Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6
[linux-2.6] / fs / xfs / xfs_vnodeops.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18
19 #include "xfs.h"
20 #include "xfs_fs.h"
21 #include "xfs_types.h"
22 #include "xfs_bit.h"
23 #include "xfs_log.h"
24 #include "xfs_inum.h"
25 #include "xfs_trans.h"
26 #include "xfs_sb.h"
27 #include "xfs_ag.h"
28 #include "xfs_dir2.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_da_btree.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_inode_item.h"
40 #include "xfs_itable.h"
41 #include "xfs_btree.h"
42 #include "xfs_ialloc.h"
43 #include "xfs_alloc.h"
44 #include "xfs_bmap.h"
45 #include "xfs_attr.h"
46 #include "xfs_rw.h"
47 #include "xfs_error.h"
48 #include "xfs_quota.h"
49 #include "xfs_utils.h"
50 #include "xfs_rtalloc.h"
51 #include "xfs_trans_space.h"
52 #include "xfs_log_priv.h"
53 #include "xfs_filestream.h"
54 #include "xfs_vnodeops.h"
55
56 int
57 xfs_open(
58         xfs_inode_t     *ip)
59 {
60         int             mode;
61
62         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
63                 return XFS_ERROR(EIO);
64
65         /*
66          * If it's a directory with any blocks, read-ahead block 0
67          * as we're almost certain to have the next operation be a read there.
68          */
69         if (S_ISDIR(ip->i_d.di_mode) && ip->i_d.di_nextents > 0) {
70                 mode = xfs_ilock_map_shared(ip);
71                 if (ip->i_d.di_nextents > 0)
72                         (void)xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
73                 xfs_iunlock(ip, mode);
74         }
75         return 0;
76 }
77
78 int
79 xfs_setattr(
80         struct xfs_inode        *ip,
81         struct iattr            *iattr,
82         int                     flags,
83         cred_t                  *credp)
84 {
85         xfs_mount_t             *mp = ip->i_mount;
86         struct inode            *inode = VFS_I(ip);
87         int                     mask = iattr->ia_valid;
88         xfs_trans_t             *tp;
89         int                     code;
90         uint                    lock_flags;
91         uint                    commit_flags=0;
92         uid_t                   uid=0, iuid=0;
93         gid_t                   gid=0, igid=0;
94         int                     timeflags = 0;
95         struct xfs_dquot        *udqp, *gdqp, *olddquot1, *olddquot2;
96         int                     file_owner;
97         int                     need_iolock = 1;
98
99         xfs_itrace_entry(ip);
100
101         if (mp->m_flags & XFS_MOUNT_RDONLY)
102                 return XFS_ERROR(EROFS);
103
104         if (XFS_FORCED_SHUTDOWN(mp))
105                 return XFS_ERROR(EIO);
106
107         olddquot1 = olddquot2 = NULL;
108         udqp = gdqp = NULL;
109
110         /*
111          * If disk quotas is on, we make sure that the dquots do exist on disk,
112          * before we start any other transactions. Trying to do this later
113          * is messy. We don't care to take a readlock to look at the ids
114          * in inode here, because we can't hold it across the trans_reserve.
115          * If the IDs do change before we take the ilock, we're covered
116          * because the i_*dquot fields will get updated anyway.
117          */
118         if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
119                 uint    qflags = 0;
120
121                 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
122                         uid = iattr->ia_uid;
123                         qflags |= XFS_QMOPT_UQUOTA;
124                 } else {
125                         uid = ip->i_d.di_uid;
126                 }
127                 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
128                         gid = iattr->ia_gid;
129                         qflags |= XFS_QMOPT_GQUOTA;
130                 }  else {
131                         gid = ip->i_d.di_gid;
132                 }
133
134                 /*
135                  * We take a reference when we initialize udqp and gdqp,
136                  * so it is important that we never blindly double trip on
137                  * the same variable. See xfs_create() for an example.
138                  */
139                 ASSERT(udqp == NULL);
140                 ASSERT(gdqp == NULL);
141                 code = XFS_QM_DQVOPALLOC(mp, ip, uid, gid, ip->i_d.di_projid,
142                                          qflags, &udqp, &gdqp);
143                 if (code)
144                         return code;
145         }
146
147         /*
148          * For the other attributes, we acquire the inode lock and
149          * first do an error checking pass.
150          */
151         tp = NULL;
152         lock_flags = XFS_ILOCK_EXCL;
153         if (flags & XFS_ATTR_NOLOCK)
154                 need_iolock = 0;
155         if (!(mask & ATTR_SIZE)) {
156                 if ((mask != (ATTR_CTIME|ATTR_ATIME|ATTR_MTIME)) ||
157                     (mp->m_flags & XFS_MOUNT_WSYNC)) {
158                         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
159                         commit_flags = 0;
160                         if ((code = xfs_trans_reserve(tp, 0,
161                                                      XFS_ICHANGE_LOG_RES(mp), 0,
162                                                      0, 0))) {
163                                 lock_flags = 0;
164                                 goto error_return;
165                         }
166                 }
167         } else {
168                 if (DM_EVENT_ENABLED(ip, DM_EVENT_TRUNCATE) &&
169                     !(flags & XFS_ATTR_DMI)) {
170                         int dmflags = AT_DELAY_FLAG(flags) | DM_SEM_FLAG_WR;
171                         code = XFS_SEND_DATA(mp, DM_EVENT_TRUNCATE, ip,
172                                 iattr->ia_size, 0, dmflags, NULL);
173                         if (code) {
174                                 lock_flags = 0;
175                                 goto error_return;
176                         }
177                 }
178                 if (need_iolock)
179                         lock_flags |= XFS_IOLOCK_EXCL;
180         }
181
182         xfs_ilock(ip, lock_flags);
183
184         /* boolean: are we the file owner? */
185         file_owner = (current_fsuid() == ip->i_d.di_uid);
186
187         /*
188          * Change various properties of a file.
189          * Only the owner or users with CAP_FOWNER
190          * capability may do these things.
191          */
192         if (mask & (ATTR_MODE|ATTR_UID|ATTR_GID)) {
193                 /*
194                  * CAP_FOWNER overrides the following restrictions:
195                  *
196                  * The user ID of the calling process must be equal
197                  * to the file owner ID, except in cases where the
198                  * CAP_FSETID capability is applicable.
199                  */
200                 if (!file_owner && !capable(CAP_FOWNER)) {
201                         code = XFS_ERROR(EPERM);
202                         goto error_return;
203                 }
204
205                 /*
206                  * CAP_FSETID overrides the following restrictions:
207                  *
208                  * The effective user ID of the calling process shall match
209                  * the file owner when setting the set-user-ID and
210                  * set-group-ID bits on that file.
211                  *
212                  * The effective group ID or one of the supplementary group
213                  * IDs of the calling process shall match the group owner of
214                  * the file when setting the set-group-ID bit on that file
215                  */
216                 if (mask & ATTR_MODE) {
217                         mode_t m = 0;
218
219                         if ((iattr->ia_mode & S_ISUID) && !file_owner)
220                                 m |= S_ISUID;
221                         if ((iattr->ia_mode & S_ISGID) &&
222                             !in_group_p((gid_t)ip->i_d.di_gid))
223                                 m |= S_ISGID;
224 #if 0
225                         /* Linux allows this, Irix doesn't. */
226                         if ((iattr->ia_mode & S_ISVTX) && !S_ISDIR(ip->i_d.di_mode))
227                                 m |= S_ISVTX;
228 #endif
229                         if (m && !capable(CAP_FSETID))
230                                 iattr->ia_mode &= ~m;
231                 }
232         }
233
234         /*
235          * Change file ownership.  Must be the owner or privileged.
236          * If the system was configured with the "restricted_chown"
237          * option, the owner is not permitted to give away the file,
238          * and can change the group id only to a group of which he
239          * or she is a member.
240          */
241         if (mask & (ATTR_UID|ATTR_GID)) {
242                 /*
243                  * These IDs could have changed since we last looked at them.
244                  * But, we're assured that if the ownership did change
245                  * while we didn't have the inode locked, inode's dquot(s)
246                  * would have changed also.
247                  */
248                 iuid = ip->i_d.di_uid;
249                 igid = ip->i_d.di_gid;
250                 gid = (mask & ATTR_GID) ? iattr->ia_gid : igid;
251                 uid = (mask & ATTR_UID) ? iattr->ia_uid : iuid;
252
253                 /*
254                  * CAP_CHOWN overrides the following restrictions:
255                  *
256                  * If _POSIX_CHOWN_RESTRICTED is defined, this capability
257                  * shall override the restriction that a process cannot
258                  * change the user ID of a file it owns and the restriction
259                  * that the group ID supplied to the chown() function
260                  * shall be equal to either the group ID or one of the
261                  * supplementary group IDs of the calling process.
262                  */
263                 if (restricted_chown &&
264                     (iuid != uid || (igid != gid &&
265                                      !in_group_p((gid_t)gid))) &&
266                     !capable(CAP_CHOWN)) {
267                         code = XFS_ERROR(EPERM);
268                         goto error_return;
269                 }
270                 /*
271                  * Do a quota reservation only if uid/gid is actually
272                  * going to change.
273                  */
274                 if ((XFS_IS_UQUOTA_ON(mp) && iuid != uid) ||
275                     (XFS_IS_GQUOTA_ON(mp) && igid != gid)) {
276                         ASSERT(tp);
277                         code = XFS_QM_DQVOPCHOWNRESV(mp, tp, ip, udqp, gdqp,
278                                                 capable(CAP_FOWNER) ?
279                                                 XFS_QMOPT_FORCE_RES : 0);
280                         if (code)       /* out of quota */
281                                 goto error_return;
282                 }
283         }
284
285         /*
286          * Truncate file.  Must have write permission and not be a directory.
287          */
288         if (mask & ATTR_SIZE) {
289                 /* Short circuit the truncate case for zero length files */
290                 if (iattr->ia_size == 0 &&
291                     ip->i_size == 0 && ip->i_d.di_nextents == 0) {
292                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
293                         lock_flags &= ~XFS_ILOCK_EXCL;
294                         if (mask & ATTR_CTIME)
295                                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
296                         code = 0;
297                         goto error_return;
298                 }
299
300                 if (S_ISDIR(ip->i_d.di_mode)) {
301                         code = XFS_ERROR(EISDIR);
302                         goto error_return;
303                 } else if (!S_ISREG(ip->i_d.di_mode)) {
304                         code = XFS_ERROR(EINVAL);
305                         goto error_return;
306                 }
307                 /*
308                  * Make sure that the dquots are attached to the inode.
309                  */
310                 if ((code = XFS_QM_DQATTACH(mp, ip, XFS_QMOPT_ILOCKED)))
311                         goto error_return;
312         }
313
314         /*
315          * Change file access or modified times.
316          */
317         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
318                 if (!file_owner) {
319                         if ((mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)) &&
320                             !capable(CAP_FOWNER)) {
321                                 code = XFS_ERROR(EPERM);
322                                 goto error_return;
323                         }
324                 }
325         }
326
327         /*
328          * Now we can make the changes.  Before we join the inode
329          * to the transaction, if ATTR_SIZE is set then take care of
330          * the part of the truncation that must be done without the
331          * inode lock.  This needs to be done before joining the inode
332          * to the transaction, because the inode cannot be unlocked
333          * once it is a part of the transaction.
334          */
335         if (mask & ATTR_SIZE) {
336                 code = 0;
337                 if (iattr->ia_size > ip->i_size) {
338                         /*
339                          * Do the first part of growing a file: zero any data
340                          * in the last block that is beyond the old EOF.  We
341                          * need to do this before the inode is joined to the
342                          * transaction to modify the i_size.
343                          */
344                         code = xfs_zero_eof(ip, iattr->ia_size, ip->i_size);
345                 }
346                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
347
348                 /*
349                  * We are going to log the inode size change in this
350                  * transaction so any previous writes that are beyond the on
351                  * disk EOF and the new EOF that have not been written out need
352                  * to be written here. If we do not write the data out, we
353                  * expose ourselves to the null files problem.
354                  *
355                  * Only flush from the on disk size to the smaller of the in
356                  * memory file size or the new size as that's the range we
357                  * really care about here and prevents waiting for other data
358                  * not within the range we care about here.
359                  */
360                 if (!code &&
361                     ip->i_size != ip->i_d.di_size &&
362                     iattr->ia_size > ip->i_d.di_size) {
363                         code = xfs_flush_pages(ip,
364                                         ip->i_d.di_size, iattr->ia_size,
365                                         XFS_B_ASYNC, FI_NONE);
366                 }
367
368                 /* wait for all I/O to complete */
369                 vn_iowait(ip);
370
371                 if (!code)
372                         code = xfs_itruncate_data(ip, iattr->ia_size);
373                 if (code) {
374                         ASSERT(tp == NULL);
375                         lock_flags &= ~XFS_ILOCK_EXCL;
376                         ASSERT(lock_flags == XFS_IOLOCK_EXCL);
377                         goto error_return;
378                 }
379                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
380                 if ((code = xfs_trans_reserve(tp, 0,
381                                              XFS_ITRUNCATE_LOG_RES(mp), 0,
382                                              XFS_TRANS_PERM_LOG_RES,
383                                              XFS_ITRUNCATE_LOG_COUNT))) {
384                         xfs_trans_cancel(tp, 0);
385                         if (need_iolock)
386                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
387                         return code;
388                 }
389                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
390                 xfs_ilock(ip, XFS_ILOCK_EXCL);
391         }
392
393         if (tp) {
394                 xfs_trans_ijoin(tp, ip, lock_flags);
395                 xfs_trans_ihold(tp, ip);
396         }
397
398         /*
399          * Truncate file.  Must have write permission and not be a directory.
400          */
401         if (mask & ATTR_SIZE) {
402                 /*
403                  * Only change the c/mtime if we are changing the size
404                  * or we are explicitly asked to change it. This handles
405                  * the semantic difference between truncate() and ftruncate()
406                  * as implemented in the VFS.
