xfs: clear XFS_IDIRTY_RELEASE on truncate down
[linux-2.6.git] / fs / xfs / xfs_vnodeops.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18
19 #include "xfs.h"
20 #include "xfs_fs.h"
21 #include "xfs_types.h"
22 #include "xfs_bit.h"
23 #include "xfs_log.h"
24 #include "xfs_inum.h"
25 #include "xfs_trans.h"
26 #include "xfs_sb.h"
27 #include "xfs_ag.h"
28 #include "xfs_dir2.h"
29 #include "xfs_mount.h"
30 #include "xfs_da_btree.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_dinode.h"
34 #include "xfs_inode.h"
35 #include "xfs_inode_item.h"
36 #include "xfs_itable.h"
37 #include "xfs_ialloc.h"
38 #include "xfs_alloc.h"
39 #include "xfs_bmap.h"
40 #include "xfs_acl.h"
41 #include "xfs_attr.h"
42 #include "xfs_rw.h"
43 #include "xfs_error.h"
44 #include "xfs_quota.h"
45 #include "xfs_utils.h"
46 #include "xfs_rtalloc.h"
47 #include "xfs_trans_space.h"
48 #include "xfs_log_priv.h"
49 #include "xfs_filestream.h"
50 #include "xfs_vnodeops.h"
51 #include "xfs_trace.h"
52
53 int
54 xfs_setattr(
55         struct xfs_inode        *ip,
56         struct iattr            *iattr,
57         int                     flags)
58 {
59         xfs_mount_t             *mp = ip->i_mount;
60         struct inode            *inode = VFS_I(ip);
61         int                     mask = iattr->ia_valid;
62         xfs_trans_t             *tp;
63         int                     code;
64         uint                    lock_flags;
65         uint                    commit_flags=0;
66         uid_t                   uid=0, iuid=0;
67         gid_t                   gid=0, igid=0;
68         struct xfs_dquot        *udqp, *gdqp, *olddquot1, *olddquot2;
69         int                     need_iolock = 1;
70
71         trace_xfs_setattr(ip);
72
73         if (mp->m_flags & XFS_MOUNT_RDONLY)
74                 return XFS_ERROR(EROFS);
75
76         if (XFS_FORCED_SHUTDOWN(mp))
77                 return XFS_ERROR(EIO);
78
79         code = -inode_change_ok(inode, iattr);
80         if (code)
81                 return code;
82
83         olddquot1 = olddquot2 = NULL;
84         udqp = gdqp = NULL;
85
86         /*
87          * If disk quotas is on, we make sure that the dquots do exist on disk,
88          * before we start any other transactions. Trying to do this later
89          * is messy. We don't care to take a readlock to look at the ids
90          * in inode here, because we can't hold it across the trans_reserve.
91          * If the IDs do change before we take the ilock, we're covered
92          * because the i_*dquot fields will get updated anyway.
93          */
94         if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
95                 uint    qflags = 0;
96
97                 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
98                         uid = iattr->ia_uid;
99                         qflags |= XFS_QMOPT_UQUOTA;
100                 } else {
101                         uid = ip->i_d.di_uid;
102                 }
103                 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
104                         gid = iattr->ia_gid;
105                         qflags |= XFS_QMOPT_GQUOTA;
106                 }  else {
107                         gid = ip->i_d.di_gid;
108                 }
109
110                 /*
111                  * We take a reference when we initialize udqp and gdqp,
112                  * so it is important that we never blindly double trip on
113                  * the same variable. See xfs_create() for an example.
114                  */
115                 ASSERT(udqp == NULL);
116                 ASSERT(gdqp == NULL);
117                 code = xfs_qm_vop_dqalloc(ip, uid, gid, xfs_get_projid(ip),
118                                          qflags, &udqp, &gdqp);
119                 if (code)
120                         return code;
121         }
122
123         /*
124          * For the other attributes, we acquire the inode lock and
125          * first do an error checking pass.
126          */
127         tp = NULL;
128         lock_flags = XFS_ILOCK_EXCL;
129         if (flags & XFS_ATTR_NOLOCK)
130                 need_iolock = 0;
131         if (!(mask & ATTR_SIZE)) {
132                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
133                 commit_flags = 0;
134                 code = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES(mp),
135                                          0, 0, 0);
136                 if (code) {
137                         lock_flags = 0;
138                         goto error_return;
139                 }
140         } else {
141                 if (need_iolock)
142                         lock_flags |= XFS_IOLOCK_EXCL;
143         }
144
145         xfs_ilock(ip, lock_flags);
146
147         /*
148          * Change file ownership.  Must be the owner or privileged.
149          */
150         if (mask & (ATTR_UID|ATTR_GID)) {
151                 /*
152                  * These IDs could have changed since we last looked at them.
153                  * But, we're assured that if the ownership did change
154                  * while we didn't have the inode locked, inode's dquot(s)
155                  * would have changed also.
156                  */
157                 iuid = ip->i_d.di_uid;
158                 igid = ip->i_d.di_gid;
159                 gid = (mask & ATTR_GID) ? iattr->ia_gid : igid;
160                 uid = (mask & ATTR_UID) ? iattr->ia_uid : iuid;
161
162                 /*
163                  * Do a quota reservation only if uid/gid is actually
164                  * going to change.
165                  */
166                 if (XFS_IS_QUOTA_RUNNING(mp) &&
167                     ((XFS_IS_UQUOTA_ON(mp) && iuid != uid) ||
168                      (XFS_IS_GQUOTA_ON(mp) && igid != gid))) {
169                         ASSERT(tp);
170                         code = xfs_qm_vop_chown_reserve(tp, ip, udqp, gdqp,
171                                                 capable(CAP_FOWNER) ?
172                                                 XFS_QMOPT_FORCE_RES : 0);
173                         if (code)       /* out of quota */
174                                 goto error_return;
175                 }
176         }
177
178         /*
179          * Truncate file.  Must have write permission and not be a directory.
180          */
181         if (mask & ATTR_SIZE) {
182                 /* Short circuit the truncate case for zero length files */
183                 if (iattr->ia_size == 0 &&
184                     ip->i_size == 0 && ip->i_d.di_nextents == 0) {
185                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
186                         lock_flags &= ~XFS_ILOCK_EXCL;
187                         if (mask & ATTR_CTIME) {
188                                 inode->i_mtime = inode->i_ctime =
189                                                 current_fs_time(inode->i_sb);
190                                 xfs_mark_inode_dirty_sync(ip);
191                         }
192                         code = 0;
193                         goto error_return;
194                 }
195
196                 if (S_ISDIR(ip->i_d.di_mode)) {
197                         code = XFS_ERROR(EISDIR);
198                         goto error_return;
199                 } else if (!S_ISREG(ip->i_d.di_mode)) {
200                         code = XFS_ERROR(EINVAL);
201                         goto error_return;
202                 }
203
204                 /*
205                  * Make sure that the dquots are attached to the inode.
206                  */
207                 code = xfs_qm_dqattach_locked(ip, 0);
208                 if (code)
209                         goto error_return;
210
211                 /*
212                  * Now we can make the changes.  Before we join the inode
213                  * to the transaction, if ATTR_SIZE is set then take care of
214                  * the part of the truncation that must be done without the
215                  * inode lock.  This needs to be done before joining the inode
216                  * to the transaction, because the inode cannot be unlocked
217                  * once it is a part of the transaction.
218                  */
219                 if (iattr->ia_size > ip->i_size) {
220                         /*
221                          * Do the first part of growing a file: zero any data
222                          * in the last block that is beyond the old EOF.  We
223                          * need to do this before the inode is joined to the
224                          * transaction to modify the i_size.
225                          */
226                         code = xfs_zero_eof(ip, iattr->ia_size, ip->i_size);
227                         if (code)
228                                 goto error_return;
229                 }
230                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
231                 lock_flags &= ~XFS_ILOCK_EXCL;
232
233                 /*
234                  * We are going to log the inode size change in this
235                  * transaction so any previous writes that are beyond the on
236                  * disk EOF and the new EOF that have not been written out need
237                  * to be written here. If we do not write the data out, we
238                  * expose ourselves to the null files problem.
239                  *
240                  * Only flush from the on disk size to the smaller of the in
241                  * memory file size or the new size as that's the range we
242                  * really care about here and prevents waiting for other data
243                  * not within the range we care about here.
244                  */
245                 if (ip->i_size != ip->i_d.di_size &&
246                     iattr->ia_size > ip->i_d.di_size) {
247                         code = xfs_flush_pages(ip,
248                                         ip->i_d.di_size, iattr->ia_size,
249                                         XBF_ASYNC, FI_NONE);
250                         if (code)
251                                 goto error_return;
252                 }
253
254                 /* wait for all I/O to complete */
255                 xfs_ioend_wait(ip);
256
257                 code = -block_truncate_page(inode->i_mapping, iattr->ia_size,
258                                             xfs_get_blocks);
259                 if (code)
260                         goto error_return;
261
262                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
263                 code = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
264                                          XFS_TRANS_PERM_LOG_RES,
265                                          XFS_ITRUNCATE_LOG_COUNT);
266                 if (code)
267                         goto error_return;
268
269                 truncate_setsize(inode, iattr->ia_size);
270
271                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
272                 lock_flags |= XFS_ILOCK_EXCL;
273
274                 xfs_ilock(ip, XFS_ILOCK_EXCL);
275
276                 xfs_trans_ijoin(tp, ip);
277
278                 /*
279                  * Only change the c/mtime if we are changing the size
280                  * or we are explicitly asked to change it. This handles
281                  * the semantic difference between truncate() and ftruncate()
282                  * as implemented in the VFS.
283                  *
284                  * The regular truncate() case without ATTR_CTIME and ATTR_MTIME
285                  * is a special case where we need to update the times despite
286                  * not having these flags set.  For all other operations the
287                  * VFS set these flags explicitly if it wants a timestamp
288                  * update.
289                  */
290                 if (iattr->ia_size != ip->i_size &&
291                     (!(mask & (ATTR_CTIME | ATTR_MTIME)))) {
292                         iattr->ia_ctime = iattr->ia_mtime =
293                                 current_fs_time(inode->i_sb);
294                         mask |= ATTR_CTIME | ATTR_MTIME;
295                 }
296
297                 if (iattr->ia_size > ip->i_size) {
298                         ip->i_d.di_size = iattr->ia_size;
299                         ip->i_size = iattr->ia_size;
300                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
301                 } else if (iattr->ia_size <= ip->i_size ||
302                            (iattr->ia_size == 0 && ip->i_d.di_nextents)) {
303                         /*
304                          * signal a sync transaction unless
305                          * we're truncating an already unlinked
306                          * file on a wsync filesystem
307                          */
308                         code = xfs_itruncate_finish(&tp, ip, iattr->ia_size,
309                                             XFS_DATA_FORK,
310                                             ((ip->i_d.di_nlink != 0 ||
311                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
312                                              ? 1 : 0));
313                         if (code)
314                                 goto abort_return;
315                         /*
316                          * Truncated "down", so we're removing references
317                          * to old data here - if we now delay flushing for
318                          * a long time, we expose ourselves unduly to the
319                          * notorious NULL files problem.  So, we mark this
320                          * vnode and flush it when the file is closed, and
321                          * do not wait the usual (long) time for writeout.
