832 #ifdef QUOTADEBUG
833         if (locked) {
834                 ASSERT(XFS_DQ_IS_LOCKED(udq));
835                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
836         }
837 #endif
838         if (! locked)
839                 xfs_dqlock(udq);
840
841         if ((tmp = udq->q_gdquot)) {
842                 if (tmp == gdq) {
843                         if (! locked)
844                                 xfs_dqunlock(udq);
845                         return;
846                 }
847
848                 udq->q_gdquot = NULL;
849                 /*
850                  * We can't keep any dqlocks when calling dqrele,
851                  * because the freelist lock comes before dqlocks.
852                  */
853                 xfs_dqunlock(udq);
854                 if (locked)
855                         xfs_dqunlock(gdq);
856                 /*
857                  * we took a hard reference once upon a time in dqget,
858                  * so give it back when the udquot no longer points at it
859                  * dqput() does the unlocking of the dquot.
860                  */
861                 xfs_qm_dqrele(tmp);
862
863                 xfs_dqlock(udq);
864                 xfs_dqlock(gdq);
865
866         } else {
867                 ASSERT(XFS_DQ_IS_LOCKED(udq));
868                 if (! locked) {
869                         xfs_dqlock(gdq);
870                 }
871         }
872
873         ASSERT(XFS_DQ_IS_LOCKED(udq));
874         ASSERT(XFS_DQ_IS_LOCKED(gdq));
875         /*
876          * Somebody could have attached a gdquot here,
877          * when we dropped the uqlock. If so, just do nothing.
878          */
879         if (udq->q_gdquot == NULL) {
880                 XFS_DQHOLD(gdq);
881                 udq->q_gdquot = gdq;
882         }
883         if (! locked) {
884                 xfs_dqunlock(gdq);
885                 xfs_dqunlock(udq);
886         }
887 }
888
889
890 /*
891  * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
892  * into account.
893  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
894  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
895  * much made this code a complete mess, but it has been pretty useful.
896  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
897  * Inode may get unlocked and relocked in here, and the caller must deal with
898  * the consequences.
899  */
900 int
901 xfs_qm_dqattach(
902         xfs_inode_t     *ip,
903         uint            flags)
904 {
905         xfs_mount_t     *mp = ip->i_mount;
906         uint            nquotas = 0;
907         int             error = 0;
908
909         if ((! XFS_IS_QUOTA_ON(mp)) ||
910             (! XFS_NOT_DQATTACHED(mp, ip)) ||
911             (ip->i_ino == mp->m_sb.sb_uquotino) ||
912             (ip->i_ino == mp->m_sb.sb_gquotino))
913                 return 0;
914
915         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
916                XFS_ISLOCKED_INODE_EXCL(ip));
917
918         if (! (flags & XFS_QMOPT_ILOCKED))
919                 xfs_ilock(ip, XFS_ILOCK_EXCL);
920
921         if (XFS_IS_UQUOTA_ON(mp)) {
922                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
923                                                 flags & XFS_QMOPT_DQALLOC,
924                                                 flags & XFS_QMOPT_DQLOCK,
925                                                 NULL, &ip->i_udquot);
926                 if (error)
927                         goto done;
928                 nquotas++;
929         }
930         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
931         if (XFS_IS_OQUOTA_ON(mp)) {
932                 error = XFS_IS_GQUOTA_ON(mp) ?
933                         xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
934                                                 flags & XFS_QMOPT_DQALLOC,
935                                                 flags & XFS_QMOPT_DQLOCK,
936                                                 ip->i_udquot, &ip->i_gdquot) :
937                         xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
938                                                 flags & XFS_QMOPT_DQALLOC,
939                                                 flags & XFS_QMOPT_DQLOCK,
940                                                 ip->i_udquot, &ip->i_gdquot);
941                 /*
942                  * Don't worry about the udquot that we may have
943                  * attached above. It'll get detached, if not already.
944                  */
945                 if (error)
946                         goto done;
947                 nquotas++;
948         }
949
950         /*
951          * Attach this group quota to the user quota as a hint.
952          * This WON'T, in general, result in a thrash.
953          */
954         if (nquotas == 2) {
955                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
956                 ASSERT(ip->i_udquot);
957                 ASSERT(ip->i_gdquot);
958
959                 /*
960                  * We may or may not have the i_udquot locked at this point,
961                  * but this check is OK since we don't depend on the i_gdquot to
962                  * be accurate 100% all the time. It is just a hint, and this
963                  * will succeed in general.
964                  */
965                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
966                         goto done;
967                 /*
968                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
969                  */
970                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
971                                          flags & XFS_QMOPT_DQLOCK);
972         }
973
974       done:
975
976 #ifdef QUOTADEBUG
977         if (! error) {
978                 if (ip->i_udquot) {
979                         if (flags & XFS_QMOPT_DQLOCK)
980                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
981                 }
982                 if (ip->i_gdquot) {
983                         if (flags & XFS_QMOPT_DQLOCK)
984                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
985                 }
986                 if (XFS_IS_UQUOTA_ON(mp))
987                         ASSERT(ip->i_udquot);
988                 if (XFS_IS_OQUOTA_ON(mp))
989                         ASSERT(ip->i_gdquot);
990         }
991 #endif
992
993         if (! (flags & XFS_QMOPT_ILOCKED))
994                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
995
996 #ifdef QUOTADEBUG
997         else
998                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
999 #endif
1000         return error;
1001 }
1002
1003 /*
1004  * Release dquots (and their references) if any.
1005  * The inode should be locked EXCL except when this's called by
1006  * xfs_ireclaim.
1007  */
1008 void
1009 xfs_qm_dqdetach(
1010         xfs_inode_t     *ip)
1011 {
1012         if (!(ip->i_udquot || ip->i_gdquot))
1013                 return;
1014
1015         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1016         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1017         if (ip->i_udquot) {
1018                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1019                 xfs_qm_dqrele(ip->i_udquot);
1020                 ip->i_udquot = NULL;
1021         }
1022         if (ip->i_gdquot) {
1023                 xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
1024                 xfs_qm_dqrele(ip->i_gdquot);
1025                 ip->i_gdquot = NULL;
1026         }
1027 }
1028
1029 /*
1030  * This is called by VFS_SYNC and flags arg determines the caller,
1031  * and its motives, as done in xfs_sync.
1032  *
1033  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1034  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1035  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1036  */
1037
1038 int
1039 xfs_qm_sync(
1040         xfs_mount_t     *mp,
1041         short           flags)
1042 {
1043         int             recl, restarts;
1044         xfs_dquot_t     *dqp;
1045         uint            flush_flags;
1046         boolean_t       nowait;
1047         int             error;
1048
1049         restarts = 0;
1050         /*
1051          * We won't block unless we are asked to.
1052          */
1053         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1054
1055   again:
1056         xfs_qm_mplist_lock(mp);
1057         /*
1058          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1059          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1060          * when we have the mplist lock, we know that dquots will be consistent
1061          * as long as we have it locked.
1062          */
1063         if (! XFS_IS_QUOTA_ON(mp)) {
1064                 xfs_qm_mplist_unlock(mp);
1065                 return 0;
1066         }
1067         FOREACH_DQUOT_IN_MP(dqp, mp) {
1068                 /*
1069                  * If this is vfs_sync calling, then skip the dquots that
1070                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1071                  * This is very similar to what xfs_sync does with inodes.
1072                  */
1073                 if (flags & SYNC_BDFLUSH) {
1074                         if (! XFS_DQ_IS_DIRTY(dqp))
1075                                 continue;
1076                 }
1077
1078                 if (nowait) {
1079                         /*
1080                          * Try to acquire the dquot lock. We are NOT out of
1081                          * lock order, but we just don't want to wait for this
1082                          * lock, unless somebody wanted us to.
1083                          */
1084                         if (! xfs_qm_dqlock_nowait(dqp))
1085                                 continue;
1086                 } else {
1087                         xfs_dqlock(dqp);
1088                 }
1089
1090                 /*
1091                  * Now, find out for sure if this dquot is dirty or not.