407                  */
408                 if (iattr->ia_size != ip->i_size || (mask & ATTR_CTIME))
409                         timeflags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
410
411                 if (iattr->ia_size > ip->i_size) {
412                         ip->i_d.di_size = iattr->ia_size;
413                         ip->i_size = iattr->ia_size;
414                         if (!(flags & XFS_ATTR_DMI))
415                                 xfs_ichgtime(ip, XFS_ICHGTIME_CHG);
416                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
417                 } else if (iattr->ia_size <= ip->i_size ||
418                            (iattr->ia_size == 0 && ip->i_d.di_nextents)) {
419                         /*
420                          * signal a sync transaction unless
421                          * we're truncating an already unlinked
422                          * file on a wsync filesystem
423                          */
424                         code = xfs_itruncate_finish(&tp, ip, iattr->ia_size,
425                                             XFS_DATA_FORK,
426                                             ((ip->i_d.di_nlink != 0 ||
427                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
428                                              ? 1 : 0));
429                         if (code)
430                                 goto abort_return;
431                         /*
432                          * Truncated "down", so we're removing references
433                          * to old data here - if we now delay flushing for
434                          * a long time, we expose ourselves unduly to the
435                          * notorious NULL files problem.  So, we mark this
436                          * vnode and flush it when the file is closed, and
437                          * do not wait the usual (long) time for writeout.
438                          */
439                         xfs_iflags_set(ip, XFS_ITRUNCATED);
440                 }
441         }
442
443         /*
444          * Change file access modes.
445          */
446         if (mask & ATTR_MODE) {
447                 ip->i_d.di_mode &= S_IFMT;
448                 ip->i_d.di_mode |= iattr->ia_mode & ~S_IFMT;
449
450                 inode->i_mode &= S_IFMT;
451                 inode->i_mode |= iattr->ia_mode & ~S_IFMT;
452
453                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
454                 timeflags |= XFS_ICHGTIME_CHG;
455         }
456
457         /*
458          * Change file ownership.  Must be the owner or privileged.
459          * If the system was configured with the "restricted_chown"
460          * option, the owner is not permitted to give away the file,
461          * and can change the group id only to a group of which he
462          * or she is a member.
463          */
464         if (mask & (ATTR_UID|ATTR_GID)) {
465                 /*
466                  * CAP_FSETID overrides the following restrictions:
467                  *
468                  * The set-user-ID and set-group-ID bits of a file will be
469                  * cleared upon successful return from chown()
470                  */
471                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
472                     !capable(CAP_FSETID)) {
473                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
474                 }
475
476                 /*
477                  * Change the ownerships and register quota modifications
478                  * in the transaction.
479                  */
480                 if (iuid != uid) {
481                         if (XFS_IS_UQUOTA_ON(mp)) {
482                                 ASSERT(mask & ATTR_UID);
483                                 ASSERT(udqp);
484                                 olddquot1 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
485                                                         &ip->i_udquot, udqp);
486                         }
487                         ip->i_d.di_uid = uid;
488                         inode->i_uid = uid;
489                 }
490                 if (igid != gid) {
491                         if (XFS_IS_GQUOTA_ON(mp)) {
492                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
493                                 ASSERT(mask & ATTR_GID);
494                                 ASSERT(gdqp);
495                                 olddquot2 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
496                                                         &ip->i_gdquot, gdqp);
497                         }
498                         ip->i_d.di_gid = gid;
499                         inode->i_gid = gid;
500                 }
501
502                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
503                 timeflags |= XFS_ICHGTIME_CHG;
504         }
505
506
507         /*
508          * Change file access or modified times.
509          */
510         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
511                 if (mask & ATTR_ATIME) {
512                         inode->i_atime = iattr->ia_atime;
513                         ip->i_d.di_atime.t_sec = iattr->ia_atime.tv_sec;
514                         ip->i_d.di_atime.t_nsec = iattr->ia_atime.tv_nsec;
515                         ip->i_update_core = 1;
516                 }
517                 if (mask & ATTR_MTIME) {
518                         inode->i_mtime = iattr->ia_mtime;
519                         ip->i_d.di_mtime.t_sec = iattr->ia_mtime.tv_sec;
520                         ip->i_d.di_mtime.t_nsec = iattr->ia_mtime.tv_nsec;
521                         timeflags &= ~XFS_ICHGTIME_MOD;
522                         timeflags |= XFS_ICHGTIME_CHG;
523                 }
524                 if (tp && (mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)))
525                         xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
526         }
527
528         /*
529          * Change file inode change time only if ATTR_CTIME set
530          * AND we have been called by a DMI function.
531          */
532
533         if ((flags & XFS_ATTR_DMI) && (mask & ATTR_CTIME)) {
534                 inode->i_ctime = iattr->ia_ctime;
535                 ip->i_d.di_ctime.t_sec = iattr->ia_ctime.tv_sec;
536                 ip->i_d.di_ctime.t_nsec = iattr->ia_ctime.tv_nsec;
537                 ip->i_update_core = 1;
538                 timeflags &= ~XFS_ICHGTIME_CHG;
539         }
540
541         /*
542          * Send out timestamp changes that need to be set to the
543          * current time.  Not done when called by a DMI function.
544          */
545         if (timeflags && !(flags & XFS_ATTR_DMI))
546                 xfs_ichgtime(ip, timeflags);
547
548         XFS_STATS_INC(xs_ig_attrchg);
549
550         /*
551          * If this is a synchronous mount, make sure that the
552          * transaction goes to disk before returning to the user.
553          * This is slightly sub-optimal in that truncates require
554          * two sync transactions instead of one for wsync filesystems.
555          * One for the truncate and one for the timestamps since we
556          * don't want to change the timestamps unless we're sure the
557          * truncate worked.  Truncates are less than 1% of the laddis
558          * mix so this probably isn't worth the trouble to optimize.
559          */
560         code = 0;
561         if (tp) {
562                 if (mp->m_flags & XFS_MOUNT_WSYNC)
563                         xfs_trans_set_sync(tp);
564
565                 code = xfs_trans_commit(tp, commit_flags);
566         }
567
568         xfs_iunlock(ip, lock_flags);
569
570         /*
571          * Release any dquot(s) the inode had kept before chown.
572          */
573         XFS_QM_DQRELE(mp, olddquot1);
574         XFS_QM_DQRELE(mp, olddquot2);
575         XFS_QM_DQRELE(mp, udqp);
576         XFS_QM_DQRELE(mp, gdqp);
577
578         if (code) {
579                 return code;
580         }
581
582         if (DM_EVENT_ENABLED(ip, DM_EVENT_ATTRIBUTE) &&
583             !(flags & XFS_ATTR_DMI)) {
584                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_ATTRIBUTE, ip, DM_RIGHT_NULL,
585                                         NULL, DM_RIGHT_NULL, NULL, NULL,
586                                         0, 0, AT_DELAY_FLAG(flags));
587         }
588         return 0;
589
590  abort_return:
591         commit_flags |= XFS_TRANS_ABORT;
592         /* FALLTHROUGH */
593  error_return:
594         XFS_QM_DQRELE(mp, udqp);
595         XFS_QM_DQRELE(mp, gdqp);
596         if (tp) {
597                 xfs_trans_cancel(tp, commit_flags);
598         }
599         if (lock_flags != 0) {
600                 xfs_iunlock(ip, lock_flags);
601         }
602         return code;
603 }
604
605 /*
606  * The maximum pathlen is 1024 bytes. Since the minimum file system
607  * blocksize is 512 bytes, we can get a max of 2 extents back from
608  * bmapi.
609  */
610 #define SYMLINK_MAPS 2
611
612 STATIC int
613 xfs_readlink_bmap(
614         xfs_inode_t     *ip,
615         char            *link)
616 {
617         xfs_mount_t     *mp = ip->i_mount;
618         int             pathlen = ip->i_d.di_size;
619         int             nmaps = SYMLINK_MAPS;
620         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
621         xfs_daddr_t     d;
622         int             byte_cnt;
623         int             n;
624         xfs_buf_t       *bp;
625         int             error = 0;
626
627         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
628                         mval, &nmaps, NULL, NULL);
629         if (error)
630                 goto out;
631
632         for (n = 0; n < nmaps; n++) {
633                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
634                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
635
636                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
637                 error = XFS_BUF_GETERROR(bp);
638                 if (error) {
639                         xfs_ioerror_alert("xfs_readlink",
640                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
641                         xfs_buf_relse(bp);
642                         goto out;
643                 }
644                 if (pathlen < byte_cnt)
645                         byte_cnt = pathlen;
646                 pathlen -= byte_cnt;
647
648                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
649                 xfs_buf_relse(bp);
650         }
651
652         link[ip->i_d.di_size] = '\0';
653         error = 0;
654
655  out:
656         return error;
657 }
658
659 int
660 xfs_readlink(
661         xfs_inode_t     *ip,
662         char            *link)
663 {
664         xfs_mount_t     *mp = ip->i_mount;
665         int             pathlen;
666         int             error = 0;
667
668         xfs_itrace_entry(ip);
669
670         if (XFS_FORCED_SHUTDOWN(mp))
671                 return XFS_ERROR(EIO);
672
673         xfs_ilock(ip, XFS_ILOCK_SHARED);
674
675         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
676         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
677
678         pathlen = ip->i_d.di_size;
679         if (!pathlen)
680                 goto out;
681
682         if (ip->i_df.if_flags & XFS_IFINLINE) {
683                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
684                 link[pathlen] = '\0';
685         } else {
686                 error = xfs_readlink_bmap(ip, link);
687         }
688
689  out:
690         xfs_iunlock(ip, XFS_ILOCK_SHARED);
691         return error;
692 }
693
694 /*
695  * xfs_fsync
696  *
697  * This is called to sync the inode and its data out to disk.  We need to hold
698  * the I/O lock while flushing the data, and the inode lock while flushing the
699  * inode.  The inode lock CANNOT be held while flushing the data, so acquire
700  * after we're done with that.
701  */
702 int
703 xfs_fsync(
704         xfs_inode_t     *ip)
705 {
706         xfs_trans_t     *tp;
707         int             error;
708         int             log_flushed = 0, changed = 1;
709
710         xfs_itrace_entry(ip);
711
712         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
713                 return XFS_ERROR(EIO);
714
715         /* capture size updates in I/O completion before writing the inode. */
716         error = filemap_fdatawait(VFS_I(ip)->i_mapping);
717         if (error)
718                 return XFS_ERROR(error);
719
720         /*
721          * We always need to make sure that the required inode state is safe on
722          * disk.  The vnode might be clean but we still might need to force the
723          * log because of committed transactions that haven't hit the disk yet.
724          * Likewise, there could be unflushed non-transactional changes to the
725          * inode core that have to go to disk and this requires us to issue
726          * a synchronous transaction to capture these changes correctly.
727          *
728          * This code relies on the assumption that if the update_* fields
729          * of the inode are clear and the inode is unpinned then it is clean
730          * and no action is required.
731          */
732         xfs_ilock(ip, XFS_ILOCK_SHARED);
733
734         if (!(ip->i_update_size || ip->i_update_core)) {
735                 /*
736                  * Timestamps/size haven't changed since last inode flush or
737                  * inode transaction commit.  That means either nothing got
738                  * written or a transaction committed which caught the updates.
739                  * If the latter happened and the transaction hasn't hit the
740                  * disk yet, the inode will be still be pinned.  If it is,
741                  * force the log.
742                  */
743
744                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
745
746                 if (xfs_ipincount(ip)) {
747                         error = _xfs_log_force(ip->i_mount, (xfs_lsn_t)0,
748                                       XFS_LOG_FORCE | XFS_LOG_SYNC,
749                                       &log_flushed);
750                 } else {
751                         /*
752                          * If the inode is not pinned and nothing has changed
753                          * we don't need to flush the cache.
754                          */
755                         changed = 0;
756                 }
757         } else  {
758                 /*
759                  * Kick off a transaction to log the inode core to get the
760                  * updates.  The sync transaction will also force the log.
761                  */
762                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
763                 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
764                 error = xfs_trans_reserve(tp, 0,
765                                 XFS_FSYNC_TS_LOG_RES(ip->i_mount), 0, 0, 0);
766                 if (error) {
767                         xfs_trans_cancel(tp, 0);
768                         return error;
769                 }
770                 xfs_ilock(ip, XFS_ILOCK_EXCL);
771
772                 /*
773                  * Note - it's possible that we might have pushed ourselves out
774                  * of the way during trans_reserve which would flush the inode.
775                  * But there's no guarantee that the inode buffer has actually
776                  * gone out yet (it's delwri).  Plus the buffer could be pinned
777                  * anyway if it's part of an inode in another recent
778                  * transaction.  So we play it safe and fire off the
779                  * transaction anyway.
780                  */
781                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
782                 xfs_trans_ihold(tp, ip);
783                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
784                 xfs_trans_set_sync(tp);
785                 error = _xfs_trans_commit(tp, 0, &log_flushed);
786
787                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
788         }
789
790         if ((ip->i_mount->m_flags & XFS_MOUNT_BARRIER) && changed) {
791                 /*
792                  * If the log write didn't issue an ordered tag we need
793                  * to flush the disk cache for the data device now.
794                  */
795                 if (!log_flushed)
796                         xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
797
798                 /*
799                  * If this inode is on the RT dev we need to flush that
800                  * cache as well.
801                  */
802                 if (XFS_IS_REALTIME_INODE(ip))
803                         xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
804         }
805
806         return error;
807 }
808
809 /*
810  * This is called by xfs_inactive to free any blocks beyond eof
811  * when the link count isn't zero and by xfs_dm_punch_hole() when
812  * punching a hole to EOF.
813  */
814 int
815 xfs_free_eofblocks(
816         xfs_mount_t     *mp,
817         xfs_inode_t     *ip,
818         int             flags)
819 {
820         xfs_trans_t     *tp;
821         int             error;
822         xfs_fileoff_t   end_fsb;
823         xfs_fileoff_t   last_fsb;
824         xfs_filblks_t   map_len;
825         int             nimaps;
826         xfs_bmbt_irec_t imap;
827         int             use_iolock = (flags & XFS_FREE_EOF_LOCK);
828
829         /*
830          * Figure out if there are any blocks beyond the end
831          * of the file.  If not, then there is nothing to do.
832          */
833         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
834         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
835         map_len = last_fsb - end_fsb;
836         if (map_len <= 0)
837                 return 0;
838
839         nimaps = 1;
840         xfs_ilock(ip, XFS_ILOCK_SHARED);
841         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
842                           NULL, 0, &imap, &nimaps, NULL, NULL);
843         xfs_iunlock(ip, XFS_ILOCK_SHARED);
844
845         if (!error && (nimaps != 0) &&
846             (imap.br_startblock != HOLESTARTBLOCK ||
847              ip->i_delayed_blks)) {
848                 /*
849                  * Attach the dquots to the inode up front.