322                          */
323                         xfs_iflags_set(ip, XFS_ITRUNCATED);
324                 }
325         } else if (tp) {
326                 xfs_trans_ijoin(tp, ip);
327         }
328
329         /*
330          * Change file ownership.  Must be the owner or privileged.
331          */
332         if (mask & (ATTR_UID|ATTR_GID)) {
333                 /*
334                  * CAP_FSETID overrides the following restrictions:
335                  *
336                  * The set-user-ID and set-group-ID bits of a file will be
337                  * cleared upon successful return from chown()
338                  */
339                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
340                     !capable(CAP_FSETID)) {
341                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
342                 }
343
344                 /*
345                  * Change the ownerships and register quota modifications
346                  * in the transaction.
347                  */
348                 if (iuid != uid) {
349                         if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_UQUOTA_ON(mp)) {
350                                 ASSERT(mask & ATTR_UID);
351                                 ASSERT(udqp);
352                                 olddquot1 = xfs_qm_vop_chown(tp, ip,
353                                                         &ip->i_udquot, udqp);
354                         }
355                         ip->i_d.di_uid = uid;
356                         inode->i_uid = uid;
357                 }
358                 if (igid != gid) {
359                         if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_GQUOTA_ON(mp)) {
360                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
361                                 ASSERT(mask & ATTR_GID);
362                                 ASSERT(gdqp);
363                                 olddquot2 = xfs_qm_vop_chown(tp, ip,
364                                                         &ip->i_gdquot, gdqp);
365                         }
366                         ip->i_d.di_gid = gid;
367                         inode->i_gid = gid;
368                 }
369         }
370
371         /*
372          * Change file access modes.
373          */
374         if (mask & ATTR_MODE) {
375                 umode_t mode = iattr->ia_mode;
376
377                 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
378                         mode &= ~S_ISGID;
379
380                 ip->i_d.di_mode &= S_IFMT;
381                 ip->i_d.di_mode |= mode & ~S_IFMT;
382
383                 inode->i_mode &= S_IFMT;
384                 inode->i_mode |= mode & ~S_IFMT;
385         }
386
387         /*
388          * Change file access or modified times.
389          */
390         if (mask & ATTR_ATIME) {
391                 inode->i_atime = iattr->ia_atime;
392                 ip->i_d.di_atime.t_sec = iattr->ia_atime.tv_sec;
393                 ip->i_d.di_atime.t_nsec = iattr->ia_atime.tv_nsec;
394                 ip->i_update_core = 1;
395         }
396         if (mask & ATTR_CTIME) {
397                 inode->i_ctime = iattr->ia_ctime;
398                 ip->i_d.di_ctime.t_sec = iattr->ia_ctime.tv_sec;
399                 ip->i_d.di_ctime.t_nsec = iattr->ia_ctime.tv_nsec;
400                 ip->i_update_core = 1;
401         }
402         if (mask & ATTR_MTIME) {
403                 inode->i_mtime = iattr->ia_mtime;
404                 ip->i_d.di_mtime.t_sec = iattr->ia_mtime.tv_sec;
405                 ip->i_d.di_mtime.t_nsec = iattr->ia_mtime.tv_nsec;
406                 ip->i_update_core = 1;
407         }
408
409         /*
410          * And finally, log the inode core if any attribute in it
411          * has been changed.
412          */
413         if (mask & (ATTR_UID|ATTR_GID|ATTR_MODE|
414                     ATTR_ATIME|ATTR_CTIME|ATTR_MTIME))
415                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
416
417         XFS_STATS_INC(xs_ig_attrchg);
418
419         /*
420          * If this is a synchronous mount, make sure that the
421          * transaction goes to disk before returning to the user.
422          * This is slightly sub-optimal in that truncates require
423          * two sync transactions instead of one for wsync filesystems.
424          * One for the truncate and one for the timestamps since we
425          * don't want to change the timestamps unless we're sure the
426          * truncate worked.  Truncates are less than 1% of the laddis
427          * mix so this probably isn't worth the trouble to optimize.
428          */
429         code = 0;
430         if (mp->m_flags & XFS_MOUNT_WSYNC)
431                 xfs_trans_set_sync(tp);
432
433         code = xfs_trans_commit(tp, commit_flags);
434
435         xfs_iunlock(ip, lock_flags);
436
437         /*
438          * Release any dquot(s) the inode had kept before chown.
439          */
440         xfs_qm_dqrele(olddquot1);
441         xfs_qm_dqrele(olddquot2);
442         xfs_qm_dqrele(udqp);
443         xfs_qm_dqrele(gdqp);
444
445         if (code)
446                 return code;
447
448         /*
449          * XXX(hch): Updating the ACL entries is not atomic vs the i_mode
450          *           update.  We could avoid this with linked transactions
451          *           and passing down the transaction pointer all the way
452          *           to attr_set.  No previous user of the generic
453          *           Posix ACL code seems to care about this issue either.
454          */
455         if ((mask & ATTR_MODE) && !(flags & XFS_ATTR_NOACL)) {
456                 code = -xfs_acl_chmod(inode);
457                 if (code)
458                         return XFS_ERROR(code);
459         }
460
461         return 0;
462
463  abort_return:
464         commit_flags |= XFS_TRANS_ABORT;
465  error_return:
466         xfs_qm_dqrele(udqp);
467         xfs_qm_dqrele(gdqp);
468         if (tp) {
469                 xfs_trans_cancel(tp, commit_flags);
470         }
471         if (lock_flags != 0) {
472                 xfs_iunlock(ip, lock_flags);
473         }
474         return code;
475 }
476
477 /*
478  * The maximum pathlen is 1024 bytes. Since the minimum file system
479  * blocksize is 512 bytes, we can get a max of 2 extents back from
480  * bmapi.
481  */
482 #define SYMLINK_MAPS 2
483
484 STATIC int
485 xfs_readlink_bmap(
486         xfs_inode_t     *ip,
487         char            *link)
488 {
489         xfs_mount_t     *mp = ip->i_mount;
490         int             pathlen = ip->i_d.di_size;
491         int             nmaps = SYMLINK_MAPS;
492         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
493         xfs_daddr_t     d;
494         int             byte_cnt;
495         int             n;
496         xfs_buf_t       *bp;
497         int             error = 0;
498
499         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
500                         mval, &nmaps, NULL);
501         if (error)
502                 goto out;
503
504         for (n = 0; n < nmaps; n++) {
505                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
506                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
507
508                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt),
509                                   XBF_LOCK | XBF_MAPPED | XBF_DONT_BLOCK);
510                 error = XFS_BUF_GETERROR(bp);
511                 if (error) {
512                         xfs_ioerror_alert("xfs_readlink",
513                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
514                         xfs_buf_relse(bp);
515                         goto out;
516                 }
517                 if (pathlen < byte_cnt)
518                         byte_cnt = pathlen;
519                 pathlen -= byte_cnt;
520
521                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
522                 xfs_buf_relse(bp);
523         }
524
525         link[ip->i_d.di_size] = '\0';
526         error = 0;
527
528  out:
529         return error;
530 }
531
532 int
533 xfs_readlink(
534         xfs_inode_t     *ip,
535         char            *link)
536 {
537         xfs_mount_t     *mp = ip->i_mount;
538         int             pathlen;
539         int             error = 0;
540
541         trace_xfs_readlink(ip);
542
543         if (XFS_FORCED_SHUTDOWN(mp))
544                 return XFS_ERROR(EIO);
545
546         xfs_ilock(ip, XFS_ILOCK_SHARED);
547
548         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
549         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
550
551         pathlen = ip->i_d.di_size;
552         if (!pathlen)
553                 goto out;
554
555         if (ip->i_df.if_flags & XFS_IFINLINE) {
556                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
557                 link[pathlen] = '\0';
558         } else {
559                 error = xfs_readlink_bmap(ip, link);
560         }
561
562  out:
563         xfs_iunlock(ip, XFS_ILOCK_SHARED);
564         return error;
565 }
566
567 /*
568  * Flags for xfs_free_eofblocks
569  */
570 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
571
572 /*
573  * This is called by xfs_inactive to free any blocks beyond eof
574  * when the link count isn't zero and by xfs_dm_punch_hole() when
575  * punching a hole to EOF.
576  */
577 STATIC int
578 xfs_free_eofblocks(
579         xfs_mount_t     *mp,
580         xfs_inode_t     *ip,
581         int             flags)
582 {
583         xfs_trans_t     *tp;
584         int             error;
585         xfs_fileoff_t   end_fsb;
586         xfs_fileoff_t   last_fsb;
587         xfs_filblks_t   map_len;
588         int             nimaps;
589         xfs_bmbt_irec_t imap;
590
591         /*
592          * Figure out if there are any blocks beyond the end
593          * of the file.  If not, then there is nothing to do.
594          */
595         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
596         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
597         if (last_fsb <= end_fsb)
598                 return 0;
599         map_len = last_fsb - end_fsb;
600
601         nimaps = 1;
602         xfs_ilock(ip, XFS_ILOCK_SHARED);
603         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
604                           NULL, 0, &imap, &nimaps, NULL);
605         xfs_iunlock(ip, XFS_ILOCK_SHARED);
606
607         if (!error && (nimaps != 0) &&
608             (imap.br_startblock != HOLESTARTBLOCK ||
609              ip->i_delayed_blks)) {
610                 /*
611                  * Attach the dquots to the inode up front.
612                  */
613                 error = xfs_qm_dqattach(ip, 0);
614                 if (error)
615                         return error;
616
617                 /*
618                  * There are blocks after the end of file.
619                  * Free them up now by truncating the file to
620                  * its current size.
621                  */
622                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
623
624                 /*
625                  * Do the xfs_itruncate_start() call before
626                  * reserving any log space because
627                  * itruncate_start will call into the buffer
628                  * cache and we can't
629                  * do that within a transaction.
