vserver 1.9.3
[linux-2.6.git] / fs / xfs / quota / xfs_qm.c
1 /*
2  * Copyright (c) 2000-2002 Silicon Graphics, Inc.  All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11  *
12  * Further, this software is distributed without any warranty that it is
13  * free of the rightful claim of any third person regarding infringement
14  * or the like.  Any license provided herein, whether implied or
15  * otherwise, applies only to this software file.  Patent licenses, if
16  * any, provided herein do not apply to combinations of this program with
17  * other software, or any other product whatsoever.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, write the Free Software Foundation, Inc., 59
21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22  *
23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24  * Mountain View, CA  94043, or:
25  *
26  * http://www.sgi.com
27  *
28  * For further information regarding this notice, see:
29  *
30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31  */
32
33 #include "xfs.h"
34 #include "xfs_fs.h"
35 #include "xfs_inum.h"
36 #include "xfs_log.h"
37 #include "xfs_clnt.h"
38 #include "xfs_trans.h"
39 #include "xfs_sb.h"
40 #include "xfs_ag.h"
41 #include "xfs_dir.h"
42 #include "xfs_dir2.h"
43 #include "xfs_alloc.h"
44 #include "xfs_dmapi.h"
45 #include "xfs_quota.h"
46 #include "xfs_mount.h"
47 #include "xfs_alloc_btree.h"
48 #include "xfs_bmap_btree.h"
49 #include "xfs_ialloc_btree.h"
50 #include "xfs_btree.h"
51 #include "xfs_ialloc.h"
52 #include "xfs_attr_sf.h"
53 #include "xfs_dir_sf.h"
54 #include "xfs_dir2_sf.h"
55 #include "xfs_dinode.h"
56 #include "xfs_inode.h"
57 #include "xfs_bmap.h"
58 #include "xfs_bit.h"
59 #include "xfs_rtalloc.h"
60 #include "xfs_error.h"
61 #include "xfs_itable.h"
62 #include "xfs_rw.h"
63 #include "xfs_acl.h"
64 #include "xfs_cap.h"
65 #include "xfs_mac.h"
66 #include "xfs_attr.h"
67 #include "xfs_buf_item.h"
68 #include "xfs_trans_space.h"
69 #include "xfs_utils.h"
70
71 #include "xfs_qm.h"
72
73 /*
74  * The global quota manager. There is only one of these for the entire
75  * system, _not_ one per file system. XQM keeps track of the overall
76  * quota functionality, including maintaining the freelist and hash
77  * tables of dquots.
78  */
79 mutex_t xfs_Gqm_lock;
80 struct xfs_qm   *xfs_Gqm;
81
82 kmem_zone_t     *qm_dqzone;
83 kmem_zone_t     *qm_dqtrxzone;
84 kmem_shaker_t   xfs_qm_shaker;
85
86 STATIC void     xfs_qm_list_init(xfs_dqlist_t *, char *, int);
87 STATIC void     xfs_qm_list_destroy(xfs_dqlist_t *);
88 STATIC int      xfs_qm_quotacheck(xfs_mount_t *);
89
90 STATIC int      xfs_qm_init_quotainos(xfs_mount_t *);
91 STATIC int      xfs_qm_shake(int, unsigned int);
92
93 #ifdef DEBUG
94 extern mutex_t  qcheck_lock;
95 #endif
96
97 #ifdef QUOTADEBUG
98 #define XQM_LIST_PRINT(l, NXT, title) \
99 { \
100         xfs_dquot_t     *dqp; int i = 0; \
101         cmn_err(CE_DEBUG, "%s (#%d)", title, (int) (l)->qh_nelems); \
102         for (dqp = (l)->qh_next; dqp != NULL; dqp = dqp->NXT) { \
103                 cmn_err(CE_DEBUG, "   %d.  \"%d (%s)\"   " \
104                                   "bcnt = %d, icnt = %d, refs = %d", \
105                         ++i, (int) INT_GET(dqp->q_core.d_id, ARCH_CONVERT), \
106                         DQFLAGTO_TYPESTR(dqp),       \
107                         (int) INT_GET(dqp->q_core.d_bcount, ARCH_CONVERT), \
108                         (int) INT_GET(dqp->q_core.d_icount, ARCH_CONVERT), \
109                         (int) dqp->q_nrefs);  } \
110 }
111 #else
112 #define XQM_LIST_PRINT(l, NXT, title) do { } while (0)
113 #endif
114
115 /*
116  * Initialize the XQM structure.
117  * Note that there is not one quota manager per file system.
118  */
119 STATIC struct xfs_qm *
120 xfs_Gqm_init(void)
121 {
122         xfs_qm_t                *xqm;
123         int                     hsize, i;
124
125         xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
126         ASSERT(xqm);
127
128         /*
129          * Initialize the dquot hash tables.
130          */
131         hsize = (DQUOT_HASH_HEURISTIC < XFS_QM_NCSIZE_THRESHOLD) ?
132                 XFS_QM_HASHSIZE_LOW : XFS_QM_HASHSIZE_HIGH;
133         xqm->qm_dqhashmask = hsize - 1;
134
135         xqm->qm_usr_dqhtable = (xfs_dqhash_t *)kmem_zalloc(hsize *
136                                                       sizeof(xfs_dqhash_t),
137                                                       KM_SLEEP);
138         xqm->qm_grp_dqhtable = (xfs_dqhash_t *)kmem_zalloc(hsize *
139                                                       sizeof(xfs_dqhash_t),
140                                                       KM_SLEEP);
141         ASSERT(xqm->qm_usr_dqhtable != NULL);
142         ASSERT(xqm->qm_grp_dqhtable != NULL);
143
144         for (i = 0; i < hsize; i++) {
145                 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i);
146                 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i);
147         }
148
149         /*
150          * Freelist of all dquots of all file systems
151          */
152         xfs_qm_freelist_init(&(xqm->qm_dqfreelist));
153
154         /*
155          * dquot zone. we register our own low-memory callback.
156          */
157         if (!qm_dqzone) {
158                 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t),
159                                                 "xfs_dquots");
160                 qm_dqzone = xqm->qm_dqzone;
161         } else
162                 xqm->qm_dqzone = qm_dqzone;
163
164         xfs_qm_shaker = kmem_shake_register(xfs_qm_shake);
165
166         /*
167          * The t_dqinfo portion of transactions.
168          */
169         if (!qm_dqtrxzone) {
170                 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t),
171                                                    "xfs_dqtrx");
172                 qm_dqtrxzone = xqm->qm_dqtrxzone;
173         } else
174                 xqm->qm_dqtrxzone = qm_dqtrxzone;
175
176         atomic_set(&xqm->qm_totaldquots, 0);
177         xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
178         xqm->qm_nrefs = 0;
179 #ifdef DEBUG
180         mutex_init(&qcheck_lock, MUTEX_DEFAULT, "qchk");
181 #endif
182         return xqm;
183 }
184
185 /*
186  * Destroy the global quota manager when its reference count goes to zero.
187  */
188 void
189 xfs_qm_destroy(
190         struct xfs_qm   *xqm)
191 {
192         int             hsize, i;
193
194         ASSERT(xqm != NULL);
195         ASSERT(xqm->qm_nrefs == 0);
196         kmem_shake_deregister(xfs_qm_shaker);
197         hsize = xqm->qm_dqhashmask + 1;
198         for (i = 0; i < hsize; i++) {
199                 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i]));
200                 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i]));
201         }
202         kmem_free(xqm->qm_usr_dqhtable, hsize * sizeof(xfs_dqhash_t));
203         kmem_free(xqm->qm_grp_dqhtable, hsize * sizeof(xfs_dqhash_t));
204         xqm->qm_usr_dqhtable = NULL;
205         xqm->qm_grp_dqhtable = NULL;
206         xqm->qm_dqhashmask = 0;
207         xfs_qm_freelist_destroy(&(xqm->qm_dqfreelist));
208 #ifdef DEBUG
209         mutex_destroy(&qcheck_lock);
210 #endif
211         kmem_free(xqm, sizeof(xfs_qm_t));
212 }
213
214 /*
215  * Called at mount time to let XQM know that another file system is
216  * starting quotas. This isn't crucial information as the individual mount
217  * structures are pretty independent, but it helps the XQM keep a
218  * global view of what's going on.
219  */
220 /* ARGSUSED */
221 STATIC int
222 xfs_qm_hold_quotafs_ref(
223         struct xfs_mount *mp)
224 {
225         /*
226          * Need to lock the xfs_Gqm structure for things like this. For example,
227          * the structure could disappear between the entry to this routine and
228          * a HOLD operation if not locked.
229          */
230         XFS_QM_LOCK(xfs_Gqm);
231
232         if (xfs_Gqm == NULL)
233                 xfs_Gqm = xfs_Gqm_init();
234         /*
235          * We can keep a list of all filesystems with quotas mounted for
236          * debugging and statistical purposes, but ...
237          * Just take a reference and get out.
238          */
239         XFS_QM_HOLD(xfs_Gqm);
240         XFS_QM_UNLOCK(xfs_Gqm);
241
242         return 0;
243 }
244
245
246 /*
247  * Release the reference that a filesystem took at mount time,
248  * so that we know when we need to destroy the entire quota manager.
249  */
250 /* ARGSUSED */
251 STATIC void
252 xfs_qm_rele_quotafs_ref(
253         struct xfs_mount *mp)
254 {
255         xfs_dquot_t     *dqp, *nextdqp;
256
257         ASSERT(xfs_Gqm);
258         ASSERT(xfs_Gqm->qm_nrefs > 0);
259
260         /*
261          * Go thru the freelist and destroy all inactive dquots.
262          */
263         xfs_qm_freelist_lock(xfs_Gqm);
264
265         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
266              dqp != (xfs_dquot_t *)&(xfs_Gqm->qm_dqfreelist); ) {
267                 xfs_dqlock(dqp);
268                 nextdqp = dqp->dq_flnext;
269                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
270                         ASSERT(dqp->q_mount == NULL);
271                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
272                         ASSERT(dqp->HL_PREVP == NULL);
273                         ASSERT(dqp->MPL_PREVP == NULL);
274                         XQM_FREELIST_REMOVE(dqp);
275                         xfs_dqunlock(dqp);
276                         xfs_qm_dqdestroy(dqp);
277                 } else {
278                         xfs_dqunlock(dqp);
279                 }
280                 dqp = nextdqp;
281         }
282         xfs_qm_freelist_unlock(xfs_Gqm);
283
284         /*
285          * Destroy the entire XQM. If somebody mounts with quotaon, this'll
286          * be restarted.
287          */
288         XFS_QM_LOCK(xfs_Gqm);
289         XFS_QM_RELE(xfs_Gqm);
290         if (xfs_Gqm->qm_nrefs == 0) {
291                 xfs_qm_destroy(xfs_Gqm);
292                 xfs_Gqm = NULL;
293         }
294         XFS_QM_UNLOCK(xfs_Gqm);
295 }
296
297 /*
298  * This is called at mount time from xfs_mountfs to initialize the quotainfo
299  * structure and start the global quotamanager (xfs_Gqm) if it hasn't done
300  * so already.  Note that the superblock has not been read in yet.
301  */
302 void
303 xfs_qm_mount_quotainit(
304         xfs_mount_t     *mp,
305         uint            flags)
306 {
307         /*
308          * User or group quotas has to be on.
309          */
310         ASSERT(flags & (XFSMNT_UQUOTA | XFSMNT_GQUOTA));
311
312         /*
313          * Initialize the flags in the mount structure. From this point
314          * onwards we look at m_qflags to figure out if quotas's ON/OFF, etc.
315          * Note that we enforce nothing if accounting is off.
316          * ie.  XFSMNT_*QUOTA must be ON for XFSMNT_*QUOTAENF.
317          * It isn't necessary to take the quotaoff lock to do this; this is
318          * called from mount.
319          */
320         if (flags & XFSMNT_UQUOTA) {
321                 mp->m_qflags |= (XFS_UQUOTA_ACCT | XFS_UQUOTA_ACTIVE);
322                 if (flags & XFSMNT_UQUOTAENF)
323                         mp->m_qflags |= XFS_UQUOTA_ENFD;
324         }
325         if (flags & XFSMNT_GQUOTA) {
326                 mp->m_qflags |= (XFS_GQUOTA_ACCT | XFS_GQUOTA_ACTIVE);
327                 if (flags & XFSMNT_GQUOTAENF)
328                         mp->m_qflags |= XFS_GQUOTA_ENFD;
329         }
330 }
331
332 /*
333  * Just destroy the quotainfo structure.
334  */
335 void
336 xfs_qm_unmount_quotadestroy(
337         xfs_mount_t     *mp)
338 {
339         if (mp->m_quotainfo)
340                 xfs_qm_destroy_quotainfo(mp);
341 }
342
343
344 /*
345  * This is called from xfs_mountfs to start quotas and initialize all
346  * necessary data structures like quotainfo.  This is also responsible for
347  * running a quotacheck as necessary.  We are guaranteed that the superblock
348  * is consistently read in at this point.