850                  */
851                 if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
852                         return error;
853
854                 /*
855                  * There are blocks after the end of file.
856                  * Free them up now by truncating the file to
857                  * its current size.
858                  */
859                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
860
861                 /*
862                  * Do the xfs_itruncate_start() call before
863                  * reserving any log space because
864                  * itruncate_start will call into the buffer
865                  * cache and we can't
866                  * do that within a transaction.
867                  */
868                 if (use_iolock)
869                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
870                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
871                                     ip->i_size);
872                 if (error) {
873                         xfs_trans_cancel(tp, 0);
874                         if (use_iolock)
875                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
876                         return error;
877                 }
878
879                 error = xfs_trans_reserve(tp, 0,
880                                           XFS_ITRUNCATE_LOG_RES(mp),
881                                           0, XFS_TRANS_PERM_LOG_RES,
882                                           XFS_ITRUNCATE_LOG_COUNT);
883                 if (error) {
884                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
885                         xfs_trans_cancel(tp, 0);
886                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
887                         return error;
888                 }
889
890                 xfs_ilock(ip, XFS_ILOCK_EXCL);
891                 xfs_trans_ijoin(tp, ip,
892                                 XFS_IOLOCK_EXCL |
893                                 XFS_ILOCK_EXCL);
894                 xfs_trans_ihold(tp, ip);
895
896                 error = xfs_itruncate_finish(&tp, ip,
897                                              ip->i_size,
898                                              XFS_DATA_FORK,
899                                              0);
900                 /*
901                  * If we get an error at this point we
902                  * simply don't bother truncating the file.
903                  */
904                 if (error) {
905                         xfs_trans_cancel(tp,
906                                          (XFS_TRANS_RELEASE_LOG_RES |
907                                           XFS_TRANS_ABORT));
908                 } else {
909                         error = xfs_trans_commit(tp,
910                                                 XFS_TRANS_RELEASE_LOG_RES);
911                 }
912                 xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
913                                             : XFS_ILOCK_EXCL));
914         }
915         return error;
916 }
917
918 /*
919  * Free a symlink that has blocks associated with it.
920  */
921 STATIC int
922 xfs_inactive_symlink_rmt(
923         xfs_inode_t     *ip,
924         xfs_trans_t     **tpp)
925 {
926         xfs_buf_t       *bp;
927         int             committed;
928         int             done;
929         int             error;
930         xfs_fsblock_t   first_block;
931         xfs_bmap_free_t free_list;
932         int             i;
933         xfs_mount_t     *mp;
934         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
935         int             nmaps;
936         xfs_trans_t     *ntp;
937         int             size;
938         xfs_trans_t     *tp;
939
940         tp = *tpp;
941         mp = ip->i_mount;
942         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
943         /*
944          * We're freeing a symlink that has some
945          * blocks allocated to it.  Free the
946          * blocks here.  We know that we've got
947          * either 1 or 2 extents and that we can
948          * free them all in one bunmapi call.
949          */
950         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
951         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
952                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
953                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
954                 xfs_trans_cancel(tp, 0);
955                 *tpp = NULL;
956                 return error;
957         }
958         /*
959          * Lock the inode, fix the size, and join it to the transaction.
960          * Hold it so in the normal path, we still have it locked for
961          * the second transaction.  In the error paths we need it
962          * held so the cancel won't rele it, see below.
963          */
964         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
965         size = (int)ip->i_d.di_size;
966         ip->i_d.di_size = 0;
967         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
968         xfs_trans_ihold(tp, ip);
969         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
970         /*
971          * Find the block(s) so we can inval and unmap them.
972          */
973         done = 0;
974         XFS_BMAP_INIT(&free_list, &first_block);
975         nmaps = ARRAY_SIZE(mval);
976         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
977                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
978                         &free_list, NULL)))
979                 goto error0;
980         /*
981          * Invalidate the block(s).
982          */
983         for (i = 0; i < nmaps; i++) {
984                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
985                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
986                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
987                 xfs_trans_binval(tp, bp);
988         }
989         /*
990          * Unmap the dead block(s) to the free_list.
991          */
992         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
993                         &first_block, &free_list, NULL, &done)))
994                 goto error1;
995         ASSERT(done);
996         /*
997          * Commit the first transaction.  This logs the EFI and the inode.
998          */
999         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
1000                 goto error1;
1001         /*
1002          * The transaction must have been committed, since there were
1003          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
1004          * The new tp has the extent freeing and EFDs.
1005          */
1006         ASSERT(committed);
1007         /*
1008          * The first xact was committed, so add the inode to the new one.
1009          * Mark it dirty so it will be logged and moved forward in the log as
1010          * part of every commit.
1011          */
1012         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1013         xfs_trans_ihold(tp, ip);
1014         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1015         /*
1016          * Get a new, empty transaction to return to our caller.
1017          */
1018         ntp = xfs_trans_dup(tp);
1019         /*
1020          * Commit the transaction containing extent freeing and EFDs.
1021          * If we get an error on the commit here or on the reserve below,
1022          * we need to unlock the inode since the new transaction doesn't
1023          * have the inode attached.
1024          */
1025         error = xfs_trans_commit(tp, 0);
1026         tp = ntp;
1027         if (error) {
1028                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1029                 goto error0;
1030         }
1031         /*
1032          * Remove the memory for extent descriptions (just bookkeeping).
1033          */
1034         if (ip->i_df.if_bytes)
1035                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
1036         ASSERT(ip->i_df.if_bytes == 0);
1037         /*
1038          * Put an itruncate log reservation in the new transaction
1039          * for our caller.
1040          */
1041         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
1042                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
1043                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1044                 goto error0;
1045         }
1046         /*
1047          * Return with the inode locked but not joined to the transaction.
1048          */
1049         *tpp = tp;
1050         return 0;
1051
1052  error1:
1053         xfs_bmap_cancel(&free_list);
1054  error0:
1055         /*
1056          * Have to come here with the inode locked and either
1057          * (held and in the transaction) or (not in the transaction).
1058          * If the inode isn't held then cancel would iput it, but
1059          * that's wrong since this is inactive and the vnode ref
1060          * count is 0 already.
1061          * Cancel won't do anything to the inode if held, but it still
1062          * needs to be locked until the cancel is done, if it was
1063          * joined to the transaction.
1064          */
1065         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1066         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1067         *tpp = NULL;
1068         return error;
1069
1070 }
1071
1072 STATIC int
1073 xfs_inactive_symlink_local(
1074         xfs_inode_t     *ip,
1075         xfs_trans_t     **tpp)
1076 {
1077         int             error;
1078
1079         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
1080         /*
1081          * We're freeing a symlink which fit into
1082          * the inode.  Just free the memory used
1083          * to hold the old symlink.
1084          */
1085         error = xfs_trans_reserve(*tpp, 0,
1086                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
1087                                   0, XFS_TRANS_PERM_LOG_RES,
1088                                   XFS_ITRUNCATE_LOG_COUNT);
1089
1090         if (error) {
1091                 xfs_trans_cancel(*tpp, 0);
1092                 *tpp = NULL;
1093                 return error;
1094         }
1095         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1096
1097         /*
1098          * Zero length symlinks _can_ exist.
1099          */
1100         if (ip->i_df.if_bytes > 0) {
1101                 xfs_idata_realloc(ip,
1102                                   -(ip->i_df.if_bytes),
1103                                   XFS_DATA_FORK);
1104                 ASSERT(ip->i_df.if_bytes == 0);
1105         }
1106         return 0;
1107 }
1108
1109 STATIC int
1110 xfs_inactive_attrs(
1111         xfs_inode_t     *ip,
1112         xfs_trans_t     **tpp)
1113 {
1114         xfs_trans_t     *tp;
1115         int             error;
1116         xfs_mount_t     *mp;
1117
1118         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1119         tp = *tpp;
1120         mp = ip->i_mount;
1121         ASSERT(ip->i_d.di_forkoff != 0);
1122         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1123         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1124         if (error)
1125                 goto error_unlock;
1126
1127         error = xfs_attr_inactive(ip);
1128         if (error)
1129                 goto error_unlock;
1130
1131         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1132         error = xfs_trans_reserve(tp, 0,
1133                                   XFS_IFREE_LOG_RES(mp),
1134                                   0, XFS_TRANS_PERM_LOG_RES,
1135                                   XFS_INACTIVE_LOG_COUNT);
1136         if (error)
1137                 goto error_cancel;
1138
1139         xfs_ilock(ip, XFS_ILOCK_EXCL);
1140         xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1141         xfs_trans_ihold(tp, ip);
1142         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1143
1144         ASSERT(ip->i_d.di_anextents == 0);
1145
1146         *tpp = tp;
1147         return 0;
1148
1149 error_cancel:
1150         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1151         xfs_trans_cancel(tp, 0);
1152 error_unlock:
1153         *tpp = NULL;
1154         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1155         return error;
1156 }
1157
1158 int
1159 xfs_release(
1160         xfs_inode_t     *ip)
1161 {
1162         xfs_mount_t     *mp = ip->i_mount;
1163         int             error;
1164
1165         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
1166                 return 0;
1167
1168         /* If this is a read-only mount, don't do this (would generate I/O) */
1169         if (mp->m_flags & XFS_MOUNT_RDONLY)
1170                 return 0;
1171
1172         if (!XFS_FORCED_SHUTDOWN(mp)) {
1173                 int truncated;
1174
1175                 /*
1176                  * If we are using filestreams, and we have an unlinked
1177                  * file that we are processing the last close on, then nothing
1178                  * will be able to reopen and write to this file. Purge this
1179                  * inode from the filestreams cache so that it doesn't delay
1180                  * teardown of the inode.
1181                  */
1182                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
1183                         xfs_filestream_deassociate(ip);
1184
1185                 /*
1186                  * If we previously truncated this file and removed old data
1187                  * in the process, we want to initiate "early" writeout on
1188                  * the last close.  This is an attempt to combat the notorious
1189                  * NULL files problem which is particularly noticable from a
1190                  * truncate down, buffered (re-)write (delalloc), followed by
1191                  * a crash.  What we are effectively doing here is
1192                  * significantly reducing the time window where we'd otherwise
1193                  * be exposed to that problem.
1194                  */
1195                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1196                 if (truncated && VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
1197                         xfs_flush_pages(ip, 0, -1, XFS_B_ASYNC, FI_NONE);
1198         }
1199
1200         if (ip->i_d.di_nlink != 0) {
1201                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1202                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1203                        ip->i_delayed_blks > 0)) &&
1204                      (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
1205                     (!(ip->i_d.di_flags &
1206                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
1207                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1208                         if (error)
1209                                 return error;
1210                 }
1211         }
1212
1213         return 0;
1214 }
1215
1216 /*
1217  * xfs_inactive
1218  *
1219  * This is called when the vnode reference count for the vnode
1220  * goes to zero.  If the file has been unlinked, then it must
1221  * now be truncated.  Also, we clear all of the read-ahead state
1222  * kept for the inode here since the file is now closed.
1223  */
1224 int
1225 xfs_inactive(
1226         xfs_inode_t     *ip)
1227 {
1228         xfs_bmap_free_t free_list;
1229         xfs_fsblock_t   first_block;
1230         int             committed;
1231         xfs_trans_t     *tp;
1232         xfs_mount_t     *mp;
1233         int             error;
1234         int             truncate;
1235
1236         xfs_itrace_entry(ip);
1237
1238         /*
1239          * If the inode is already free, then there can be nothing
1240          * to clean up here.
1241          */
1242         if (ip->i_d.di_mode == 0 || VN_BAD(VFS_I(ip))) {
1243                 ASSERT(ip->i_df.if_real_bytes == 0);
1244                 ASSERT(ip->i_df.if_broot_bytes == 0);
1245                 return VN_INACTIVE_CACHE;
1246         }
1247
1248         /*
1249          * Only do a truncate if it's a regular file with
1250          * some actual space in it.  It's OK to look at the
1251          * inode's fields without the lock because we're the
1252          * only one with a reference to the inode.
1253          */
1254         truncate = ((ip->i_d.di_nlink == 0) &&
1255             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1256              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1257             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1258
1259         mp = ip->i_mount;
1260
1261         if (ip->i_d.di_nlink == 0 && DM_EVENT_ENABLED(ip, DM_EVENT_DESTROY))
1262                 XFS_SEND_DESTROY(mp, ip, DM_RIGHT_NULL);
1263
1264         error = 0;
1265
1266         /* If this is a read-only mount, don't do this (would generate I/O) */
1267         if (mp->m_flags & XFS_MOUNT_RDONLY)
1268                 goto out;
1269
1270         if (ip->i_d.di_nlink != 0) {
1271                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1272                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1273                        ip->i_delayed_blks > 0)) &&
1274                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1275                      (!(ip->i_d.di_flags &
1276                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1277                       (ip->i_delayed_blks != 0)))) {
1278                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1279                         if (error)
1280                                 return VN_INACTIVE_CACHE;
1281                 }
1282                 goto out;
1283         }
1284
1285         ASSERT(ip->i_d.di_nlink == 0);
1286
1287         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
1288                 return VN_INACTIVE_CACHE;
1289
1290         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1291         if (truncate) {
1292                 /*
1293                  * Do the xfs_itruncate_start() call before
1294                  * reserving any log space because itruncate_start
1295                  * will call into the buffer cache and we can't
1296                  * do that within a transaction.
1297                  */
1298                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1299
1300                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1301                 if (error) {
1302                         xfs_trans_cancel(tp, 0);
1303                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1304                         return VN_INACTIVE_CACHE;
1305                 }
1306
1307                 error = xfs_trans_reserve(tp, 0,
1308                                           XFS_ITRUNCATE_LOG_RES(mp),
1309                                           0, XFS_TRANS_PERM_LOG_RES,
1310                                           XFS_ITRUNCATE_LOG_COUNT);
1311                 if (error) {
1312                         /* Don't call itruncate_cleanup */
1313                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1314                         xfs_trans_cancel(tp, 0);
1315                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1316                         return VN_INACTIVE_CACHE;
1317                 }
1318
1319                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1320                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1321                 xfs_trans_ihold(tp, ip);
1322
1323                 /*
1324                  * normally, we have to run xfs_itruncate_finish sync.