630                  */
631                 if (flags & XFS_FREE_EOF_TRYLOCK) {
632                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
633                                 xfs_trans_cancel(tp, 0);
634                                 return 0;
635                         }
636                 } else {
637                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
638                 }
639                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
640                                     ip->i_size);
641                 if (error) {
642                         xfs_trans_cancel(tp, 0);
643                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
644                         return error;
645                 }
646
647                 error = xfs_trans_reserve(tp, 0,
648                                           XFS_ITRUNCATE_LOG_RES(mp),
649                                           0, XFS_TRANS_PERM_LOG_RES,
650                                           XFS_ITRUNCATE_LOG_COUNT);
651                 if (error) {
652                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
653                         xfs_trans_cancel(tp, 0);
654                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
655                         return error;
656                 }
657
658                 xfs_ilock(ip, XFS_ILOCK_EXCL);
659                 xfs_trans_ijoin(tp, ip);
660
661                 error = xfs_itruncate_finish(&tp, ip,
662                                              ip->i_size,
663                                              XFS_DATA_FORK,
664                                              0);
665                 /*
666                  * If we get an error at this point we
667                  * simply don't bother truncating the file.
668                  */
669                 if (error) {
670                         xfs_trans_cancel(tp,
671                                          (XFS_TRANS_RELEASE_LOG_RES |
672                                           XFS_TRANS_ABORT));
673                 } else {
674                         error = xfs_trans_commit(tp,
675                                                 XFS_TRANS_RELEASE_LOG_RES);
676                 }
677                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
678         }
679         return error;
680 }
681
682 /*
683  * Free a symlink that has blocks associated with it.
684  */
685 STATIC int
686 xfs_inactive_symlink_rmt(
687         xfs_inode_t     *ip,
688         xfs_trans_t     **tpp)
689 {
690         xfs_buf_t       *bp;
691         int             committed;
692         int             done;
693         int             error;
694         xfs_fsblock_t   first_block;
695         xfs_bmap_free_t free_list;
696         int             i;
697         xfs_mount_t     *mp;
698         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
699         int             nmaps;
700         xfs_trans_t     *ntp;
701         int             size;
702         xfs_trans_t     *tp;
703
704         tp = *tpp;
705         mp = ip->i_mount;
706         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
707         /*
708          * We're freeing a symlink that has some
709          * blocks allocated to it.  Free the
710          * blocks here.  We know that we've got
711          * either 1 or 2 extents and that we can
712          * free them all in one bunmapi call.
713          */
714         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
715         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
716                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
717                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
718                 xfs_trans_cancel(tp, 0);
719                 *tpp = NULL;
720                 return error;
721         }
722         /*
723          * Lock the inode, fix the size, and join it to the transaction.
724          * Hold it so in the normal path, we still have it locked for
725          * the second transaction.  In the error paths we need it
726          * held so the cancel won't rele it, see below.
727          */
728         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
729         size = (int)ip->i_d.di_size;
730         ip->i_d.di_size = 0;
731         xfs_trans_ijoin(tp, ip);
732         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
733         /*
734          * Find the block(s) so we can inval and unmap them.
735          */
736         done = 0;
737         xfs_bmap_init(&free_list, &first_block);
738         nmaps = ARRAY_SIZE(mval);
739         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
740                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
741                         &free_list)))
742                 goto error0;
743         /*
744          * Invalidate the block(s).
745          */
746         for (i = 0; i < nmaps; i++) {
747                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
748                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
749                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
750                 xfs_trans_binval(tp, bp);
751         }
752         /*
753          * Unmap the dead block(s) to the free_list.
754          */
755         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
756                         &first_block, &free_list, &done)))
757                 goto error1;
758         ASSERT(done);
759         /*
760          * Commit the first transaction.  This logs the EFI and the inode.
761          */
762         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
763                 goto error1;
764         /*
765          * The transaction must have been committed, since there were
766          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
767          * The new tp has the extent freeing and EFDs.
768          */
769         ASSERT(committed);
770         /*
771          * The first xact was committed, so add the inode to the new one.
772          * Mark it dirty so it will be logged and moved forward in the log as
773          * part of every commit.
774          */
775         xfs_trans_ijoin(tp, ip);
776         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
777         /*
778          * Get a new, empty transaction to return to our caller.
779          */
780         ntp = xfs_trans_dup(tp);
781         /*
782          * Commit the transaction containing extent freeing and EFDs.
783          * If we get an error on the commit here or on the reserve below,
784          * we need to unlock the inode since the new transaction doesn't
785          * have the inode attached.
786          */
787         error = xfs_trans_commit(tp, 0);
788         tp = ntp;
789         if (error) {
790                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
791                 goto error0;
792         }
793         /*
794          * transaction commit worked ok so we can drop the extra ticket
795          * reference that we gained in xfs_trans_dup()
796          */
797         xfs_log_ticket_put(tp->t_ticket);
798
799         /*
800          * Remove the memory for extent descriptions (just bookkeeping).
801          */
802         if (ip->i_df.if_bytes)
803                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
804         ASSERT(ip->i_df.if_bytes == 0);
805         /*
806          * Put an itruncate log reservation in the new transaction
807          * for our caller.
808          */
809         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
810                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
811                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
812                 goto error0;
813         }
814         /*
815          * Return with the inode locked but not joined to the transaction.
816          */
817         *tpp = tp;
818         return 0;
819
820  error1:
821         xfs_bmap_cancel(&free_list);
822  error0:
823         /*
824          * Have to come here with the inode locked and either
825          * (held and in the transaction) or (not in the transaction).
826          * If the inode isn't held then cancel would iput it, but
827          * that's wrong since this is inactive and the vnode ref
828          * count is 0 already.
829          * Cancel won't do anything to the inode if held, but it still
830          * needs to be locked until the cancel is done, if it was
831          * joined to the transaction.
832          */
833         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
834         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
835         *tpp = NULL;
836         return error;
837
838 }
839
840 STATIC int
841 xfs_inactive_symlink_local(
842         xfs_inode_t     *ip,
843         xfs_trans_t     **tpp)
844 {
845         int             error;
846
847         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
848         /*
849          * We're freeing a symlink which fit into
850          * the inode.  Just free the memory used
851          * to hold the old symlink.
852          */
853         error = xfs_trans_reserve(*tpp, 0,
854                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
855                                   0, XFS_TRANS_PERM_LOG_RES,
856                                   XFS_ITRUNCATE_LOG_COUNT);
857
858         if (error) {
859                 xfs_trans_cancel(*tpp, 0);
860                 *tpp = NULL;
861                 return error;
862         }
863         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
864
865         /*
866          * Zero length symlinks _can_ exist.
867          */
868         if (ip->i_df.if_bytes > 0) {
869                 xfs_idata_realloc(ip,
870                                   -(ip->i_df.if_bytes),
871                                   XFS_DATA_FORK);
872                 ASSERT(ip->i_df.if_bytes == 0);
873         }
874         return 0;
875 }
876
877 STATIC int
878 xfs_inactive_attrs(
879         xfs_inode_t     *ip,
880         xfs_trans_t     **tpp)
881 {
882         xfs_trans_t     *tp;
883         int             error;
884         xfs_mount_t     *mp;
885
886         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
887         tp = *tpp;
888         mp = ip->i_mount;
889         ASSERT(ip->i_d.di_forkoff != 0);
890         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
891         xfs_iunlock(ip, XFS_ILOCK_EXCL);
892         if (error)
893                 goto error_unlock;
894
895         error = xfs_attr_inactive(ip);
896         if (error)
897                 goto error_unlock;
898
899         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
900         error = xfs_trans_reserve(tp, 0,
901                                   XFS_IFREE_LOG_RES(mp),
902                                   0, XFS_TRANS_PERM_LOG_RES,
903                                   XFS_INACTIVE_LOG_COUNT);
904         if (error)
905                 goto error_cancel;
906
907         xfs_ilock(ip, XFS_ILOCK_EXCL);
908         xfs_trans_ijoin(tp, ip);
909         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
910
911         ASSERT(ip->i_d.di_anextents == 0);
912
913         *tpp = tp;
914         return 0;
915
916 error_cancel:
917         ASSERT(XFS_FORCED_SHUTDOWN(mp));
918         xfs_trans_cancel(tp, 0);
919 error_unlock:
920         *tpp = NULL;
921         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
922         return error;
923 }
924
925 int
926 xfs_release(
927         xfs_inode_t     *ip)
928 {
929         xfs_mount_t     *mp = ip->i_mount;
930         int             error;
931
932         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
933                 return 0;
934
935         /* If this is a read-only mount, don't do this (would generate I/O) */
936         if (mp->m_flags & XFS_MOUNT_RDONLY)
937                 return 0;
938
939         if (!XFS_FORCED_SHUTDOWN(mp)) {
940                 int truncated;
941
942                 /*
943                  * If we are using filestreams, and we have an unlinked
944                  * file that we are processing the last close on, then nothing
945                  * will be able to reopen and write to this file. Purge this
946                  * inode from the filestreams cache so that it doesn't delay
947                  * teardown of the inode.
948                  */
949                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
950                         xfs_filestream_deassociate(ip);
951
952                 /*
953                  * If we previously truncated this file and removed old data
954                  * in the process, we want to initiate "early" writeout on
955                  * the last close.  This is an attempt to combat the notorious
956                  * NULL files problem which is particularly noticeable from a
957                  * truncate down, buffered (re-)write (delalloc), followed by
958                  * a crash.  What we are effectively doing here is
959                  * significantly reducing the time window where we'd otherwise
960                  * be exposed to that problem.
961                  */
962                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
963                 if (truncated) {
964                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
965                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
966                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
967                 }
968         }
969
970         if (ip->i_d.di_nlink == 0)
971                 return 0;
972
973         if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
974              ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
975                ip->i_delayed_blks > 0)) &&
976              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
977             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
978
979                 /*
980                  * If we can't get the iolock just skip truncating the blocks
981                  * past EOF because we could deadlock with the mmap_sem
982                  * otherwise.  We'll get another chance to drop them once the
983                  * last reference to the inode is dropped, so we'll never leak
984                  * blocks permanently.
985                  *
986                  * Further, check if the inode is being opened, written and
987                  * closed frequently and we have delayed allocation blocks
988                  * outstanding (e.g. streaming writes from the NFS server),
989                  * truncating the blocks past EOF will cause fragmentation to
990                  * occur.
991                  *
992                  * In this case don't do the truncation, either, but we have to
993                  * be careful how we detect this case. Blocks beyond EOF show
994                  * up as i_delayed_blks even when the inode is clean, so we
995                  * need to truncate them away first before checking for a dirty
996                  * release. Hence on the first dirty close we will still remove
997                  * the speculative allocation, but after that we will leave it
998                  * in place.
999                  */
1000                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1001                         return 0;
1002
1003                 error = xfs_free_eofblocks(mp, ip,
1004                                            XFS_FREE_EOF_TRYLOCK);
1005                 if (error)
1006                         return error;
1007
1008                 /* delalloc blocks after truncation means it really is dirty */
1009                 if (ip->i_delayed_blks)
1010                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1011         }
1012         return 0;
1013 }
1014
1015 /*
1016  * xfs_inactive
1017  *
1018  * This is called when the vnode reference count for the vnode
1019  * goes to zero.  If the file has been unlinked, then it must
1020  * now be truncated.  Also, we clear all of the read-ahead state
1021  * kept for the inode here since the file is now closed.
