xfs: Convert remaining cmn_err() callers to new API
[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 noticable 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 && VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
964                         xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
965         }
966
967         if (ip->i_d.di_nlink == 0)
968                 return 0;
969
970         if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
971              ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
972                ip->i_delayed_blks > 0)) &&
973              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
974             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
975
976                 /*
977                  * If we can't get the iolock just skip truncating the blocks
978                  * past EOF because we could deadlock with the mmap_sem
979                  * otherwise.  We'll get another chance to drop them once the
980                  * last reference to the inode is dropped, so we'll never leak
981                  * blocks permanently.
982                  *
983                  * Further, check if the inode is being opened, written and
984                  * closed frequently and we have delayed allocation blocks
985                  * oustanding (e.g. streaming writes from the NFS server),
986                  * truncating the blocks past EOF will cause fragmentation to
987                  * occur.
988                  *
989                  * In this case don't do the truncation, either, but we have to
990                  * be careful how we detect this case. Blocks beyond EOF show
991                  * up as i_delayed_blks even when the inode is clean, so we
992                  * need to truncate them away first before checking for a dirty
993                  * release. Hence on the first dirty close we will still remove
994                  * the speculative allocation, but after that we will leave it
995                  * in place.
996                  */
997                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
998                         return 0;
999
1000                 error = xfs_free_eofblocks(mp, ip,
1001                                            XFS_FREE_EOF_TRYLOCK);
1002                 if (error)
1003                         return error;
1004
1005                 /* delalloc blocks after truncation means it really is dirty */
1006                 if (ip->i_delayed_blks)
1007                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1008         }
1009         return 0;
1010 }
1011
1012 /*
1013  * xfs_inactive
1014  *
1015  * This is called when the vnode reference count for the vnode
1016  * goes to zero.  If the file has been unlinked, then it must
1017  * now be truncated.  Also, we clear all of the read-ahead state
1018  * kept for the inode here since the file is now closed.
1019  */
1020 int
1021 xfs_inactive(
1022         xfs_inode_t     *ip)
1023 {
1024         xfs_bmap_free_t free_list;
1025         xfs_fsblock_t   first_block;
1026         int             committed;
1027         xfs_trans_t     *tp;
1028         xfs_mount_t     *mp;
1029         int             error;
1030         int             truncate;
1031
1032         /*
1033          * If the inode is already free, then there can be nothing
1034          * to clean up here.
1035          */
1036         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
1037                 ASSERT(ip->i_df.if_real_bytes == 0);
1038                 ASSERT(ip->i_df.if_broot_bytes == 0);
1039                 return VN_INACTIVE_CACHE;
1040         }
1041
1042         /*
1043          * Only do a truncate if it's a regular file with
1044          * some actual space in it.  It's OK to look at the
1045          * inode's fields without the lock because we're the
1046          * only one with a reference to the inode.
1047          */
1048         truncate = ((ip->i_d.di_nlink == 0) &&
1049             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1050              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1051             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1052
1053         mp = ip->i_mount;
1054
1055         error = 0;
1056
1057         /* If this is a read-only mount, don't do this (would generate I/O) */
1058         if (mp->m_flags & XFS_MOUNT_RDONLY)
1059                 goto out;
1060
1061         if (ip->i_d.di_nlink != 0) {
1062                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1063                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1064                        ip->i_delayed_blks > 0)) &&
1065                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1066                      (!(ip->i_d.di_flags &
1067                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1068                       (ip->i_delayed_blks != 0)))) {
1069                         error = xfs_free_eofblocks(mp, ip, 0);
1070                         if (error)
1071                                 return VN_INACTIVE_CACHE;
1072                 }
1073                 goto out;
1074         }
1075
1076         ASSERT(ip->i_d.di_nlink == 0);
1077
1078         error = xfs_qm_dqattach(ip, 0);
1079         if (error)
1080                 return VN_INACTIVE_CACHE;
1081
1082         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1083         if (truncate) {
1084                 /*
1085                  * Do the xfs_itruncate_start() call before
1086                  * reserving any log space because itruncate_start
1087                  * will call into the buffer cache and we can't
1088                  * do that within a transaction.
1089                  */
1090                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1091
1092                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1093                 if (error) {
1094                         xfs_trans_cancel(tp, 0);
1095                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1096                         return VN_INACTIVE_CACHE;
1097                 }
1098
1099                 error = xfs_trans_reserve(tp, 0,
1100                                           XFS_ITRUNCATE_LOG_RES(mp),
1101                                           0, XFS_TRANS_PERM_LOG_RES,
1102                                           XFS_ITRUNCATE_LOG_COUNT);
1103                 if (error) {
1104                         /* Don't call itruncate_cleanup */
1105                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1106                         xfs_trans_cancel(tp, 0);
1107                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1108                         return VN_INACTIVE_CACHE;
1109                 }
1110
1111                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1112                 xfs_trans_ijoin(tp, ip);
1113
1114                 /*
1115                  * normally, we have to run xfs_itruncate_finish sync.