1325                  * But if filesystem is wsync and we're in the inactive
1326                  * path, then we know that nlink == 0, and that the
1327                  * xaction that made nlink == 0 is permanently committed
1328                  * since xfs_remove runs as a synchronous transaction.
1329                  */
1330                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1331                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1332
1333                 if (error) {
1334                         xfs_trans_cancel(tp,
1335                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1336                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1337                         return VN_INACTIVE_CACHE;
1338                 }
1339         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1340
1341                 /*
1342                  * If we get an error while cleaning up a
1343                  * symlink we bail out.
1344                  */
1345                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1346                         xfs_inactive_symlink_rmt(ip, &tp) :
1347                         xfs_inactive_symlink_local(ip, &tp);
1348
1349                 if (error) {
1350                         ASSERT(tp == NULL);
1351                         return VN_INACTIVE_CACHE;
1352                 }
1353
1354                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1355                 xfs_trans_ihold(tp, ip);
1356         } else {
1357                 error = xfs_trans_reserve(tp, 0,
1358                                           XFS_IFREE_LOG_RES(mp),
1359                                           0, XFS_TRANS_PERM_LOG_RES,
1360                                           XFS_INACTIVE_LOG_COUNT);
1361                 if (error) {
1362                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1363                         xfs_trans_cancel(tp, 0);
1364                         return VN_INACTIVE_CACHE;
1365                 }
1366
1367                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1368                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1369                 xfs_trans_ihold(tp, ip);
1370         }
1371
1372         /*
1373          * If there are attributes associated with the file
1374          * then blow them away now.  The code calls a routine
1375          * that recursively deconstructs the attribute fork.
1376          * We need to just commit the current transaction
1377          * because we can't use it for xfs_attr_inactive().
1378          */
1379         if (ip->i_d.di_anextents > 0) {
1380                 error = xfs_inactive_attrs(ip, &tp);
1381                 /*
1382                  * If we got an error, the transaction is already
1383                  * cancelled, and the inode is unlocked. Just get out.
1384                  */
1385                  if (error)
1386                          return VN_INACTIVE_CACHE;
1387         } else if (ip->i_afp) {
1388                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1389         }
1390
1391         /*
1392          * Free the inode.
1393          */
1394         XFS_BMAP_INIT(&free_list, &first_block);
1395         error = xfs_ifree(tp, ip, &free_list);
1396         if (error) {
1397                 /*
1398                  * If we fail to free the inode, shut down.  The cancel
1399                  * might do that, we need to make sure.  Otherwise the
1400                  * inode might be lost for a long time or forever.
1401                  */
1402                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1403                         cmn_err(CE_NOTE,
1404                 "xfs_inactive:  xfs_ifree() returned an error = %d on %s",
1405                                 error, mp->m_fsname);
1406                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1407                 }
1408                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1409         } else {
1410                 /*
1411                  * Credit the quota account(s). The inode is gone.
1412                  */
1413                 XFS_TRANS_MOD_DQUOT_BYINO(mp, tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1414
1415                 /*
1416                  * Just ignore errors at this point.  There is nothing we can
1417                  * do except to try to keep going. Make sure it's not a silent
1418                  * error.
1419                  */
1420                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1421                 if (error)
1422                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1423                                 "xfs_bmap_finish() returned error %d", error);
1424                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1425                 if (error)
1426                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1427                                 "xfs_trans_commit() returned error %d", error);
1428         }
1429         /*
1430          * Release the dquots held by inode, if any.
1431          */
1432         XFS_QM_DQDETACH(mp, ip);
1433
1434         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1435
1436  out:
1437         return VN_INACTIVE_CACHE;
1438 }
1439
1440 /*
1441  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1442  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1443  * ci_name->name will point to a the actual name (caller must free) or
1444  * will be set to NULL if an exact match is found.
1445  */
1446 int
1447 xfs_lookup(
1448         xfs_inode_t             *dp,
1449         struct xfs_name         *name,
1450         xfs_inode_t             **ipp,
1451         struct xfs_name         *ci_name)
1452 {
1453         xfs_ino_t               inum;
1454         int                     error;
1455         uint                    lock_mode;
1456
1457         xfs_itrace_entry(dp);
1458
1459         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1460                 return XFS_ERROR(EIO);
1461
1462         lock_mode = xfs_ilock_map_shared(dp);
1463         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1464         xfs_iunlock_map_shared(dp, lock_mode);
1465
1466         if (error)
1467                 goto out;
1468
1469         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp, 0);
1470         if (error)
1471                 goto out_free_name;
1472
1473         xfs_itrace_ref(*ipp);
1474         return 0;
1475
1476 out_free_name:
1477         if (ci_name)
1478                 kmem_free(ci_name->name);
1479 out:
1480         *ipp = NULL;
1481         return error;
1482 }
1483
1484 int
1485 xfs_create(
1486         xfs_inode_t             *dp,
1487         struct xfs_name         *name,
1488         mode_t                  mode,
1489         xfs_dev_t               rdev,
1490         xfs_inode_t             **ipp,
1491         cred_t                  *credp)
1492 {
1493         xfs_mount_t             *mp = dp->i_mount;
1494         xfs_inode_t             *ip;
1495         xfs_trans_t             *tp;
1496         int                     error;
1497         xfs_bmap_free_t         free_list;
1498         xfs_fsblock_t           first_block;
1499         boolean_t               unlock_dp_on_error = B_FALSE;
1500         int                     dm_event_sent = 0;
1501         uint                    cancel_flags;
1502         int                     committed;
1503         xfs_prid_t              prid;
1504         struct xfs_dquot        *udqp, *gdqp;
1505         uint                    resblks;
1506
1507         ASSERT(!*ipp);
1508         xfs_itrace_entry(dp);
1509
1510         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
1511                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
1512                                 dp, DM_RIGHT_NULL, NULL,
1513                                 DM_RIGHT_NULL, name->name, NULL,
1514                                 mode, 0, 0);
1515
1516                 if (error)
1517                         return error;
1518                 dm_event_sent = 1;
1519         }
1520
1521         if (XFS_FORCED_SHUTDOWN(mp))
1522                 return XFS_ERROR(EIO);
1523
1524         /* Return through std_return after this point. */
1525
1526         udqp = gdqp = NULL;
1527         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1528                 prid = dp->i_d.di_projid;
1529         else
1530                 prid = (xfs_prid_t)dfltprid;
1531
1532         /*
1533          * Make sure that we have allocated dquot(s) on disk.
1534          */
1535         error = XFS_QM_DQVOPALLOC(mp, dp,
1536                         current_fsuid(), current_fsgid(), prid,
1537                         XFS_QMOPT_QUOTALL|XFS_QMOPT_INHERIT, &udqp, &gdqp);
1538         if (error)
1539                 goto std_return;
1540
1541         ip = NULL;
1542
1543         tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1544         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1545         resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1546         /*
1547          * Initially assume that the file does not exist and
1548          * reserve the resources for that case.  If that is not
1549          * the case we'll drop the one we have and get a more
1550          * appropriate transaction later.
1551          */
1552         error = xfs_trans_reserve(tp, resblks, XFS_CREATE_LOG_RES(mp), 0,
1553                         XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1554         if (error == ENOSPC) {
1555                 resblks = 0;
1556                 error = xfs_trans_reserve(tp, 0, XFS_CREATE_LOG_RES(mp), 0,
1557                                 XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1558         }
1559         if (error) {
1560                 cancel_flags = 0;
1561                 goto error_return;
1562         }
1563
1564         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1565         unlock_dp_on_error = B_TRUE;
1566
1567         XFS_BMAP_INIT(&free_list, &first_block);
1568
1569         ASSERT(ip == NULL);
1570
1571         /*
1572          * Reserve disk quota and the inode.
1573          */
1574         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
1575         if (error)
1576                 goto error_return;
1577
1578         error = xfs_dir_canenter(tp, dp, name, resblks);
1579         if (error)
1580                 goto error_return;
1581         error = xfs_dir_ialloc(&tp, dp, mode, 1,
1582                         rdev, credp, prid, resblks > 0,
1583                         &ip, &committed);
1584         if (error) {
1585                 if (error == ENOSPC)
1586                         goto error_return;
1587                 goto abort_return;
1588         }
1589         xfs_itrace_ref(ip);
1590
1591         /*
1592          * At this point, we've gotten a newly allocated inode.
1593          * It is locked (and joined to the transaction).
1594          */
1595
1596         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1597
1598         /*
1599          * Now we join the directory inode to the transaction.  We do not do it
1600          * earlier because xfs_dir_ialloc might commit the previous transaction
1601          * (and release all the locks).  An error from here on will result in
1602          * the transaction cancel unlocking dp so don't do it explicitly in the
1603          * error path.
1604          */
1605         IHOLD(dp);
1606         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1607         unlock_dp_on_error = B_FALSE;
1608
1609         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1610                                         &first_block, &free_list, resblks ?
1611                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1612         if (error) {
1613                 ASSERT(error != ENOSPC);
1614                 goto abort_return;
1615         }
1616         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1617         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1618
1619         /*
1620          * If this is a synchronous mount, make sure that the
1621          * create transaction goes to disk before returning to
1622          * the user.
1623          */
1624         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1625                 xfs_trans_set_sync(tp);
1626         }
1627
1628         dp->i_gen++;
1629
1630         /*
1631          * Attach the dquot(s) to the inodes and modify them incore.
1632          * These ids of the inode couldn't have changed since the new
1633          * inode has been locked ever since it was created.
1634          */
1635         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
1636
1637         /*
1638          * xfs_trans_commit normally decrements the vnode ref count
1639          * when it unlocks the inode. Since we want to return the
1640          * vnode to the caller, we bump the vnode ref count now.
1641          */
1642         IHOLD(ip);
1643
1644         error = xfs_bmap_finish(&tp, &free_list, &committed);
1645         if (error) {
1646                 xfs_bmap_cancel(&free_list);
1647                 goto abort_rele;
1648         }
1649
1650         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1651         if (error) {
1652                 IRELE(ip);
1653                 tp = NULL;
1654                 goto error_return;
1655         }
1656
1657         XFS_QM_DQRELE(mp, udqp);
1658         XFS_QM_DQRELE(mp, gdqp);
1659
1660         *ipp = ip;
1661
1662         /* Fallthrough to std_return with error = 0  */
1663
1664 std_return:
1665         if ((*ipp || (error != 0 && dm_event_sent != 0)) &&
1666             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
1667                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
1668                         dp, DM_RIGHT_NULL,
1669                         *ipp ? ip : NULL,
1670                         DM_RIGHT_NULL, name->name, NULL,
1671                         mode, error, 0);
1672         }
1673         return error;
1674
1675  abort_return:
1676         cancel_flags |= XFS_TRANS_ABORT;
1677         /* FALLTHROUGH */
1678
1679  error_return:
1680         if (tp != NULL)
1681                 xfs_trans_cancel(tp, cancel_flags);
1682
1683         XFS_QM_DQRELE(mp, udqp);
1684         XFS_QM_DQRELE(mp, gdqp);
1685
1686         if (unlock_dp_on_error)
1687                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1688
1689         goto std_return;
1690
1691  abort_rele:
1692         /*
1693          * Wait until after the current transaction is aborted to
1694          * release the inode.  This prevents recursive transactions
1695          * and deadlocks from xfs_inactive.
1696          */
1697         cancel_flags |= XFS_TRANS_ABORT;
1698         xfs_trans_cancel(tp, cancel_flags);
1699         IRELE(ip);
1700
1701         XFS_QM_DQRELE(mp, udqp);
1702         XFS_QM_DQRELE(mp, gdqp);
1703
1704         goto std_return;
1705 }
1706
1707 #ifdef DEBUG
1708 int xfs_locked_n;
1709 int xfs_small_retries;
1710 int xfs_middle_retries;
1711 int xfs_lots_retries;
1712 int xfs_lock_delays;
1713 #endif
1714
1715 /*
1716  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1717  * a different value
1718  */
1719 static inline int
1720 xfs_lock_inumorder(int lock_mode, int subclass)
1721 {
1722         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1723                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1724         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1725                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1726
1727         return lock_mode;
1728 }
1729
1730 /*
1731  * The following routine will lock n inodes in exclusive mode.
1732  * We assume the caller calls us with the inodes in i_ino order.
1733  *
1734  * We need to detect deadlock where an inode that we lock
1735  * is in the AIL and we start waiting for another inode that is locked
1736  * by a thread in a long running transaction (such as truncate). This can
1737  * result in deadlock since the long running trans might need to wait
1738  * for the inode we just locked in order to push the tail and free space
1739  * in the log.
1740  */
1741 void
1742 xfs_lock_inodes(
1743         xfs_inode_t     **ips,
1744         int             inodes,
1745         uint            lock_mode)
1746 {
1747         int             attempts = 0, i, j, try_lock;
1748         xfs_log_item_t  *lp;
1749
1750         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1751
1752         try_lock = 0;
1753         i = 0;
1754
1755 again:
1756         for (; i < inodes; i++) {
1757                 ASSERT(ips[i]);
1758
1759                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1760                         continue;
1761
1762                 /*
1763                  * If try_lock is not set yet, make sure all locked inodes
1764                  * are not in the AIL.
1765                  * If any are, set try_lock to be used later.
1766                  */
1767
1768                 if (!try_lock) {
1769                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1770                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1771                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1772                                         try_lock++;
1773                                 }
1774                         }
1775                 }
1776
1777                 /*
1778                  * If any of the previous locks we have locked is in the AIL,
1779                  * we must TRY to get the second and subsequent locks. If
1780                  * we can't get any, we must release all we have
1781                  * and try again.
1782                  */
1783
1784                 if (try_lock) {
1785                         /* try_lock must be 0 if i is 0. */
1786                         /*
1787                          * try_lock means we have an inode locked
1788                          * that is in the AIL.
1789                          */
1790                         ASSERT(i != 0);
1791                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1792                                 attempts++;
1793
1794                                 /*
1795                                  * Unlock all previous guys and try again.
1796                                  * xfs_iunlock will try to push the tail
1797                                  * if the inode is in the AIL.