349  */
350 int
351 xfs_qm_mount_quotas(
352         xfs_mount_t     *mp)
353 {
354         unsigned long   s;
355         int             error = 0;
356         uint            sbf;
357
358         /*
359          * If a file system had quotas running earlier, but decided to
360          * mount without -o quota/uquota/gquota options, revoke the
361          * quotachecked license, and bail out.
362          */
363         if (! XFS_IS_QUOTA_ON(mp) &&
364             (mp->m_sb.sb_qflags & (XFS_UQUOTA_ACCT|XFS_GQUOTA_ACCT))) {
365                 mp->m_qflags = 0;
366                 goto write_changes;
367         }
368
369         /*
370          * If quotas on realtime volumes is not supported, we disable
371          * quotas immediately.
372          */
373         if (mp->m_sb.sb_rextents) {
374                 cmn_err(CE_NOTE,
375                         "Cannot turn on quotas for realtime filesystem %s",
376                         mp->m_fsname);
377                 mp->m_qflags = 0;
378                 goto write_changes;
379         }
380
381 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
382         cmn_err(CE_NOTE, "Attempting to turn on disk quotas.");
383 #endif
384
385         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
386         /*
387          * Allocate the quotainfo structure inside the mount struct, and
388          * create quotainode(s), and change/rev superblock if necessary.
389          */
390         if ((error = xfs_qm_init_quotainfo(mp))) {
391                 /*
392                  * We must turn off quotas.
393                  */
394                 ASSERT(mp->m_quotainfo == NULL);
395                 mp->m_qflags = 0;
396                 goto write_changes;
397         }
398         /*
399          * If any of the quotas are not consistent, do a quotacheck.
400          */
401         if (XFS_QM_NEED_QUOTACHECK(mp)) {
402 #ifdef DEBUG
403                 cmn_err(CE_NOTE, "Doing a quotacheck. Please wait.");
404 #endif
405                 if ((error = xfs_qm_quotacheck(mp))) {
406                         cmn_err(CE_WARN, "Quotacheck unsuccessful (Error %d): "
407                                 "Disabling quotas.",
408                                 error);
409                         /*
410                          * We must turn off quotas.
411                          */
412                         ASSERT(mp->m_quotainfo != NULL);
413                         ASSERT(xfs_Gqm != NULL);
414                         xfs_qm_destroy_quotainfo(mp);
415                         mp->m_qflags = 0;
416                         goto write_changes;
417                 }
418 #ifdef DEBUG
419                 cmn_err(CE_NOTE, "Done quotacheck.");
420 #endif
421         }
422  write_changes:
423         /*
424          * We actually don't have to acquire the SB_LOCK at all.
425          * This can only be called from mount, and that's single threaded. XXX
426          */
427         s = XFS_SB_LOCK(mp);
428         sbf = mp->m_sb.sb_qflags;
429         mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL;
430         XFS_SB_UNLOCK(mp, s);
431
432         if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) {
433                 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) {
434                         /*
435                          * We could only have been turning quotas off.
436                          * We aren't in very good shape actually because
437                          * the incore structures are convinced that quotas are
438                          * off, but the on disk superblock doesn't know that !
439                          */
440                         ASSERT(!(XFS_IS_QUOTA_RUNNING(mp)));
441                         xfs_fs_cmn_err(CE_ALERT, mp,
442                                 "XFS mount_quotas: Superblock update failed!");
443                 }
444         }
445
446         if (error) {
447                 xfs_fs_cmn_err(CE_WARN, mp,
448                         "Failed to initialize disk quotas.");
449         }
450         return XFS_ERROR(error);
451 }
452
453 /*
454  * Called from the vfsops layer.
455  */
456 int
457 xfs_qm_unmount_quotas(
458         xfs_mount_t     *mp)
459 {
460         xfs_inode_t     *uqp, *gqp;
461         int             error = 0;
462
463         /*
464          * Release the dquots that root inode, et al might be holding,
465          * before we flush quotas and blow away the quotainfo structure.
466          */
467         ASSERT(mp->m_rootip);
468         xfs_qm_dqdetach(mp->m_rootip);
469         if (mp->m_rbmip)
470                 xfs_qm_dqdetach(mp->m_rbmip);
471         if (mp->m_rsumip)
472                 xfs_qm_dqdetach(mp->m_rsumip);
473
474         /*
475          * Flush out the quota inodes.
476          */
477         uqp = gqp = NULL;
478         if (mp->m_quotainfo) {
479                 if ((uqp = mp->m_quotainfo->qi_uquotaip) != NULL) {
480                         xfs_ilock(uqp, XFS_ILOCK_EXCL);
481                         xfs_iflock(uqp);
482                         error = xfs_iflush(uqp, XFS_IFLUSH_SYNC);
483                         xfs_iunlock(uqp, XFS_ILOCK_EXCL);
484                         if (unlikely(error == EFSCORRUPTED)) {
485                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(1)",
486                                                  XFS_ERRLEVEL_LOW, mp);
487                                 goto out;
488                         }
489                 }
490                 if ((gqp = mp->m_quotainfo->qi_gquotaip) != NULL) {
491                         xfs_ilock(gqp, XFS_ILOCK_EXCL);
492                         xfs_iflock(gqp);
493                         error = xfs_iflush(gqp, XFS_IFLUSH_SYNC);
494                         xfs_iunlock(gqp, XFS_ILOCK_EXCL);
495                         if (unlikely(error == EFSCORRUPTED)) {
496                                 XFS_ERROR_REPORT("xfs_qm_unmount_quotas(2)",
497                                                  XFS_ERRLEVEL_LOW, mp);
498                                 goto out;
499                         }
500                 }
501         }
502         if (uqp) {
503                  XFS_PURGE_INODE(uqp);
504                  mp->m_quotainfo->qi_uquotaip = NULL;
505         }
506         if (gqp) {
507                 XFS_PURGE_INODE(gqp);
508                 mp->m_quotainfo->qi_gquotaip = NULL;
509         }
510 out:
511         return XFS_ERROR(error);
512 }
513
514 /*
515  * Flush all dquots of the given file system to disk. The dquots are
516  * _not_ purged from memory here, just their data written to disk.
517  */
518 int
519 xfs_qm_dqflush_all(
520         xfs_mount_t     *mp,
521         int             flags)
522 {
523         int             recl;
524         xfs_dquot_t     *dqp;
525         int             niters;
526         int             error;
527
528         if (mp->m_quotainfo == NULL)
529                 return (0);
530         niters = 0;
531 again:
532         xfs_qm_mplist_lock(mp);
533         FOREACH_DQUOT_IN_MP(dqp, mp) {
534                 xfs_dqlock(dqp);
535                 if (! XFS_DQ_IS_DIRTY(dqp)) {
536                         xfs_dqunlock(dqp);
537                         continue;
538                 }
539                 xfs_dqtrace_entry(dqp, "FLUSHALL: DQDIRTY");
540                 /* XXX a sentinel would be better */
541                 recl = XFS_QI_MPLRECLAIMS(mp);
542                 if (! xfs_qm_dqflock_nowait(dqp)) {
543                         /*
544                          * If we can't grab the flush lock then check
545                          * to see if the dquot has been flushed delayed
546                          * write.  If so, grab its buffer and send it
547                          * out immediately.  We'll be able to acquire
548                          * the flush lock when the I/O completes.
549                          */
550                         xfs_qm_dqflock_pushbuf_wait(dqp);
551                 }
552                 /*
553                  * Let go of the mplist lock. We don't want to hold it
554                  * across a disk write.
555                  */
556                 xfs_qm_mplist_unlock(mp);
557                 error = xfs_qm_dqflush(dqp, flags);
558                 xfs_dqunlock(dqp);
559                 if (error)
560                         return (error);
561
562                 xfs_qm_mplist_lock(mp);
563                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
564                         xfs_qm_mplist_unlock(mp);
565                         /* XXX restart limit */
566                         goto again;
567                 }
568         }
569
570         xfs_qm_mplist_unlock(mp);
571         /* return ! busy */
572         return (0);
573 }
574 /*
575  * Release the group dquot pointers the user dquots may be
576  * carrying around as a hint. mplist is locked on entry and exit.
577  */
578 STATIC void
579 xfs_qm_detach_gdquots(
580         xfs_mount_t     *mp)
581 {
582         xfs_dquot_t     *dqp, *gdqp;
583         int             nrecl;
584
585  again:
586         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
587         dqp = XFS_QI_MPLNEXT(mp);
588         while (dqp) {
589                 xfs_dqlock(dqp);
590                 if ((gdqp = dqp->q_gdquot)) {
591                         xfs_dqlock(gdqp);
592                         dqp->q_gdquot = NULL;
593                 }
594                 xfs_dqunlock(dqp);
595
596                 if (gdqp) {
597                         /*
598                          * Can't hold the mplist lock across a dqput.
599                          * XXXmust convert to marker based iterations here.
600                          */
601                         nrecl = XFS_QI_MPLRECLAIMS(mp);
602                         xfs_qm_mplist_unlock(mp);
603                         xfs_qm_dqput(gdqp);
604
605                         xfs_qm_mplist_lock(mp);
606                         if (nrecl != XFS_QI_MPLRECLAIMS(mp))
607                                 goto again;
608                 }
609                 dqp = dqp->MPL_NEXT;
610         }
611 }
612
613 /*
614  * Go through all the incore dquots of this file system and take them
615  * off the mplist and hashlist, if the dquot type matches the dqtype
616  * parameter. This is used when turning off quota accounting for
617  * users and/or groups, as well as when the filesystem is unmounting.
618  */
619 STATIC int
620 xfs_qm_dqpurge_int(
621         xfs_mount_t     *mp,
622         uint            flags) /* QUOTAOFF/UMOUNTING/UQUOTA/GQUOTA */
623 {
624         xfs_dquot_t     *dqp;
625         uint            dqtype;
626         int             nrecl;
627         xfs_dquot_t     *nextdqp;
628         int             nmisses;
629
630         if (mp->m_quotainfo == NULL)
631                 return (0);
632
633         dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
634         dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;
635
636         xfs_qm_mplist_lock(mp);
637
638         /*
639          * In the first pass through all incore dquots of this filesystem,
640          * we release the group dquot pointers the user dquots may be
641          * carrying around as a hint. We need to do this irrespective of
642          * what's being turned off.
643          */
644         xfs_qm_detach_gdquots(mp);
645
646       again:
647         nmisses = 0;
648         ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
649         /*
650          * Try to get rid of all of the unwanted dquots. The idea is to
651          * get them off mplist and hashlist, but leave them on freelist.
652          */
653         dqp = XFS_QI_MPLNEXT(mp);
654         while (dqp) {
655                 /*
656                  * It's OK to look at the type without taking dqlock here.
657                  * We're holding the mplist lock here, and that's needed for
658                  * a dqreclaim.
659                  */
660                 if ((dqp->dq_flags & dqtype) == 0) {
661                         dqp = dqp->MPL_NEXT;
662                         continue;
663                 }
664
665                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
666                         nrecl = XFS_QI_MPLRECLAIMS(mp);
667                         xfs_qm_mplist_unlock(mp);
668                         XFS_DQ_HASH_LOCK(dqp->q_hash);
669                         xfs_qm_mplist_lock(mp);
670
671                         /*
672                          * XXXTheoretically, we can get into a very long
673                          * ping pong game here.
674                          * No one can be adding dquots to the mplist at
675                          * this point, but somebody might be taking things off.
676                          */
677                         if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
678                                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
679                                 goto again;
680                         }
681                 }
682
683                 /*
684                  * Take the dquot off the mplist and hashlist. It may remain on
685                  * freelist in INACTIVE state.
686                  */
687                 nextdqp = dqp->MPL_NEXT;
688                 nmisses += xfs_qm_dqpurge(dqp, flags);
689                 dqp = nextdqp;
690         }
691         xfs_qm_mplist_unlock(mp);
692         return nmisses;
693 }
694
695 int
696 xfs_qm_dqpurge_all(
697         xfs_mount_t     *mp,
698         uint            flags)
699 {
700         int             ndquots;
701
702         /*
703          * Purge the dquot cache.
704          * None of the dquots should really be busy at this point.
705          */
706         if (mp->m_quotainfo) {
707                 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
708                         delay(ndquots * 10);
709                 }
710         }
711         return 0;
712 }
713
714 STATIC int
715 xfs_qm_dqattach_one(
716         xfs_inode_t     *ip,
717         xfs_dqid_t      id,
718         uint            type,
719         uint            doalloc,
720         uint            dolock,
721         xfs_dquot_t     *udqhint, /* hint */
722         xfs_dquot_t     **IO_idqpp)
723 {
724         xfs_dquot_t     *dqp;
725         int             error;
726
727         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
728         error = 0;
729         /*
730          * See if we already have it in the inode itself. IO_idqpp is
731          * &i_udquot or &i_gdquot. This made the code look weird, but
732          * made the logic a lot simpler.