1092                  */
1093                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1094                         xfs_dqunlock(dqp);
1095                         continue;
1096                 }
1097
1098                 /* XXX a sentinel would be better */
1099                 recl = XFS_QI_MPLRECLAIMS(mp);
1100                 if (! xfs_qm_dqflock_nowait(dqp)) {
1101                         if (nowait) {
1102                                 xfs_dqunlock(dqp);
1103                                 continue;
1104                         }
1105                         /*
1106                          * If we can't grab the flush lock then if the caller
1107                          * really wanted us to give this our best shot, so
1108                          * see if we can give a push to the buffer before we wait
1109                          * on the flush lock. At this point, we know that
1110                          * even though the dquot is being flushed,
1111                          * it has (new) dirty data.
1112                          */
1113                         xfs_qm_dqflock_pushbuf_wait(dqp);
1114                 }
1115                 /*
1116                  * Let go of the mplist lock. We don't want to hold it
1117                  * across a disk write
1118                  */
1119                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1120                 xfs_qm_mplist_unlock(mp);
1121                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1122                 error = xfs_qm_dqflush(dqp, flush_flags);
1123                 xfs_dqunlock(dqp);
1124                 if (error && XFS_FORCED_SHUTDOWN(mp))
1125                         return 0;       /* Need to prevent umount failure */
1126                 else if (error)
1127                         return error;
1128
1129                 xfs_qm_mplist_lock(mp);
1130                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1131                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1132                                 break;
1133
1134                         xfs_qm_mplist_unlock(mp);
1135                         goto again;
1136                 }
1137         }
1138
1139         xfs_qm_mplist_unlock(mp);
1140         return 0;
1141 }
1142
1143
1144 /*
1145  * This initializes all the quota information that's kept in the
1146  * mount structure
1147  */
1148 STATIC int
1149 xfs_qm_init_quotainfo(
1150         xfs_mount_t     *mp)
1151 {
1152         xfs_quotainfo_t *qinf;
1153         int             error;
1154         xfs_dquot_t     *dqp;
1155
1156         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1157
1158         /*
1159          * Tell XQM that we exist as soon as possible.
1160          */
1161         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1162                 return error;
1163         }
1164
1165         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1166
1167         /*
1168          * See if quotainodes are setup, and if not, allocate them,
1169          * and change the superblock accordingly.
1170          */
1171         if ((error = xfs_qm_init_quotainos(mp))) {
1172                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1173                 mp->m_quotainfo = NULL;
1174                 return error;
1175         }
1176
1177         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1178         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1179         qinf->qi_dqreclaims = 0;
1180
1181         /* mutex used to serialize quotaoffs */
1182         mutex_init(&qinf->qi_quotaofflock);
1183
1184         /* Precalc some constants */
1185         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1186         ASSERT(qinf->qi_dqchunklen);
1187         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1188         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1189
1190         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1191
1192         /*
1193          * We try to get the limits from the superuser's limits fields.
1194          * This is quite hacky, but it is standard quota practice.
1195          * We look at the USR dquot with id == 0 first, but if user quotas
1196          * are not enabled we goto the GRP dquot with id == 0.
1197          * We don't really care to keep separate default limits for user
1198          * and group quotas, at least not at this point.
1199          */
1200         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1201                              XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 
1202                              (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP :
1203                                 XFS_DQ_PROJ),
1204                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1205                              &dqp);
1206         if (! error) {
1207                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1208
1209                 /*
1210                  * The warnings and timers set the grace period given to
1211                  * a user or group before he or she can not perform any
1212                  * more writing. If it is zero, a default is used.
1213                  */
1214                 qinf->qi_btimelimit = ddqp->d_btimer ?
1215                         be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT;
1216                 qinf->qi_itimelimit = ddqp->d_itimer ?
1217                         be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT;
1218                 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ?
1219                         be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT;
1220                 qinf->qi_bwarnlimit = ddqp->d_bwarns ?
1221                         be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT;
1222                 qinf->qi_iwarnlimit = ddqp->d_iwarns ?
1223                         be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT;
1224                 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ?
1225                         be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT;
1226                 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit);
1227                 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit);
1228                 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit);
1229                 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit);
1230                 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit);
1231                 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit);
1232  
1233                 /*
1234                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1235                  * we don't want this dquot cached. We haven't done a
1236                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1237                  */
1238                 xfs_qm_dqdestroy(dqp);
1239         } else {
1240                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1241                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1242                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1243                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1244                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1245                 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT;
1246         }
1247
1248         return 0;
1249 }
1250
1251
1252 /*
1253  * Gets called when unmounting a filesystem or when all quotas get
1254  * turned off.
1255  * This purges the quota inodes, destroys locks and frees itself.
1256  */
1257 void
1258 xfs_qm_destroy_quotainfo(
1259         xfs_mount_t     *mp)
1260 {
1261         xfs_quotainfo_t *qi;
1262
1263         qi = mp->m_quotainfo;
1264         ASSERT(qi != NULL);
1265         ASSERT(xfs_Gqm != NULL);
1266
1267         /*
1268          * Release the reference that XQM kept, so that we know
1269          * when the XQM structure should be freed. We cannot assume
1270          * that xfs_Gqm is non-null after this point.
1271          */
1272         xfs_qm_rele_quotafs_ref(mp);
1273
1274         spinlock_destroy(&qi->qi_pinlock);
1275         xfs_qm_list_destroy(&qi->qi_dqlist);
1276
1277         if (qi->qi_uquotaip) {
1278                 XFS_PURGE_INODE(qi->qi_uquotaip);
1279                 qi->qi_uquotaip = NULL; /* paranoia */
1280         }
1281         if (qi->qi_gquotaip) {
1282                 XFS_PURGE_INODE(qi->qi_gquotaip);
1283                 qi->qi_gquotaip = NULL;
1284         }
1285         mutex_destroy(&qi->qi_quotaofflock);
1286         kmem_free(qi, sizeof(xfs_quotainfo_t));
1287         mp->m_quotainfo = NULL;
1288 }
1289
1290
1291
1292 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1293
1294 /* ARGSUSED */
1295 STATIC void
1296 xfs_qm_list_init(
1297         xfs_dqlist_t    *list,
1298         char            *str,
1299         int             n)
1300 {
1301         mutex_init(&list->qh_lock);
1302         list->qh_next = NULL;
1303         list->qh_version = 0;
1304         list->qh_nelems = 0;
1305 }
1306
1307 STATIC void
1308 xfs_qm_list_destroy(
1309         xfs_dqlist_t    *list)
1310 {
1311         mutex_destroy(&(list->qh_lock));
1312 }
1313
1314
1315 /*
1316  * Stripped down version of dqattach. This doesn't attach, or even look at the
1317  * dquots attached to the inode. The rationale is that there won't be any
1318  * attached at the time this is called from quotacheck.
1319  */
1320 STATIC int
1321 xfs_qm_dqget_noattach(
1322         xfs_inode_t     *ip,
1323         xfs_dquot_t     **O_udqpp,
1324         xfs_dquot_t     **O_gdqpp)
1325 {
1326         int             error;
1327         xfs_mount_t     *mp;
1328         xfs_dquot_t     *udqp, *gdqp;
1329
1330         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1331         mp = ip->i_mount;
1332         udqp = NULL;
1333         gdqp = NULL;
1334
1335         if (XFS_IS_UQUOTA_ON(mp)) {
1336                 ASSERT(ip->i_udquot == NULL);
1337                 /*
1338                  * We want the dquot allocated if it doesn't exist.
1339                  */
1340                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1341                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1342                                          &udqp))) {
1343                         /*
1344                          * Shouldn't be able to turn off quotas here.
1345                          */
1346                         ASSERT(error != ESRCH);
1347                         ASSERT(error != ENOENT);
1348                         return error;
1349                 }
1350                 ASSERT(udqp);
1351         }
1352
1353         if (XFS_IS_OQUOTA_ON(mp)) {
1354                 ASSERT(ip->i_gdquot == NULL);
1355                 if (udqp)
1356                         xfs_dqunlock(udqp);
1357                 error = XFS_IS_GQUOTA_ON(mp) ?