1798                                  */
1799
1800                                 for(j = i - 1; j >= 0; j--) {
1801
1802                                         /*
1803                                          * Check to see if we've already
1804                                          * unlocked this one.
1805                                          * Not the first one going back,
1806                                          * and the inode ptr is the same.
1807                                          */
1808                                         if ((j != (i - 1)) && ips[j] ==
1809                                                                 ips[j+1])
1810                                                 continue;
1811
1812                                         xfs_iunlock(ips[j], lock_mode);
1813                                 }
1814
1815                                 if ((attempts % 5) == 0) {
1816                                         delay(1); /* Don't just spin the CPU */
1817 #ifdef DEBUG
1818                                         xfs_lock_delays++;
1819 #endif
1820                                 }
1821                                 i = 0;
1822                                 try_lock = 0;
1823                                 goto again;
1824                         }
1825                 } else {
1826                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1827                 }
1828         }
1829
1830 #ifdef DEBUG
1831         if (attempts) {
1832                 if (attempts < 5) xfs_small_retries++;
1833                 else if (attempts < 100) xfs_middle_retries++;
1834                 else xfs_lots_retries++;
1835         } else {
1836                 xfs_locked_n++;
1837         }
1838 #endif
1839 }
1840
1841 void
1842 xfs_lock_two_inodes(
1843         xfs_inode_t             *ip0,
1844         xfs_inode_t             *ip1,
1845         uint                    lock_mode)
1846 {
1847         xfs_inode_t             *temp;
1848         int                     attempts = 0;
1849         xfs_log_item_t          *lp;
1850
1851         ASSERT(ip0->i_ino != ip1->i_ino);
1852
1853         if (ip0->i_ino > ip1->i_ino) {
1854                 temp = ip0;
1855                 ip0 = ip1;
1856                 ip1 = temp;
1857         }
1858
1859  again:
1860         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1861
1862         /*
1863          * If the first lock we have locked is in the AIL, we must TRY to get
1864          * the second lock. If we can't get it, we must release the first one
1865          * and try again.
1866          */
1867         lp = (xfs_log_item_t *)ip0->i_itemp;
1868         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1869                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1870                         xfs_iunlock(ip0, lock_mode);
1871                         if ((++attempts % 5) == 0)
1872                                 delay(1); /* Don't just spin the CPU */
1873                         goto again;
1874                 }
1875         } else {
1876                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1877         }
1878 }
1879
1880 int
1881 xfs_remove(
1882         xfs_inode_t             *dp,
1883         struct xfs_name         *name,
1884         xfs_inode_t             *ip)
1885 {
1886         xfs_mount_t             *mp = dp->i_mount;
1887         xfs_trans_t             *tp = NULL;
1888         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1889         int                     error = 0;
1890         xfs_bmap_free_t         free_list;
1891         xfs_fsblock_t           first_block;
1892         int                     cancel_flags;
1893         int                     committed;
1894         int                     link_zero;
1895         uint                    resblks;
1896         uint                    log_count;
1897
1898         xfs_itrace_entry(dp);
1899         xfs_itrace_entry(ip);
1900
1901         if (XFS_FORCED_SHUTDOWN(mp))
1902                 return XFS_ERROR(EIO);
1903
1904         if (DM_EVENT_ENABLED(dp, DM_EVENT_REMOVE)) {
1905                 error = XFS_SEND_NAMESP(mp, DM_EVENT_REMOVE, dp, DM_RIGHT_NULL,
1906                                         NULL, DM_RIGHT_NULL, name->name, NULL,
1907                                         ip->i_d.di_mode, 0, 0);
1908                 if (error)
1909                         return error;
1910         }
1911
1912         error = XFS_QM_DQATTACH(mp, dp, 0);
1913         if (error)
1914                 goto std_return;
1915
1916         error = XFS_QM_DQATTACH(mp, ip, 0);
1917         if (error)
1918                 goto std_return;
1919
1920         if (is_dir) {
1921                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1922                 log_count = XFS_DEFAULT_LOG_COUNT;
1923         } else {
1924                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1925                 log_count = XFS_REMOVE_LOG_COUNT;
1926         }
1927         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1928
1929         /*
1930          * We try to get the real space reservation first,
1931          * allowing for directory btree deletion(s) implying
1932          * possible bmap insert(s).  If we can't get the space
1933          * reservation then we use 0 instead, and avoid the bmap
1934          * btree insert(s) in the directory code by, if the bmap
1935          * insert tries to happen, instead trimming the LAST
1936          * block from the directory.
1937          */
1938         resblks = XFS_REMOVE_SPACE_RES(mp);
1939         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1940                                   XFS_TRANS_PERM_LOG_RES, log_count);
1941         if (error == ENOSPC) {
1942                 resblks = 0;
1943                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1944                                           XFS_TRANS_PERM_LOG_RES, log_count);
1945         }
1946         if (error) {
1947                 ASSERT(error != ENOSPC);
1948                 cancel_flags = 0;
1949                 goto out_trans_cancel;
1950         }
1951
1952         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1953
1954         /*
1955          * At this point, we've gotten both the directory and the entry
1956          * inodes locked.
1957          */
1958         IHOLD(ip);
1959         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1960
1961         IHOLD(dp);
1962         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1963
1964         /*
1965          * If we're removing a directory perform some additional validation.
1966          */
1967         if (is_dir) {
1968                 ASSERT(ip->i_d.di_nlink >= 2);
1969                 if (ip->i_d.di_nlink != 2) {
1970                         error = XFS_ERROR(ENOTEMPTY);
1971                         goto out_trans_cancel;
1972                 }
1973                 if (!xfs_dir_isempty(ip)) {
1974                         error = XFS_ERROR(ENOTEMPTY);
1975                         goto out_trans_cancel;
1976                 }
1977         }
1978
1979         XFS_BMAP_INIT(&free_list, &first_block);
1980         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1981                                         &first_block, &free_list, resblks);
1982         if (error) {
1983                 ASSERT(error != ENOENT);
1984                 goto out_bmap_cancel;
1985         }
1986         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1987
1988         /*
1989          * Bump the in memory generation count on the parent
1990          * directory so that other can know that it has changed.
1991          */
1992         dp->i_gen++;
1993         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1994
1995         if (is_dir) {
1996                 /*
1997                  * Drop the link from ip's "..".
1998                  */
1999                 error = xfs_droplink(tp, dp);
2000                 if (error)
2001                         goto out_bmap_cancel;
2002
2003                 /*
2004                  * Drop the link from dp to ip.
2005                  */
2006                 error = xfs_droplink(tp, ip);
2007                 if (error)
2008                         goto out_bmap_cancel;
2009         } else {
2010                 /*
2011                  * When removing a non-directory we need to log the parent
2012                  * inode here for the i_gen update.  For a directory this is
2013                  * done implicitly by the xfs_droplink call for the ".." entry.
2014                  */
2015                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2016         }
2017
2018         /*
2019          * Drop the "." link from ip to self.
2020          */
2021         error = xfs_droplink(tp, ip);
2022         if (error)
2023                 goto out_bmap_cancel;
2024
2025         /*
2026          * Determine if this is the last link while
2027          * we are in the transaction.
2028          */
2029         link_zero = (ip->i_d.di_nlink == 0);
2030
2031         /*
2032          * If this is a synchronous mount, make sure that the
2033          * remove transaction goes to disk before returning to
2034          * the user.
2035          */
2036         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2037                 xfs_trans_set_sync(tp);
2038
2039         error = xfs_bmap_finish(&tp, &free_list, &committed);
2040         if (error)
2041                 goto out_bmap_cancel;
2042
2043         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2044         if (error)
2045                 goto std_return;
2046
2047         /*
2048          * If we are using filestreams, kill the stream association.
2049          * If the file is still open it may get a new one but that
2050          * will get killed on last close in xfs_close() so we don't
2051          * have to worry about that.
2052          */
2053         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
2054                 xfs_filestream_deassociate(ip);
2055
2056         xfs_itrace_exit(ip);
2057         xfs_itrace_exit(dp);
2058
2059  std_return:
2060         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTREMOVE)) {
2061                 XFS_SEND_NAMESP(mp, DM_EVENT_POSTREMOVE, dp, DM_RIGHT_NULL,
2062                                 NULL, DM_RIGHT_NULL, name->name, NULL,
2063                                 ip->i_d.di_mode, error, 0);
2064         }
2065
2066         return error;
2067
2068  out_bmap_cancel:
2069         xfs_bmap_cancel(&free_list);
2070         cancel_flags |= XFS_TRANS_ABORT;
2071  out_trans_cancel:
2072         xfs_trans_cancel(tp, cancel_flags);
2073         goto std_return;
2074 }
2075
2076 int
2077 xfs_link(
2078         xfs_inode_t             *tdp,
2079         xfs_inode_t             *sip,
2080         struct xfs_name         *target_name)
2081 {
2082         xfs_mount_t             *mp = tdp->i_mount;
2083         xfs_trans_t             *tp;
2084         int                     error;
2085         xfs_bmap_free_t         free_list;
2086         xfs_fsblock_t           first_block;
2087         int                     cancel_flags;
2088         int                     committed;
2089         int                     resblks;
2090
2091         xfs_itrace_entry(tdp);
2092         xfs_itrace_entry(sip);
2093
2094         ASSERT(!S_ISDIR(sip->i_d.di_mode));
2095
2096         if (XFS_FORCED_SHUTDOWN(mp))
2097                 return XFS_ERROR(EIO);
2098
2099         if (DM_EVENT_ENABLED(tdp, DM_EVENT_LINK)) {
2100                 error = XFS_SEND_NAMESP(mp, DM_EVENT_LINK,
2101                                         tdp, DM_RIGHT_NULL,
2102                                         sip, DM_RIGHT_NULL,
2103                                         target_name->name, NULL, 0, 0, 0);
2104                 if (error)
2105                         return error;
2106         }
2107
2108         /* Return through std_return after this point. */
2109
2110         error = XFS_QM_DQATTACH(mp, sip, 0);
2111         if (!error && sip != tdp)
2112                 error = XFS_QM_DQATTACH(mp, tdp, 0);
2113         if (error)
2114                 goto std_return;
2115
2116         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
2117         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2118         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
2119         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
2120                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2121         if (error == ENOSPC) {
2122                 resblks = 0;
2123                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
2124                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2125         }
2126         if (error) {
2127                 cancel_flags = 0;
2128                 goto error_return;
2129         }
2130
2131         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
2132
2133         /*
2134          * Increment vnode ref counts since xfs_trans_commit &
2135          * xfs_trans_cancel will both unlock the inodes and
2136          * decrement the associated ref counts.
2137          */
2138         IHOLD(sip);
2139         IHOLD(tdp);
2140         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
2141         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
2142
2143         /*
2144          * If the source has too many links, we can't make any more to it.
2145          */
2146         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
2147                 error = XFS_ERROR(EMLINK);
2148                 goto error_return;
2149         }
2150
2151         /*
2152          * If we are using project inheritance, we only allow hard link
2153          * creation in our tree when the project IDs are the same; else
2154          * the tree quota mechanism could be circumvented.
2155          */
2156         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2157                      (tdp->i_d.di_projid != sip->i_d.di_projid))) {
2158                 error = XFS_ERROR(EXDEV);
2159                 goto error_return;
2160         }
2161
2162         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
2163         if (error)
2164                 goto error_return;
2165
2166         XFS_BMAP_INIT(&free_list, &first_block);
2167
2168         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
2169                                         &first_block, &free_list, resblks);
2170         if (error)
2171                 goto abort_return;
2172         xfs_ichgtime(tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2173         tdp->i_gen++;
2174         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
2175
2176         error = xfs_bumplink(tp, sip);
2177         if (error)
2178                 goto abort_return;
2179
2180         /*
2181          * If this is a synchronous mount, make sure that the
2182          * link transaction goes to disk before returning to
2183          * the user.
2184          */
2185         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2186                 xfs_trans_set_sync(tp);
2187         }
2188
2189         error = xfs_bmap_finish (&tp, &free_list, &committed);
2190         if (error) {
2191                 xfs_bmap_cancel(&free_list);
2192                 goto abort_return;
2193         }
2194
2195         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2196         if (error)
2197                 goto std_return;
2198
2199         /* Fall through to std_return with error = 0. */
2200 std_return:
2201         if (DM_EVENT_ENABLED(sip, DM_EVENT_POSTLINK)) {
2202                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTLINK,
2203                                 tdp, DM_RIGHT_NULL,
2204                                 sip, DM_RIGHT_NULL,
2205                                 target_name->name, NULL, 0, error, 0);
2206         }
2207         return error;
2208
2209  abort_return:
2210         cancel_flags |= XFS_TRANS_ABORT;
2211         /* FALLTHROUGH */
2212
2213  error_return:
2214         xfs_trans_cancel(tp, cancel_flags);
2215         goto std_return;
2216 }
2217
2218
2219 int
2220 xfs_mkdir(
2221         xfs_inode_t             *dp,
2222         struct xfs_name         *dir_name,
2223         mode_t                  mode,
2224         xfs_inode_t             **ipp,
2225         cred_t                  *credp)
2226 {
2227         xfs_mount_t             *mp = dp->i_mount;
2228         xfs_inode_t             *cdp;   /* inode of created dir */
2229         xfs_trans_t             *tp;
2230         int                     cancel_flags;
2231         int                     error;
2232         int                     committed;
2233         xfs_bmap_free_t         free_list;
2234         xfs_fsblock_t           first_block;
2235         boolean_t               unlock_dp_on_error = B_FALSE;
2236         boolean_t               created = B_FALSE;
2237         int                     dm_event_sent = 0;
2238         xfs_prid_t              prid;
2239         struct xfs_dquot        *udqp, *gdqp;
2240         uint                    resblks;
2241
2242         if (XFS_FORCED_SHUTDOWN(mp))
2243                 return XFS_ERROR(EIO);
2244
2245         tp = NULL;
2246
2247         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
2248                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
2249                                         dp, DM_RIGHT_NULL, NULL,
2250                                         DM_RIGHT_NULL, dir_name->name, NULL,
2251                                         mode, 0, 0);
2252                 if (error)
2253                         return error;
2254                 dm_event_sent = 1;
2255         }
2256
2257         /* Return through std_return after this point. */
2258
2259         xfs_itrace_entry(dp);
2260
2261         mp = dp->i_mount;
2262         udqp = gdqp = NULL;
2263         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2264                 prid = dp->i_d.di_projid;
2265         else
2266                 prid = (xfs_prid_t)dfltprid;
2267
2268         /*
2269          * Make sure that we have allocated dquot(s) on disk.