733          */
734         if ((dqp = *IO_idqpp)) {
735                 if (dolock)
736                         xfs_dqlock(dqp);
737                 xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
738                 goto done;
739         }
740
741         /*
742          * udqhint is the i_udquot field in inode, and is non-NULL only
743          * when the type arg is XFS_DQ_GROUP. Its purpose is to save a
744          * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
745          * the user dquot.
746          */
747         ASSERT(!udqhint || type == XFS_DQ_GROUP);
748         if (udqhint && !dolock)
749                 xfs_dqlock(udqhint);
750
751         /*
752          * No need to take dqlock to look at the id.
753          * The ID can't change until it gets reclaimed, and it won't
754          * be reclaimed as long as we have a ref from inode and we hold
755          * the ilock.
756          */
757         if (udqhint &&
758             (dqp = udqhint->q_gdquot) &&
759             (INT_GET(dqp->q_core.d_id, ARCH_CONVERT) == id)) {
760                 ASSERT(XFS_DQ_IS_LOCKED(udqhint));
761                 xfs_dqlock(dqp);
762                 XFS_DQHOLD(dqp);
763                 ASSERT(*IO_idqpp == NULL);
764                 *IO_idqpp = dqp;
765                 if (!dolock) {
766                         xfs_dqunlock(dqp);
767                         xfs_dqunlock(udqhint);
768                 }
769                 goto done;
770         }
771         /*
772          * We can't hold a dquot lock when we call the dqget code.
773          * We'll deadlock in no time, because of (not conforming to)
774          * lock ordering - the inodelock comes before any dquot lock,
775          * and we may drop and reacquire the ilock in xfs_qm_dqget().
776          */
777         if (udqhint)
778                 xfs_dqunlock(udqhint);
779         /*
780          * Find the dquot from somewhere. This bumps the
781          * reference count of dquot and returns it locked.
782          * This can return ENOENT if dquot didn't exist on
783          * disk and we didn't ask it to allocate;
784          * ESRCH if quotas got turned off suddenly.
785          */
786         if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
787                                  doalloc|XFS_QMOPT_DOWARN, &dqp))) {
788                 if (udqhint && dolock)
789                         xfs_dqlock(udqhint);
790                 goto done;
791         }
792
793         xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
794         /*
795          * dqget may have dropped and re-acquired the ilock, but it guarantees
796          * that the dquot returned is the one that should go in the inode.
797          */
798         *IO_idqpp = dqp;
799         ASSERT(dqp);
800         ASSERT(XFS_DQ_IS_LOCKED(dqp));
801         if (! dolock) {
802                 xfs_dqunlock(dqp);
803                 goto done;
804         }
805         if (! udqhint)
806                 goto done;
807
808         ASSERT(udqhint);
809         ASSERT(dolock);
810         ASSERT(XFS_DQ_IS_LOCKED(dqp));
811         if (! xfs_qm_dqlock_nowait(udqhint)) {
812                 xfs_dqunlock(dqp);
813                 xfs_dqlock(udqhint);
814                 xfs_dqlock(dqp);
815         }
816       done:
817 #ifdef QUOTADEBUG
818         if (udqhint) {
819                 if (dolock)
820                         ASSERT(XFS_DQ_IS_LOCKED(udqhint));
821         }
822         if (! error) {
823                 if (dolock)
824                         ASSERT(XFS_DQ_IS_LOCKED(dqp));
825         }
826 #endif
827         return (error);
828 }
829
830
831 /*
832  * Given a udquot and gdquot, attach a ptr to the group dquot in the
833  * udquot as a hint for future lookups. The idea sounds simple, but the
834  * execution isn't, because the udquot might have a group dquot attached
835  * already and getting rid of that gets us into lock ordering contraints.
836  * The process is complicated more by the fact that the dquots may or may not
837  * be locked on entry.
838  */
839 STATIC void
840 xfs_qm_dqattach_grouphint(
841         xfs_dquot_t     *udq,
842         xfs_dquot_t     *gdq,
843         uint            locked)
844 {
845         xfs_dquot_t     *tmp;
846
847 #ifdef QUOTADEBUG
848         if (locked) {
849                 ASSERT(XFS_DQ_IS_LOCKED(udq));
850                 ASSERT(XFS_DQ_IS_LOCKED(gdq));
851         }
852 #endif
853         if (! locked)
854                 xfs_dqlock(udq);
855
856         if ((tmp = udq->q_gdquot)) {
857                 if (tmp == gdq) {
858                         if (! locked)
859                                 xfs_dqunlock(udq);
860                         return;
861                 }
862
863                 udq->q_gdquot = NULL;
864                 /*
865                  * We can't keep any dqlocks when calling dqrele,
866                  * because the freelist lock comes before dqlocks.
867                  */
868                 xfs_dqunlock(udq);
869                 if (locked)
870                         xfs_dqunlock(gdq);
871                 /*
872                  * we took a hard reference once upon a time in dqget,
873                  * so give it back when the udquot no longer points at it
874                  * dqput() does the unlocking of the dquot.
875                  */
876                 xfs_qm_dqrele(tmp);
877
878                 xfs_dqlock(udq);
879                 xfs_dqlock(gdq);
880
881         } else {
882                 ASSERT(XFS_DQ_IS_LOCKED(udq));
883                 if (! locked) {
884                         xfs_dqlock(gdq);
885                 }
886         }
887
888         ASSERT(XFS_DQ_IS_LOCKED(udq));
889         ASSERT(XFS_DQ_IS_LOCKED(gdq));
890         /*
891          * Somebody could have attached a gdquot here,
892          * when we dropped the uqlock. If so, just do nothing.
893          */
894         if (udq->q_gdquot == NULL) {
895                 XFS_DQHOLD(gdq);
896                 udq->q_gdquot = gdq;
897         }
898         if (! locked) {
899                 xfs_dqunlock(gdq);
900                 xfs_dqunlock(udq);
901         }
902 }
903
904
905 /*
906  * Given a locked inode, attach dquot(s) to it, taking UQUOTAON / GQUOTAON
907  * in to account.
908  * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
909  * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
910  * much made this code a complete mess, but it has been pretty useful.
911  * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
912  * Inode may get unlocked and relocked in here, and the caller must deal with
913  * the consequences.
914  */
915 int
916 xfs_qm_dqattach(
917         xfs_inode_t     *ip,
918         uint            flags)
919 {
920         xfs_mount_t     *mp = ip->i_mount;
921         uint            nquotas = 0;
922         int             error = 0;
923
924         if ((! XFS_IS_QUOTA_ON(mp)) ||
925             (! XFS_NOT_DQATTACHED(mp, ip)) ||
926             (ip->i_ino == mp->m_sb.sb_uquotino) ||
927             (ip->i_ino == mp->m_sb.sb_gquotino))
928                 return (0);
929
930         ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
931                XFS_ISLOCKED_INODE_EXCL(ip));
932
933         if (! (flags & XFS_QMOPT_ILOCKED))
934                 xfs_ilock(ip, XFS_ILOCK_EXCL);
935
936         if (XFS_IS_UQUOTA_ON(mp)) {
937                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
938                                                 flags & XFS_QMOPT_DQALLOC,
939                                                 flags & XFS_QMOPT_DQLOCK,
940                                                 NULL, &ip->i_udquot);
941                 if (error)
942                         goto done;
943                 nquotas++;
944         }
945         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
946         if (XFS_IS_GQUOTA_ON(mp)) {
947                 error = xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
948                                                 flags & XFS_QMOPT_DQALLOC,
949                                                 flags & XFS_QMOPT_DQLOCK,
950                                                 ip->i_udquot, &ip->i_gdquot);
951                 /*
952                  * Don't worry about the udquot that we may have
953                  * attached above. It'll get detached, if not already.
954                  */
955                 if (error)
956                         goto done;
957                 nquotas++;
958         }
959
960         /*
961          * Attach this group quota to the user quota as a hint.
962          * This WON'T, in general, result in a thrash.
963          */
964         if (nquotas == 2) {
965                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
966                 ASSERT(ip->i_udquot);
967                 ASSERT(ip->i_gdquot);
968
969                 /*
970                  * We may or may not have the i_udquot locked at this point,
971                  * but this check is OK since we don't depend on the i_gdquot to
972                  * be accurate 100% all the time. It is just a hint, and this
973                  * will succeed in general.
974                  */
975                 if (ip->i_udquot->q_gdquot == ip->i_gdquot)
976                         goto done;
977                 /*
978                  * Attach i_gdquot to the gdquot hint inside the i_udquot.
979                  */
980                 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
981                                          flags & XFS_QMOPT_DQLOCK);
982         }
983
984       done:
985
986 #ifdef QUOTADEBUG
987         if (! error) {
988                 if (ip->i_udquot) {
989                         if (flags & XFS_QMOPT_DQLOCK)
990                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
991                 }
992                 if (ip->i_gdquot) {
993                         if (flags & XFS_QMOPT_DQLOCK)
994                                 ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
995                 }
996                 if (XFS_IS_UQUOTA_ON(mp))
997                         ASSERT(ip->i_udquot);
998                 if (XFS_IS_GQUOTA_ON(mp))
999                         ASSERT(ip->i_gdquot);
1000         }
1001 #endif
1002
1003         if (! (flags & XFS_QMOPT_ILOCKED))
1004                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1005
1006 #ifdef QUOTADEBUG
1007         else
1008                 ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1009 #endif
1010         return (error);
1011 }
1012
1013 /*
1014  * Release dquots (and their references) if any.
1015  * The inode should be locked EXCL except when this's called by
1016  * xfs_ireclaim.
1017  */
1018 void
1019 xfs_qm_dqdetach(
1020         xfs_inode_t     *ip)
1021 {
1022         if (!(ip->i_udquot || ip->i_gdquot))
1023                 return;
1024
1025         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
1026         ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
1027         if (ip->i_udquot)
1028                 xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
1029         if (ip->i_udquot) {
1030                 xfs_qm_dqrele(ip->i_udquot);
1031                 ip->i_udquot = NULL;
1032         }
1033         if (ip->i_gdquot) {
1034                 xfs_qm_dqrele(ip->i_gdquot);
1035                 ip->i_gdquot = NULL;
1036         }
1037 }
1038
1039 /*
1040  * This is called by VFS_SYNC and flags arg determines the caller,
1041  * and its motives, as done in xfs_sync.
1042  *
1043  * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
1044  * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
1045  * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
1046  */
1047
1048 int
1049 xfs_qm_sync(
1050         xfs_mount_t     *mp,
1051         short           flags)
1052 {
1053         int             recl, restarts;
1054         xfs_dquot_t     *dqp;
1055         uint            flush_flags;
1056         boolean_t       nowait;
1057         int             error;
1058
1059         restarts = 0;
1060         /*
1061          * We won't block unless we are asked to.
1062          */
1063         nowait = (boolean_t)(flags & SYNC_BDFLUSH || (flags & SYNC_WAIT) == 0);
1064
1065   again:
1066         xfs_qm_mplist_lock(mp);
1067         /*
1068          * dqpurge_all() also takes the mplist lock and iterate thru all dquots
1069          * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared
1070          * when we have the mplist lock, we know that dquots will be consistent
1071          * as long as we have it locked.
1072          */
1073         if (! XFS_IS_QUOTA_ON(mp)) {
1074                 xfs_qm_mplist_unlock(mp);
1075                 return (0);
1076         }
1077         FOREACH_DQUOT_IN_MP(dqp, mp) {
1078                 /*
1079                  * If this is vfs_sync calling, then skip the dquots that
1080                  * don't 'seem' to be dirty. ie. don't acquire dqlock.
1081                  * This is very similar to what xfs_sync does with inodes.
1082                  */
1083                 if (flags & SYNC_BDFLUSH) {
1084                         if (! XFS_DQ_IS_DIRTY(dqp))
1085                                 continue;
1086                 }
1087
1088                 if (nowait) {
1089                         /*
1090                          * Try to acquire the dquot lock. We are NOT out of
1091                          * lock order, but we just don't want to wait for this
1092                          * lock, unless somebody wanted us to.
1093                          */
1094                         if (! xfs_qm_dqlock_nowait(dqp))
1095                                 continue;
1096                 } else {
1097                         xfs_dqlock(dqp);
1098                 }
1099
1100                 /*
1101                  * Now, find out for sure if this dquot is dirty or not.
1102                  */
1103                 if (! XFS_DQ_IS_DIRTY(dqp)) {
1104                         xfs_dqunlock(dqp);
1105                         continue;
1106                 }
1107
1108                 /* XXX a sentinel would be better */
1109                 recl = XFS_QI_MPLRECLAIMS(mp);
1110                 if (! xfs_qm_dqflock_nowait(dqp)) {
1111                         if (nowait) {
1112                                 xfs_dqunlock(dqp);
1113                                 continue;
1114                         }
1115                         /*
1116                          * If we can't grab the flush lock then if the caller
1117                          * really wanted us to give this our best shot,
1118                          * see if we can give a push to the buffer before we wait
1119                          * on the flush lock. At this point, we know that
1120                          * eventhough the dquot is being flushed,
1121                          * it has (new) dirty data.