1358                                 xfs_qm_dqget(mp, ip,
1359                                              ip->i_d.di_gid, XFS_DQ_GROUP,
1360                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1361                                              &gdqp) :
1362                                 xfs_qm_dqget(mp, ip,
1363                                              ip->i_d.di_projid, XFS_DQ_PROJ,
1364                                              XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1365                                              &gdqp);
1366                 if (error) {
1367                         if (udqp)
1368                                 xfs_qm_dqrele(udqp);
1369                         ASSERT(error != ESRCH);
1370                         ASSERT(error != ENOENT);
1371                         return error;
1372                 }
1373                 ASSERT(gdqp);
1374
1375                 /* Reacquire the locks in the right order */
1376                 if (udqp) {
1377                         if (! xfs_qm_dqlock_nowait(udqp)) {
1378                                 xfs_dqunlock(gdqp);
1379                                 xfs_dqlock(udqp);
1380                                 xfs_dqlock(gdqp);
1381                         }
1382                 }
1383         }
1384
1385         *O_udqpp = udqp;
1386         *O_gdqpp = gdqp;
1387
1388 #ifdef QUOTADEBUG
1389         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1390         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1391 #endif
1392         return 0;
1393 }
1394
1395 /*
1396  * Create an inode and return with a reference already taken, but unlocked
1397  * This is how we create quota inodes
1398  */
1399 STATIC int
1400 xfs_qm_qino_alloc(
1401         xfs_mount_t     *mp,
1402         xfs_inode_t     **ip,
1403         __int64_t       sbfields,
1404         uint            flags)
1405 {
1406         xfs_trans_t     *tp;
1407         int             error;
1408         unsigned long   s;
1409         int             committed;
1410
1411         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE);
1412         if ((error = xfs_trans_reserve(tp,
1413                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1414                                       XFS_CREATE_LOG_RES(mp), 0,
1415                                       XFS_TRANS_PERM_LOG_RES,
1416                                       XFS_CREATE_LOG_COUNT))) {
1417                 xfs_trans_cancel(tp, 0);
1418                 return error;
1419         }
1420
1421         if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0,
1422                                    &xfs_zerocr, 0, 1, ip, &committed))) {
1423                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1424                                  XFS_TRANS_ABORT);
1425                 return error;
1426         }
1427
1428         /*
1429          * Keep an extra reference to this quota inode. This inode is
1430          * locked exclusively and joined to the transaction already.
1431          */
1432         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1433         VN_HOLD(XFS_ITOV((*ip)));
1434
1435         /*
1436          * Make the changes in the superblock, and log those too.
1437          * sbfields arg may contain fields other than *QUOTINO;
1438          * VERSIONNUM for example.
1439          */
1440         s = XFS_SB_LOCK(mp);
1441         if (flags & XFS_QMOPT_SBVERSION) {
1442 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1443                 unsigned oldv = mp->m_sb.sb_versionnum;
1444 #endif
1445                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1446                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1447                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1448                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1449                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1450
1451                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1452                 mp->m_sb.sb_uquotino = NULLFSINO;
1453                 mp->m_sb.sb_gquotino = NULLFSINO;
1454
1455                 /* qflags will get updated _after_ quotacheck */
1456                 mp->m_sb.sb_qflags = 0;
1457 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1458                 cmn_err(CE_NOTE,
1459                         "Old superblock version %x, converting to %x.",
1460                         oldv, mp->m_sb.sb_versionnum);
1461 #endif
1462         }
1463         if (flags & XFS_QMOPT_UQUOTA)
1464                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1465         else
1466                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1467         XFS_SB_UNLOCK(mp, s);
1468         xfs_mod_sb(tp, sbfields);
1469
1470         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) {
1471                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1472                 return error;
1473         }
1474         return 0;
1475 }
1476
1477
1478 STATIC int
1479 xfs_qm_reset_dqcounts(
1480         xfs_mount_t     *mp,
1481         xfs_buf_t       *bp,
1482         xfs_dqid_t      id,
1483         uint            type)
1484 {
1485         xfs_disk_dquot_t        *ddq;
1486         int                     j;
1487
1488         xfs_buftrace("RESET DQUOTS", bp);
1489         /*
1490          * Reset all counters and timers. They'll be
1491          * started afresh by xfs_qm_quotacheck.
1492          */
1493 #ifdef DEBUG
1494         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1495         do_div(j, sizeof(xfs_dqblk_t));
1496         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1497 #endif
1498         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1499         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1500                 /*
1501                  * Do a sanity check, and if needed, repair the dqblk. Don't
1502                  * output any warnings because it's perfectly possible to
1503                  * find uninitialised dquot blks. See comment in xfs_qm_dqcheck.
1504                  */
1505                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1506                                       "xfs_quotacheck");
1507                 ddq->d_bcount = 0;
1508                 ddq->d_icount = 0;
1509                 ddq->d_rtbcount = 0;
1510                 ddq->d_btimer = 0;
1511                 ddq->d_itimer = 0;
1512                 ddq->d_rtbtimer = 0;
1513                 ddq->d_bwarns = 0;
1514                 ddq->d_iwarns = 0;
1515                 ddq->d_rtbwarns = 0;
1516                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1517         }
1518
1519         return 0;
1520 }
1521
1522 STATIC int
1523 xfs_qm_dqiter_bufs(
1524         xfs_mount_t     *mp,
1525         xfs_dqid_t      firstid,
1526         xfs_fsblock_t   bno,
1527         xfs_filblks_t   blkcnt,
1528         uint            flags)
1529 {
1530         xfs_buf_t       *bp;
1531         int             error;
1532         int             notcommitted;
1533         int             incr;
1534         int             type;
1535
1536         ASSERT(blkcnt > 0);
1537         notcommitted = 0;
1538         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1539                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1540         type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER :
1541                 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP);
1542         error = 0;
1543
1544         /*
1545          * Blkcnt arg can be a very big number, and might even be
1546          * larger than the log itself. So, we have to break it up into
1547          * manageable-sized transactions.
1548          * Note that we don't start a permanent transaction here; we might
1549          * not be able to get a log reservation for the whole thing up front,
1550          * and we don't really care to either, because we just discard
1551          * everything if we were to crash in the middle of this loop.
1552          */
1553         while (blkcnt--) {
1554                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1555                               XFS_FSB_TO_DADDR(mp, bno),
1556                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1557                 if (error)
1558                         break;
1559
1560                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid, type);
1561                 xfs_bdwrite(mp, bp);
1562                 /*
1563                  * goto the next block.
1564                  */
1565                 bno++;
1566                 firstid += XFS_QM_DQPERBLK(mp);
1567         }
1568         return error;
1569 }
1570
1571 /*
1572  * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a
1573  * caller supplied function for every chunk of dquots that we find.
1574  */
1575 STATIC int
1576 xfs_qm_dqiterate(
1577         xfs_mount_t     *mp,
1578         xfs_inode_t     *qip,
1579         uint            flags)
1580 {
1581         xfs_bmbt_irec_t         *map;
1582         int                     i, nmaps;       /* number of map entries */
1583         int                     error;          /* return value */
1584         xfs_fileoff_t           lblkno;
1585         xfs_filblks_t           maxlblkcnt;
1586         xfs_dqid_t              firstid;
1587         xfs_fsblock_t           rablkno;
1588         xfs_filblks_t           rablkcnt;
1589
1590         error = 0;
1591         /*
1592          * This looks racy, but we can't keep an inode lock across a
1593          * trans_reserve. But, this gets called during quotacheck, and that
1594          * happens only at mount time which is single threaded.
1595          */
1596         if (qip->i_d.di_nblocks == 0)
1597                 return 0;
1598
1599         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1600
1601         lblkno = 0;
1602         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1603         do {
1604                 nmaps = XFS_DQITER_MAP_SIZE;
1605                 /*
1606                  * We aren't changing the inode itself. Just changing
1607                  * some of its data. No new blocks are added here, and
1608                  * the inode is never added to the transaction.