2270          */
2271         error = XFS_QM_DQVOPALLOC(mp, dp,
2272                         current_fsuid(), current_fsgid(), prid,
2273                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2274         if (error)
2275                 goto std_return;
2276
2277         tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
2278         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2279         resblks = XFS_MKDIR_SPACE_RES(mp, dir_name->len);
2280         error = xfs_trans_reserve(tp, resblks, XFS_MKDIR_LOG_RES(mp), 0,
2281                                   XFS_TRANS_PERM_LOG_RES, XFS_MKDIR_LOG_COUNT);
2282         if (error == ENOSPC) {
2283                 resblks = 0;
2284                 error = xfs_trans_reserve(tp, 0, XFS_MKDIR_LOG_RES(mp), 0,
2285                                           XFS_TRANS_PERM_LOG_RES,
2286                                           XFS_MKDIR_LOG_COUNT);
2287         }
2288         if (error) {
2289                 cancel_flags = 0;
2290                 goto error_return;
2291         }
2292
2293         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2294         unlock_dp_on_error = B_TRUE;
2295
2296         /*
2297          * Check for directory link count overflow.
2298          */
2299         if (dp->i_d.di_nlink >= XFS_MAXLINK) {
2300                 error = XFS_ERROR(EMLINK);
2301                 goto error_return;
2302         }
2303
2304         /*
2305          * Reserve disk quota and the inode.
2306          */
2307         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2308         if (error)
2309                 goto error_return;
2310
2311         error = xfs_dir_canenter(tp, dp, dir_name, resblks);
2312         if (error)
2313                 goto error_return;
2314         /*
2315          * create the directory inode.
2316          */
2317         error = xfs_dir_ialloc(&tp, dp, mode, 2,
2318                         0, credp, prid, resblks > 0,
2319                 &cdp, NULL);
2320         if (error) {
2321                 if (error == ENOSPC)
2322                         goto error_return;
2323                 goto abort_return;
2324         }
2325         xfs_itrace_ref(cdp);
2326
2327         /*
2328          * Now we add the directory inode to the transaction.
2329          * We waited until now since xfs_dir_ialloc might start
2330          * a new transaction.  Had we joined the transaction
2331          * earlier, the locks might have gotten released. An error
2332          * from here on will result in the transaction cancel
2333          * unlocking dp so don't do it explicitly in the error path.
2334          */
2335         IHOLD(dp);
2336         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2337         unlock_dp_on_error = B_FALSE;
2338
2339         XFS_BMAP_INIT(&free_list, &first_block);
2340
2341         error = xfs_dir_createname(tp, dp, dir_name, cdp->i_ino,
2342                                         &first_block, &free_list, resblks ?
2343                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
2344         if (error) {
2345                 ASSERT(error != ENOSPC);
2346                 goto error1;
2347         }
2348         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2349
2350         /*
2351          * Bump the in memory version number of the parent directory
2352          * so that other processes accessing it will recognize that
2353          * the directory has changed.
2354          */
2355         dp->i_gen++;
2356
2357         error = xfs_dir_init(tp, cdp, dp);
2358         if (error)
2359                 goto error2;
2360
2361         cdp->i_gen = 1;
2362         error = xfs_bumplink(tp, dp);
2363         if (error)
2364                 goto error2;
2365
2366         created = B_TRUE;
2367
2368         *ipp = cdp;
2369         IHOLD(cdp);
2370
2371         /*
2372          * Attach the dquots to the new inode and modify the icount incore.
2373          */
2374         XFS_QM_DQVOPCREATE(mp, tp, cdp, udqp, gdqp);
2375
2376         /*
2377          * If this is a synchronous mount, make sure that the
2378          * mkdir transaction goes to disk before returning to
2379          * the user.
2380          */
2381         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2382                 xfs_trans_set_sync(tp);
2383         }
2384
2385         error = xfs_bmap_finish(&tp, &free_list, &committed);
2386         if (error) {
2387                 IRELE(cdp);
2388                 goto error2;
2389         }
2390
2391         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2392         XFS_QM_DQRELE(mp, udqp);
2393         XFS_QM_DQRELE(mp, gdqp);
2394         if (error) {
2395                 IRELE(cdp);
2396         }
2397
2398         /* Fall through to std_return with error = 0 or errno from
2399          * xfs_trans_commit. */
2400
2401 std_return:
2402         if ((created || (error != 0 && dm_event_sent != 0)) &&
2403             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
2404                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
2405                                         dp, DM_RIGHT_NULL,
2406                                         created ? cdp : NULL,
2407                                         DM_RIGHT_NULL,
2408                                         dir_name->name, NULL,
2409                                         mode, error, 0);
2410         }
2411         return error;
2412
2413  error2:
2414  error1:
2415         xfs_bmap_cancel(&free_list);
2416  abort_return:
2417         cancel_flags |= XFS_TRANS_ABORT;
2418  error_return:
2419         xfs_trans_cancel(tp, cancel_flags);
2420         XFS_QM_DQRELE(mp, udqp);
2421         XFS_QM_DQRELE(mp, gdqp);
2422
2423         if (unlock_dp_on_error)
2424                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2425
2426         goto std_return;
2427 }
2428
2429 int
2430 xfs_symlink(
2431         xfs_inode_t             *dp,
2432         struct xfs_name         *link_name,
2433         const char              *target_path,
2434         mode_t                  mode,
2435         xfs_inode_t             **ipp,
2436         cred_t                  *credp)
2437 {
2438         xfs_mount_t             *mp = dp->i_mount;
2439         xfs_trans_t             *tp;
2440         xfs_inode_t             *ip;
2441         int                     error;
2442         int                     pathlen;
2443         xfs_bmap_free_t         free_list;
2444         xfs_fsblock_t           first_block;
2445         boolean_t               unlock_dp_on_error = B_FALSE;
2446         uint                    cancel_flags;
2447         int                     committed;
2448         xfs_fileoff_t           first_fsb;
2449         xfs_filblks_t           fs_blocks;
2450         int                     nmaps;
2451         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
2452         xfs_daddr_t             d;
2453         const char              *cur_chunk;
2454         int                     byte_cnt;
2455         int                     n;
2456         xfs_buf_t               *bp;
2457         xfs_prid_t              prid;
2458         struct xfs_dquot        *udqp, *gdqp;
2459         uint                    resblks;
2460
2461         *ipp = NULL;
2462         error = 0;
2463         ip = NULL;
2464         tp = NULL;
2465
2466         xfs_itrace_entry(dp);
2467
2468         if (XFS_FORCED_SHUTDOWN(mp))
2469                 return XFS_ERROR(EIO);
2470
2471         /*
2472          * Check component lengths of the target path name.
2473          */
2474         pathlen = strlen(target_path);
2475         if (pathlen >= MAXPATHLEN)      /* total string too long */
2476                 return XFS_ERROR(ENAMETOOLONG);
2477
2478         if (DM_EVENT_ENABLED(dp, DM_EVENT_SYMLINK)) {
2479                 error = XFS_SEND_NAMESP(mp, DM_EVENT_SYMLINK, dp,
2480                                         DM_RIGHT_NULL, NULL, DM_RIGHT_NULL,
2481                                         link_name->name, target_path, 0, 0, 0);
2482                 if (error)
2483                         return error;
2484         }
2485
2486         /* Return through std_return after this point. */
2487
2488         udqp = gdqp = NULL;
2489         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2490                 prid = dp->i_d.di_projid;
2491         else
2492                 prid = (xfs_prid_t)dfltprid;
2493
2494         /*
2495          * Make sure that we have allocated dquot(s) on disk.
2496          */
2497         error = XFS_QM_DQVOPALLOC(mp, dp,
2498                         current_fsuid(), current_fsgid(), prid,
2499                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2500         if (error)
2501                 goto std_return;
2502
2503         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
2504         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2505         /*
2506          * The symlink will fit into the inode data fork?
2507          * There can't be any attributes so we get the whole variable part.
2508          */
2509         if (pathlen <= XFS_LITINO(mp))
2510                 fs_blocks = 0;
2511         else
2512                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2513         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2514         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2515                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2516         if (error == ENOSPC && fs_blocks == 0) {
2517                 resblks = 0;
2518                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2519                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2520         }
2521         if (error) {
2522                 cancel_flags = 0;
2523                 goto error_return;
2524         }
2525
2526         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2527         unlock_dp_on_error = B_TRUE;
2528
2529         /*
2530          * Check whether the directory allows new symlinks or not.
2531          */
2532         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2533                 error = XFS_ERROR(EPERM);
2534                 goto error_return;
2535         }
2536
2537         /*
2538          * Reserve disk quota : blocks and inode.
2539          */
2540         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2541         if (error)
2542                 goto error_return;
2543
2544         /*
2545          * Check for ability to enter directory entry, if no space reserved.
2546          */
2547         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2548         if (error)
2549                 goto error_return;
2550         /*
2551          * Initialize the bmap freelist prior to calling either
2552          * bmapi or the directory create code.
2553          */
2554         XFS_BMAP_INIT(&free_list, &first_block);
2555
2556         /*
2557          * Allocate an inode for the symlink.
2558          */
2559         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT),
2560                                1, 0, credp, prid, resblks > 0, &ip, NULL);
2561         if (error) {
2562                 if (error == ENOSPC)
2563                         goto error_return;
2564                 goto error1;
2565         }
2566         xfs_itrace_ref(ip);
2567
2568         /*
2569          * An error after we've joined dp to the transaction will result in the
2570          * transaction cancel unlocking dp so don't do it explicitly in the
2571          * error path.
2572          */
2573         IHOLD(dp);
2574         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2575         unlock_dp_on_error = B_FALSE;
2576
2577         /*
2578          * Also attach the dquot(s) to it, if applicable.
2579          */
2580         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
2581
2582         if (resblks)
2583                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2584         /*
2585          * If the symlink will fit into the inode, write it inline.
2586          */
2587         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2588                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2589                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2590                 ip->i_d.di_size = pathlen;
2591
2592                 /*
2593                  * The inode was initially created in extent format.
2594                  */
2595                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2596                 ip->i_df.if_flags |= XFS_IFINLINE;
2597
2598                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2599                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2600
2601         } else {
2602                 first_fsb = 0;
2603                 nmaps = SYMLINK_MAPS;
2604
2605                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2606                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2607                                   &first_block, resblks, mval, &nmaps,
2608                                   &free_list, NULL);
2609                 if (error) {
2610                         goto error1;
2611                 }
2612
2613                 if (resblks)
2614                         resblks -= fs_blocks;
2615                 ip->i_d.di_size = pathlen;
2616                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2617
2618                 cur_chunk = target_path;
2619                 for (n = 0; n < nmaps; n++) {
2620                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2621                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2622                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2623                                                BTOBB(byte_cnt), 0);
2624                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2625                         if (pathlen < byte_cnt) {
2626                                 byte_cnt = pathlen;
2627                         }
2628                         pathlen -= byte_cnt;
2629
2630                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2631                         cur_chunk += byte_cnt;
2632
2633                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2634                 }
2635         }
2636
2637         /*
2638          * Create the directory entry for the symlink.
2639          */
2640         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2641                                         &first_block, &free_list, resblks);
2642         if (error)
2643                 goto error1;
2644         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2645         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2646
2647         /*
2648          * Bump the in memory version number of the parent directory
2649          * so that other processes accessing it will recognize that
2650          * the directory has changed.
2651          */
2652         dp->i_gen++;
2653
2654         /*
2655          * If this is a synchronous mount, make sure that the
2656          * symlink transaction goes to disk before returning to
2657          * the user.
2658          */
2659         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2660                 xfs_trans_set_sync(tp);
2661         }
2662
2663         /*
2664          * xfs_trans_commit normally decrements the vnode ref count
2665          * when it unlocks the inode. Since we want to return the
2666          * vnode to the caller, we bump the vnode ref count now.
2667          */
2668         IHOLD(ip);
2669
2670         error = xfs_bmap_finish(&tp, &free_list, &committed);
2671         if (error) {
2672                 goto error2;
2673         }
2674         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2675         XFS_QM_DQRELE(mp, udqp);
2676         XFS_QM_DQRELE(mp, gdqp);
2677
2678         /* Fall through to std_return with error = 0 or errno from
2679          * xfs_trans_commit     */
2680 std_return:
2681         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTSYMLINK)) {
2682                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTSYMLINK,
2683                                         dp, DM_RIGHT_NULL,
2684                                         error ? NULL : ip,
2685                                         DM_RIGHT_NULL, link_name->name,
2686                                         target_path, 0, error, 0);
2687         }
2688
2689         if (!error)
2690                 *ipp = ip;
2691         return error;
2692
2693  error2:
2694         IRELE(ip);
2695  error1:
2696         xfs_bmap_cancel(&free_list);
2697         cancel_flags |= XFS_TRANS_ABORT;
2698  error_return:
2699         xfs_trans_cancel(tp, cancel_flags);
2700         XFS_QM_DQRELE(mp, udqp);
2701         XFS_QM_DQRELE(mp, gdqp);
2702
2703         if (unlock_dp_on_error)
2704                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2705
2706         goto std_return;
2707 }
2708
2709 int
2710 xfs_inode_flush(
2711         xfs_inode_t     *ip,
2712         int             flags)
2713 {
2714         xfs_mount_t     *mp = ip->i_mount;
2715         int             error = 0;
2716
2717         if (XFS_FORCED_SHUTDOWN(mp))
2718                 return XFS_ERROR(EIO);
2719
2720         /*
2721          * Bypass inodes which have already been cleaned by
2722          * the inode flush clustering code inside xfs_iflush
2723          */
2724         if (xfs_inode_clean(ip))
2725                 return 0;
2726
2727         /*
2728          * We make this non-blocking if the inode is contended,
2729          * return EAGAIN to indicate to the caller that they
2730          * did not succeed. This prevents the flush path from
2731          * blocking on inodes inside another operation right
2732          * now, they get caught later by xfs_sync.