1122                          */
1123                         xfs_qm_dqflock_pushbuf_wait(dqp);
1124                 }
1125                 /*
1126                  * Let go of the mplist lock. We don't want to hold it
1127                  * across a disk write
1128                  */
1129                 flush_flags = (nowait) ? XFS_QMOPT_DELWRI : XFS_QMOPT_SYNC;
1130                 xfs_qm_mplist_unlock(mp);
1131                 xfs_dqtrace_entry(dqp, "XQM_SYNC: DQFLUSH");
1132                 error = xfs_qm_dqflush(dqp, flush_flags);
1133                 xfs_dqunlock(dqp);
1134                 if (error && XFS_FORCED_SHUTDOWN(mp))
1135                         return(0);      /* Need to prevent umount failure */
1136                 else if (error)
1137                         return (error);
1138
1139                 xfs_qm_mplist_lock(mp);
1140                 if (recl != XFS_QI_MPLRECLAIMS(mp)) {
1141                         if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS)
1142                                 break;
1143
1144                         xfs_qm_mplist_unlock(mp);
1145                         goto again;
1146                 }
1147         }
1148
1149         xfs_qm_mplist_unlock(mp);
1150         return (0);
1151 }
1152
1153
1154 /*
1155  * This initializes all the quota information that's kept in the
1156  * mount structure
1157  */
1158 int
1159 xfs_qm_init_quotainfo(
1160         xfs_mount_t     *mp)
1161 {
1162         xfs_quotainfo_t *qinf;
1163         int             error;
1164         xfs_dquot_t     *dqp;
1165
1166         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1167
1168         /*
1169          * Tell XQM that we exist as soon as possible.
1170          */
1171         if ((error = xfs_qm_hold_quotafs_ref(mp))) {
1172                 return (error);
1173         }
1174
1175         qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP);
1176
1177         /*
1178          * See if quotainodes are setup, and if not, allocate them,
1179          * and change the superblock accordingly.
1180          */
1181         if ((error = xfs_qm_init_quotainos(mp))) {
1182                 kmem_free(qinf, sizeof(xfs_quotainfo_t));
1183                 mp->m_quotainfo = NULL;
1184                 return (error);
1185         }
1186
1187         spinlock_init(&qinf->qi_pinlock, "xfs_qinf_pin");
1188         xfs_qm_list_init(&qinf->qi_dqlist, "mpdqlist", 0);
1189         qinf->qi_dqreclaims = 0;
1190
1191         /* mutex used to serialize quotaoffs */
1192         mutex_init(&qinf->qi_quotaofflock, MUTEX_DEFAULT, "qoff");
1193
1194         /* Precalc some constants */
1195         qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1196         ASSERT(qinf->qi_dqchunklen);
1197         qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen);
1198         do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t));
1199
1200         mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD);
1201
1202         /*
1203          * We try to get the limits from the superuser's limits fields.
1204          * This is quite hacky, but it is standard quota practice.
1205          * We look at the USR dquot with id == 0 first, but if user quotas
1206          * are not enabled we goto the GRP dquot with id == 0.
1207          * We don't really care to keep separate default limits for user
1208          * and group quotas, at least not at this point.
1209          */
1210         error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0,
1211                              (XFS_IS_UQUOTA_RUNNING(mp)) ?
1212                              XFS_DQ_USER : XFS_DQ_GROUP,
1213                              XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN,
1214                              &dqp);
1215         if (! error) {
1216                 xfs_disk_dquot_t        *ddqp = &dqp->q_core;
1217
1218                 /*
1219                  * The warnings and timers set the grace period given to
1220                  * a user or group before he or she can not perform any
1221                  * more writing. If it is zero, a default is used.
1222                  */
1223                 qinf->qi_btimelimit =
1224                                 INT_GET(ddqp->d_btimer, ARCH_CONVERT) ?
1225                                 INT_GET(ddqp->d_btimer, ARCH_CONVERT) :
1226                                 XFS_QM_BTIMELIMIT;
1227                 qinf->qi_itimelimit =
1228                                 INT_GET(ddqp->d_itimer, ARCH_CONVERT) ?
1229                                 INT_GET(ddqp->d_itimer, ARCH_CONVERT) :
1230                                 XFS_QM_ITIMELIMIT;
1231                 qinf->qi_rtbtimelimit =
1232                                 INT_GET(ddqp->d_rtbtimer, ARCH_CONVERT) ?
1233                                 INT_GET(ddqp->d_rtbtimer, ARCH_CONVERT) :
1234                                 XFS_QM_RTBTIMELIMIT;
1235                 qinf->qi_bwarnlimit =
1236                                 INT_GET(ddqp->d_bwarns, ARCH_CONVERT) ?
1237                                 INT_GET(ddqp->d_bwarns, ARCH_CONVERT) :
1238                                 XFS_QM_BWARNLIMIT;
1239                 qinf->qi_iwarnlimit =
1240                                 INT_GET(ddqp->d_iwarns, ARCH_CONVERT) ?
1241                                 INT_GET(ddqp->d_iwarns, ARCH_CONVERT) :
1242                                 XFS_QM_IWARNLIMIT;
1243                 qinf->qi_bhardlimit =
1244                                 INT_GET(ddqp->d_blk_hardlimit, ARCH_CONVERT);
1245                 qinf->qi_bsoftlimit =
1246                                 INT_GET(ddqp->d_blk_softlimit, ARCH_CONVERT);
1247                 qinf->qi_ihardlimit =
1248                                 INT_GET(ddqp->d_ino_hardlimit, ARCH_CONVERT);
1249                 qinf->qi_isoftlimit =
1250                                 INT_GET(ddqp->d_ino_softlimit, ARCH_CONVERT);
1251                 qinf->qi_rtbhardlimit =
1252                                 INT_GET(ddqp->d_rtb_hardlimit, ARCH_CONVERT);
1253                 qinf->qi_rtbsoftlimit =
1254                                 INT_GET(ddqp->d_rtb_softlimit, ARCH_CONVERT);
1255  
1256                 /*
1257                  * We sent the XFS_QMOPT_DQSUSER flag to dqget because
1258                  * we don't want this dquot cached. We haven't done a
1259                  * quotacheck yet, and quotacheck doesn't like incore dquots.
1260                  */
1261                 xfs_qm_dqdestroy(dqp);
1262         } else {
1263                 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT;
1264                 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT;
1265                 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT;
1266                 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT;
1267                 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT;
1268         }
1269
1270         return (0);
1271 }
1272
1273
1274 /*
1275  * Gets called when unmounting a filesystem or when all quotas get
1276  * turned off.
1277  * This purges the quota inodes, destroys locks and frees itself.
1278  */
1279 void
1280 xfs_qm_destroy_quotainfo(
1281         xfs_mount_t     *mp)
1282 {
1283         xfs_quotainfo_t *qi;
1284
1285         qi = mp->m_quotainfo;
1286         ASSERT(qi != NULL);
1287         ASSERT(xfs_Gqm != NULL);
1288
1289         /*
1290          * Release the reference that XQM kept, so that we know
1291          * when the XQM structure should be freed. We cannot assume
1292          * that xfs_Gqm is non-null after this point.
1293          */
1294         xfs_qm_rele_quotafs_ref(mp);
1295
1296         spinlock_destroy(&qi->qi_pinlock);
1297         xfs_qm_list_destroy(&qi->qi_dqlist);
1298
1299         if (qi->qi_uquotaip) {
1300                 XFS_PURGE_INODE(qi->qi_uquotaip);
1301                 qi->qi_uquotaip = NULL; /* paranoia */
1302         }
1303         if (qi->qi_gquotaip) {
1304                 XFS_PURGE_INODE(qi->qi_gquotaip);
1305                 qi->qi_gquotaip = NULL;
1306         }
1307         mutex_destroy(&qi->qi_quotaofflock);
1308         kmem_free(qi, sizeof(xfs_quotainfo_t));
1309         mp->m_quotainfo = NULL;
1310 }
1311
1312
1313
1314 /* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */
1315
1316 /* ARGSUSED */
1317 STATIC void
1318 xfs_qm_list_init(
1319         xfs_dqlist_t    *list,
1320         char            *str,
1321         int             n)
1322 {
1323         mutex_init(&list->qh_lock, MUTEX_DEFAULT, str);
1324         list->qh_next = NULL;
1325         list->qh_version = 0;
1326         list->qh_nelems = 0;
1327 }
1328
1329 STATIC void
1330 xfs_qm_list_destroy(
1331         xfs_dqlist_t    *list)
1332 {
1333         mutex_destroy(&(list->qh_lock));
1334 }
1335
1336
1337 /*
1338  * Stripped down version of dqattach. This doesn't attach, or even look at the
1339  * dquots attached to the inode. The rationale is that there won't be any
1340  * attached at the time this is called from quotacheck.
1341  */
1342 STATIC int
1343 xfs_qm_dqget_noattach(
1344         xfs_inode_t     *ip,
1345         xfs_dquot_t     **O_udqpp,
1346         xfs_dquot_t     **O_gdqpp)
1347 {
1348         int             error;
1349         xfs_mount_t     *mp;
1350         xfs_dquot_t     *udqp, *gdqp;
1351
1352         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
1353         mp = ip->i_mount;
1354         udqp = NULL;
1355         gdqp = NULL;
1356
1357         if (XFS_IS_UQUOTA_ON(mp)) {
1358                 ASSERT(ip->i_udquot == NULL);
1359                 /*
1360                  * We want the dquot allocated if it doesn't exist.
1361                  */
1362                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER,
1363                                          XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN,
1364                                          &udqp))) {
1365                         /*
1366                          * Shouldn't be able to turn off quotas here.
1367                          */
1368                         ASSERT(error != ESRCH);
1369                         ASSERT(error != ENOENT);
1370                         return (error);
1371                 }
1372                 ASSERT(udqp);
1373         }
1374
1375         if (XFS_IS_GQUOTA_ON(mp)) {
1376                 ASSERT(ip->i_gdquot == NULL);
1377                 if (udqp)
1378                         xfs_dqunlock(udqp);
1379                 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_gid, XFS_DQ_GROUP,
1380                                          XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN,
1381                                          &gdqp))) {
1382                         if (udqp)
1383                                 xfs_qm_dqrele(udqp);
1384                         ASSERT(error != ESRCH);
1385                         ASSERT(error != ENOENT);
1386                         return (error);
1387                 }
1388                 ASSERT(gdqp);
1389
1390                 /* Reacquire the locks in the right order */
1391                 if (udqp) {
1392                         if (! xfs_qm_dqlock_nowait(udqp)) {
1393                                 xfs_dqunlock(gdqp);
1394                                 xfs_dqlock(udqp);
1395                                 xfs_dqlock(gdqp);
1396                         }
1397                 }
1398         }
1399
1400         *O_udqpp = udqp;
1401         *O_gdqpp = gdqp;
1402
1403 #ifdef QUOTADEBUG
1404         if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp));
1405         if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp));
1406 #endif
1407         return (0);
1408 }
1409
1410 /*
1411  * Create an inode and return with a reference already taken, but unlocked
1412  * This is how we create quota inodes
1413  */
1414 STATIC int
1415 xfs_qm_qino_alloc(
1416         xfs_mount_t     *mp,
1417         xfs_inode_t     **ip,
1418         __int64_t       sbfields,
1419         uint            flags)
1420 {
1421         xfs_trans_t     *tp;
1422         int             error;
1423         unsigned long s;
1424         cred_t          zerocr;
1425         int             committed;
1426
1427         tp = xfs_trans_alloc(mp,XFS_TRANS_QM_QINOCREATE);
1428         if ((error = xfs_trans_reserve(tp,
1429                                       XFS_QM_QINOCREATE_SPACE_RES(mp),
1430                                       XFS_CREATE_LOG_RES(mp), 0,
1431                                       XFS_TRANS_PERM_LOG_RES,
1432                                       XFS_CREATE_LOG_COUNT))) {
1433                 xfs_trans_cancel(tp, 0);
1434                 return (error);
1435         }
1436         memset(&zerocr, 0, sizeof(zerocr));
1437
1438         if ((error = xfs_dir_ialloc(&tp, mp->m_rootip, S_IFREG, 1, 0,
1439                                    &zerocr, 0, 1, ip, &committed))) {
1440                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES |
1441                                  XFS_TRANS_ABORT);
1442                 return (error);
1443         }
1444
1445         /*
1446          * Keep an extra reference to this quota inode. This inode is
1447          * locked exclusively and joined to the transaction already.
1448          */
1449         ASSERT(XFS_ISLOCKED_INODE_EXCL(*ip));
1450         VN_HOLD(XFS_ITOV((*ip)));
1451
1452         /*
1453          * Make the changes in the superblock, and log those too.
1454          * sbfields arg may contain fields other than *QUOTINO;
1455          * VERSIONNUM for example.