1609                  */
1610                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1611                 error = xfs_bmapi(NULL, qip, lblkno,
1612                                   maxlblkcnt - lblkno,
1613                                   XFS_BMAPI_METADATA,
1614                                   NULL,
1615                                   0, map, &nmaps, NULL, NULL);
1616                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1617                 if (error)
1618                         break;
1619
1620                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1621                 for (i = 0; i < nmaps; i++) {
1622                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1623                         ASSERT(map[i].br_blockcount);
1624
1625
1626                         lblkno += map[i].br_blockcount;
1627
1628                         if (map[i].br_startblock == HOLESTARTBLOCK)
1629                                 continue;
1630
1631                         firstid = (xfs_dqid_t) map[i].br_startoff *
1632                                 XFS_QM_DQPERBLK(mp);
1633                         /*
1634                          * Do a read-ahead on the next extent.
1635                          */
1636                         if ((i+1 < nmaps) &&
1637                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1638                                 rablkcnt =  map[i+1].br_blockcount;
1639                                 rablkno = map[i+1].br_startblock;
1640                                 while (rablkcnt--) {
1641                                         xfs_baread(mp->m_ddev_targp,
1642                                                XFS_FSB_TO_DADDR(mp, rablkno),
1643                                                (int)XFS_QI_DQCHUNKLEN(mp));
1644                                         rablkno++;
1645                                 }
1646                         }
1647                         /*
1648                          * Iterate thru all the blks in the extent and
1649                          * reset the counters of all the dquots inside them.
1650                          */
1651                         if ((error = xfs_qm_dqiter_bufs(mp,
1652                                                        firstid,
1653                                                        map[i].br_startblock,
1654                                                        map[i].br_blockcount,
1655                                                        flags))) {
1656                                 break;
1657                         }
1658                 }
1659
1660                 if (error)
1661                         break;
1662         } while (nmaps > 0);
1663
1664         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1665
1666         return error;
1667 }
1668
1669 /*
1670  * Called by dqusage_adjust in doing a quotacheck.
1671  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1672  * this updates its incore copy as well as the buffer copy. This is
1673  * so that once the quotacheck is done, we can just log all the buffers,
1674  * as opposed to logging numerous updates to individual dquots.
1675  */
1676 STATIC void
1677 xfs_qm_quotacheck_dqadjust(
1678         xfs_dquot_t             *dqp,
1679         xfs_qcnt_t              nblks,
1680         xfs_qcnt_t              rtblks)
1681 {
1682         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1683         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1684         /*
1685          * Adjust the inode count and the block count to reflect this inode's
1686          * resource usage.
1687          */
1688         be64_add(&dqp->q_core.d_icount, 1);
1689         dqp->q_res_icount++;
1690         if (nblks) {
1691                 be64_add(&dqp->q_core.d_bcount, nblks);
1692                 dqp->q_res_bcount += nblks;
1693         }
1694         if (rtblks) {
1695                 be64_add(&dqp->q_core.d_rtbcount, rtblks);
1696                 dqp->q_res_rtbcount += rtblks;
1697         }
1698
1699         /*
1700          * Set default limits, adjust timers (since we changed usages)
1701          */
1702         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1703                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1704                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1705         }
1706
1707         dqp->dq_flags |= XFS_DQ_DIRTY;
1708 }
1709
1710 STATIC int
1711 xfs_qm_get_rtblks(
1712         xfs_inode_t     *ip,
1713         xfs_qcnt_t      *O_rtblks)
1714 {
1715         xfs_filblks_t   rtblks;                 /* total rt blks */
1716         xfs_extnum_t    idx;                    /* extent record index */
1717         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1718         xfs_extnum_t    nextents;               /* number of extent entries */
1719         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1720         int             error;
1721
1722         ASSERT(XFS_IS_REALTIME_INODE(ip));
1723         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1724         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1725                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1726                         return error;
1727         }
1728         rtblks = 0;
1729         nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
1730         for (idx = 0; idx < nextents; idx++) {
1731                 ep = xfs_iext_get_ext(ifp, idx);
1732                 rtblks += xfs_bmbt_get_blockcount(ep);
1733         }
1734         *O_rtblks = (xfs_qcnt_t)rtblks;
1735         return 0;
1736 }
1737
1738 /*
1739  * callback routine supplied to bulkstat(). Given an inumber, find its
1740  * dquots and update them to account for resources taken by that inode.
1741  */
1742 /* ARGSUSED */
1743 STATIC int
1744 xfs_qm_dqusage_adjust(
1745         xfs_mount_t     *mp,            /* mount point for filesystem */
1746         xfs_ino_t       ino,            /* inode number to get data for */
1747         void            __user *buffer, /* not used */
1748         int             ubsize,         /* not used */
1749         void            *private_data,  /* not used */
1750         xfs_daddr_t     bno,            /* starting block of inode cluster */
1751         int             *ubused,        /* not used */
1752         void            *dip,           /* on-disk inode pointer (not used) */
1753         int             *res)           /* result code value */
1754 {
1755         xfs_inode_t     *ip;
1756         xfs_dquot_t     *udqp, *gdqp;
1757         xfs_qcnt_t      nblks, rtblks;
1758         int             error;
1759
1760         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1761
1762         /*
1763          * rootino must have its resources accounted for, not so with the quota
1764          * inodes.
1765          */
1766         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1767                 *res = BULKSTAT_RV_NOTHING;
1768                 return XFS_ERROR(EINVAL);
1769         }
1770
1771         /*
1772          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1773          * interface expects the inode to be exclusively locked because that's
1774          * the case in all other instances. It's OK that we do this because
1775          * quotacheck is done only at mount time.
1776          */
1777         if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip, bno))) {
1778                 *res = BULKSTAT_RV_NOTHING;
1779                 return error;
1780         }
1781
1782         if (ip->i_d.di_mode == 0) {
1783                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1784                 *res = BULKSTAT_RV_NOTHING;
1785                 return XFS_ERROR(ENOENT);
1786         }
1787
1788         /*
1789          * Obtain the locked dquots. In case of an error (eg. allocation
1790          * fails for ENOSPC), we return the negative of the error number
1791          * to bulkstat, so that it can get propagated to quotacheck() and
1792          * making us disable quotas for the file system.
1793          */
1794         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1795                 xfs_iput(ip, XFS_ILOCK_EXCL);
1796                 *res = BULKSTAT_RV_GIVEUP;
1797                 return error;
1798         }
1799
1800         rtblks = 0;
1801         if (! XFS_IS_REALTIME_INODE(ip)) {
1802                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1803         } else {
1804                 /*
1805                  * Walk thru the extent list and count the realtime blocks.
1806                  */
1807                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1808                         xfs_iput(ip, XFS_ILOCK_EXCL);
1809                         if (udqp)
1810                                 xfs_qm_dqput(udqp);
1811                         if (gdqp)
1812                                 xfs_qm_dqput(gdqp);
1813                         *res = BULKSTAT_RV_GIVEUP;
1814                         return error;
1815                 }
1816                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1817         }
1818         ASSERT(ip->i_delayed_blks == 0);
1819
1820         /*
1821          * We can't release the inode while holding its dquot locks.
1822          * The inode can go into inactive and might try to acquire the dquotlocks.
1823          * So, just unlock here and do a vn_rele at the end.
1824          */
1825         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1826
1827         /*
1828          * Add the (disk blocks and inode) resources occupied by this
1829          * inode to its dquots. We do this adjustment in the incore dquot,
1830          * and also copy the changes to its buffer.
1831          * We don't care about putting these changes in a transaction
1832          * envelope because if we crash in the middle of a 'quotacheck'
1833          * we have to start from the beginning anyway.
1834          * Once we're done, we'll log all the dquot bufs.
1835          *
1836          * The *QUOTA_ON checks below may look pretty racy, but quotachecks
1837          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1838          */
1839         if (XFS_IS_UQUOTA_ON(mp)) {
1840                 ASSERT(udqp);
1841                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1842                 xfs_qm_dqput(udqp);
1843         }
1844         if (XFS_IS_OQUOTA_ON(mp)) {
1845                 ASSERT(gdqp);
1846                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1847                 xfs_qm_dqput(gdqp);
1848         }
1849         /*
1850          * Now release the inode. This will send it to 'inactive', and
1851          * possibly even free blocks.