2733          */
2734         if (flags & FLUSH_SYNC) {
2735                 xfs_ilock(ip, XFS_ILOCK_SHARED);
2736                 xfs_iflock(ip);
2737         } else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2738                 if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
2739                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2740                         return EAGAIN;
2741                 }
2742         } else {
2743                 return EAGAIN;
2744         }
2745
2746         error = xfs_iflush(ip, (flags & FLUSH_SYNC) ? XFS_IFLUSH_SYNC
2747                                                     : XFS_IFLUSH_ASYNC_NOBLOCK);
2748         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2749
2750         return error;
2751 }
2752
2753
2754 int
2755 xfs_set_dmattrs(
2756         xfs_inode_t     *ip,
2757         u_int           evmask,
2758         u_int16_t       state)
2759 {
2760         xfs_mount_t     *mp = ip->i_mount;
2761         xfs_trans_t     *tp;
2762         int             error;
2763
2764         if (!capable(CAP_SYS_ADMIN))
2765                 return XFS_ERROR(EPERM);
2766
2767         if (XFS_FORCED_SHUTDOWN(mp))
2768                 return XFS_ERROR(EIO);
2769
2770         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2771         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2772         if (error) {
2773                 xfs_trans_cancel(tp, 0);
2774                 return error;
2775         }
2776         xfs_ilock(ip, XFS_ILOCK_EXCL);
2777         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2778
2779         ip->i_d.di_dmevmask = evmask;
2780         ip->i_d.di_dmstate  = state;
2781
2782         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2783         IHOLD(ip);
2784         error = xfs_trans_commit(tp, 0);
2785
2786         return error;
2787 }
2788
2789 int
2790 xfs_reclaim(
2791         xfs_inode_t     *ip)
2792 {
2793
2794         xfs_itrace_entry(ip);
2795
2796         ASSERT(!VN_MAPPED(VFS_I(ip)));
2797
2798         /* bad inode, get out here ASAP */
2799         if (VN_BAD(VFS_I(ip))) {
2800                 xfs_ireclaim(ip);
2801                 return 0;
2802         }
2803
2804         vn_iowait(ip);
2805
2806         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0);
2807
2808         /*
2809          * Make sure the atime in the XFS inode is correct before freeing the
2810          * Linux inode.
2811          */
2812         xfs_synchronize_atime(ip);
2813
2814         /*
2815          * If we have nothing to flush with this inode then complete the
2816          * teardown now, otherwise break the link between the xfs inode and the
2817          * linux inode and clean up the xfs inode later. This avoids flushing
2818          * the inode to disk during the delete operation itself.
2819          *
2820          * When breaking the link, we need to set the XFS_IRECLAIMABLE flag
2821          * first to ensure that xfs_iunpin() will never see an xfs inode
2822          * that has a linux inode being reclaimed. Synchronisation is provided
2823          * by the i_flags_lock.
2824          */
2825         if (!ip->i_update_core && (ip->i_itemp == NULL)) {
2826                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2827                 xfs_iflock(ip);
2828                 return xfs_finish_reclaim(ip, 1, XFS_IFLUSH_DELWRI_ELSE_SYNC);
2829         } else {
2830                 xfs_mount_t     *mp = ip->i_mount;
2831
2832                 /* Protect sync and unpin from us */
2833                 XFS_MOUNT_ILOCK(mp);
2834                 spin_lock(&ip->i_flags_lock);
2835                 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
2836                 VFS_I(ip)->i_private = NULL;
2837                 ip->i_vnode = NULL;
2838                 spin_unlock(&ip->i_flags_lock);
2839                 list_add_tail(&ip->i_reclaim, &mp->m_del_inodes);
2840                 XFS_MOUNT_IUNLOCK(mp);
2841         }
2842         return 0;
2843 }
2844
2845 int
2846 xfs_finish_reclaim(
2847         xfs_inode_t     *ip,
2848         int             locked,
2849         int             sync_mode)
2850 {
2851         xfs_perag_t     *pag = xfs_get_perag(ip->i_mount, ip->i_ino);
2852         struct inode    *vp = VFS_I(ip);
2853
2854         if (vp && VN_BAD(vp))
2855                 goto reclaim;
2856
2857         /* The hash lock here protects a thread in xfs_iget_core from
2858          * racing with us on linking the inode back with a vnode.
2859          * Once we have the XFS_IRECLAIM flag set it will not touch
2860          * us.
2861          */
2862         write_lock(&pag->pag_ici_lock);
2863         spin_lock(&ip->i_flags_lock);
2864         if (__xfs_iflags_test(ip, XFS_IRECLAIM) ||
2865             (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) && vp == NULL)) {
2866                 spin_unlock(&ip->i_flags_lock);
2867                 write_unlock(&pag->pag_ici_lock);
2868                 if (locked) {
2869                         xfs_ifunlock(ip);
2870                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2871                 }
2872                 return 1;
2873         }
2874         __xfs_iflags_set(ip, XFS_IRECLAIM);
2875         spin_unlock(&ip->i_flags_lock);
2876         write_unlock(&pag->pag_ici_lock);
2877         xfs_put_perag(ip->i_mount, pag);
2878
2879         /*
2880          * If the inode is still dirty, then flush it out.  If the inode
2881          * is not in the AIL, then it will be OK to flush it delwri as
2882          * long as xfs_iflush() does not keep any references to the inode.
2883          * We leave that decision up to xfs_iflush() since it has the
2884          * knowledge of whether it's OK to simply do a delwri flush of
2885          * the inode or whether we need to wait until the inode is
2886          * pulled from the AIL.
2887          * We get the flush lock regardless, though, just to make sure
2888          * we don't free it while it is being flushed.
2889          */
2890         if (!locked) {
2891                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2892                 xfs_iflock(ip);
2893         }
2894
2895         /*
2896          * In the case of a forced shutdown we rely on xfs_iflush() to
2897          * wait for the inode to be unpinned before returning an error.
2898          */
2899         if (xfs_iflush(ip, sync_mode) == 0) {
2900                 /* synchronize with xfs_iflush_done */
2901                 xfs_iflock(ip);
2902                 xfs_ifunlock(ip);
2903         }
2904
2905         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2906
2907  reclaim:
2908         xfs_ireclaim(ip);
2909         return 0;
2910 }
2911
2912 int
2913 xfs_finish_reclaim_all(xfs_mount_t *mp, int noblock)
2914 {
2915         int             purged;
2916         xfs_inode_t     *ip, *n;
2917         int             done = 0;
2918
2919         while (!done) {
2920                 purged = 0;
2921                 XFS_MOUNT_ILOCK(mp);
2922                 list_for_each_entry_safe(ip, n, &mp->m_del_inodes, i_reclaim) {
2923                         if (noblock) {
2924                                 if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0)
2925                                         continue;
2926                                 if (xfs_ipincount(ip) ||
2927                                     !xfs_iflock_nowait(ip)) {
2928                                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2929                                         continue;
2930                                 }
2931                         }
2932                         XFS_MOUNT_IUNLOCK(mp);
2933                         if (xfs_finish_reclaim(ip, noblock,
2934                                         XFS_IFLUSH_DELWRI_ELSE_ASYNC))
2935                                 delay(1);
2936                         purged = 1;
2937                         break;
2938                 }
2939
2940                 done = !purged;
2941         }
2942
2943         XFS_MOUNT_IUNLOCK(mp);
2944         return 0;
2945 }
2946
2947 /*
2948  * xfs_alloc_file_space()
2949  *      This routine allocates disk space for the given file.
2950  *
2951  *      If alloc_type == 0, this request is for an ALLOCSP type
2952  *      request which will change the file size.  In this case, no
2953  *      DMAPI event will be generated by the call.  A TRUNCATE event
2954  *      will be generated later by xfs_setattr.
2955  *
2956  *      If alloc_type != 0, this request is for a RESVSP type
2957  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2958  *      lower block boundary byte address is less than the file's
2959  *      length.
2960  *
2961  * RETURNS:
2962  *       0 on success
2963  *      errno on error
2964  *
2965  */
2966 STATIC int
2967 xfs_alloc_file_space(
2968         xfs_inode_t             *ip,
2969         xfs_off_t               offset,
2970         xfs_off_t               len,
2971         int                     alloc_type,
2972         int                     attr_flags)
2973 {
2974         xfs_mount_t             *mp = ip->i_mount;
2975         xfs_off_t               count;
2976         xfs_filblks_t           allocated_fsb;
2977         xfs_filblks_t           allocatesize_fsb;
2978         xfs_extlen_t            extsz, temp;
2979         xfs_fileoff_t           startoffset_fsb;
2980         xfs_fsblock_t           firstfsb;
2981         int                     nimaps;
2982         int                     bmapi_flag;
2983         int                     quota_flag;
2984         int                     rt;
2985         xfs_trans_t             *tp;
2986         xfs_bmbt_irec_t         imaps[1], *imapp;
2987         xfs_bmap_free_t         free_list;
2988         uint                    qblocks, resblks, resrtextents;
2989         int                     committed;
2990         int                     error;
2991
2992         xfs_itrace_entry(ip);
2993
2994         if (XFS_FORCED_SHUTDOWN(mp))
2995                 return XFS_ERROR(EIO);
2996
2997         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
2998                 return error;
2999
3000         if (len <= 0)
3001                 return XFS_ERROR(EINVAL);
3002
3003         rt = XFS_IS_REALTIME_INODE(ip);
3004         extsz = xfs_get_extsz_hint(ip);
3005
3006         count = len;
3007         imapp = &imaps[0];
3008         nimaps = 1;
3009         bmapi_flag = XFS_BMAPI_WRITE | (alloc_type ? XFS_BMAPI_PREALLOC : 0);
3010         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
3011         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
3012
3013         /*      Generate a DMAPI event if needed.       */
3014         if (alloc_type != 0 && offset < ip->i_size &&
3015                         (attr_flags & XFS_ATTR_DMI) == 0  &&
3016                         DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3017                 xfs_off_t           end_dmi_offset;
3018
3019                 end_dmi_offset = offset+len;
3020                 if (end_dmi_offset > ip->i_size)
3021                         end_dmi_offset = ip->i_size;
3022                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, offset,
3023                                       end_dmi_offset - offset, 0, NULL);
3024                 if (error)
3025                         return error;
3026         }
3027
3028         /*
3029          * Allocate file space until done or until there is an error
3030          */
3031 retry:
3032         while (allocatesize_fsb && !error) {
3033                 xfs_fileoff_t   s, e;
3034
3035                 /*
3036                  * Determine space reservations for data/realtime.
3037                  */
3038                 if (unlikely(extsz)) {
3039                         s = startoffset_fsb;
3040                         do_div(s, extsz);
3041                         s *= extsz;
3042                         e = startoffset_fsb + allocatesize_fsb;
3043                         if ((temp = do_mod(startoffset_fsb, extsz)))
3044                                 e += temp;
3045                         if ((temp = do_mod(e, extsz)))
3046                                 e += extsz - temp;
3047                 } else {
3048                         s = 0;
3049                         e = allocatesize_fsb;
3050                 }
3051
3052                 if (unlikely(rt)) {
3053                         resrtextents = qblocks = (uint)(e - s);
3054                         resrtextents /= mp->m_sb.sb_rextsize;
3055                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3056                         quota_flag = XFS_QMOPT_RES_RTBLKS;
3057                 } else {
3058                         resrtextents = 0;
3059                         resblks = qblocks = \
3060                                 XFS_DIOSTRAT_SPACE_RES(mp, (uint)(e - s));
3061                         quota_flag = XFS_QMOPT_RES_REGBLKS;
3062                 }
3063
3064                 /*
3065                  * Allocate and setup the transaction.
3066                  */
3067                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3068                 error = xfs_trans_reserve(tp, resblks,
3069                                           XFS_WRITE_LOG_RES(mp), resrtextents,
3070                                           XFS_TRANS_PERM_LOG_RES,
3071                                           XFS_WRITE_LOG_COUNT);
3072                 /*
3073                  * Check for running out of space
3074                  */
3075                 if (error) {
3076                         /*
3077                          * Free the transaction structure.
3078                          */
3079                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3080                         xfs_trans_cancel(tp, 0);
3081                         break;
3082                 }
3083                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3084                 error = XFS_TRANS_RESERVE_QUOTA_NBLKS(mp, tp, ip,
3085                                                       qblocks, 0, quota_flag);
3086                 if (error)
3087                         goto error1;
3088
3089                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3090                 xfs_trans_ihold(tp, ip);
3091
3092                 /*
3093                  * Issue the xfs_bmapi() call to allocate the blocks
3094                  */
3095                 XFS_BMAP_INIT(&free_list, &firstfsb);
3096                 error = xfs_bmapi(tp, ip, startoffset_fsb,
3097                                   allocatesize_fsb, bmapi_flag,
3098                                   &firstfsb, 0, imapp, &nimaps,
3099                                   &free_list, NULL);
3100                 if (error) {
3101                         goto error0;
3102                 }
3103
3104                 /*
3105                  * Complete the transaction
3106                  */
3107                 error = xfs_bmap_finish(&tp, &free_list, &committed);
3108                 if (error) {
3109                         goto error0;
3110                 }
3111
3112                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3113                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3114                 if (error) {
3115                         break;
3116                 }
3117
3118                 allocated_fsb = imapp->br_blockcount;
3119
3120                 if (nimaps == 0) {
3121                         error = XFS_ERROR(ENOSPC);
3122                         break;
3123                 }
3124
3125                 startoffset_fsb += allocated_fsb;
3126                 allocatesize_fsb -= allocated_fsb;
3127         }
3128 dmapi_enospc_check:
3129         if (error == ENOSPC && (attr_flags & XFS_ATTR_DMI) == 0 &&
3130             DM_EVENT_ENABLED(ip, DM_EVENT_NOSPACE)) {
3131                 error = XFS_SEND_NAMESP(mp, DM_EVENT_NOSPACE,
3132                                 ip, DM_RIGHT_NULL,
3133                                 ip, DM_RIGHT_NULL,
3134                                 NULL, NULL, 0, 0, 0); /* Delay flag intentionally unused */
3135                 if (error == 0)
3136                         goto retry;     /* Maybe DMAPI app. has made space */
3137                 /* else fall through with error from XFS_SEND_DATA */
3138         }
3139
3140         return error;
3141
3142 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
3143         xfs_bmap_cancel(&free_list);
3144         XFS_TRANS_UNRESERVE_QUOTA_NBLKS(mp, tp, ip, qblocks, 0, quota_flag);
3145
3146 error1: /* Just cancel transaction */
3147         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
3148         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3149         goto dmapi_enospc_check;
3150 }
3151
3152 /*
3153  * Zero file bytes between startoff and endoff inclusive.