1022  */
1023 int
1024 xfs_inactive(
1025         xfs_inode_t     *ip)
1026 {
1027         xfs_bmap_free_t free_list;
1028         xfs_fsblock_t   first_block;
1029         int             committed;
1030         xfs_trans_t     *tp;
1031         xfs_mount_t     *mp;
1032         int             error;
1033         int             truncate;
1034
1035         /*
1036          * If the inode is already free, then there can be nothing
1037          * to clean up here.
1038          */
1039         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
1040                 ASSERT(ip->i_df.if_real_bytes == 0);
1041                 ASSERT(ip->i_df.if_broot_bytes == 0);
1042                 return VN_INACTIVE_CACHE;
1043         }
1044
1045         /*
1046          * Only do a truncate if it's a regular file with
1047          * some actual space in it.  It's OK to look at the
1048          * inode's fields without the lock because we're the
1049          * only one with a reference to the inode.
1050          */
1051         truncate = ((ip->i_d.di_nlink == 0) &&
1052             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1053              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1054             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1055
1056         mp = ip->i_mount;
1057
1058         error = 0;
1059
1060         /* If this is a read-only mount, don't do this (would generate I/O) */
1061         if (mp->m_flags & XFS_MOUNT_RDONLY)
1062                 goto out;
1063
1064         if (ip->i_d.di_nlink != 0) {
1065                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1066                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1067                        ip->i_delayed_blks > 0)) &&
1068                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1069                      (!(ip->i_d.di_flags &
1070                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1071                       (ip->i_delayed_blks != 0)))) {
1072                         error = xfs_free_eofblocks(mp, ip, 0);
1073                         if (error)
1074                                 return VN_INACTIVE_CACHE;
1075                 }
1076                 goto out;
1077         }
1078
1079         ASSERT(ip->i_d.di_nlink == 0);
1080
1081         error = xfs_qm_dqattach(ip, 0);
1082         if (error)
1083                 return VN_INACTIVE_CACHE;
1084
1085         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1086         if (truncate) {
1087                 /*
1088                  * Do the xfs_itruncate_start() call before
1089                  * reserving any log space because itruncate_start
1090                  * will call into the buffer cache and we can't
1091                  * do that within a transaction.
1092                  */
1093                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1094
1095                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1096                 if (error) {
1097                         xfs_trans_cancel(tp, 0);
1098                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1099                         return VN_INACTIVE_CACHE;
1100                 }
1101
1102                 error = xfs_trans_reserve(tp, 0,
1103                                           XFS_ITRUNCATE_LOG_RES(mp),
1104                                           0, XFS_TRANS_PERM_LOG_RES,
1105                                           XFS_ITRUNCATE_LOG_COUNT);
1106                 if (error) {
1107                         /* Don't call itruncate_cleanup */
1108                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1109                         xfs_trans_cancel(tp, 0);
1110                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1111                         return VN_INACTIVE_CACHE;
1112                 }
1113
1114                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1115                 xfs_trans_ijoin(tp, ip);
1116
1117                 /*
1118                  * normally, we have to run xfs_itruncate_finish sync.
1119                  * But if filesystem is wsync and we're in the inactive
1120                  * path, then we know that nlink == 0, and that the
1121                  * xaction that made nlink == 0 is permanently committed
1122                  * since xfs_remove runs as a synchronous transaction.
1123                  */
1124                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1125                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1126
1127                 if (error) {
1128                         xfs_trans_cancel(tp,
1129                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1130                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1131                         return VN_INACTIVE_CACHE;
1132                 }
1133         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1134
1135                 /*
1136                  * If we get an error while cleaning up a
1137                  * symlink we bail out.
1138                  */
1139                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1140                         xfs_inactive_symlink_rmt(ip, &tp) :
1141                         xfs_inactive_symlink_local(ip, &tp);
1142
1143                 if (error) {
1144                         ASSERT(tp == NULL);
1145                         return VN_INACTIVE_CACHE;
1146                 }
1147
1148                 xfs_trans_ijoin(tp, ip);
1149         } else {
1150                 error = xfs_trans_reserve(tp, 0,
1151                                           XFS_IFREE_LOG_RES(mp),
1152                                           0, XFS_TRANS_PERM_LOG_RES,
1153                                           XFS_INACTIVE_LOG_COUNT);
1154                 if (error) {
1155                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1156                         xfs_trans_cancel(tp, 0);
1157                         return VN_INACTIVE_CACHE;
1158                 }
1159
1160                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1161                 xfs_trans_ijoin(tp, ip);
1162         }
1163
1164         /*
1165          * If there are attributes associated with the file
1166          * then blow them away now.  The code calls a routine
1167          * that recursively deconstructs the attribute fork.
1168          * We need to just commit the current transaction
1169          * because we can't use it for xfs_attr_inactive().
1170          */
1171         if (ip->i_d.di_anextents > 0) {
1172                 error = xfs_inactive_attrs(ip, &tp);
1173                 /*
1174                  * If we got an error, the transaction is already
1175                  * cancelled, and the inode is unlocked. Just get out.
1176                  */
1177                  if (error)
1178                          return VN_INACTIVE_CACHE;
1179         } else if (ip->i_afp) {
1180                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1181         }
1182
1183         /*
1184          * Free the inode.
1185          */
1186         xfs_bmap_init(&free_list, &first_block);
1187         error = xfs_ifree(tp, ip, &free_list);
1188         if (error) {
1189                 /*
1190                  * If we fail to free the inode, shut down.  The cancel
1191                  * might do that, we need to make sure.  Otherwise the
1192                  * inode might be lost for a long time or forever.
1193                  */
1194                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1195                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
1196                                 __func__, error);
1197                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1198                 }
1199                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1200         } else {
1201                 /*
1202                  * Credit the quota account(s). The inode is gone.
1203                  */
1204                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1205
1206                 /*
1207                  * Just ignore errors at this point.  There is nothing we can
1208                  * do except to try to keep going. Make sure it's not a silent
1209                  * error.
1210                  */
1211                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1212                 if (error)
1213                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
1214                                 __func__, error);
1215                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1216                 if (error)
1217                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1218                                 __func__, error);
1219         }
1220
1221         /*
1222          * Release the dquots held by inode, if any.
1223          */
1224         xfs_qm_dqdetach(ip);
1225         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1226
1227  out:
1228         return VN_INACTIVE_CACHE;
1229 }
1230
1231 /*
1232  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1233  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1234  * ci_name->name will point to a the actual name (caller must free) or
1235  * will be set to NULL if an exact match is found.
1236  */
1237 int
1238 xfs_lookup(
1239         xfs_inode_t             *dp,
1240         struct xfs_name         *name,
1241         xfs_inode_t             **ipp,
1242         struct xfs_name         *ci_name)
1243 {
1244         xfs_ino_t               inum;
1245         int                     error;
1246         uint                    lock_mode;
1247
1248         trace_xfs_lookup(dp, name);
1249
1250         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1251                 return XFS_ERROR(EIO);
1252
1253         lock_mode = xfs_ilock_map_shared(dp);
1254         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1255         xfs_iunlock_map_shared(dp, lock_mode);
1256
1257         if (error)
1258                 goto out;
1259
1260         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
1261         if (error)
1262                 goto out_free_name;
1263
1264         return 0;
1265
1266 out_free_name:
1267         if (ci_name)
1268                 kmem_free(ci_name->name);
1269 out:
1270         *ipp = NULL;
1271         return error;
1272 }
1273
1274 int
1275 xfs_create(
1276         xfs_inode_t             *dp,
1277         struct xfs_name         *name,
1278         mode_t                  mode,
1279         xfs_dev_t               rdev,
1280         xfs_inode_t             **ipp)
1281 {
1282         int                     is_dir = S_ISDIR(mode);
1283         struct xfs_mount        *mp = dp->i_mount;
1284         struct xfs_inode        *ip = NULL;
1285         struct xfs_trans        *tp = NULL;
1286         int                     error;
1287         xfs_bmap_free_t         free_list;
1288         xfs_fsblock_t           first_block;
1289         boolean_t               unlock_dp_on_error = B_FALSE;
1290         uint                    cancel_flags;
1291         int                     committed;
1292         prid_t                  prid;
1293         struct xfs_dquot        *udqp = NULL;
1294         struct xfs_dquot        *gdqp = NULL;
1295         uint                    resblks;
1296         uint                    log_res;
1297         uint                    log_count;
1298
1299         trace_xfs_create(dp, name);
1300
1301         if (XFS_FORCED_SHUTDOWN(mp))
1302                 return XFS_ERROR(EIO);
1303
1304         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1305                 prid = xfs_get_projid(dp);
1306         else
1307                 prid = XFS_PROJID_DEFAULT;
1308
1309         /*
1310          * Make sure that we have allocated dquot(s) on disk.
1311          */
1312         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1313                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1314         if (error)
1315                 return error;
1316
1317         if (is_dir) {
1318                 rdev = 0;
1319                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1320                 log_res = XFS_MKDIR_LOG_RES(mp);
1321                 log_count = XFS_MKDIR_LOG_COUNT;
1322                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
1323         } else {
1324                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1325                 log_res = XFS_CREATE_LOG_RES(mp);
1326                 log_count = XFS_CREATE_LOG_COUNT;
1327                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1328         }
1329
1330         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1331
1332         /*
1333          * Initially assume that the file does not exist and
1334          * reserve the resources for that case.  If that is not
1335          * the case we'll drop the one we have and get a more
1336          * appropriate transaction later.
1337          */
1338         error = xfs_trans_reserve(tp, resblks, log_res, 0,
1339                         XFS_TRANS_PERM_LOG_RES, log_count);
1340         if (error == ENOSPC) {
1341                 /* flush outstanding delalloc blocks and retry */
1342                 xfs_flush_inodes(dp);
1343                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
1344                                 XFS_TRANS_PERM_LOG_RES, log_count);
1345         }
1346         if (error == ENOSPC) {
1347                 /* No space at all so try a "no-allocation" reservation */
1348                 resblks = 0;
1349                 error = xfs_trans_reserve(tp, 0, log_res, 0,
1350                                 XFS_TRANS_PERM_LOG_RES, log_count);
1351         }
1352         if (error) {
1353                 cancel_flags = 0;
1354                 goto out_trans_cancel;
1355         }
1356
1357         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1358         unlock_dp_on_error = B_TRUE;
1359
1360         /*
1361          * Check for directory link count overflow.
1362          */
1363         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
1364                 error = XFS_ERROR(EMLINK);
1365                 goto out_trans_cancel;
1366         }
1367
1368         xfs_bmap_init(&free_list, &first_block);
1369
1370         /*
1371          * Reserve disk quota and the inode.