1456          */
1457         s = XFS_SB_LOCK(mp);
1458         if (flags & XFS_QMOPT_SBVERSION) {
1459 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1460                 unsigned oldv = mp->m_sb.sb_versionnum;
1461 #endif
1462                 ASSERT(!XFS_SB_VERSION_HASQUOTA(&mp->m_sb));
1463                 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1464                                    XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) ==
1465                        (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
1466                         XFS_SB_GQUOTINO | XFS_SB_QFLAGS));
1467
1468                 XFS_SB_VERSION_ADDQUOTA(&mp->m_sb);
1469                 mp->m_sb.sb_uquotino = NULLFSINO;
1470                 mp->m_sb.sb_gquotino = NULLFSINO;
1471
1472                 /* qflags will get updated _after_ quotacheck */
1473                 mp->m_sb.sb_qflags = 0;
1474 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
1475                 cmn_err(CE_NOTE,
1476                         "Old superblock version %x, converting to %x.",
1477                         oldv, mp->m_sb.sb_versionnum);
1478 #endif
1479         }
1480         if (flags & XFS_QMOPT_UQUOTA)
1481                 mp->m_sb.sb_uquotino = (*ip)->i_ino;
1482         else
1483                 mp->m_sb.sb_gquotino = (*ip)->i_ino;
1484         XFS_SB_UNLOCK(mp, s);
1485         xfs_mod_sb(tp, sbfields);
1486
1487         if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES,
1488                                      NULL))) {
1489                 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!");
1490                 return (error);
1491         }
1492         return (0);
1493 }
1494
1495
1496 STATIC int
1497 xfs_qm_reset_dqcounts(
1498         xfs_mount_t     *mp,
1499         xfs_buf_t       *bp,
1500         xfs_dqid_t      id,
1501         uint            type)
1502 {
1503         xfs_disk_dquot_t        *ddq;
1504         int                     j;
1505
1506         xfs_buftrace("RESET DQUOTS", bp);
1507         /*
1508          * Reset all counters and timers. They'll be
1509          * started afresh by xfs_qm_quotacheck.
1510          */
1511 #ifdef DEBUG
1512         j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB);
1513         do_div(j, sizeof(xfs_dqblk_t));
1514         ASSERT(XFS_QM_DQPERBLK(mp) == j);
1515 #endif
1516         ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp);
1517         for (j = 0; j < XFS_QM_DQPERBLK(mp); j++) {
1518                 /*
1519                  * Do a sanity check, and if needed, repair the dqblk. Don't
1520                  * output any warnings because it's perfectly possible to
1521                  * find unitialized dquot blks. See comment in xfs_qm_dqcheck.
1522                  */
1523                 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR,
1524                                       "xfs_quotacheck");
1525                 INT_SET(ddq->d_bcount, ARCH_CONVERT, 0ULL);
1526                 INT_SET(ddq->d_icount, ARCH_CONVERT, 0ULL);
1527                 INT_SET(ddq->d_rtbcount, ARCH_CONVERT, 0ULL);
1528                 INT_SET(ddq->d_btimer, ARCH_CONVERT, (time_t)0);
1529                 INT_SET(ddq->d_itimer, ARCH_CONVERT, (time_t)0);
1530                 INT_SET(ddq->d_bwarns, ARCH_CONVERT, 0UL);
1531                 INT_SET(ddq->d_iwarns, ARCH_CONVERT, 0UL);
1532                 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1);
1533         }
1534
1535         return (0);
1536 }
1537
1538 STATIC int
1539 xfs_qm_dqiter_bufs(
1540         xfs_mount_t     *mp,
1541         xfs_dqid_t      firstid,
1542         xfs_fsblock_t   bno,
1543         xfs_filblks_t   blkcnt,
1544         uint            flags)
1545 {
1546         xfs_buf_t       *bp;
1547         int             error;
1548         int             notcommitted;
1549         int             incr;
1550
1551         ASSERT(blkcnt > 0);
1552         notcommitted = 0;
1553         incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ?
1554                 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt;
1555         error = 0;
1556
1557         /*
1558          * Blkcnt arg can be a very big number, and might even be
1559          * larger than the log itself. So, we have to break it up into
1560          * manageable-sized transactions.
1561          * Note that we don't start a permanent transaction here; we might
1562          * not be able to get a log reservation for the whole thing up front,
1563          * and we don't really care to either, because we just discard
1564          * everything if we were to crash in the middle of this loop.
1565          */
1566         while (blkcnt--) {
1567                 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp,
1568                               XFS_FSB_TO_DADDR(mp, bno),
1569                               (int)XFS_QI_DQCHUNKLEN(mp), 0, &bp);
1570                 if (error)
1571                         break;
1572
1573                 (void) xfs_qm_reset_dqcounts(mp, bp, firstid,
1574                                              flags & XFS_QMOPT_UQUOTA ?
1575                                              XFS_DQ_USER : XFS_DQ_GROUP);
1576                 xfs_bdwrite(mp, bp);
1577                 /*
1578                  * goto the next block.
1579                  */
1580                 bno++;
1581                 firstid += XFS_QM_DQPERBLK(mp);
1582         }
1583         return (error);
1584 }
1585
1586 /*
1587  * Iterate over all allocated USR/GRP dquots in the system, calling a
1588  * caller supplied function for every chunk of dquots that we find.
1589  */
1590 STATIC int
1591 xfs_qm_dqiterate(
1592         xfs_mount_t     *mp,
1593         xfs_inode_t     *qip,
1594         uint            flags)
1595 {
1596         xfs_bmbt_irec_t         *map;
1597         int                     i, nmaps;       /* number of map entries */
1598         int                     error;          /* return value */
1599         xfs_fileoff_t           lblkno;
1600         xfs_filblks_t           maxlblkcnt;
1601         xfs_dqid_t              firstid;
1602         xfs_fsblock_t           rablkno;
1603         xfs_filblks_t           rablkcnt;
1604
1605         error = 0;
1606         /*
1607          * This looks racey, but we can't keep an inode lock across a
1608          * trans_reserve. But, this gets called during quotacheck, and that
1609          * happens only at mount time which is single threaded.
1610          */
1611         if (qip->i_d.di_nblocks == 0)
1612                 return (0);
1613
1614         map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP);
1615
1616         lblkno = 0;
1617         maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1618         do {
1619                 nmaps = XFS_DQITER_MAP_SIZE;
1620                 /*
1621                  * We aren't changing the inode itself. Just changing
1622                  * some of its data. No new blocks are added here, and
1623                  * the inode is never added to the transaction.
1624                  */
1625                 xfs_ilock(qip, XFS_ILOCK_SHARED);
1626                 error = xfs_bmapi(NULL, qip, lblkno,
1627                                   maxlblkcnt - lblkno,
1628                                   XFS_BMAPI_METADATA,
1629                                   NULL,
1630                                   0, map, &nmaps, NULL);
1631                 xfs_iunlock(qip, XFS_ILOCK_SHARED);
1632                 if (error)
1633                         break;
1634
1635                 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE);
1636                 for (i = 0; i < nmaps; i++) {
1637                         ASSERT(map[i].br_startblock != DELAYSTARTBLOCK);
1638                         ASSERT(map[i].br_blockcount);
1639
1640
1641                         lblkno += map[i].br_blockcount;
1642
1643                         if (map[i].br_startblock == HOLESTARTBLOCK)
1644                                 continue;
1645
1646                         firstid = (xfs_dqid_t) map[i].br_startoff *
1647                                 XFS_QM_DQPERBLK(mp);
1648                         /*
1649                          * Do a read-ahead on the next extent.
1650                          */
1651                         if ((i+1 < nmaps) &&
1652                             (map[i+1].br_startblock != HOLESTARTBLOCK)) {
1653                                 rablkcnt =  map[i+1].br_blockcount;
1654                                 rablkno = map[i+1].br_startblock;
1655                                 while (rablkcnt--) {
1656                                         xfs_baread(mp->m_ddev_targp,
1657                                                XFS_FSB_TO_DADDR(mp, rablkno),
1658                                                (int)XFS_QI_DQCHUNKLEN(mp));
1659                                         rablkno++;
1660                                 }
1661                         }
1662                         /*
1663                          * Iterate thru all the blks in the extent and
1664                          * reset the counters of all the dquots inside them.
1665                          */
1666                         if ((error = xfs_qm_dqiter_bufs(mp,
1667                                                        firstid,
1668                                                        map[i].br_startblock,
1669                                                        map[i].br_blockcount,
1670                                                        flags))) {
1671                                 break;
1672                         }
1673                 }
1674
1675                 if (error)
1676                         break;
1677         } while (nmaps > 0);
1678
1679         kmem_free(map, XFS_DQITER_MAP_SIZE * sizeof(*map));
1680
1681         return (error);
1682 }
1683
1684 /*
1685  * Called by dqusage_adjust in doing a quotacheck.
1686  * Given the inode, and a dquot (either USR or GRP, doesn't matter),
1687  * this updates its incore copy as well as the buffer copy. This is
1688  * so that once the quotacheck is done, we can just log all the buffers,
1689  * as opposed to logging numerous updates to individual dquots.
1690  */
1691 STATIC void
1692 xfs_qm_quotacheck_dqadjust(
1693         xfs_dquot_t             *dqp,
1694         xfs_qcnt_t              nblks,
1695         xfs_qcnt_t              rtblks)
1696 {
1697         ASSERT(XFS_DQ_IS_LOCKED(dqp));
1698         xfs_dqtrace_entry(dqp, "QCHECK DQADJUST");
1699         /*
1700          * Adjust the inode count and the block count to reflect this inode's
1701          * resource usage.
1702          */
1703         INT_MOD(dqp->q_core.d_icount, ARCH_CONVERT, +1);
1704         dqp->q_res_icount++;
1705         if (nblks) {
1706                 INT_MOD(dqp->q_core.d_bcount, ARCH_CONVERT, nblks);
1707                 dqp->q_res_bcount += nblks;
1708         }
1709         if (rtblks) {
1710                 INT_MOD(dqp->q_core.d_rtbcount, ARCH_CONVERT, rtblks);
1711                 dqp->q_res_rtbcount += rtblks;
1712         }
1713
1714         /*
1715          * Set default limits, adjust timers (since we changed usages)
1716          */
1717         if (! XFS_IS_SUSER_DQUOT(dqp)) {
1718                 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core);
1719                 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core);
1720         }
1721
1722         dqp->dq_flags |= XFS_DQ_DIRTY;
1723 }
1724
1725 STATIC int
1726 xfs_qm_get_rtblks(
1727         xfs_inode_t     *ip,
1728         xfs_qcnt_t      *O_rtblks)
1729 {
1730         xfs_filblks_t   rtblks;                 /* total rt blks */
1731         xfs_ifork_t     *ifp;                   /* inode fork pointer */
1732         xfs_extnum_t    nextents;               /* number of extent entries */
1733         xfs_bmbt_rec_t  *base;                  /* base of extent array */
1734         xfs_bmbt_rec_t  *ep;                    /* pointer to an extent entry */
1735         int             error;
1736
1737         ASSERT(XFS_IS_REALTIME_INODE(ip));
1738         ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1739         if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1740                 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK)))
1741                         return (error);
1742         }
1743         rtblks = 0;
1744         nextents = ifp->if_bytes / sizeof(xfs_bmbt_rec_t);
1745         base = &ifp->if_u1.if_extents[0];
1746         for (ep = base; ep < &base[nextents]; ep++)
1747                 rtblks += xfs_bmbt_get_blockcount(ep);
1748         *O_rtblks = (xfs_qcnt_t)rtblks;
1749         return (0);
1750 }
1751
1752 /*
1753  * callback routine supplied to bulkstat(). Given an inumber, find its
1754  * dquots and update them to account for resources taken by that inode.
1755  */
1756 /* ARGSUSED */
1757 STATIC int
1758 xfs_qm_dqusage_adjust(
1759         xfs_mount_t     *mp,            /* mount point for filesystem */
1760         xfs_ino_t       ino,            /* inode number to get data for */
1761         void            __user *buffer, /* not used */
1762         int             ubsize,         /* not used */
1763         void            *private_data,  /* not used */
1764         xfs_daddr_t     bno,            /* starting block of inode cluster */
1765         int             *ubused,        /* not used */
1766         void            *dip,           /* on-disk inode pointer (not used) */
1767         int             *res)           /* result code value */
1768 {
1769         xfs_inode_t     *ip;
1770         xfs_dquot_t     *udqp, *gdqp;
1771         xfs_qcnt_t      nblks, rtblks;
1772         int             error;
1773
1774         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1775
1776         /*
1777          * rootino must have its resources accounted for, not so with the quota
1778          * inodes.
1779          */
1780         if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) {
1781                 *res = BULKSTAT_RV_NOTHING;
1782                 return XFS_ERROR(EINVAL);
1783         }
1784
1785         /*
1786          * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget
1787          * interface expects the inode to be exclusively locked because that's
1788          * the case in all other instances. It's OK that we do this because
1789          * quotacheck is done only at mount time.