3154  * The iolock is held exclusive and no blocks are buffered.
3155  */
3156 STATIC int
3157 xfs_zero_remaining_bytes(
3158         xfs_inode_t             *ip,
3159         xfs_off_t               startoff,
3160         xfs_off_t               endoff)
3161 {
3162         xfs_bmbt_irec_t         imap;
3163         xfs_fileoff_t           offset_fsb;
3164         xfs_off_t               lastoffset;
3165         xfs_off_t               offset;
3166         xfs_buf_t               *bp;
3167         xfs_mount_t             *mp = ip->i_mount;
3168         int                     nimap;
3169         int                     error = 0;
3170
3171         bp = xfs_buf_get_noaddr(mp->m_sb.sb_blocksize,
3172                                 XFS_IS_REALTIME_INODE(ip) ?
3173                                 mp->m_rtdev_targp : mp->m_ddev_targp);
3174
3175         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
3176                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
3177                 nimap = 1;
3178                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
3179                         NULL, 0, &imap, &nimap, NULL, NULL);
3180                 if (error || nimap < 1)
3181                         break;
3182                 ASSERT(imap.br_blockcount >= 1);
3183                 ASSERT(imap.br_startoff == offset_fsb);
3184                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
3185                 if (lastoffset > endoff)
3186                         lastoffset = endoff;
3187                 if (imap.br_startblock == HOLESTARTBLOCK)
3188                         continue;
3189                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3190                 if (imap.br_state == XFS_EXT_UNWRITTEN)
3191                         continue;
3192                 XFS_BUF_UNDONE(bp);
3193                 XFS_BUF_UNWRITE(bp);
3194                 XFS_BUF_READ(bp);
3195                 XFS_BUF_SET_ADDR(bp, XFS_FSB_TO_DB(ip, imap.br_startblock));
3196                 xfsbdstrat(mp, bp);
3197                 error = xfs_iowait(bp);
3198                 if (error) {
3199                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
3200                                           mp, bp, XFS_BUF_ADDR(bp));
3201                         break;
3202                 }
3203                 memset(XFS_BUF_PTR(bp) +
3204                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
3205                       0, lastoffset - offset + 1);
3206                 XFS_BUF_UNDONE(bp);
3207                 XFS_BUF_UNREAD(bp);
3208                 XFS_BUF_WRITE(bp);
3209                 xfsbdstrat(mp, bp);
3210                 error = xfs_iowait(bp);
3211                 if (error) {
3212                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
3213                                           mp, bp, XFS_BUF_ADDR(bp));
3214                         break;
3215                 }
3216         }
3217         xfs_buf_free(bp);
3218         return error;
3219 }
3220
3221 /*
3222  * xfs_free_file_space()
3223  *      This routine frees disk space for the given file.
3224  *
3225  *      This routine is only called by xfs_change_file_space
3226  *      for an UNRESVSP type call.
3227  *
3228  * RETURNS:
3229  *       0 on success
3230  *      errno on error
3231  *
3232  */
3233 STATIC int
3234 xfs_free_file_space(
3235         xfs_inode_t             *ip,
3236         xfs_off_t               offset,
3237         xfs_off_t               len,
3238         int                     attr_flags)
3239 {
3240         int                     committed;
3241         int                     done;
3242         xfs_off_t               end_dmi_offset;
3243         xfs_fileoff_t           endoffset_fsb;
3244         int                     error;
3245         xfs_fsblock_t           firstfsb;
3246         xfs_bmap_free_t         free_list;
3247         xfs_bmbt_irec_t         imap;
3248         xfs_off_t               ioffset;
3249         xfs_extlen_t            mod=0;
3250         xfs_mount_t             *mp;
3251         int                     nimap;
3252         uint                    resblks;
3253         uint                    rounding;
3254         int                     rt;
3255         xfs_fileoff_t           startoffset_fsb;
3256         xfs_trans_t             *tp;
3257         int                     need_iolock = 1;
3258
3259         mp = ip->i_mount;
3260
3261         xfs_itrace_entry(ip);
3262
3263         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
3264                 return error;
3265
3266         error = 0;
3267         if (len <= 0)   /* if nothing being freed */
3268                 return error;
3269         rt = XFS_IS_REALTIME_INODE(ip);
3270         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
3271         end_dmi_offset = offset + len;
3272         endoffset_fsb = XFS_B_TO_FSBT(mp, end_dmi_offset);
3273
3274         if (offset < ip->i_size && (attr_flags & XFS_ATTR_DMI) == 0 &&
3275             DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3276                 if (end_dmi_offset > ip->i_size)
3277                         end_dmi_offset = ip->i_size;
3278                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip,
3279                                 offset, end_dmi_offset - offset,
3280                                 AT_DELAY_FLAG(attr_flags), NULL);
3281                 if (error)
3282                         return error;
3283         }
3284
3285         if (attr_flags & XFS_ATTR_NOLOCK)
3286                 need_iolock = 0;
3287         if (need_iolock) {
3288                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
3289                 vn_iowait(ip);  /* wait for the completion of any pending DIOs */
3290         }
3291
3292         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
3293         ioffset = offset & ~(rounding - 1);
3294
3295         if (VN_CACHED(VFS_I(ip)) != 0) {
3296                 xfs_inval_cached_trace(ip, ioffset, -1, ioffset, -1);
3297                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
3298                 if (error)
3299                         goto out_unlock_iolock;
3300         }
3301
3302         /*
3303          * Need to zero the stuff we're not freeing, on disk.
3304          * If its a realtime file & can't use unwritten extents then we
3305          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
3306          * will take care of it for us.
3307          */
3308         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
3309                 nimap = 1;
3310                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
3311                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3312                 if (error)
3313                         goto out_unlock_iolock;
3314                 ASSERT(nimap == 0 || nimap == 1);
3315                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3316                         xfs_daddr_t     block;
3317
3318                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3319                         block = imap.br_startblock;
3320                         mod = do_div(block, mp->m_sb.sb_rextsize);
3321                         if (mod)
3322                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
3323                 }
3324                 nimap = 1;
3325                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
3326                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3327                 if (error)
3328                         goto out_unlock_iolock;
3329                 ASSERT(nimap == 0 || nimap == 1);
3330                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3331                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3332                         mod++;
3333                         if (mod && (mod != mp->m_sb.sb_rextsize))
3334                                 endoffset_fsb -= mod;
3335                 }
3336         }
3337         if ((done = (endoffset_fsb <= startoffset_fsb)))
3338                 /*
3339                  * One contiguous piece to clear
3340                  */
3341                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
3342         else {
3343                 /*
3344                  * Some full blocks, possibly two pieces to clear
3345                  */
3346                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
3347                         error = xfs_zero_remaining_bytes(ip, offset,
3348                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
3349                 if (!error &&
3350                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
3351                         error = xfs_zero_remaining_bytes(ip,
3352                                 XFS_FSB_TO_B(mp, endoffset_fsb),
3353                                 offset + len - 1);
3354         }
3355
3356         /*
3357          * free file space until done or until there is an error
3358          */
3359         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3360         while (!error && !done) {
3361
3362                 /*
3363                  * allocate and setup the transaction. Allow this
3364                  * transaction to dip into the reserve blocks to ensure
3365                  * the freeing of the space succeeds at ENOSPC.
3366                  */
3367                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3368                 tp->t_flags |= XFS_TRANS_RESERVE;
3369                 error = xfs_trans_reserve(tp,
3370                                           resblks,
3371                                           XFS_WRITE_LOG_RES(mp),
3372                                           0,
3373                                           XFS_TRANS_PERM_LOG_RES,
3374                                           XFS_WRITE_LOG_COUNT);
3375
3376                 /*
3377                  * check for running out of space
3378                  */
3379                 if (error) {
3380                         /*
3381                          * Free the transaction structure.
3382                          */
3383                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3384                         xfs_trans_cancel(tp, 0);
3385                         break;
3386                 }
3387                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3388                 error = XFS_TRANS_RESERVE_QUOTA(mp, tp,
3389                                 ip->i_udquot, ip->i_gdquot, resblks, 0,
3390                                 XFS_QMOPT_RES_REGBLKS);
3391                 if (error)
3392                         goto error1;
3393
3394                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3395                 xfs_trans_ihold(tp, ip);
3396
3397                 /*
3398                  * issue the bunmapi() call to free the blocks
3399                  */
3400                 XFS_BMAP_INIT(&free_list, &firstfsb);
3401                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
3402                                   endoffset_fsb - startoffset_fsb,
3403                                   0, 2, &firstfsb, &free_list, NULL, &done);
3404                 if (error) {
3405                         goto error0;
3406                 }
3407
3408                 /*
3409                  * complete the transaction
3410                  */
3411                 error = xfs_bmap_finish(&tp, &free_list, &committed);
3412                 if (error) {
3413                         goto error0;
3414                 }
3415
3416                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3417                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3418         }
3419
3420  out_unlock_iolock:
3421         if (need_iolock)
3422                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
3423         return error;
3424
3425  error0:
3426         xfs_bmap_cancel(&free_list);
3427  error1:
3428         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
3429         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
3430                     XFS_ILOCK_EXCL);
3431         return error;
3432 }
3433
3434 /*
3435  * xfs_change_file_space()
3436  *      This routine allocates or frees disk space for the given file.
3437  *      The user specified parameters are checked for alignment and size
3438  *      limitations.
3439  *
3440  * RETURNS:
3441  *       0 on success
3442  *      errno on error
3443  *
3444  */
3445 int
3446 xfs_change_file_space(
3447         xfs_inode_t     *ip,
3448         int             cmd,
3449         xfs_flock64_t   *bf,
3450         xfs_off_t       offset,
3451         cred_t          *credp,
3452         int             attr_flags)
3453 {
3454         xfs_mount_t     *mp = ip->i_mount;
3455         int             clrprealloc;
3456         int             error;
3457         xfs_fsize_t     fsize;
3458         int             setprealloc;
3459         xfs_off_t       startoffset;
3460         xfs_off_t       llen;
3461         xfs_trans_t     *tp;
3462         struct iattr    iattr;
3463
3464         xfs_itrace_entry(ip);
3465
3466         if (!S_ISREG(ip->i_d.di_mode))
3467                 return XFS_ERROR(EINVAL);
3468
3469         switch (bf->l_whence) {
3470         case 0: /*SEEK_SET*/
3471                 break;
3472         case 1: /*SEEK_CUR*/
3473                 bf->l_start += offset;
3474                 break;
3475         case 2: /*SEEK_END*/
3476                 bf->l_start += ip->i_size;
3477                 break;
3478         default:
3479                 return XFS_ERROR(EINVAL);
3480         }
3481
3482         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
3483
3484         if (   (bf->l_start < 0)
3485             || (bf->l_start > XFS_MAXIOFFSET(mp))
3486             || (bf->l_start + llen < 0)
3487             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
3488                 return XFS_ERROR(EINVAL);
3489
3490         bf->l_whence = 0;
3491
3492         startoffset = bf->l_start;
3493         fsize = ip->i_size;
3494
3495         /*
3496          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
3497          * file space.
3498          * These calls do NOT zero the data space allocated to the file,
3499          * nor do they change the file size.
3500          *
3501          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
3502          * space.
3503          * These calls cause the new file data to be zeroed and the file
3504          * size to be changed.
3505          */
3506         setprealloc = clrprealloc = 0;
3507
3508         switch (cmd) {
3509         case XFS_IOC_RESVSP:
3510         case XFS_IOC_RESVSP64:
3511                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
3512                                                                 1, attr_flags);
3513                 if (error)
3514                         return error;
3515                 setprealloc = 1;
3516                 break;
3517
3518         case XFS_IOC_UNRESVSP:
3519         case XFS_IOC_UNRESVSP64:
3520                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
3521                                                                 attr_flags)))
3522                         return error;
3523                 break;
3524
3525         case XFS_IOC_ALLOCSP:
3526         case XFS_IOC_ALLOCSP64:
3527         case XFS_IOC_FREESP:
3528         case XFS_IOC_FREESP64:
3529                 if (startoffset > fsize) {
3530                         error = xfs_alloc_file_space(ip, fsize,
3531                                         startoffset - fsize, 0, attr_flags);
3532                         if (error)
3533                                 break;
3534                 }
3535
3536                 iattr.ia_valid = ATTR_SIZE;
3537                 iattr.ia_size = startoffset;
3538
3539                 error = xfs_setattr(ip, &iattr, attr_flags, credp);
3540
3541                 if (error)
3542                         return error;
3543
3544                 clrprealloc = 1;
3545                 break;
3546
3547         default:
3548                 ASSERT(0);
3549                 return XFS_ERROR(EINVAL);
3550         }
3551
3552         /*
3553          * update the inode timestamp, mode, and prealloc flag bits
3554          */
3555         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
3556
3557         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
3558                                       0, 0, 0))) {
3559                 /* ASSERT(0); */
3560                 xfs_trans_cancel(tp, 0);
3561                 return error;
3562         }
3563
3564         xfs_ilock(ip, XFS_ILOCK_EXCL);
3565
3566         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3567         xfs_trans_ihold(tp, ip);
3568
3569         if ((attr_flags & XFS_ATTR_DMI) == 0) {
3570                 ip->i_d.di_mode &= ~S_ISUID;
3571
3572                 /*
3573                  * Note that we don't have to worry about mandatory
3574                  * file locking being disabled here because we only
3575                  * clear the S_ISGID bit if the Group execute bit is
3576                  * on, but if it was on then mandatory locking wouldn't
3577                  * have been enabled.
3578                  */
3579                 if (ip->i_d.di_mode & S_IXGRP)
3580                         ip->i_d.di_mode &= ~S_ISGID;
3581
3582                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3583         }
3584         if (setprealloc)
3585                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
3586         else if (clrprealloc)
3587                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
3588
3589         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
3590         xfs_trans_set_sync(tp);
3591
3592         error = xfs_trans_commit(tp, 0);
3593
3594         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3595
3596         return error;
3597 }