1116                  * But if filesystem is wsync and we're in the inactive
1117                  * path, then we know that nlink == 0, and that the
1118                  * xaction that made nlink == 0 is permanently committed
1119                  * since xfs_remove runs as a synchronous transaction.
1120                  */
1121                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1122                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1123
1124                 if (error) {
1125                         xfs_trans_cancel(tp,
1126                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1127                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1128                         return VN_INACTIVE_CACHE;
1129                 }
1130         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1131
1132                 /*
1133                  * If we get an error while cleaning up a
1134                  * symlink we bail out.
1135                  */
1136                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1137                         xfs_inactive_symlink_rmt(ip, &tp) :
1138                         xfs_inactive_symlink_local(ip, &tp);
1139
1140                 if (error) {
1141                         ASSERT(tp == NULL);
1142                         return VN_INACTIVE_CACHE;
1143                 }
1144
1145                 xfs_trans_ijoin(tp, ip);
1146         } else {
1147                 error = xfs_trans_reserve(tp, 0,
1148                                           XFS_IFREE_LOG_RES(mp),
1149                                           0, XFS_TRANS_PERM_LOG_RES,
1150                                           XFS_INACTIVE_LOG_COUNT);
1151                 if (error) {
1152                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1153                         xfs_trans_cancel(tp, 0);
1154                         return VN_INACTIVE_CACHE;
1155                 }
1156
1157                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1158                 xfs_trans_ijoin(tp, ip);
1159         }
1160
1161         /*
1162          * If there are attributes associated with the file
1163          * then blow them away now.  The code calls a routine
1164          * that recursively deconstructs the attribute fork.
1165          * We need to just commit the current transaction
1166          * because we can't use it for xfs_attr_inactive().
1167          */
1168         if (ip->i_d.di_anextents > 0) {
1169                 error = xfs_inactive_attrs(ip, &tp);
1170                 /*
1171                  * If we got an error, the transaction is already
1172                  * cancelled, and the inode is unlocked. Just get out.
1173                  */
1174                  if (error)
1175                          return VN_INACTIVE_CACHE;
1176         } else if (ip->i_afp) {
1177                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1178         }
1179
1180         /*
1181          * Free the inode.
1182          */
1183         xfs_bmap_init(&free_list, &first_block);
1184         error = xfs_ifree(tp, ip, &free_list);
1185         if (error) {
1186                 /*
1187                  * If we fail to free the inode, shut down.  The cancel
1188                  * might do that, we need to make sure.  Otherwise the
1189                  * inode might be lost for a long time or forever.
1190                  */
1191                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1192                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
1193                                 __func__, error);
1194                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1195                 }
1196                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1197         } else {
1198                 /*
1199                  * Credit the quota account(s). The inode is gone.
1200                  */
1201                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1202
1203                 /*
1204                  * Just ignore errors at this point.  There is nothing we can
1205                  * do except to try to keep going. Make sure it's not a silent
1206                  * error.
1207                  */
1208                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1209                 if (error)
1210                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
1211                                 __func__, error);
1212                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1213                 if (error)
1214                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1215                                 __func__, error);
1216         }
1217
1218         /*
1219          * Release the dquots held by inode, if any.
1220          */
1221         xfs_qm_dqdetach(ip);
1222         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1223
1224  out:
1225         return VN_INACTIVE_CACHE;
1226 }
1227
1228 /*
1229  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1230  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1231  * ci_name->name will point to a the actual name (caller must free) or
1232  * will be set to NULL if an exact match is found.
1233  */
1234 int
1235 xfs_lookup(
1236         xfs_inode_t             *dp,
1237         struct xfs_name         *name,
1238         xfs_inode_t             **ipp,
1239         struct xfs_name         *ci_name)
1240 {
1241         xfs_ino_t               inum;
1242         int                     error;
1243         uint                    lock_mode;
1244
1245         trace_xfs_lookup(dp, name);
1246
1247         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1248                 return XFS_ERROR(EIO);
1249
1250         lock_mode = xfs_ilock_map_shared(dp);
1251         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1252         xfs_iunlock_map_shared(dp, lock_mode);
1253
1254         if (error)
1255                 goto out;
1256
1257         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
1258         if (error)
1259                 goto out_free_name;
1260
1261         return 0;
1262
1263 out_free_name:
1264         if (ci_name)
1265                 kmem_free(ci_name->name);
1266 out:
1267         *ipp = NULL;
1268         return error;
1269 }
1270
1271 int
1272 xfs_create(
1273         xfs_inode_t             *dp,
1274         struct xfs_name         *name,
1275         mode_t                  mode,
1276         xfs_dev_t               rdev,
1277         xfs_inode_t             **ipp)
1278 {
1279         int                     is_dir = S_ISDIR(mode);
1280         struct xfs_mount        *mp = dp->i_mount;
1281         struct xfs_inode        *ip = NULL;
1282         struct xfs_trans        *tp = NULL;
1283         int                     error;
1284         xfs_bmap_free_t         free_list;
1285         xfs_fsblock_t           first_block;
1286         boolean_t               unlock_dp_on_error = B_FALSE;
1287         uint                    cancel_flags;
1288         int                     committed;
1289         prid_t                  prid;
1290         struct xfs_dquot        *udqp = NULL;
1291         struct xfs_dquot        *gdqp = NULL;
1292         uint                    resblks;
1293         uint                    log_res;
1294         uint                    log_count;
1295
1296         trace_xfs_create(dp, name);
1297
1298         if (XFS_FORCED_SHUTDOWN(mp))
1299                 return XFS_ERROR(EIO);
1300
1301         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1302                 prid = xfs_get_projid(dp);
1303         else
1304                 prid = XFS_PROJID_DEFAULT;
1305
1306         /*
1307          * Make sure that we have allocated dquot(s) on disk.