1852          */
1853         VN_RELE(XFS_ITOV(ip));
1854
1855         /*
1856          * Goto next inode.
1857          */
1858         *res = BULKSTAT_RV_DIDONE;
1859         return 0;
1860 }
1861
1862 /*
1863  * Walk thru all the filesystem inodes and construct a consistent view
1864  * of the disk quota world. If the quotacheck fails, disable quotas.
1865  */
1866 int
1867 xfs_qm_quotacheck(
1868         xfs_mount_t     *mp)
1869 {
1870         int             done, count, error;
1871         xfs_ino_t       lastino;
1872         size_t          structsz;
1873         xfs_inode_t     *uip, *gip;
1874         uint            flags;
1875
1876         count = INT_MAX;
1877         structsz = 1;
1878         lastino = 0;
1879         flags = 0;
1880
1881         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1882         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1883
1884         /*
1885          * There should be no cached dquots. The (simplistic) quotacheck
1886          * algorithm doesn't like that.
1887          */
1888         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1889
1890         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1891
1892         /*
1893          * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset
1894          * their counters to zero. We need a clean slate.
1895          * We don't log our changes till later.
1896          */
1897         if ((uip = XFS_QI_UQIP(mp))) {
1898                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1899                         goto error_return;
1900                 flags |= XFS_UQUOTA_CHKD;
1901         }
1902
1903         if ((gip = XFS_QI_GQIP(mp))) {
1904                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ?
1905                                         XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA)))
1906                         goto error_return;
1907                 flags |= XFS_OQUOTA_CHKD;
1908         }
1909
1910         do {
1911                 /*
1912                  * Iterate thru all the inodes in the file system,
1913                  * adjusting the corresponding dquot counters in core.
1914                  */
1915                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1916                                      xfs_qm_dqusage_adjust, NULL,
1917                                      structsz, NULL, BULKSTAT_FG_IGET, &done)))
1918                         break;
1919
1920         } while (! done);
1921
1922         /*
1923          * We can get this error if we couldn't do a dquot allocation inside
1924          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1925          * dirty dquots that might be cached, we just want to get rid of them
1926          * and turn quotaoff. The dquots won't be attached to any of the inodes
1927          * at this point (because we intentionally didn't in dqget_noattach).
1928          */
1929         if (error) {
1930                 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL | XFS_QMOPT_QUOTAOFF);
1931                 goto error_return;
1932         }
1933         /*
1934          * We've made all the changes that we need to make incore.
1935          * Now flush_them down to disk buffers.
1936          */
1937         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1938
1939         /*
1940          * We didn't log anything, because if we crashed, we'll have to
1941          * start the quotacheck from scratch anyway. However, we must make
1942          * sure that our dquot changes are secure before we put the
1943          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1944          * flush.
1945          */
1946         XFS_bflush(mp->m_ddev_targp);
1947
1948         /*
1949          * If one type of quotas is off, then it will lose its
1950          * quotachecked status, since we won't be doing accounting for
1951          * that type anymore.
1952          */
1953         mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD);
1954         mp->m_qflags |= flags;
1955
1956         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1957
1958  error_return:
1959         if (error) {
1960                 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): "
1961                         "Disabling quotas.",
1962                         mp->m_fsname, error);
1963                 /*
1964                  * We must turn off quotas.
1965                  */
1966                 ASSERT(mp->m_quotainfo != NULL);
1967                 ASSERT(xfs_Gqm != NULL);
1968                 xfs_qm_destroy_quotainfo(mp);
1969                 (void)xfs_mount_reset_sbqflags(mp);
1970         } else {
1971                 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1972         }
1973         return (error);
1974 }
1975
1976 /*
1977  * This is called after the superblock has been read in and we're ready to
1978  * iget the quota inodes.
1979  */
1980 STATIC int
1981 xfs_qm_init_quotainos(
1982         xfs_mount_t     *mp)
1983 {
1984         xfs_inode_t     *uip, *gip;
1985         int             error;
1986         __int64_t       sbflags;
1987         uint            flags;
1988
1989         ASSERT(mp->m_quotainfo);
1990         uip = gip = NULL;
1991         sbflags = 0;
1992         flags = 0;
1993
1994         /*
1995          * Get the uquota and gquota inodes
1996          */
1997         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
1998                 if (XFS_IS_UQUOTA_ON(mp) &&
1999                     mp->m_sb.sb_uquotino != NULLFSINO) {
2000                         ASSERT(mp->m_sb.sb_uquotino > 0);
2001                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
2002                                              0, 0, &uip, 0)))
2003                                 return XFS_ERROR(error);
2004                 }
2005                 if (XFS_IS_OQUOTA_ON(mp) &&
2006                     mp->m_sb.sb_gquotino != NULLFSINO) {
2007                         ASSERT(mp->m_sb.sb_gquotino > 0);
2008                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2009                                              0, 0, &gip, 0))) {
2010                                 if (uip)
2011                                         VN_RELE(XFS_ITOV(uip));
2012                                 return XFS_ERROR(error);
2013                         }
2014                 }
2015         } else {
2016                 flags |= XFS_QMOPT_SBVERSION;
2017                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2018                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2019         }
2020
2021         /*
2022          * Create the two inodes, if they don't exist already. The changes
2023          * made above will get added to a transaction and logged in one of
2024          * the qino_alloc calls below.  If the device is readonly,
2025          * temporarily switch to read-write to do this.
2026          */
2027         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2028                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2029                                               sbflags | XFS_SB_UQUOTINO,
2030                                               flags | XFS_QMOPT_UQUOTA)))
2031                         return XFS_ERROR(error);
2032
2033                 flags &= ~XFS_QMOPT_SBVERSION;
2034         }
2035         if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) {
2036                 flags |= (XFS_IS_GQUOTA_ON(mp) ?
2037                                 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA);
2038                 error = xfs_qm_qino_alloc(mp, &gip,
2039                                           sbflags | XFS_SB_GQUOTINO, flags);
2040                 if (error) {
2041                         if (uip)
2042                                 VN_RELE(XFS_ITOV(uip));
2043
2044                         return XFS_ERROR(error);
2045                 }
2046         }
2047
2048         XFS_QI_UQIP(mp) = uip;
2049         XFS_QI_GQIP(mp) = gip;
2050
2051         return 0;
2052 }
2053
2054
2055 /*
2056  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2057  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2058  * favor the lookup function ...
2059  * XXXsup merge this with qm_reclaim_one().
2060  */
2061 STATIC int
2062 xfs_qm_shake_freelist(
2063         int howmany)
2064 {
2065         int             nreclaimed;
2066         xfs_dqhash_t    *hash;
2067         xfs_dquot_t     *dqp, *nextdqp;
2068         int             restarts;
2069         int             nflushes;
2070
2071         if (howmany <= 0)
2072                 return 0;
2073
2074         nreclaimed = 0;
2075         restarts = 0;
2076         nflushes = 0;
2077
2078 #ifdef QUOTADEBUG
2079         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2080 #endif
2081         /* lock order is : hashchainlock, freelistlock, mplistlock */
2082  tryagain:
2083         xfs_qm_freelist_lock(xfs_Gqm);
2084
2085         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2086              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2087               nreclaimed < howmany); ) {
2088                 xfs_dqlock(dqp);
2089
2090                 /*
2091                  * We are racing with dqlookup here. Naturally we don't
2092                  * want to reclaim a dquot that lookup wants.
2093                  */
2094                 if (dqp->dq_flags & XFS_DQ_WANT) {
2095                         xfs_dqunlock(dqp);
2096                         xfs_qm_freelist_unlock(xfs_Gqm);
2097                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2098                                 return nreclaimed;
2099                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2100                         goto tryagain;
2101                 }
2102
2103                 /*
2104                  * If the dquot is inactive, we are assured that it is
2105                  * not on the mplist or the hashlist, and that makes our
2106                  * life easier.