1790          */
1791         if ((error = xfs_iget(mp, NULL, ino, XFS_ILOCK_EXCL, &ip, bno))) {
1792                 *res = BULKSTAT_RV_NOTHING;
1793                 return (error);
1794         }
1795
1796         if (ip->i_d.di_mode == 0) {
1797                 xfs_iput_new(ip, XFS_ILOCK_EXCL);
1798                 *res = BULKSTAT_RV_NOTHING;
1799                 return XFS_ERROR(ENOENT);
1800         }
1801
1802         /*
1803          * Obtain the locked dquots. In case of an error (eg. allocation
1804          * fails for ENOSPC), we return the negative of the error number
1805          * to bulkstat, so that it can get propagated to quotacheck() and
1806          * making us disable quotas for the file system.
1807          */
1808         if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) {
1809                 xfs_iput(ip, XFS_ILOCK_EXCL);
1810                 *res = BULKSTAT_RV_GIVEUP;
1811                 return (error);
1812         }
1813
1814         rtblks = 0;
1815         if (! XFS_IS_REALTIME_INODE(ip)) {
1816                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks;
1817         } else {
1818                 /*
1819                  * Walk thru the extent list and count the realtime blocks.
1820                  */
1821                 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) {
1822                         xfs_iput(ip, XFS_ILOCK_EXCL);
1823                         if (udqp)
1824                                 xfs_qm_dqput(udqp);
1825                         if (gdqp)
1826                                 xfs_qm_dqput(gdqp);
1827                         *res = BULKSTAT_RV_GIVEUP;
1828                         return (error);
1829                 }
1830                 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks;
1831         }
1832         ASSERT(ip->i_delayed_blks == 0);
1833
1834         /*
1835          * We can't release the inode while holding its dquot locks.
1836          * The inode can go into inactive and might try to acquire the dquotlocks.
1837          * So, just unlock here and do a vn_rele at the end.
1838          */
1839         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1840
1841         /*
1842          * Add the (disk blocks and inode) resources occupied by this
1843          * inode to its dquots. We do this adjustment in the incore dquot,
1844          * and also copy the changes to its buffer.
1845          * We don't care about putting these changes in a transaction
1846          * envelope because if we crash in the middle of a 'quotacheck'
1847          * we have to start from the beginning anyway.
1848          * Once we're done, we'll log all the dquot bufs.
1849          *
1850          * The *QUOTA_ON checks below may look pretty racey, but quotachecks
1851          * and quotaoffs don't race. (Quotachecks happen at mount time only).
1852          */
1853         if (XFS_IS_UQUOTA_ON(mp)) {
1854                 ASSERT(udqp);
1855                 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks);
1856                 xfs_qm_dqput(udqp);
1857         }
1858         if (XFS_IS_GQUOTA_ON(mp)) {
1859                 ASSERT(gdqp);
1860                 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks);
1861                 xfs_qm_dqput(gdqp);
1862         }
1863         /*
1864          * Now release the inode. This will send it to 'inactive', and
1865          * possibly even free blocks.
1866          */
1867         VN_RELE(XFS_ITOV(ip));
1868
1869         /*
1870          * Goto next inode.
1871          */
1872         *res = BULKSTAT_RV_DIDONE;
1873         return (0);
1874 }
1875
1876 /*
1877  * Walk thru all the filesystem inodes and construct a consistent view
1878  * of the disk quota world.
1879  */
1880 STATIC int
1881 xfs_qm_quotacheck(
1882         xfs_mount_t     *mp)
1883 {
1884         int             done, count, error;
1885         xfs_ino_t       lastino;
1886         size_t          structsz;
1887         xfs_inode_t     *uip, *gip;
1888         uint            flags;
1889
1890         count = INT_MAX;
1891         structsz = 1;
1892         lastino = 0;
1893         flags = 0;
1894
1895         ASSERT(XFS_QI_UQIP(mp) || XFS_QI_GQIP(mp));
1896         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
1897
1898         /*
1899          * There should be no cached dquots. The (simplistic) quotacheck
1900          * algorithm doesn't like that.
1901          */
1902         ASSERT(XFS_QI_MPLNDQUOTS(mp) == 0);
1903
1904         cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname);
1905
1906         /*
1907          * First we go thru all the dquots on disk, USR and GRP, and reset
1908          * their counters to zero. We need a clean slate.
1909          * We don't log our changes till later.
1910          */
1911         if ((uip = XFS_QI_UQIP(mp))) {
1912                 if ((error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA)))
1913                         goto error_return;
1914                 flags |= XFS_UQUOTA_CHKD;
1915         }
1916
1917         if ((gip = XFS_QI_GQIP(mp))) {
1918                 if ((error = xfs_qm_dqiterate(mp, gip, XFS_QMOPT_GQUOTA)))
1919                         goto error_return;
1920                 flags |= XFS_GQUOTA_CHKD;
1921         }
1922
1923         do {
1924                 /*
1925                  * Iterate thru all the inodes in the file system,
1926                  * adjusting the corresponding dquot counters in core.
1927                  */
1928                 if ((error = xfs_bulkstat(mp, &lastino, &count,
1929                                      xfs_qm_dqusage_adjust, NULL,
1930                                      structsz, NULL,
1931                                      BULKSTAT_FG_IGET|BULKSTAT_FG_VFSLOCKED,
1932                                      &done)))
1933                         break;
1934
1935         } while (! done);
1936
1937         /*
1938          * We can get this error if we couldn't do a dquot allocation inside
1939          * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the
1940          * dirty dquots that might be cached, we just want to get rid of them
1941          * and turn quotaoff. The dquots won't be attached to any of the inodes
1942          * at this point (because we intentionally didn't in dqget_noattach).
1943          */
1944         if (error) {
1945                 xfs_qm_dqpurge_all(mp,
1946                                    XFS_QMOPT_UQUOTA|XFS_QMOPT_GQUOTA|
1947                                    XFS_QMOPT_QUOTAOFF);
1948                 goto error_return;
1949         }
1950         /*
1951          * We've made all the changes that we need to make incore.
1952          * Now flush_them down to disk buffers.
1953          */
1954         xfs_qm_dqflush_all(mp, XFS_QMOPT_DELWRI);
1955
1956         /*
1957          * We didn't log anything, because if we crashed, we'll have to
1958          * start the quotacheck from scratch anyway. However, we must make
1959          * sure that our dquot changes are secure before we put the
1960          * quotacheck'd stamp on the superblock. So, here we do a synchronous
1961          * flush.
1962          */
1963         XFS_bflush(mp->m_ddev_targp);
1964
1965         /*
1966          * If one type of quotas is off, then it will lose its
1967          * quotachecked status, since we won't be doing accounting for
1968          * that type anymore.
1969          */
1970         mp->m_qflags &= ~(XFS_GQUOTA_CHKD | XFS_UQUOTA_CHKD);
1971         mp->m_qflags |= flags;
1972
1973         XQM_LIST_PRINT(&(XFS_QI_MPL_LIST(mp)), MPL_NEXT, "++++ Mp list +++");
1974
1975  error_return:
1976         cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname);
1977         return (error);
1978 }
1979
1980 /*
1981  * This is called after the superblock has been read in and we're ready to
1982  * iget the quota inodes.
1983  */
1984 STATIC int
1985 xfs_qm_init_quotainos(
1986         xfs_mount_t     *mp)
1987 {
1988         xfs_inode_t     *uip, *gip;
1989         int             error;
1990         __int64_t       sbflags;
1991         uint            flags;
1992
1993         ASSERT(mp->m_quotainfo);
1994         uip = gip = NULL;
1995         sbflags = 0;
1996         flags = 0;
1997
1998         /*
1999          * Get the uquota and gquota inodes
2000          */
2001         if (XFS_SB_VERSION_HASQUOTA(&mp->m_sb)) {
2002                 if (XFS_IS_UQUOTA_ON(mp) &&
2003                     mp->m_sb.sb_uquotino != NULLFSINO) {
2004                         ASSERT(mp->m_sb.sb_uquotino > 0);
2005                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino,
2006                                              0, &uip, 0)))
2007                                 return XFS_ERROR(error);
2008                 }
2009                 if (XFS_IS_GQUOTA_ON(mp) &&
2010                     mp->m_sb.sb_gquotino != NULLFSINO) {
2011                         ASSERT(mp->m_sb.sb_gquotino > 0);
2012                         if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino,
2013                                              0, &gip, 0))) {
2014                                 if (uip)
2015                                         VN_RELE(XFS_ITOV(uip));
2016                                 return XFS_ERROR(error);
2017                         }
2018                 }
2019         } else {
2020                 flags |= XFS_QMOPT_SBVERSION;
2021                 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO |
2022                             XFS_SB_GQUOTINO | XFS_SB_QFLAGS);
2023         }
2024
2025         /*
2026          * Create the two inodes, if they don't exist already. The changes
2027          * made above will get added to a transaction and logged in one of
2028          * the qino_alloc calls below.  If the device is readonly,
2029          * temporarily switch to read-write to do this.
2030          */
2031         if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) {
2032                 if ((error = xfs_qm_qino_alloc(mp, &uip,
2033                                               sbflags | XFS_SB_UQUOTINO,
2034                                               flags | XFS_QMOPT_UQUOTA)))
2035                         return XFS_ERROR(error);
2036
2037                 flags &= ~XFS_QMOPT_SBVERSION;
2038         }
2039         if (XFS_IS_GQUOTA_ON(mp) && gip == NULL) {
2040                 if ((error = xfs_qm_qino_alloc(mp, &gip,
2041                                               sbflags | XFS_SB_GQUOTINO,
2042                                               flags | XFS_QMOPT_GQUOTA))) {
2043                         if (uip)
2044                                 VN_RELE(XFS_ITOV(uip));
2045
2046                         return XFS_ERROR(error);
2047                 }
2048         }
2049
2050         XFS_QI_UQIP(mp) = uip;
2051         XFS_QI_GQIP(mp) = gip;
2052
2053         return (0);
2054 }
2055
2056
2057 /*
2058  * Traverse the freelist of dquots and attempt to reclaim a maximum of
2059  * 'howmany' dquots. This operation races with dqlookup(), and attempts to
2060  * favor the lookup function ...
2061  * XXXsup merge this with qm_reclaim_one().
2062  */
2063 STATIC int
2064 xfs_qm_shake_freelist(
2065         int howmany)
2066 {
2067         int             nreclaimed;
2068         xfs_dqhash_t    *hash;
2069         xfs_dquot_t     *dqp, *nextdqp;
2070         int             restarts;
2071         int             nflushes;
2072
2073         if (howmany <= 0)
2074                 return (0);
2075
2076         nreclaimed = 0;
2077         restarts = 0;
2078         nflushes = 0;
2079
2080 #ifdef QUOTADEBUG
2081         cmn_err(CE_DEBUG, "Shake free 0x%x", howmany);
2082 #endif
2083         /* lock order is : hashchainlock, freelistlock, mplistlock */
2084  tryagain:
2085         xfs_qm_freelist_lock(xfs_Gqm);
2086
2087         for (dqp = xfs_Gqm->qm_dqfreelist.qh_next;
2088              ((dqp != (xfs_dquot_t *) &xfs_Gqm->qm_dqfreelist) &&
2089               nreclaimed < howmany); ) {
2090                 xfs_dqlock(dqp);
2091
2092                 /*
2093                  * We are racing with dqlookup here. Naturally we don't
2094                  * want to reclaim a dquot that lookup wants.
2095                  */
2096                 if (dqp->dq_flags & XFS_DQ_WANT) {
2097                         xfs_dqunlock(dqp);
2098                         xfs_qm_freelist_unlock(xfs_Gqm);
2099                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2100                                 return (nreclaimed);
2101                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2102                         goto tryagain;
2103                 }
2104
2105                 /*
2106                  * If the dquot is inactive, we are assured that it is
2107                  * not on the mplist or the hashlist, and that makes our
2108                  * life easier.
2109                  */
2110                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2111                         ASSERT(dqp->q_mount == NULL);
2112                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2113                         ASSERT(dqp->HL_PREVP == NULL);
2114                         ASSERT(dqp->MPL_PREVP == NULL);
2115                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2116                         nextdqp = dqp->dq_flnext;
2117                         goto off_freelist;
2118                 }
2119
2120                 ASSERT(dqp->MPL_PREVP);
2121                 /*
2122                  * Try to grab the flush lock. If this dquot is in the process of
2123                  * getting flushed to disk, we don't want to reclaim it.
2124                  */
2125                 if (! xfs_qm_dqflock_nowait(dqp)) {
2126                         xfs_dqunlock(dqp);
2127                         dqp = dqp->dq_flnext;
2128                         continue;
2129                 }
2130
2131                 /*
2132                  * We have the flush lock so we know that this is not in the
2133                  * process of being flushed. So, if this is dirty, flush it
2134                  * DELWRI so that we don't get a freelist infested with
2135                  * dirty dquots.
2136                  */
2137                 if (XFS_DQ_IS_DIRTY(dqp)) {
2138                         xfs_dqtrace_entry(dqp, "DQSHAKE: DQDIRTY");
2139                         /*
2140                          * We flush it delayed write, so don't bother
2141                          * releasing the mplock.
