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