2107                  */
2108                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2109                         ASSERT(dqp->q_mount == NULL);
2110                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2111                         ASSERT(dqp->HL_PREVP == NULL);
2112                         ASSERT(dqp->MPL_PREVP == NULL);
2113                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2114                         nextdqp = dqp->dq_flnext;
2115                         goto off_freelist;
2116                 }
2117
2118                 ASSERT(dqp->MPL_PREVP);
2119                 /*
2120                  * Try to grab the flush lock. If this dquot is in the process of
2121                  * getting flushed to disk, we don't want to reclaim it.
2122                  */
2123                 if (! xfs_qm_dqflock_nowait(dqp)) {
2124                         xfs_dqunlock(dqp);
2125                         dqp = dqp->dq_flnext;
2126                         continue;
2127                 }
2128
2129                 /*
2130                  * We have the flush lock so we know that this is not in the
2131                  * process of being flushed. So, if this is dirty, flush it
2132                  * DELWRI so that we don't get a freelist infested with
2133                  * dirty dquots.
2134                  */
2135                 if (XFS_DQ_IS_DIRTY(dqp)) {
2136                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2137                         /*
2138                          * We flush it delayed write, so don't bother
2139                          * releasing the mplock.
2140                          */
2141                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2142                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2143                         dqp = dqp->dq_flnext;
2144                         continue;
2145                 }
2146                 /*
2147                  * We're trying to get the hashlock out of order. This races
2148                  * with dqlookup; so, we giveup and goto the next dquot if
2149                  * we couldn't get the hashlock. This way, we won't starve
2150                  * a dqlookup process that holds the hashlock that is
2151                  * waiting for the freelist lock.
2152                  */
2153                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2154                         xfs_dqfunlock(dqp);
2155                         xfs_dqunlock(dqp);
2156                         dqp = dqp->dq_flnext;
2157                         continue;
2158                 }
2159                 /*
2160                  * This races with dquot allocation code as well as dqflush_all
2161                  * and reclaim code. So, if we failed to grab the mplist lock,
2162                  * giveup everything and start over.
2163                  */
2164                 hash = dqp->q_hash;
2165                 ASSERT(hash);
2166                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2167                         /* XXX put a sentinel so that we can come back here */
2168                         xfs_dqfunlock(dqp);
2169                         xfs_dqunlock(dqp);
2170                         XFS_DQ_HASH_UNLOCK(hash);
2171                         xfs_qm_freelist_unlock(xfs_Gqm);
2172                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2173                                 return nreclaimed;
2174                         goto tryagain;
2175                 }
2176                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2177 #ifdef QUOTADEBUG
2178                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2179                         dqp, be32_to_cpu(dqp->q_core.d_id));
2180 #endif
2181                 ASSERT(dqp->q_nrefs == 0);
2182                 nextdqp = dqp->dq_flnext;
2183                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2184                 XQM_HASHLIST_REMOVE(hash, dqp);
2185                 xfs_dqfunlock(dqp);
2186                 xfs_qm_mplist_unlock(dqp->q_mount);
2187                 XFS_DQ_HASH_UNLOCK(hash);
2188
2189  off_freelist:
2190                 XQM_FREELIST_REMOVE(dqp);
2191                 xfs_dqunlock(dqp);
2192                 nreclaimed++;
2193                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2194                 xfs_qm_dqdestroy(dqp);
2195                 dqp = nextdqp;
2196         }
2197         xfs_qm_freelist_unlock(xfs_Gqm);
2198         return nreclaimed;
2199 }
2200
2201
2202 /*
2203  * The kmem_shake interface is invoked when memory is running low.
2204  */
2205 /* ARGSUSED */
2206 STATIC int
2207 xfs_qm_shake(int nr_to_scan, gfp_t gfp_mask)
2208 {
2209         int     ndqused, nfree, n;
2210
2211         if (!kmem_shake_allow(gfp_mask))
2212                 return 0;
2213         if (!xfs_Gqm)
2214                 return 0;
2215
2216         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2217         /* incore dquots in all f/s's */
2218         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2219
2220         ASSERT(ndqused >= 0);
2221
2222         if (nfree <= ndqused && nfree < ndquot)
2223                 return 0;
2224
2225         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2226         n = nfree - ndqused - ndquot;           /* # over target */
2227
2228         return xfs_qm_shake_freelist(MAX(nfree, n));
2229 }
2230
2231
2232 /*
2233  * Just pop the least recently used dquot off the freelist and
2234  * recycle it. The returned dquot is locked.
2235  */
2236 STATIC xfs_dquot_t *
2237 xfs_qm_dqreclaim_one(void)
2238 {
2239         xfs_dquot_t     *dqpout;
2240         xfs_dquot_t     *dqp;
2241         int             restarts;
2242         int             nflushes;
2243
2244         restarts = 0;
2245         dqpout = NULL;
2246         nflushes = 0;
2247
2248         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2249  startagain:
2250         xfs_qm_freelist_lock(xfs_Gqm);
2251
2252         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2253                 xfs_dqlock(dqp);
2254
2255                 /*
2256                  * We are racing with dqlookup here. Naturally we don't
2257                  * want to reclaim a dquot that lookup wants. We release the
2258                  * freelist lock and start over, so that lookup will grab
2259                  * both the dquot and the freelistlock.
2260                  */
2261                 if (dqp->dq_flags & XFS_DQ_WANT) {
2262                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2263                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2264                         xfs_dqunlock(dqp);
2265                         xfs_qm_freelist_unlock(xfs_Gqm);
2266                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2267                                 return NULL;
2268                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2269                         goto startagain;
2270                 }
2271
2272                 /*
2273                  * If the dquot is inactive, we are assured that it is
2274                  * not on the mplist or the hashlist, and that makes our
2275                  * life easier.
2276                  */
2277                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2278                         ASSERT(dqp->q_mount == NULL);
2279                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2280                         ASSERT(dqp->HL_PREVP == NULL);
2281                         ASSERT(dqp->MPL_PREVP == NULL);
2282                         XQM_FREELIST_REMOVE(dqp);
2283                         xfs_dqunlock(dqp);
2284                         dqpout = dqp;
2285                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2286                         break;
2287                 }
2288
2289                 ASSERT(dqp->q_hash);
2290                 ASSERT(dqp->MPL_PREVP);
2291
2292                 /*
2293                  * Try to grab the flush lock. If this dquot is in the process of
2294                  * getting flushed to disk, we don't want to reclaim it.
2295                  */
2296                 if (! xfs_qm_dqflock_nowait(dqp)) {
2297                         xfs_dqunlock(dqp);
2298                         continue;
2299                 }
2300
2301                 /*
2302                  * We have the flush lock so we know that this is not in the
2303                  * process of being flushed. So, if this is dirty, flush it
2304                  * DELWRI so that we don't get a freelist infested with
2305                  * dirty dquots.
2306                  */
2307                 if (XFS_DQ_IS_DIRTY(dqp)) {
2308                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2309                         /*
2310                          * We flush it delayed write, so don't bother
2311                          * releasing the freelist lock.
2312                          */
2313                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2314                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2315                         continue;
2316                 }
2317
2318                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2319                         xfs_dqfunlock(dqp);
2320                         xfs_dqunlock(dqp);
2321                         continue;
2322                 }
2323
2324                 if (! xfs_qm_dqhashlock_nowait(dqp))
2325                         goto mplistunlock;
2326
2327                 ASSERT(dqp->q_nrefs == 0);
2328                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2329                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2330                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2331                 XQM_FREELIST_REMOVE(dqp);
2332                 dqpout = dqp;
2333                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2334  mplistunlock:
2335                 xfs_qm_mplist_unlock(dqp->q_mount);
2336                 xfs_dqfunlock(dqp);
2337                 xfs_dqunlock(dqp);
2338                 if (dqpout)
2339                         break;
2340         }
2341
2342         xfs_qm_freelist_unlock(xfs_Gqm);
2343         return dqpout;
2344 }
2345
2346
2347 /*------------------------------------------------------------------*/
2348
2349 /*
2350  * Return a new incore dquot. Depending on the number of
2351  * dquots in the system, we either allocate a new one on the kernel heap,
2352  * or reclaim a free one.
2353  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2354  * to reclaim an existing one from the freelist.