2142                          */
2143                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2144                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2145                         dqp = dqp->dq_flnext;
2146                         continue;
2147                 }
2148                 /*
2149                  * We're trying to get the hashlock out of order. This races
2150                  * with dqlookup; so, we giveup and goto the next dquot if
2151                  * we couldn't get the hashlock. This way, we won't starve
2152                  * a dqlookup process that holds the hashlock that is
2153                  * waiting for the freelist lock.
2154                  */
2155                 if (! xfs_qm_dqhashlock_nowait(dqp)) {
2156                         xfs_dqfunlock(dqp);
2157                         xfs_dqunlock(dqp);
2158                         dqp = dqp->dq_flnext;
2159                         continue;
2160                 }
2161                 /*
2162                  * This races with dquot allocation code as well as dqflush_all
2163                  * and reclaim code. So, if we failed to grab the mplist lock,
2164                  * giveup everything and start over.
2165                  */
2166                 hash = dqp->q_hash;
2167                 ASSERT(hash);
2168                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2169                         /* XXX put a sentinel so that we can come back here */
2170                         xfs_dqfunlock(dqp);
2171                         xfs_dqunlock(dqp);
2172                         XFS_DQ_HASH_UNLOCK(hash);
2173                         xfs_qm_freelist_unlock(xfs_Gqm);
2174                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2175                                 return (nreclaimed);
2176                         goto tryagain;
2177                 }
2178                 xfs_dqtrace_entry(dqp, "DQSHAKE: UNLINKING");
2179 #ifdef QUOTADEBUG
2180                 cmn_err(CE_DEBUG, "Shake 0x%p, ID 0x%x\n",
2181                         dqp, INT_GET(dqp->q_core.d_id, ARCH_CONVERT));
2182 #endif
2183                 ASSERT(dqp->q_nrefs == 0);
2184                 nextdqp = dqp->dq_flnext;
2185                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2186                 XQM_HASHLIST_REMOVE(hash, dqp);
2187                 xfs_dqfunlock(dqp);
2188                 xfs_qm_mplist_unlock(dqp->q_mount);
2189                 XFS_DQ_HASH_UNLOCK(hash);
2190
2191  off_freelist:
2192                 XQM_FREELIST_REMOVE(dqp);
2193                 xfs_dqunlock(dqp);
2194                 nreclaimed++;
2195                 XQM_STATS_INC(xqmstats.xs_qm_dqshake_reclaims);
2196                 xfs_qm_dqdestroy(dqp);
2197                 dqp = nextdqp;
2198         }
2199         xfs_qm_freelist_unlock(xfs_Gqm);
2200         return (nreclaimed);
2201 }
2202
2203
2204 /*
2205  * The kmem_shake interface is invoked when memory is running low.
2206  */
2207 /* ARGSUSED */
2208 STATIC int
2209 xfs_qm_shake(int nr_to_scan, unsigned int gfp_mask)
2210 {
2211         int     ndqused, nfree, n;
2212
2213         if (!kmem_shake_allow(gfp_mask))
2214                 return (0);
2215         if (!xfs_Gqm)
2216                 return (0);
2217
2218         nfree = xfs_Gqm->qm_dqfreelist.qh_nelems; /* free dquots */
2219         /* incore dquots in all f/s's */
2220         ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
2221
2222         ASSERT(ndqused >= 0);
2223
2224         if (nfree <= ndqused && nfree < ndquot)
2225                 return (0);
2226
2227         ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
2228         n = nfree - ndqused - ndquot;           /* # over target */
2229
2230         return xfs_qm_shake_freelist(MAX(nfree, n));
2231 }
2232
2233
2234 /*
2235  * Just pop the least recently used dquot off the freelist and
2236  * recycle it. The returned dquot is locked.
2237  */
2238 STATIC xfs_dquot_t *
2239 xfs_qm_dqreclaim_one(void)
2240 {
2241         xfs_dquot_t     *dqpout;
2242         xfs_dquot_t     *dqp;
2243         int             restarts;
2244         int             nflushes;
2245
2246         restarts = 0;
2247         dqpout = NULL;
2248         nflushes = 0;
2249
2250         /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */
2251  startagain:
2252         xfs_qm_freelist_lock(xfs_Gqm);
2253
2254         FOREACH_DQUOT_IN_FREELIST(dqp, &(xfs_Gqm->qm_dqfreelist)) {
2255                 xfs_dqlock(dqp);
2256
2257                 /*
2258                  * We are racing with dqlookup here. Naturally we don't
2259                  * want to reclaim a dquot that lookup wants. We release the
2260                  * freelist lock and start over, so that lookup will grab
2261                  * both the dquot and the freelistlock.
2262                  */
2263                 if (dqp->dq_flags & XFS_DQ_WANT) {
2264                         ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE));
2265                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQWANT");
2266                         xfs_dqunlock(dqp);
2267                         xfs_qm_freelist_unlock(xfs_Gqm);
2268                         if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
2269                                 return (NULL);
2270                         XQM_STATS_INC(xqmstats.xs_qm_dqwants);
2271                         goto startagain;
2272                 }
2273
2274                 /*
2275                  * If the dquot is inactive, we are assured that it is
2276                  * not on the mplist or the hashlist, and that makes our
2277                  * life easier.
2278                  */
2279                 if (dqp->dq_flags & XFS_DQ_INACTIVE) {
2280                         ASSERT(dqp->q_mount == NULL);
2281                         ASSERT(! XFS_DQ_IS_DIRTY(dqp));
2282                         ASSERT(dqp->HL_PREVP == NULL);
2283                         ASSERT(dqp->MPL_PREVP == NULL);
2284                         XQM_FREELIST_REMOVE(dqp);
2285                         xfs_dqunlock(dqp);
2286                         dqpout = dqp;
2287                         XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims);
2288                         break;
2289                 }
2290
2291                 ASSERT(dqp->q_hash);
2292                 ASSERT(dqp->MPL_PREVP);
2293
2294                 /*
2295                  * Try to grab the flush lock. If this dquot is in the process of
2296                  * getting flushed to disk, we don't want to reclaim it.
2297                  */
2298                 if (! xfs_qm_dqflock_nowait(dqp)) {
2299                         xfs_dqunlock(dqp);
2300                         continue;
2301                 }
2302
2303                 /*
2304                  * We have the flush lock so we know that this is not in the
2305                  * process of being flushed. So, if this is dirty, flush it
2306                  * DELWRI so that we don't get a freelist infested with
2307                  * dirty dquots.
2308                  */
2309                 if (XFS_DQ_IS_DIRTY(dqp)) {
2310                         xfs_dqtrace_entry(dqp, "DQRECLAIM: DQDIRTY");
2311                         /*
2312                          * We flush it delayed write, so don't bother
2313                          * releasing the freelist lock.
2314                          */
2315                         (void) xfs_qm_dqflush(dqp, XFS_QMOPT_DELWRI);
2316                         xfs_dqunlock(dqp); /* dqflush unlocks dqflock */
2317                         continue;
2318                 }
2319
2320                 if (! xfs_qm_mplist_nowait(dqp->q_mount)) {
2321                         xfs_dqfunlock(dqp);
2322                         xfs_dqunlock(dqp);
2323                         continue;
2324                 }
2325
2326                 if (! xfs_qm_dqhashlock_nowait(dqp))
2327                         goto mplistunlock;
2328
2329                 ASSERT(dqp->q_nrefs == 0);
2330                 xfs_dqtrace_entry(dqp, "DQRECLAIM: UNLINKING");
2331                 XQM_MPLIST_REMOVE(&(XFS_QI_MPL_LIST(dqp->q_mount)), dqp);
2332                 XQM_HASHLIST_REMOVE(dqp->q_hash, dqp);
2333                 XQM_FREELIST_REMOVE(dqp);
2334                 dqpout = dqp;
2335                 XFS_DQ_HASH_UNLOCK(dqp->q_hash);
2336  mplistunlock:
2337                 xfs_qm_mplist_unlock(dqp->q_mount);
2338                 xfs_dqfunlock(dqp);
2339                 xfs_dqunlock(dqp);
2340                 if (dqpout)
2341                         break;
2342         }
2343
2344         xfs_qm_freelist_unlock(xfs_Gqm);
2345         return (dqpout);
2346 }
2347
2348
2349 /*------------------------------------------------------------------*/
2350
2351 /*
2352  * Return a new incore dquot. Depending on the number of
2353  * dquots in the system, we either allocate a new one on the kernel heap,
2354  * or reclaim a free one.
2355  * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
2356  * to reclaim an existing one from the freelist.
2357  */
2358 boolean_t
2359 xfs_qm_dqalloc_incore(
2360         xfs_dquot_t **O_dqpp)
2361 {
2362         xfs_dquot_t     *dqp;
2363
2364         /*
2365          * Check against high water mark to see if we want to pop
2366          * a nincompoop dquot off the freelist.
2367          */
2368         if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
2369                 /*
2370                  * Try to recycle a dquot from the freelist.
2371                  */
2372                 if ((dqp = xfs_qm_dqreclaim_one())) {
2373                         XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
2374                         /*
2375                          * Just zero the core here. The rest will get
2376                          * reinitialized by caller. XXX we shouldn't even
2377                          * do this zero ...
2378                          */
2379                         memset(&dqp->q_core, 0, sizeof(dqp->q_core));
2380                         *O_dqpp = dqp;
2381                         return (B_FALSE);
2382                 }
2383                 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
2384         }
2385
2386         /*
2387          * Allocate a brand new dquot on the kernel heap and return it
2388          * to the caller to initialize.
2389          */
2390         ASSERT(xfs_Gqm->qm_dqzone != NULL);
2391         *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
2392         atomic_inc(&xfs_Gqm->qm_totaldquots);
2393
2394         return (B_TRUE);
2395 }
2396
2397
2398 /*
2399  * Start a transaction and write the incore superblock changes to
2400  * disk. flags parameter indicates which fields have changed.
2401  */
2402 int
2403 xfs_qm_write_sb_changes(
2404         xfs_mount_t     *mp,
2405         __int64_t       flags)
2406 {
2407         xfs_trans_t     *tp;
2408         int             error;
2409
2410 #ifdef QUOTADEBUG
2411         cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname);
2412 #endif
2413         tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
2414         if ((error = xfs_trans_reserve(tp, 0,
2415                                       mp->m_sb.sb_sectsize + 128, 0,
2416                                       0,
2417                                       XFS_DEFAULT_LOG_COUNT))) {
2418                 xfs_trans_cancel(tp, 0);
2419                 return (error);
2420         }
2421
2422         xfs_mod_sb(tp, flags);
2423         (void) xfs_trans_commit(tp, 0, NULL);
2424
2425         return (0);
2426 }
2427
2428
2429 /* --------------- utility functions for vnodeops ---------------- */
2430
2431
2432 /*
2433  * Given an inode, a uid and gid (from cred_t) make sure that we have
2434  * allocated relevant dquot(s) on disk, and that we won't exceed inode
2435  * quotas by creating this file.
2436  * This also attaches dquot(s) to the given inode after locking it,
2437  * and returns the dquots corresponding to the uid and/or gid.
2438  *
2439  * in   : inode (unlocked)
2440  * out  : udquot, gdquot with references taken and unlocked
2441  */
2442 int
2443 xfs_qm_vop_dqalloc(
2444         xfs_mount_t     *mp,
2445         xfs_inode_t     *ip,
2446         uid_t           uid,
2447         gid_t           gid,
2448         uint            flags,
2449         xfs_dquot_t     **O_udqpp,
2450         xfs_dquot_t     **O_gdqpp)
2451 {
2452         int             error;
2453         xfs_dquot_t     *uq, *gq;
2454         uint            lockflags;
2455
2456         if (!XFS_IS_QUOTA_ON(mp))
2457                 return 0;
2458
2459         lockflags = XFS_ILOCK_EXCL;
2460         xfs_ilock(ip, lockflags);
2461
2462         if ((flags & XFS_QMOPT_INHERIT) &&
2463             XFS_INHERIT_GID(ip, XFS_MTOVFS(mp)))
2464                 gid = ip->i_d.di_gid;
2465
2466         /*
2467          * Attach the dquot(s) to this inode, doing a dquot allocation
2468          * if necessary. The dquot(s) will not be locked.
2469          */
2470         if (XFS_NOT_DQATTACHED(mp, ip)) {
2471                 if ((error = xfs_qm_dqattach(ip, XFS_QMOPT_DQALLOC |
2472                                             XFS_QMOPT_ILOCKED))) {
2473                         xfs_iunlock(ip, lockflags);
2474                         return (error);
2475                 }
2476         }
2477
2478         uq = gq = NULL;
2479         if ((flags & XFS_QMOPT_UQUOTA) &&
2480             XFS_IS_UQUOTA_ON(mp)) {
2481                 if (ip->i_d.di_uid != uid) {
2482                         /*
2483                          * What we need is the dquot that has this uid, and
2484                          * if we send the inode to dqget, the uid of the inode
2485                          * takes priority over what's sent in the uid argument.
2486                          * We must unlock inode here before calling dqget if
2487                          * we're not sending the inode, because otherwise
2488                          * we'll deadlock by doing trans_reserve while
2489                          * holding ilock.
