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