1372          */
1373         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1374         if (error)
1375                 goto out_trans_cancel;
1376
1377         error = xfs_dir_canenter(tp, dp, name, resblks);
1378         if (error)
1379                 goto out_trans_cancel;
1380
1381         /*
1382          * A newly created regular or special file just has one directory
1383          * entry pointing to them, but a directory also the "." entry
1384          * pointing to itself.
1385          */
1386         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1387                                prid, resblks > 0, &ip, &committed);
1388         if (error) {
1389                 if (error == ENOSPC)
1390                         goto out_trans_cancel;
1391                 goto out_trans_abort;
1392         }
1393
1394         /*
1395          * Now we join the directory inode to the transaction.  We do not do it
1396          * earlier because xfs_dir_ialloc might commit the previous transaction
1397          * (and release all the locks).  An error from here on will result in
1398          * the transaction cancel unlocking dp so don't do it explicitly in the
1399          * error path.
1400          */
1401         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1402         unlock_dp_on_error = B_FALSE;
1403
1404         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1405                                         &first_block, &free_list, resblks ?
1406                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1407         if (error) {
1408                 ASSERT(error != ENOSPC);
1409                 goto out_trans_abort;
1410         }
1411         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1412         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1413
1414         if (is_dir) {
1415                 error = xfs_dir_init(tp, ip, dp);
1416                 if (error)
1417                         goto out_bmap_cancel;
1418
1419                 error = xfs_bumplink(tp, dp);
1420                 if (error)
1421                         goto out_bmap_cancel;
1422         }
1423
1424         /*
1425          * If this is a synchronous mount, make sure that the
1426          * create transaction goes to disk before returning to
1427          * the user.
1428          */
1429         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1430                 xfs_trans_set_sync(tp);
1431
1432         /*
1433          * Attach the dquot(s) to the inodes and modify them incore.
1434          * These ids of the inode couldn't have changed since the new
1435          * inode has been locked ever since it was created.
1436          */
1437         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1438
1439         error = xfs_bmap_finish(&tp, &free_list, &committed);
1440         if (error)
1441                 goto out_bmap_cancel;
1442
1443         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1444         if (error)
1445                 goto out_release_inode;
1446
1447         xfs_qm_dqrele(udqp);
1448         xfs_qm_dqrele(gdqp);
1449
1450         *ipp = ip;
1451         return 0;
1452
1453  out_bmap_cancel:
1454         xfs_bmap_cancel(&free_list);
1455  out_trans_abort:
1456         cancel_flags |= XFS_TRANS_ABORT;
1457  out_trans_cancel:
1458         xfs_trans_cancel(tp, cancel_flags);
1459  out_release_inode:
1460         /*
1461          * Wait until after the current transaction is aborted to
1462          * release the inode.  This prevents recursive transactions
1463          * and deadlocks from xfs_inactive.
1464          */
1465         if (ip)
1466                 IRELE(ip);
1467
1468         xfs_qm_dqrele(udqp);
1469         xfs_qm_dqrele(gdqp);
1470
1471         if (unlock_dp_on_error)
1472                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1473         return error;
1474 }
1475
1476 #ifdef DEBUG
1477 int xfs_locked_n;
1478 int xfs_small_retries;
1479 int xfs_middle_retries;
1480 int xfs_lots_retries;
1481 int xfs_lock_delays;
1482 #endif
1483
1484 /*
1485  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1486  * a different value
1487  */
1488 static inline int
1489 xfs_lock_inumorder(int lock_mode, int subclass)
1490 {
1491         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1492                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1493         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1494                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1495
1496         return lock_mode;
1497 }
1498
1499 /*
1500  * The following routine will lock n inodes in exclusive mode.
1501  * We assume the caller calls us with the inodes in i_ino order.
1502  *
1503  * We need to detect deadlock where an inode that we lock
1504  * is in the AIL and we start waiting for another inode that is locked
1505  * by a thread in a long running transaction (such as truncate). This can
1506  * result in deadlock since the long running trans might need to wait
1507  * for the inode we just locked in order to push the tail and free space
1508  * in the log.
1509  */
1510 void
1511 xfs_lock_inodes(
1512         xfs_inode_t     **ips,
1513         int             inodes,
1514         uint            lock_mode)
1515 {
1516         int             attempts = 0, i, j, try_lock;
1517         xfs_log_item_t  *lp;
1518
1519         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1520
1521         try_lock = 0;
1522         i = 0;
1523
1524 again:
1525         for (; i < inodes; i++) {
1526                 ASSERT(ips[i]);
1527
1528                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1529                         continue;
1530
1531                 /*
1532                  * If try_lock is not set yet, make sure all locked inodes
1533                  * are not in the AIL.
1534                  * If any are, set try_lock to be used later.
1535                  */
1536
1537                 if (!try_lock) {
1538                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1539                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1540                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1541                                         try_lock++;
1542                                 }
1543                         }
1544                 }
1545
1546                 /*
1547                  * If any of the previous locks we have locked is in the AIL,
1548                  * we must TRY to get the second and subsequent locks. If
1549                  * we can't get any, we must release all we have
1550                  * and try again.
1551                  */
1552
1553                 if (try_lock) {
1554                         /* try_lock must be 0 if i is 0. */
1555                         /*
1556                          * try_lock means we have an inode locked
1557                          * that is in the AIL.
1558                          */
1559                         ASSERT(i != 0);
1560                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1561                                 attempts++;
1562
1563                                 /*
1564                                  * Unlock all previous guys and try again.
1565                                  * xfs_iunlock will try to push the tail
1566                                  * if the inode is in the AIL.
1567                                  */
1568
1569                                 for(j = i - 1; j >= 0; j--) {
1570
1571                                         /*
1572                                          * Check to see if we've already
1573                                          * unlocked this one.
1574                                          * Not the first one going back,
1575                                          * and the inode ptr is the same.
1576                                          */
1577                                         if ((j != (i - 1)) && ips[j] ==
1578                                                                 ips[j+1])
1579                                                 continue;
1580
1581                                         xfs_iunlock(ips[j], lock_mode);
1582                                 }
1583
1584                                 if ((attempts % 5) == 0) {
1585                                         delay(1); /* Don't just spin the CPU */
1586 #ifdef DEBUG
1587                                         xfs_lock_delays++;
1588 #endif
1589                                 }
1590                                 i = 0;
1591                                 try_lock = 0;
1592                                 goto again;
1593                         }
1594                 } else {
1595                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1596                 }
1597         }
1598
1599 #ifdef DEBUG
1600         if (attempts) {
1601                 if (attempts < 5) xfs_small_retries++;
1602                 else if (attempts < 100) xfs_middle_retries++;
1603                 else xfs_lots_retries++;
1604         } else {
1605                 xfs_locked_n++;
1606         }
1607 #endif
1608 }
1609
1610 /*
1611  * xfs_lock_two_inodes() can only be used to lock one type of lock
1612  * at a time - the iolock or the ilock, but not both at once. If
1613  * we lock both at once, lockdep will report false positives saying
1614  * we have violated locking orders.
1615  */
1616 void
1617 xfs_lock_two_inodes(
1618         xfs_inode_t             *ip0,
1619         xfs_inode_t             *ip1,
1620         uint                    lock_mode)
1621 {
1622         xfs_inode_t             *temp;
1623         int                     attempts = 0;
1624         xfs_log_item_t          *lp;
1625
1626         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1627                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1628         ASSERT(ip0->i_ino != ip1->i_ino);
1629
1630         if (ip0->i_ino > ip1->i_ino) {
1631                 temp = ip0;
1632                 ip0 = ip1;
1633                 ip1 = temp;
1634         }
1635
1636  again:
1637         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1638
1639         /*
1640          * If the first lock we have locked is in the AIL, we must TRY to get
1641          * the second lock. If we can't get it, we must release the first one
1642          * and try again.
1643          */
1644         lp = (xfs_log_item_t *)ip0->i_itemp;
1645         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1646                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1647                         xfs_iunlock(ip0, lock_mode);
1648                         if ((++attempts % 5) == 0)
1649                                 delay(1); /* Don't just spin the CPU */
1650                         goto again;
1651                 }
1652         } else {
1653                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1654         }
1655 }
1656
1657 int
1658 xfs_remove(
1659         xfs_inode_t             *dp,
1660         struct xfs_name         *name,
1661         xfs_inode_t             *ip)
1662 {
1663         xfs_mount_t             *mp = dp->i_mount;
1664         xfs_trans_t             *tp = NULL;
1665         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1666         int                     error = 0;
1667         xfs_bmap_free_t         free_list;
1668         xfs_fsblock_t           first_block;
1669         int                     cancel_flags;
1670         int                     committed;
1671         int                     link_zero;
1672         uint                    resblks;
1673         uint                    log_count;
1674
1675         trace_xfs_remove(dp, name);
1676
1677         if (XFS_FORCED_SHUTDOWN(mp))
1678                 return XFS_ERROR(EIO);
1679
1680         error = xfs_qm_dqattach(dp, 0);
1681         if (error)
1682                 goto std_return;
1683
1684         error = xfs_qm_dqattach(ip, 0);
1685         if (error)
1686                 goto std_return;
1687
1688         if (is_dir) {
1689                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1690                 log_count = XFS_DEFAULT_LOG_COUNT;
1691         } else {
1692                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1693                 log_count = XFS_REMOVE_LOG_COUNT;
1694         }
1695         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1696
1697         /*
1698          * We try to get the real space reservation first,
1699          * allowing for directory btree deletion(s) implying
1700          * possible bmap insert(s).  If we can't get the space
1701          * reservation then we use 0 instead, and avoid the bmap
1702          * btree insert(s) in the directory code by, if the bmap
1703          * insert tries to happen, instead trimming the LAST
1704          * block from the directory.
1705          */
1706         resblks = XFS_REMOVE_SPACE_RES(mp);
1707         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1708                                   XFS_TRANS_PERM_LOG_RES, log_count);
1709         if (error == ENOSPC) {
1710                 resblks = 0;
1711                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1712                                           XFS_TRANS_PERM_LOG_RES, log_count);
1713         }
1714         if (error) {
1715                 ASSERT(error != ENOSPC);
1716                 cancel_flags = 0;
1717                 goto out_trans_cancel;
1718         }
1719
1720         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1721
1722         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1723         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1724
1725         /*
1726          * If we're removing a directory perform some additional validation.
1727          */
1728         if (is_dir) {
1729                 ASSERT(ip->i_d.di_nlink >= 2);
1730                 if (ip->i_d.di_nlink != 2) {
1731                         error = XFS_ERROR(ENOTEMPTY);
1732                         goto out_trans_cancel;
1733                 }
1734                 if (!xfs_dir_isempty(ip)) {
1735                         error = XFS_ERROR(ENOTEMPTY);
1736                         goto out_trans_cancel;
1737                 }
1738         }
1739
1740         xfs_bmap_init(&free_list, &first_block);
1741         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1742                                         &first_block, &free_list, resblks);
1743         if (error) {
1744                 ASSERT(error != ENOENT);
1745                 goto out_bmap_cancel;
1746         }
1747         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1748
1749         if (is_dir) {
1750                 /*
1751                  * Drop the link from ip's "..".