2355  */
2356 boolean_t
2357 xfs_qm_dqalloc_incore(
2358         xfs_dquot_t **O_dqpp)
2359 {
2360         xfs_dquot_t     *dqp;
2361
2362         /*
2363          * Check against high water mark to see if we want to pop
2364          * a nincompoop dquot off the freelist.
2365          */
2366         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2367                 /*
2368                  * Try to recycle a dquot from the freelist.
2369                  */
2370                 if ((dqp = xfs_qm_dqreclaim_one())) {
2371                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2372                         /*
2373                          * Just zero the core here. The rest will get
2374                          * reinitialized by caller. XXX we shouldn't even
2375                          * do this zero ...
2376                          */
2377                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2378                         *O_dqpp = dqp;
2379                         return B_FALSE;
2380                 }
2381                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2382         }
2383
2384         /*
2385          * Allocate a brand new dquot on the kernel heap and return it
2386          * to the caller to initialize.
2387          */
2388         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2389         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2390         atomic_inc(&xfs_Gqm->qm_totaldquots);
2391
2392         return B_TRUE;
2393 }
2394
2395
2396 /*
2397  * Start a transaction and write the incore superblock changes to
2398  * disk. flags parameter indicates which fields have changed.
2399  */
2400 int
2401 xfs_qm_write_sb_changes(
2402         xfs_mount_t     *mp,
2403         __int64_t       flags)
2404 {
2405         xfs_trans_t     *tp;
2406         int             error;
2407
2408 #ifdef QUOTADEBUG
2409         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2410 #endif
2411         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2412         if ((error = xfs_trans_reserve(tp, 0,
2413                                       mp->m_sb.sb_sectsize + 128, 0,
2414                                       0,
2415                                       XFS_DEFAULT_LOG_COUNT))) {
2416                 xfs_trans_cancel(tp, 0);
2417                 return error;
2418         }
2419
2420         xfs_mod_sb(tp, flags);
2421         (void) xfs_trans_commit(tp, 0);
2422
2423         return 0;
2424 }
2425
2426
2427 /* --------------- utility functions for vnodeops ---------------- */
2428
2429
2430 /*
2431  * Given an inode, a uid and gid (from cred_t) make sure that we have
2432  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2433  * quotas by creating this file.
2434  * This also attaches dquot(s) to the given inode after locking it,
2435  * and returns the dquots corresponding to the uid and/or gid.
2436  *
2437  * in   : inode (unlocked)
2438  * out  : udquot, gdquot with references taken and unlocked
2439  */
2440 int
2441 xfs_qm_vop_dqalloc(
2442         xfs_mount_t     *mp,
2443         xfs_inode_t     *ip,
2444         uid_t           uid,
2445         gid_t           gid,
2446         prid_t          prid,
2447         uint            flags,
2448         xfs_dquot_t     **O_udqpp,
2449         xfs_dquot_t     **O_gdqpp)
2450 {
2451         int             error;
2452         xfs_dquot_t     *uq, *gq;
2453         uint            lockflags;
2454
2455         if (!XFS_IS_QUOTA_ON(mp))
2456                 return 0;
2457
2458         lockflags = XFS_ILOCK_EXCL;
2459         xfs_ilock(ip, lockflags);
2460
2461         if ((flags & XFS_QMOPT_INHERIT) &&
2462             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2463                 gid = ip->i_d.di_gid;
2464
2465         /*
2466          * Attach the dquot(s) to this inode, doing a dquot allocation
2467          * if necessary. The dquot(s) will not be locked.
2468          */
2469         if (XFS_NOT_DQATTACHED(mp, ip)) {
2470                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2471                                             XFS_QMOPT_ILOCKED))) {
2472                         xfs_iunlock(ip, lockflags);
2473                         return error;
2474                 }
2475         }
2476
2477         uq = gq = NULL;
2478         if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) {
2479                 if (ip->i_d.di_uid != uid) {
2480                         /*
2481                          * What we need is the dquot that has this uid, and
2482                          * if we send the inode to dqget, the uid of the inode
2483                          * takes priority over what's sent in the uid argument.
2484                          * We must unlock inode here before calling dqget if
2485                          * we're not sending the inode, because otherwise
2486                          * we'll deadlock by doing trans_reserve while
2487                          * holding ilock.
2488                          */
2489                         xfs_iunlock(ip, lockflags);
2490                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2491                                                  XFS_DQ_USER,
2492                                                  XFS_QMOPT_DQALLOC |
2493                                                  XFS_QMOPT_DOWARN,
2494                                                  &uq))) {
2495                                 ASSERT(error != ENOENT);
2496                                 return error;
2497                         }
2498                         /*
2499                          * Get the ilock in the right order.
2500                          */
2501                         xfs_dqunlock(uq);
2502                         lockflags = XFS_ILOCK_SHARED;
2503                         xfs_ilock(ip, lockflags);
2504                 } else {
2505                         /*
2506                          * Take an extra reference, because we'll return
2507                          * this to caller
2508                          */
2509                         ASSERT(ip->i_udquot);
2510                         uq = ip->i_udquot;
2511                         xfs_dqlock(uq);
2512                         XFS_DQHOLD(uq);
2513                         xfs_dqunlock(uq);
2514                 }
2515         }
2516         if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) {
2517                 if (ip->i_d.di_gid != gid) {
2518                         xfs_iunlock(ip, lockflags);
2519                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2520                                                  XFS_DQ_GROUP,
2521                                                  XFS_QMOPT_DQALLOC |
2522                                                  XFS_QMOPT_DOWARN,
2523                                                  &gq))) {
2524                                 if (uq)
2525                                         xfs_qm_dqrele(uq);
2526                                 ASSERT(error != ENOENT);
2527                                 return error;
2528                         }
2529                         xfs_dqunlock(gq);
2530                         lockflags = XFS_ILOCK_SHARED;
2531                         xfs_ilock(ip, lockflags);
2532                 } else {
2533                         ASSERT(ip->i_gdquot);
2534                         gq = ip->i_gdquot;
2535                         xfs_dqlock(gq);
2536                         XFS_DQHOLD(gq);
2537                         xfs_dqunlock(gq);
2538                 }
2539         } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) {
2540                 if (ip->i_d.di_projid != prid) {
2541                         xfs_iunlock(ip, lockflags);
2542                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid,
2543                                                  XFS_DQ_PROJ,
2544                                                  XFS_QMOPT_DQALLOC |
2545                                                  XFS_QMOPT_DOWARN,
2546                                                  &gq))) {
2547                                 if (uq)
2548                                         xfs_qm_dqrele(uq);
2549                                 ASSERT(error != ENOENT);
2550                                 return (error);
2551                         }
2552                         xfs_dqunlock(gq);
2553                         lockflags = XFS_ILOCK_SHARED;
2554                         xfs_ilock(ip, lockflags);
2555                 } else {
2556                         ASSERT(ip->i_gdquot);
2557                         gq = ip->i_gdquot;
2558                         xfs_dqlock(gq);
2559                         XFS_DQHOLD(gq);
2560                         xfs_dqunlock(gq);
2561                 }
2562         }
2563         if (uq)
2564                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2565
2566         xfs_iunlock(ip, lockflags);
2567         if (O_udqpp)
2568                 *O_udqpp = uq;
2569         else if (uq)
2570                 xfs_qm_dqrele(uq);
2571         if (O_gdqpp)
2572                 *O_gdqpp = gq;
2573         else if (gq)
2574                 xfs_qm_dqrele(gq);
2575         return 0;
2576 }
2577
2578 /*
2579  * Actually transfer ownership, and do dquot modifications.
2580  * These were already reserved.
2581  */
2582 xfs_dquot_t *
2583 xfs_qm_vop_chown(
2584         xfs_trans_t     *tp,
2585         xfs_inode_t     *ip,
2586         xfs_dquot_t     **IO_olddq,
2587         xfs_dquot_t     *newdq)
2588 {
2589         xfs_dquot_t     *prevdq;
2590         uint            bfield = XFS_IS_REALTIME_INODE(ip) ?