2490                          */
2491                         xfs_iunlock(ip, lockflags);
2492                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid,
2493                                                  XFS_DQ_USER,
2494                                                  XFS_QMOPT_DQALLOC |
2495                                                  XFS_QMOPT_DOWARN,
2496                                                  &uq))) {
2497                                 ASSERT(error != ENOENT);
2498                                 return (error);
2499                         }
2500                         /*
2501                          * Get the ilock in the right order.
2502                          */
2503                         xfs_dqunlock(uq);
2504                         lockflags = XFS_ILOCK_SHARED;
2505                         xfs_ilock(ip, lockflags);
2506                 } else {
2507                         /*
2508                          * Take an extra reference, because we'll return
2509                          * this to caller
2510                          */
2511                         ASSERT(ip->i_udquot);
2512                         uq = ip->i_udquot;
2513                         xfs_dqlock(uq);
2514                         XFS_DQHOLD(uq);
2515                         xfs_dqunlock(uq);
2516                 }
2517         }
2518         if ((flags & XFS_QMOPT_GQUOTA) &&
2519             XFS_IS_GQUOTA_ON(mp)) {
2520                 if (ip->i_d.di_gid != gid) {
2521                         xfs_iunlock(ip, lockflags);
2522                         if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid,
2523                                                  XFS_DQ_GROUP,
2524                                                  XFS_QMOPT_DQALLOC |
2525                                                  XFS_QMOPT_DOWARN,
2526                                                  &gq))) {
2527                                 if (uq)
2528                                         xfs_qm_dqrele(uq);
2529                                 ASSERT(error != ENOENT);
2530                                 return (error);
2531                         }
2532                         xfs_dqunlock(gq);
2533                         lockflags = XFS_ILOCK_SHARED;
2534                         xfs_ilock(ip, lockflags);
2535                 } else {
2536                         ASSERT(ip->i_gdquot);
2537                         gq = ip->i_gdquot;
2538                         xfs_dqlock(gq);
2539                         XFS_DQHOLD(gq);
2540                         xfs_dqunlock(gq);
2541                 }
2542         }
2543         if (uq)
2544                 xfs_dqtrace_entry_ino(uq, "DQALLOC", ip);
2545
2546         xfs_iunlock(ip, lockflags);
2547         if (O_udqpp)
2548                 *O_udqpp = uq;
2549         else if (uq)
2550                 xfs_qm_dqrele(uq);
2551         if (O_gdqpp)
2552                 *O_gdqpp = gq;
2553         else if (gq)
2554                 xfs_qm_dqrele(gq);
2555         return (0);
2556 }
2557
2558 /*
2559  * Actually transfer ownership, and do dquot modifications.
2560  * These were already reserved.
2561  */
2562 xfs_dquot_t *
2563 xfs_qm_vop_chown(
2564         xfs_trans_t     *tp,
2565         xfs_inode_t     *ip,
2566         xfs_dquot_t     **IO_olddq,
2567         xfs_dquot_t     *newdq)
2568 {
2569         xfs_dquot_t     *prevdq;
2570         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2571         ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount));
2572
2573         /* old dquot */
2574         prevdq = *IO_olddq;
2575         ASSERT(prevdq);
2576         ASSERT(prevdq != newdq);
2577
2578         xfs_trans_mod_dquot(tp, prevdq,
2579                             XFS_TRANS_DQ_BCOUNT,
2580                             -(ip->i_d.di_nblocks));
2581         xfs_trans_mod_dquot(tp, prevdq,
2582                             XFS_TRANS_DQ_ICOUNT,
2583                             -1);
2584
2585         /* the sparkling new dquot */
2586         xfs_trans_mod_dquot(tp, newdq,
2587                             XFS_TRANS_DQ_BCOUNT,
2588                             ip->i_d.di_nblocks);
2589         xfs_trans_mod_dquot(tp, newdq,
2590                             XFS_TRANS_DQ_ICOUNT,
2591                             1);
2592
2593         /*
2594          * Take an extra reference, because the inode
2595          * is going to keep this dquot pointer even
2596          * after the trans_commit.
2597          */
2598         xfs_dqlock(newdq);
2599         XFS_DQHOLD(newdq);
2600         xfs_dqunlock(newdq);
2601         *IO_olddq = newdq;
2602
2603         return (prevdq);
2604 }
2605
2606 /*
2607  * Quota reservations for setattr(AT_UID|AT_GID).
2608  */
2609 int
2610 xfs_qm_vop_chown_reserve(
2611         xfs_trans_t     *tp,
2612         xfs_inode_t     *ip,
2613         xfs_dquot_t     *udqp,
2614         xfs_dquot_t     *gdqp,
2615         uint            flags)
2616 {
2617         int             error;
2618         xfs_mount_t     *mp;
2619         uint            delblks;
2620         xfs_dquot_t     *unresudq, *unresgdq, *delblksudq, *delblksgdq;
2621
2622         ASSERT(XFS_ISLOCKED_INODE(ip));
2623         mp = ip->i_mount;
2624         ASSERT(XFS_IS_QUOTA_RUNNING(mp));
2625
2626         delblks = ip->i_delayed_blks;
2627         delblksudq = delblksgdq = unresudq = unresgdq = NULL;
2628
2629         if (XFS_IS_UQUOTA_ON(mp) && udqp &&
2630             ip->i_d.di_uid != (uid_t)INT_GET(udqp->q_core.d_id, ARCH_CONVERT)) {
2631                 delblksudq = udqp;
2632                 /*
2633                  * If there are delayed allocation blocks, then we have to
2634                  * unreserve those from the old dquot, and add them to the
2635                  * new dquot.
2636                  */
2637                 if (delblks) {
2638                         ASSERT(ip->i_udquot);
2639                         unresudq = ip->i_udquot;
2640                 }
2641         }
2642         if (XFS_IS_GQUOTA_ON(ip->i_mount) && gdqp &&
2643             ip->i_d.di_gid != INT_GET(gdqp->q_core.d_id, ARCH_CONVERT)) {
2644                 delblksgdq = gdqp;
2645                 if (delblks) {
2646                         ASSERT(ip->i_gdquot);
2647                         unresgdq = ip->i_gdquot;
2648                 }
2649         }
2650
2651         if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount,
2652                                 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1,
2653                                 flags | XFS_QMOPT_RES_REGBLKS)))
2654                 return (error);
2655
2656         /*
2657          * Do the delayed blks reservations/unreservations now. Since, these
2658          * are done without the help of a transaction, if a reservation fails
2659          * its previous reservations won't be automatically undone by trans
2660          * code. So, we have to do it manually here.
2661          */
2662         if (delblks) {
2663                 /*
2664                  * Do the reservations first. Unreservation can't fail.
2665                  */
2666                 ASSERT(delblksudq || delblksgdq);
2667                 ASSERT(unresudq || unresgdq);
2668                 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2669                                 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0,
2670                                 flags | XFS_QMOPT_RES_REGBLKS)))
2671                         return (error);
2672                 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount,
2673                                 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0,
2674                                 XFS_QMOPT_RES_REGBLKS);
2675         }
2676
2677         return (0);
2678 }
2679
2680 int
2681 xfs_qm_vop_rename_dqattach(
2682         xfs_inode_t     **i_tab)
2683 {
2684         xfs_inode_t     *ip;
2685         int             i;
2686         int             error;
2687
2688         ip = i_tab[0];
2689
2690         if (! XFS_IS_QUOTA_ON(ip->i_mount))
2691                 return (0);
2692
2693         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2694                 error = xfs_qm_dqattach(ip, 0);
2695                 if (error)
2696                         return (error);
2697         }
2698         for (i = 1; (i < 4 && i_tab[i]); i++) {
2699                 /*
2700                  * Watch out for duplicate entries in the table.
2701                  */
2702                 if ((ip = i_tab[i]) != i_tab[i-1]) {
2703                         if (XFS_NOT_DQATTACHED(ip->i_mount, ip)) {
2704                                 error = xfs_qm_dqattach(ip, 0);
2705                                 if (error)
2706                                         return (error);
2707                         }
2708                 }
2709         }
2710         return (0);
2711 }
2712
2713 void
2714 xfs_qm_vop_dqattach_and_dqmod_newinode(
2715         xfs_trans_t     *tp,
2716         xfs_inode_t     *ip,
2717         xfs_dquot_t     *udqp,
2718         xfs_dquot_t     *gdqp)
2719 {
2720         if (!XFS_IS_QUOTA_ON(tp->t_mountp))
2721                 return;
2722
2723         ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
2724         ASSERT(XFS_IS_QUOTA_RUNNING(tp->t_mountp));
2725
2726         if (udqp) {
2727                 xfs_dqlock(udqp);
2728                 XFS_DQHOLD(udqp);
2729                 xfs_dqunlock(udqp);
2730                 ASSERT(ip->i_udquot == NULL);
2731                 ip->i_udquot = udqp;
2732                 ASSERT(ip->i_d.di_uid == INT_GET(udqp->q_core.d_id, ARCH_CONVERT));
2733                 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1);
2734         }
2735         if (gdqp) {
2736                 xfs_dqlock(gdqp);
2737                 XFS_DQHOLD(gdqp);
2738                 xfs_dqunlock(gdqp);
2739                 ASSERT(ip->i_gdquot == NULL);
2740                 ip->i_gdquot = gdqp;
2741                 ASSERT(ip->i_d.di_gid == INT_GET(gdqp->q_core.d_id, ARCH_CONVERT));
2742                 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1);
2743         }
2744 }
2745
2746 /* ------------- list stuff -----------------*/
2747 void
2748 xfs_qm_freelist_init(xfs_frlist_t *ql)
2749 {
2750         ql->qh_next = ql->qh_prev = (xfs_dquot_t *) ql;
2751         mutex_init(&ql->qh_lock, MUTEX_DEFAULT, "dqf");
2752         ql->qh_version = 0;
2753         ql->qh_nelems = 0;
2754 }
2755
2756 void
2757 xfs_qm_freelist_destroy(xfs_frlist_t *ql)
2758 {
2759         xfs_dquot_t     *dqp, *nextdqp;
2760
2761         mutex_lock(&ql->qh_lock, PINOD);
2762         for (dqp = ql->qh_next;
2763              dqp != (xfs_dquot_t *)ql; ) {
2764                 xfs_dqlock(dqp);
2765                 nextdqp = dqp->dq_flnext;
2766 #ifdef QUOTADEBUG
2767                 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp);
2768 #endif
2769                 XQM_FREELIST_REMOVE(dqp);
2770                 xfs_dqunlock(dqp);
2771                 xfs_qm_dqdestroy(dqp);
2772                 dqp = nextdqp;
2773         }
2774         /*
2775          * Don't bother about unlocking.
2776          */
2777         mutex_destroy(&ql->qh_lock);
2778
2779         ASSERT(ql->qh_nelems == 0);
2780 }
2781
2782 void
2783 xfs_qm_freelist_insert(xfs_frlist_t *ql, xfs_dquot_t *dq)
2784 {
2785         dq->dq_flnext = ql->qh_next;
2786         dq->dq_flprev = (xfs_dquot_t *)ql;
2787         ql->qh_next = dq;
2788         dq->dq_flnext->dq_flprev = dq;
2789         xfs_Gqm->qm_dqfreelist.qh_nelems++;
2790         xfs_Gqm->qm_dqfreelist.qh_version++;
2791 }
2792
2793 void
2794 xfs_qm_freelist_unlink(xfs_dquot_t *dq)
2795 {
2796         xfs_dquot_t *next = dq->dq_flnext;
2797         xfs_dquot_t *prev = dq->dq_flprev;
2798
2799         next->dq_flprev = prev;
2800         prev->dq_flnext = next;
2801         dq->dq_flnext = dq->dq_flprev = dq;
2802         xfs_Gqm->qm_dqfreelist.qh_nelems--;
2803         xfs_Gqm->qm_dqfreelist.qh_version++;
2804 }
2805
2806 void
2807 xfs_qm_freelist_append(xfs_frlist_t *ql, xfs_dquot_t *dq)
2808 {
2809         xfs_qm_freelist_insert((xfs_frlist_t *)ql->qh_prev, dq);
2810 }
2811
2812 int
2813 xfs_qm_dqhashlock_nowait(
2814         xfs_dquot_t *dqp)
2815 {
2816         int locked;
2817
2818         locked = mutex_trylock(&((dqp)->q_hash->qh_lock));
2819         return (locked);
2820 }
2821
2822 int
2823 xfs_qm_freelist_lock_nowait(
2824         xfs_qm_t *xqm)
2825 {
2826         int locked;
2827
2828         locked = mutex_trylock(&(xqm->qm_dqfreelist.qh_lock));
2829         return (locked);
2830 }
2831
2832 int
2833 xfs_qm_mplist_nowait(
2834         xfs_mount_t     *mp)
2835 {
2836         int locked;
2837
2838         ASSERT(mp->m_quotainfo);
2839         locked = mutex_trylock(&(XFS_QI_MPLLOCK(mp)));
2840         return (locked);
2841 }