1752                  */
1753                 error = xfs_droplink(tp, dp);
1754                 if (error)
1755                         goto out_bmap_cancel;
1756
1757                 /*
1758                  * Drop the "." link from ip to self.
1759                  */
1760                 error = xfs_droplink(tp, ip);
1761                 if (error)
1762                         goto out_bmap_cancel;
1763         } else {
1764                 /*
1765                  * When removing a non-directory we need to log the parent
1766                  * inode here.  For a directory this is done implicitly
1767                  * by the xfs_droplink call for the ".." entry.
1768                  */
1769                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1770         }
1771
1772         /*
1773          * Drop the link from dp to ip.
1774          */
1775         error = xfs_droplink(tp, ip);
1776         if (error)
1777                 goto out_bmap_cancel;
1778
1779         /*
1780          * Determine if this is the last link while
1781          * we are in the transaction.
1782          */
1783         link_zero = (ip->i_d.di_nlink == 0);
1784
1785         /*
1786          * If this is a synchronous mount, make sure that the
1787          * remove transaction goes to disk before returning to
1788          * the user.
1789          */
1790         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1791                 xfs_trans_set_sync(tp);
1792
1793         error = xfs_bmap_finish(&tp, &free_list, &committed);
1794         if (error)
1795                 goto out_bmap_cancel;
1796
1797         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1798         if (error)
1799                 goto std_return;
1800
1801         /*
1802          * If we are using filestreams, kill the stream association.
1803          * If the file is still open it may get a new one but that
1804          * will get killed on last close in xfs_close() so we don't
1805          * have to worry about that.
1806          */
1807         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1808                 xfs_filestream_deassociate(ip);
1809
1810         return 0;
1811
1812  out_bmap_cancel:
1813         xfs_bmap_cancel(&free_list);
1814         cancel_flags |= XFS_TRANS_ABORT;
1815  out_trans_cancel:
1816         xfs_trans_cancel(tp, cancel_flags);
1817  std_return:
1818         return error;
1819 }
1820
1821 int
1822 xfs_link(
1823         xfs_inode_t             *tdp,
1824         xfs_inode_t             *sip,
1825         struct xfs_name         *target_name)
1826 {
1827         xfs_mount_t             *mp = tdp->i_mount;
1828         xfs_trans_t             *tp;
1829         int                     error;
1830         xfs_bmap_free_t         free_list;
1831         xfs_fsblock_t           first_block;
1832         int                     cancel_flags;
1833         int                     committed;
1834         int                     resblks;
1835
1836         trace_xfs_link(tdp, target_name);
1837
1838         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1839
1840         if (XFS_FORCED_SHUTDOWN(mp))
1841                 return XFS_ERROR(EIO);
1842
1843         error = xfs_qm_dqattach(sip, 0);
1844         if (error)
1845                 goto std_return;
1846
1847         error = xfs_qm_dqattach(tdp, 0);
1848         if (error)
1849                 goto std_return;
1850
1851         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1852         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1853         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1854         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1855                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1856         if (error == ENOSPC) {
1857                 resblks = 0;
1858                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1859                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1860         }
1861         if (error) {
1862                 cancel_flags = 0;
1863                 goto error_return;
1864         }
1865
1866         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1867
1868         xfs_trans_ijoin_ref(tp, sip, XFS_ILOCK_EXCL);
1869         xfs_trans_ijoin_ref(tp, tdp, XFS_ILOCK_EXCL);
1870
1871         /*
1872          * If the source has too many links, we can't make any more to it.
1873          */
1874         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1875                 error = XFS_ERROR(EMLINK);
1876                 goto error_return;
1877         }
1878
1879         /*
1880          * If we are using project inheritance, we only allow hard link
1881          * creation in our tree when the project IDs are the same; else
1882          * the tree quota mechanism could be circumvented.
1883          */
1884         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1885                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1886                 error = XFS_ERROR(EXDEV);
1887                 goto error_return;
1888         }
1889
1890         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1891         if (error)
1892                 goto error_return;
1893
1894         xfs_bmap_init(&free_list, &first_block);
1895
1896         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1897                                         &first_block, &free_list, resblks);
1898         if (error)
1899                 goto abort_return;
1900         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1901         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1902
1903         error = xfs_bumplink(tp, sip);
1904         if (error)
1905                 goto abort_return;
1906
1907         /*
1908          * If this is a synchronous mount, make sure that the
1909          * link transaction goes to disk before returning to
1910          * the user.
1911          */
1912         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1913                 xfs_trans_set_sync(tp);
1914         }
1915
1916         error = xfs_bmap_finish (&tp, &free_list, &committed);
1917         if (error) {
1918                 xfs_bmap_cancel(&free_list);
1919                 goto abort_return;
1920         }
1921
1922         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1923
1924  abort_return:
1925         cancel_flags |= XFS_TRANS_ABORT;
1926  error_return:
1927         xfs_trans_cancel(tp, cancel_flags);
1928  std_return:
1929         return error;
1930 }
1931
1932 int
1933 xfs_symlink(
1934         xfs_inode_t             *dp,
1935         struct xfs_name         *link_name,
1936         const char              *target_path,
1937         mode_t                  mode,
1938         xfs_inode_t             **ipp)
1939 {
1940         xfs_mount_t             *mp = dp->i_mount;
1941         xfs_trans_t             *tp;
1942         xfs_inode_t             *ip;
1943         int                     error;
1944         int                     pathlen;
1945         xfs_bmap_free_t         free_list;
1946         xfs_fsblock_t           first_block;
1947         boolean_t               unlock_dp_on_error = B_FALSE;
1948         uint                    cancel_flags;
1949         int                     committed;
1950         xfs_fileoff_t           first_fsb;
1951         xfs_filblks_t           fs_blocks;
1952         int                     nmaps;
1953         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1954         xfs_daddr_t             d;
1955         const char              *cur_chunk;
1956         int                     byte_cnt;
1957         int                     n;
1958         xfs_buf_t               *bp;
1959         prid_t                  prid;
1960         struct xfs_dquot        *udqp, *gdqp;
1961         uint                    resblks;
1962
1963         *ipp = NULL;
1964         error = 0;
1965         ip = NULL;
1966         tp = NULL;
1967
1968         trace_xfs_symlink(dp, link_name);
1969
1970         if (XFS_FORCED_SHUTDOWN(mp))
1971                 return XFS_ERROR(EIO);
1972
1973         /*
1974          * Check component lengths of the target path name.
1975          */
1976         pathlen = strlen(target_path);
1977         if (pathlen >= MAXPATHLEN)      /* total string too long */
1978                 return XFS_ERROR(ENAMETOOLONG);
1979
1980         udqp = gdqp = NULL;
1981         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1982                 prid = xfs_get_projid(dp);
1983         else
1984                 prid = XFS_PROJID_DEFAULT;
1985
1986         /*
1987          * Make sure that we have allocated dquot(s) on disk.
1988          */
1989         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1990                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1991         if (error)
1992                 goto std_return;
1993
1994         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1995         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1996         /*
1997          * The symlink will fit into the inode data fork?
1998          * There can't be any attributes so we get the whole variable part.
1999          */
2000         if (pathlen <= XFS_LITINO(mp))
2001                 fs_blocks = 0;
2002         else
2003                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2004         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2005         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2006                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2007         if (error == ENOSPC && fs_blocks == 0) {
2008                 resblks = 0;
2009                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2010                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2011         }
2012         if (error) {
2013                 cancel_flags = 0;
2014                 goto error_return;
2015         }
2016
2017         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2018         unlock_dp_on_error = B_TRUE;
2019
2020         /*
2021          * Check whether the directory allows new symlinks or not.
2022          */
2023         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2024                 error = XFS_ERROR(EPERM);
2025                 goto error_return;
2026         }
2027
2028         /*
2029          * Reserve disk quota : blocks and inode.
2030          */
2031         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
2032         if (error)
2033                 goto error_return;
2034
2035         /*
2036          * Check for ability to enter directory entry, if no space reserved.
2037          */
2038         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2039         if (error)
2040                 goto error_return;
2041         /*
2042          * Initialize the bmap freelist prior to calling either
2043          * bmapi or the directory create code.
2044          */
2045         xfs_bmap_init(&free_list, &first_block);
2046
2047         /*
2048          * Allocate an inode for the symlink.
2049          */
2050         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
2051                                prid, resblks > 0, &ip, NULL);
2052         if (error) {
2053                 if (error == ENOSPC)
2054                         goto error_return;
2055                 goto error1;
2056         }
2057
2058         /*
2059          * An error after we've joined dp to the transaction will result in the
2060          * transaction cancel unlocking dp so don't do it explicitly in the
2061          * error path.
2062          */
2063         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
2064         unlock_dp_on_error = B_FALSE;
2065
2066         /*
2067          * Also attach the dquot(s) to it, if applicable.
2068          */
2069         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
2070
2071         if (resblks)
2072                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2073         /*
2074          * If the symlink will fit into the inode, write it inline.
2075          */
2076         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2077                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2078                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2079                 ip->i_d.di_size = pathlen;
2080
2081                 /*
2082                  * The inode was initially created in extent format.
2083                  */
2084                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2085                 ip->i_df.if_flags |= XFS_IFINLINE;
2086
2087                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2088                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2089
2090         } else {
2091                 first_fsb = 0;
2092                 nmaps = SYMLINK_MAPS;
2093
2094                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2095                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2096                                   &first_block, resblks, mval, &nmaps,
2097                                   &free_list);
2098                 if (error)
2099                         goto error2;
2100
2101                 if (resblks)
2102                         resblks -= fs_blocks;
2103                 ip->i_d.di_size = pathlen;
2104                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2105
2106                 cur_chunk = target_path;
2107                 for (n = 0; n < nmaps; n++) {
2108                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2109                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2110                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2111                                                BTOBB(byte_cnt), 0);
2112                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2113                         if (pathlen < byte_cnt) {
2114                                 byte_cnt = pathlen;
2115                         }
2116                         pathlen -= byte_cnt;
2117
2118                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2119                         cur_chunk += byte_cnt;
2120
2121                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2122                 }
2123         }
2124
2125         /*
2126          * Create the directory entry for the symlink.
2127          */
2128         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2129                                         &first_block, &free_list, resblks);
2130         if (error)
2131                 goto error2;
2132         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2133         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2134
2135         /*
2136          * If this is a synchronous mount, make sure that the
2137          * symlink transaction goes to disk before returning to
2138          * the user.