2591                                  XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT;
2592
2593         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2594         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2595
2596         /* old dquot */
2597         prevdq = *IO_olddq;
2598         ASSERT(prevdq);
2599         ASSERT(prevdq != newdq);
2600
2601         xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks));
2602         xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1);
2603
2604         /* the sparkling new dquot */
2605         xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks);
2606         xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1);
2607
2608         /*
2609          * Take an extra reference, because the inode
2610          * is going to keep this dquot pointer even
2611          * after the trans_commit.
2612          */
2613         xfs_dqlock(newdq);
2614         XFS_DQHOLD(newdq);
2615         xfs_dqunlock(newdq);
2616         *IO_olddq = newdq;
2617
2618         return prevdq;
2619 }
2620
2621 /*
2622  * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID).
2623  */
2624 int
2625 xfs_qm_vop_chown_reserve(
2626         xfs_trans_t     *tp,
2627         xfs_inode_t     *ip,
2628         xfs_dquot_t     *udqp,
2629         xfs_dquot_t     *gdqp,
2630         uint            flags)
2631 {
2632         int             error;
2633         xfs_mount_t     *mp;
2634         uint            delblks, blkflags, prjflags = 0;
2635         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2636
2637         ASSERT(XFS_ISLOCKED_INODE(ip));
2638         mp = ip->i_mount;
2639         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2640
2641         delblks = ip->i_delayed_blks;
2642         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2643         blkflags = XFS_IS_REALTIME_INODE(ip) ?
2644                         XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS;
2645
2646         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2647             ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) {
2648                 delblksudq = udqp;
2649                 /*
2650                  * If there are delayed allocation blocks, then we have to
2651                  * unreserve those from the old dquot, and add them to the
2652                  * new dquot.
2653                  */
2654                 if (delblks) {
2655                         ASSERT(ip->i_udquot);
2656                         unresudq = ip->i_udquot;
2657                 }
2658         }
2659         if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) {
2660                 if (XFS_IS_PQUOTA_ON(ip->i_mount) &&
2661                      ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id))
2662                         prjflags = XFS_QMOPT_ENOSPC;
2663
2664                 if (prjflags ||
2665                     (XFS_IS_GQUOTA_ON(ip->i_mount) &&
2666                      ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) {
2667                         delblksgdq = gdqp;
2668                         if (delblks) {
2669                                 ASSERT(ip->i_gdquot);
2670                                 unresgdq = ip->i_gdquot;
2671                         }
2672                 }
2673         }
2674
2675         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2676                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2677                                 flags | blkflags | prjflags)))
2678                 return (error);
2679
2680         /*
2681          * Do the delayed blks reservations/unreservations now. Since, these
2682          * are done without the help of a transaction, if a reservation fails
2683          * its previous reservations won't be automatically undone by trans
2684          * code. So, we have to do it manually here.
2685          */
2686         if (delblks) {
2687                 /*
2688                  * Do the reservations first. Unreservation can't fail.
2689                  */
2690                 ASSERT(delblksudq || delblksgdq);
2691                 ASSERT(unresudq || unresgdq);
2692                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2693                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2694                                 flags | blkflags | prjflags)))
2695                         return (error);
2696                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2697                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2698                                 blkflags);
2699         }
2700
2701         return (0);
2702 }
2703
2704 int
2705 xfs_qm_vop_rename_dqattach(
2706         xfs_inode_t     **i_tab)
2707 {
2708         xfs_inode_t     *ip;
2709         int             i;
2710         int             error;
2711
2712         ip = i_tab[0];
2713
2714         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2715                 return 0;
2716
2717         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2718                 error = xfs_qm_dqattach(ip, 0);
2719                 if (error)
2720                         return error;
2721         }
2722         for (i = 1; (i < 4 && i_tab[i]); i++) {
2723                 /*
2724                  * Watch out for duplicate entries in the table.
2725                  */
2726                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2727                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2728                                 error = xfs_qm_dqattach(ip, 0);
2729                                 if (error)
2730                                         return error;
2731                         }
2732                 }
2733         }
2734         return 0;
2735 }
2736
2737 void
2738 xfs_qm_vop_dqattach_and_dqmod_newinode(
2739         xfs_trans_t     *tp,
2740         xfs_inode_t     *ip,
2741         xfs_dquot_t     *udqp,
2742         xfs_dquot_t     *gdqp)
2743 {
2744         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2745                 return;
2746
2747         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2748         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2749
2750         if (udqp) {
2751                 xfs_dqlock(udqp);
2752                 XFS_DQHOLD(udqp);
2753                 xfs_dqunlock(udqp);
2754                 ASSERT(ip->i_udquot == NULL);
2755                 ip->i_udquot = udqp;
2756                 ASSERT(XFS_IS_UQUOTA_ON(tp->t_mountp));
2757                 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id));
2758                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2759         }
2760         if (gdqp) {
2761                 xfs_dqlock(gdqp);
2762                 XFS_DQHOLD(gdqp);
2763                 xfs_dqunlock(gdqp);
2764                 ASSERT(ip->i_gdquot == NULL);
2765                 ip->i_gdquot = gdqp;
2766                 ASSERT(XFS_IS_OQUOTA_ON(tp->t_mountp));
2767                 ASSERT((XFS_IS_GQUOTA_ON(tp->t_mountp) ?
2768                         ip->i_d.di_gid : ip->i_d.di_projid) ==
2769                                 be32_to_cpu(gdqp->q_core.d_id));
2770                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2771         }
2772 }
2773
2774 /* ------------- list stuff -----------------*/
2775 STATIC void
2776 xfs_qm_freelist_init(xfs_frlist_t *ql)
2777 {
2778         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2779         mutex_init(&ql->qh_lock);
2780         ql->qh_version = 0;
2781         ql->qh_nelems = 0;
2782 }
2783
2784 STATIC void
2785 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2786 {
2787         xfs_dquot_t     *dqp, *nextdqp;
2788
2789         mutex_lock(&ql->qh_lock);
2790         for (dqp = ql->qh_next;
2791              dqp != (xfs_dquot_t *)ql; ) {
2792                 xfs_dqlock(dqp);
2793                 nextdqp = dqp->dq_flnext;
2794 #ifdef QUOTADEBUG
2795                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2796 #endif
2797                 XQM_FREELIST_REMOVE(dqp);
2798                 xfs_dqunlock(dqp);
2799                 xfs_qm_dqdestroy(dqp);
2800                 dqp = nextdqp;
2801         }
2802         mutex_unlock(&ql->qh_lock);
2803         mutex_destroy(&ql->qh_lock);
2804
2805         ASSERT(ql->qh_nelems == 0);
2806 }
2807
2808 STATIC void
2809 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2810 {
2811         dq->dq_flnext = ql->qh_next;
2812         dq->dq_flprev = (xfs_dquot_t *)ql;
2813         ql->qh_next = dq;
2814         dq->dq_flnext->dq_flprev = dq;
2815         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2816         xfs_Gqm->qm_dqfreelist.qh_version++;
2817 }
2818
2819 void
2820 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2821 {
2822         xfs_dquot_t *next = dq->dq_flnext;
2823         xfs_dquot_t *prev = dq->dq_flprev;
2824
2825         next->dq_flprev = prev;
2826         prev->dq_flnext = next;
2827         dq->dq_flnext = dq->dq_flprev = dq;
2828         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2829         xfs_Gqm->qm_dqfreelist.qh_version++;
2830 }
2831
2832 void
2833 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2834 {
2835         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2836 }
2837
2838 STATIC int
2839 xfs_qm_dqhashlock_nowait(
2840         xfs_dquot_t *dqp)
2841 {
2842         int locked;
2843
2844         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2845         return locked;
2846 }
2847
2848 int
2849 xfs_qm_freelist_lock_nowait(
2850         xfs_qm_t *xqm)
2851 {
2852         int locked;
2853
2854         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2855         return locked;
2856 }
2857
2858 STATIC int
2859 xfs_qm_mplist_nowait(
2860         xfs_mount_t     *mp)
2861 {
2862         int locked;
2863
2864         ASSERT(mp->m_quotainfo);
2865         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2866         return locked;
2867 }