2139          */
2140         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2141                 xfs_trans_set_sync(tp);
2142         }
2143
2144         error = xfs_bmap_finish(&tp, &free_list, &committed);
2145         if (error) {
2146                 goto error2;
2147         }
2148         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2149         xfs_qm_dqrele(udqp);
2150         xfs_qm_dqrele(gdqp);
2151
2152         *ipp = ip;
2153         return 0;
2154
2155  error2:
2156         IRELE(ip);
2157  error1:
2158         xfs_bmap_cancel(&free_list);
2159         cancel_flags |= XFS_TRANS_ABORT;
2160  error_return:
2161         xfs_trans_cancel(tp, cancel_flags);
2162         xfs_qm_dqrele(udqp);
2163         xfs_qm_dqrele(gdqp);
2164
2165         if (unlock_dp_on_error)
2166                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2167  std_return:
2168         return error;
2169 }
2170
2171 int
2172 xfs_set_dmattrs(
2173         xfs_inode_t     *ip,
2174         u_int           evmask,
2175         u_int16_t       state)
2176 {
2177         xfs_mount_t     *mp = ip->i_mount;
2178         xfs_trans_t     *tp;
2179         int             error;
2180
2181         if (!capable(CAP_SYS_ADMIN))
2182                 return XFS_ERROR(EPERM);
2183
2184         if (XFS_FORCED_SHUTDOWN(mp))
2185                 return XFS_ERROR(EIO);
2186
2187         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2188         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2189         if (error) {
2190                 xfs_trans_cancel(tp, 0);
2191                 return error;
2192         }
2193         xfs_ilock(ip, XFS_ILOCK_EXCL);
2194         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
2195
2196         ip->i_d.di_dmevmask = evmask;
2197         ip->i_d.di_dmstate  = state;
2198
2199         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2200         error = xfs_trans_commit(tp, 0);
2201
2202         return error;
2203 }
2204
2205 /*
2206  * xfs_alloc_file_space()
2207  *      This routine allocates disk space for the given file.
2208  *
2209  *      If alloc_type == 0, this request is for an ALLOCSP type
2210  *      request which will change the file size.  In this case, no
2211  *      DMAPI event will be generated by the call.  A TRUNCATE event
2212  *      will be generated later by xfs_setattr.
2213  *
2214  *      If alloc_type != 0, this request is for a RESVSP type
2215  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2216  *      lower block boundary byte address is less than the file's
2217  *      length.
2218  *
2219  * RETURNS:
2220  *       0 on success
2221  *      errno on error
2222  *
2223  */
2224 STATIC int
2225 xfs_alloc_file_space(
2226         xfs_inode_t             *ip,
2227         xfs_off_t               offset,
2228         xfs_off_t               len,
2229         int                     alloc_type,
2230         int                     attr_flags)
2231 {
2232         xfs_mount_t             *mp = ip->i_mount;
2233         xfs_off_t               count;
2234         xfs_filblks_t           allocated_fsb;
2235         xfs_filblks_t           allocatesize_fsb;
2236         xfs_extlen_t            extsz, temp;
2237         xfs_fileoff_t           startoffset_fsb;
2238         xfs_fsblock_t           firstfsb;
2239         int                     nimaps;
2240         int                     bmapi_flag;
2241         int                     quota_flag;
2242         int                     rt;
2243         xfs_trans_t             *tp;
2244         xfs_bmbt_irec_t         imaps[1], *imapp;
2245         xfs_bmap_free_t         free_list;
2246         uint                    qblocks, resblks, resrtextents;
2247         int                     committed;
2248         int                     error;
2249
2250         trace_xfs_alloc_file_space(ip);
2251
2252         if (XFS_FORCED_SHUTDOWN(mp))
2253                 return XFS_ERROR(EIO);
2254
2255         error = xfs_qm_dqattach(ip, 0);
2256         if (error)
2257                 return error;
2258
2259         if (len <= 0)
2260                 return XFS_ERROR(EINVAL);
2261
2262         rt = XFS_IS_REALTIME_INODE(ip);
2263         extsz = xfs_get_extsz_hint(ip);
2264
2265         count = len;
2266         imapp = &imaps[0];
2267         nimaps = 1;
2268         bmapi_flag = XFS_BMAPI_WRITE | alloc_type;
2269         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
2270         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
2271
2272         /*
2273          * Allocate file space until done or until there is an error
2274          */
2275         while (allocatesize_fsb && !error) {
2276                 xfs_fileoff_t   s, e;
2277
2278                 /*
2279                  * Determine space reservations for data/realtime.
2280                  */
2281                 if (unlikely(extsz)) {
2282                         s = startoffset_fsb;
2283                         do_div(s, extsz);
2284                         s *= extsz;
2285                         e = startoffset_fsb + allocatesize_fsb;
2286                         if ((temp = do_mod(startoffset_fsb, extsz)))
2287                                 e += temp;
2288                         if ((temp = do_mod(e, extsz)))
2289                                 e += extsz - temp;
2290                 } else {
2291                         s = 0;
2292                         e = allocatesize_fsb;
2293                 }
2294
2295                 /*
2296                  * The transaction reservation is limited to a 32-bit block
2297                  * count, hence we need to limit the number of blocks we are
2298                  * trying to reserve to avoid an overflow. We can't allocate
2299                  * more than @nimaps extents, and an extent is limited on disk
2300                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
2301                  */
2302                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
2303                 if (unlikely(rt)) {
2304                         resrtextents = qblocks = resblks;
2305                         resrtextents /= mp->m_sb.sb_rextsize;
2306                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2307                         quota_flag = XFS_QMOPT_RES_RTBLKS;
2308                 } else {
2309                         resrtextents = 0;
2310                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
2311                         quota_flag = XFS_QMOPT_RES_REGBLKS;
2312                 }
2313
2314                 /*
2315                  * Allocate and setup the transaction.
2316                  */
2317                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2318                 error = xfs_trans_reserve(tp, resblks,
2319                                           XFS_WRITE_LOG_RES(mp), resrtextents,
2320                                           XFS_TRANS_PERM_LOG_RES,
2321                                           XFS_WRITE_LOG_COUNT);
2322                 /*
2323                  * Check for running out of space
2324                  */
2325                 if (error) {
2326                         /*
2327                          * Free the transaction structure.
2328                          */
2329                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2330                         xfs_trans_cancel(tp, 0);
2331                         break;
2332                 }
2333                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2334                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
2335                                                       0, quota_flag);
2336                 if (error)
2337                         goto error1;
2338
2339                 xfs_trans_ijoin(tp, ip);
2340
2341                 /*
2342                  * Issue the xfs_bmapi() call to allocate the blocks
2343                  */
2344                 xfs_bmap_init(&free_list, &firstfsb);
2345                 error = xfs_bmapi(tp, ip, startoffset_fsb,
2346                                   allocatesize_fsb, bmapi_flag,
2347                                   &firstfsb, 0, imapp, &nimaps,
2348                                   &free_list);
2349                 if (error) {
2350                         goto error0;
2351                 }
2352
2353                 /*
2354                  * Complete the transaction
2355                  */
2356                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2357                 if (error) {
2358                         goto error0;
2359                 }
2360
2361                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2362                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2363                 if (error) {
2364                         break;
2365                 }
2366
2367                 allocated_fsb = imapp->br_blockcount;
2368
2369                 if (nimaps == 0) {
2370                         error = XFS_ERROR(ENOSPC);
2371                         break;
2372                 }
2373
2374                 startoffset_fsb += allocated_fsb;
2375                 allocatesize_fsb -= allocated_fsb;
2376         }
2377
2378         return error;
2379
2380 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
2381         xfs_bmap_cancel(&free_list);
2382         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
2383
2384 error1: /* Just cancel transaction */
2385         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2386         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2387         return error;
2388 }
2389
2390 /*
2391  * Zero file bytes between startoff and endoff inclusive.
2392  * The iolock is held exclusive and no blocks are buffered.
2393  *
2394  * This function is used by xfs_free_file_space() to zero
2395  * partial blocks when the range to free is not block aligned.
2396  * When unreserving space with boundaries that are not block
2397  * aligned we round up the start and round down the end
2398  * boundaries and then use this function to zero the parts of
2399  * the blocks that got dropped during the rounding.
2400  */
2401 STATIC int
2402 xfs_zero_remaining_bytes(
2403         xfs_inode_t             *ip,
2404         xfs_off_t               startoff,
2405         xfs_off_t               endoff)
2406 {
2407         xfs_bmbt_irec_t         imap;
2408         xfs_fileoff_t           offset_fsb;
2409         xfs_off_t               lastoffset;
2410         xfs_off_t               offset;
2411         xfs_buf_t               *bp;
2412         xfs_mount_t             *mp = ip->i_mount;
2413         int                     nimap;
2414         int                     error = 0;
2415
2416         /*
2417          * Avoid doing I/O beyond eof - it's not necessary
2418          * since nothing can read beyond eof.  The space will
2419          * be zeroed when the file is extended anyway.
2420          */
2421         if (startoff >= ip->i_size)
2422                 return 0;
2423
2424         if (endoff > ip->i_size)
2425                 endoff = ip->i_size;
2426
2427         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
2428                                         mp->m_rtdev_targp : mp->m_ddev_targp,
2429                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
2430         if (!bp)
2431                 return XFS_ERROR(ENOMEM);
2432
2433         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
2434                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
2435                 nimap = 1;
2436                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
2437                         NULL, 0, &imap, &nimap, NULL);
2438                 if (error || nimap < 1)
2439                         break;
2440                 ASSERT(imap.br_blockcount >= 1);
2441                 ASSERT(imap.br_startoff == offset_fsb);
2442                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
2443                 if (lastoffset > endoff)
2444                         lastoffset = endoff;
2445                 if (imap.br_startblock == HOLESTARTBLOCK)
2446                         continue;
2447                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2448                 if (imap.br_state == XFS_EXT_UNWRITTEN)
2449                         continue;
2450                 XFS_BUF_UNDONE(bp);
2451                 XFS_BUF_UNWRITE(bp);
2452                 XFS_BUF_READ(bp);
2453                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
2454                 xfsbdstrat(mp, bp);
2455                 error = xfs_buf_iowait(bp);
2456                 if (error) {
2457                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
2458                                           mp, bp, XFS_BUF_ADDR(bp));
2459                         break;
2460                 }
2461                 memset(XFS_BUF_PTR(bp) +
2462                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2463                       0, lastoffset - offset + 1);
2464                 XFS_BUF_UNDONE(bp);
2465                 XFS_BUF_UNREAD(bp);
2466                 XFS_BUF_WRITE(bp);
2467                 xfsbdstrat(mp, bp);
2468                 error = xfs_buf_iowait(bp);
2469                 if (error) {
2470                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
2471                                           mp, bp, XFS_BUF_ADDR(bp));
2472                         break;
2473                 }
2474         }
2475         xfs_buf_free(bp);
2476         return error;
2477 }
2478
2479 /*
2480  * xfs_free_file_space()
2481  *      This routine frees disk space for the given file.
2482  *
2483  *      This routine is only called by xfs_change_file_space
2484  *      for an UNRESVSP type call.
2485  *
2486  * RETURNS:
2487  *       0 on success
2488  *      errno on error
2489  *
2490  */
2491 STATIC int
2492 xfs_free_file_space(
2493         xfs_inode_t             *ip,
2494         xfs_off_t               offset,
2495         xfs_off_t               len,
2496         int                     attr_flags)
2497 {
2498         int                     committed;
2499         int                     done;
2500         xfs_fileoff_t           endoffset_fsb;
2501         int                     error;
2502         xfs_fsblock_t           firstfsb;
2503         xfs_bmap_free_t         free_list;
2504         xfs_bmbt_irec_t         imap;
2505         xfs_off_t               ioffset;
2506         xfs_extlen_t            mod=0;
2507         xfs_mount_t             *mp;
2508         int                     nimap;
2509         uint                    resblks;
2510         uint                    rounding;
2511         int                     rt;
2512         xfs_fileoff_t           startoffset_fsb;
2513         xfs_trans_t             *tp;
2514         int                     need_iolock = 1;
2515
2516         mp = ip->i_mount;
2517
2518         trace_xfs_free_file_space(ip);
2519
2520         error = xfs_qm_dqattach(ip, 0);
2521         if (error)
2522                 return error;
2523
2524         error = 0;
2525         if (len <= 0)   /* if nothing being freed */
2526                 return error;
2527         rt = XFS_IS_REALTIME_INODE(ip);
2528         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2529         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2530
2531         if (attr_flags & XFS_ATTR_NOLOCK)
2532                 need_iolock = 0;
2533         if (need_iolock) {
2534                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2535                 /* wait for the completion of any pending DIOs */
2536                 xfs_ioend_wait(ip);
2537         }
2538
2539         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2540         ioffset = offset & ~(rounding - 1);
2541
2542         if (VN_CACHED(VFS_I(ip)) != 0) {
2543                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2544                 if (error)
2545                         goto out_unlock_iolock;
2546         }
2547
2548         /*
2549          * Need to zero the stuff we're not freeing, on disk.
2550          * If it's a realtime file & can't use unwritten extents then we
2551          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2552          * will take care of it for us.
2553          */
2554         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2555                 nimap = 1;
2556                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
2557                         1, 0, NULL, 0, &imap, &nimap, NULL);
2558                 if (error)
2559                         goto out_unlock_iolock;
2560                 ASSERT(nimap == 0 || nimap == 1);
2561                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2562                         xfs_daddr_t     block;
2563
2564                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2565                         block = imap.br_startblock;
2566                         mod = do_div(block, mp->m_sb.sb_rextsize);
2567                         if (mod)
2568                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2569                 }
2570                 nimap = 1;
2571                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
2572                         1, 0, NULL, 0, &imap, &nimap, NULL);
2573                 if (error)
2574                         goto out_unlock_iolock;
2575                 ASSERT(nimap == 0 || nimap == 1);
2576                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2577                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2578                         mod++;
2579                         if (mod && (mod != mp->m_sb.sb_rextsize))
2580                                 endoffset_fsb -= mod;
2581                 }
2582         }
2583         if ((done = (endoffset_fsb <= startoffset_fsb)))
2584                 /*
2585                  * One contiguous piece to clear
2586                  */
2587                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2588         else {
2589                 /*
2590                  * Some full blocks, possibly two pieces to clear
2591                  */
2592                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2593                         error = xfs_zero_remaining_bytes(ip, offset,
2594                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2595                 if (!error &&
2596                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2597                         error = xfs_zero_remaining_bytes(ip,
2598                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2599                                 offset + len - 1);
2600         }
2601
2602         /*
2603          * free file space until done or until there is an error
2604          */
2605         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2606         while (!error && !done) {
2607
2608                 /*
2609                  * allocate and setup the transaction. Allow this
2610                  * transaction to dip into the reserve blocks to ensure
2611                  * the freeing of the space succeeds at ENOSPC.
2612                  */
2613                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2614                 tp->t_flags |= XFS_TRANS_RESERVE;
2615                 error = xfs_trans_reserve(tp,
2616                                           resblks,
2617                                           XFS_WRITE_LOG_RES(mp),
2618                                           0,
2619                                           XFS_TRANS_PERM_LOG_RES,
2620                                           XFS_WRITE_LOG_COUNT);
2621
2622                 /*
2623                  * check for running out of space
2624                  */
2625                 if (error) {
2626                         /*
2627                          * Free the transaction structure.
2628                          */
2629                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2630                         xfs_trans_cancel(tp, 0);
2631                         break;
2632                 }
2633                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2634                 error = xfs_trans_reserve_quota(tp, mp,
2635                                 ip->i_udquot, ip->i_gdquot,
2636                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2637                 if (error)
2638                         goto error1;
2639
2640                 xfs_trans_ijoin(tp, ip);
2641
2642                 /*
2643                  * issue the bunmapi() call to free the blocks
2644                  */
2645                 xfs_bmap_init(&free_list, &firstfsb);
2646                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2647                                   endoffset_fsb - startoffset_fsb,
2648                                   0, 2, &firstfsb, &free_list, &done);
2649                 if (error) {
2650                         goto error0;
2651                 }
2652
2653                 /*
2654                  * complete the transaction
2655                  */
2656                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2657                 if (error) {
2658                         goto error0;
2659                 }
2660
2661                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2662                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2663         }
2664
2665  out_unlock_iolock:
2666         if (need_iolock)
2667                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2668         return error;
2669
2670  error0:
2671         xfs_bmap_cancel(&free_list);
2672  error1:
2673         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2674         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2675                     XFS_ILOCK_EXCL);
2676         return error;
2677 }
2678
2679 /*
2680  * xfs_change_file_space()
2681  *      This routine allocates or frees disk space for the given file.
2682  *      The user specified parameters are checked for alignment and size
2683  *      limitations.
2684  *
2685  * RETURNS:
2686  *       0 on success
2687  *      errno on error
2688  *
2689  */
2690 int
2691 xfs_change_file_space(
2692         xfs_inode_t     *ip,
2693         int             cmd,
2694         xfs_flock64_t   *bf,
2695         xfs_off_t       offset,
2696         int             attr_flags)
2697 {
2698         xfs_mount_t     *mp = ip->i_mount;
2699         int             clrprealloc;
2700         int             error;
2701         xfs_fsize_t     fsize;
2702         int             setprealloc;
2703         xfs_off_t       startoffset;
2704         xfs_off_t       llen;
2705         xfs_trans_t     *tp;
2706         struct iattr    iattr;
2707         int             prealloc_type;
2708
2709         if (!S_ISREG(ip->i_d.di_mode))
2710                 return XFS_ERROR(EINVAL);
2711
2712         switch (bf->l_whence) {
2713         case 0: /*SEEK_SET*/
2714                 break;
2715         case 1: /*SEEK_CUR*/
2716                 bf->l_start += offset;
2717                 break;
2718         case 2: /*SEEK_END*/
2719                 bf->l_start += ip->i_size;
2720                 break;
2721         default:
2722                 return XFS_ERROR(EINVAL);
2723         }
2724
2725         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2726
2727         if (   (bf->l_start < 0)
2728             || (bf->l_start > XFS_MAXIOFFSET(mp))
2729             || (bf->l_start + llen < 0)
2730             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2731                 return XFS_ERROR(EINVAL);
2732
2733         bf->l_whence = 0;
2734
2735         startoffset = bf->l_start;
2736         fsize = ip->i_size;
2737
2738         /*
2739          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2740          * file space.
2741          * These calls do NOT zero the data space allocated to the file,
2742          * nor do they change the file size.
2743          *
2744          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2745          * space.
2746          * These calls cause the new file data to be zeroed and the file
2747          * size to be changed.
2748          */
2749         setprealloc = clrprealloc = 0;
2750         prealloc_type = XFS_BMAPI_PREALLOC;
2751
2752         switch (cmd) {
2753         case XFS_IOC_ZERO_RANGE:
2754                 prealloc_type |= XFS_BMAPI_CONVERT;
2755                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2756                 /* FALLTHRU */
2757         case XFS_IOC_RESVSP:
2758         case XFS_IOC_RESVSP64:
2759                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2760                                                 prealloc_type, attr_flags);
2761                 if (error)
2762                         return error;
2763                 setprealloc = 1;
2764                 break;
2765
2766         case XFS_IOC_UNRESVSP:
2767         case XFS_IOC_UNRESVSP64:
2768                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2769                                                                 attr_flags)))
2770                         return error;
2771                 break;
2772
2773         case XFS_IOC_ALLOCSP:
2774         case XFS_IOC_ALLOCSP64:
2775         case XFS_IOC_FREESP:
2776         case XFS_IOC_FREESP64:
2777                 if (startoffset > fsize) {
2778                         error = xfs_alloc_file_space(ip, fsize,
2779                                         startoffset - fsize, 0, attr_flags);
2780                         if (error)
2781                                 break;
2782                 }
2783
2784                 iattr.ia_valid = ATTR_SIZE;
2785                 iattr.ia_size = startoffset;
2786
2787                 error = xfs_setattr(ip, &iattr, attr_flags);
2788
2789                 if (error)
2790                         return error;
2791
2792                 clrprealloc = 1;
2793                 break;
2794
2795         default:
2796                 ASSERT(0);
2797                 return XFS_ERROR(EINVAL);
2798         }
2799
2800         /*
2801          * update the inode timestamp, mode, and prealloc flag bits
2802          */
2803         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2804
2805         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2806                                       0, 0, 0))) {
2807                 /* ASSERT(0); */
2808                 xfs_trans_cancel(tp, 0);
2809                 return error;
2810         }
2811
2812         xfs_ilock(ip, XFS_ILOCK_EXCL);
2813
2814         xfs_trans_ijoin(tp, ip);
2815
2816         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2817                 ip->i_d.di_mode &= ~S_ISUID;
2818
2819                 /*
2820                  * Note that we don't have to worry about mandatory
2821                  * file locking being disabled here because we only
2822                  * clear the S_ISGID bit if the Group execute bit is
2823                  * on, but if it was on then mandatory locking wouldn't
2824                  * have been enabled.
2825                  */
2826                 if (ip->i_d.di_mode & S_IXGRP)
2827                         ip->i_d.di_mode &= ~S_ISGID;
2828
2829                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2830         }
2831         if (setprealloc)
2832                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2833         else if (clrprealloc)
2834                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2835
2836         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2837         if (attr_flags & XFS_ATTR_SYNC)
2838                 xfs_trans_set_sync(tp);
2839
2840         error = xfs_trans_commit(tp, 0);
2841
2842         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2843
2844         return error;
2845 }