1308          */
1309         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1310                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1311         if (error)
1312                 return error;
1313
1314         if (is_dir) {
1315                 rdev = 0;
1316                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1317                 log_res = XFS_MKDIR_LOG_RES(mp);
1318                 log_count = XFS_MKDIR_LOG_COUNT;
1319                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
1320         } else {
1321                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1322                 log_res = XFS_CREATE_LOG_RES(mp);
1323                 log_count = XFS_CREATE_LOG_COUNT;
1324                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1325         }
1326
1327         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1328
1329         /*
1330          * Initially assume that the file does not exist and
1331          * reserve the resources for that case.  If that is not
1332          * the case we'll drop the one we have and get a more
1333          * appropriate transaction later.
1334          */
1335         error = xfs_trans_reserve(tp, resblks, log_res, 0,
1336                         XFS_TRANS_PERM_LOG_RES, log_count);
1337         if (error == ENOSPC) {
1338                 /* flush outstanding delalloc blocks and retry */
1339                 xfs_flush_inodes(dp);
1340                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
1341                                 XFS_TRANS_PERM_LOG_RES, log_count);
1342         }
1343         if (error == ENOSPC) {
1344                 /* No space at all so try a "no-allocation" reservation */
1345                 resblks = 0;
1346                 error = xfs_trans_reserve(tp, 0, log_res, 0,
1347                                 XFS_TRANS_PERM_LOG_RES, log_count);
1348         }
1349         if (error) {
1350                 cancel_flags = 0;
1351                 goto out_trans_cancel;
1352         }
1353
1354         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1355         unlock_dp_on_error = B_TRUE;
1356
1357         /*
1358          * Check for directory link count overflow.
1359          */
1360         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
1361                 error = XFS_ERROR(EMLINK);
1362                 goto out_trans_cancel;
1363         }
1364
1365         xfs_bmap_init(&free_list, &first_block);
1366
1367         /*
1368          * Reserve disk quota and the inode.
1369          */
1370         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1371         if (error)
1372                 goto out_trans_cancel;
1373
1374         error = xfs_dir_canenter(tp, dp, name, resblks);
1375         if (error)
1376                 goto out_trans_cancel;
1377
1378         /*
1379          * A newly created regular or special file just has one directory
1380          * entry pointing to them, but a directory also the "." entry
1381          * pointing to itself.
1382          */
1383         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1384                                prid, resblks > 0, &ip, &committed);
1385         if (error) {
1386                 if (error == ENOSPC)
1387                         goto out_trans_cancel;
1388                 goto out_trans_abort;
1389         }
1390
1391         /*
1392          * Now we join the directory inode to the transaction.  We do not do it
1393          * earlier because xfs_dir_ialloc might commit the previous transaction
1394          * (and release all the locks).  An error from here on will result in
1395          * the transaction cancel unlocking dp so don't do it explicitly in the
1396          * error path.
1397          */
1398         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1399         unlock_dp_on_error = B_FALSE;
1400
1401         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1402                                         &first_block, &free_list, resblks ?
1403                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1404         if (error) {
1405                 ASSERT(error != ENOSPC);
1406                 goto out_trans_abort;
1407         }
1408         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1409         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1410
1411         if (is_dir) {
1412                 error = xfs_dir_init(tp, ip, dp);
1413                 if (error)
1414                         goto out_bmap_cancel;
1415
1416                 error = xfs_bumplink(tp, dp);
1417                 if (error)
1418                         goto out_bmap_cancel;
1419         }
1420
1421         /*
1422          * If this is a synchronous mount, make sure that the
1423          * create transaction goes to disk before returning to
1424          * the user.
1425          */
1426         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1427                 xfs_trans_set_sync(tp);
1428
1429         /*
1430          * Attach the dquot(s) to the inodes and modify them incore.
1431          * These ids of the inode couldn't have changed since the new
1432          * inode has been locked ever since it was created.
1433          */
1434         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1435
1436         error = xfs_bmap_finish(&tp, &free_list, &committed);
1437         if (error)
1438                 goto out_bmap_cancel;
1439
1440         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1441         if (error)
1442                 goto out_release_inode;
1443
1444         xfs_qm_dqrele(udqp);
1445         xfs_qm_dqrele(gdqp);
1446
1447         *ipp = ip;
1448         return 0;
1449
1450  out_bmap_cancel:
1451         xfs_bmap_cancel(&free_list);
1452  out_trans_abort:
1453         cancel_flags |= XFS_TRANS_ABORT;
1454  out_trans_cancel:
1455         xfs_trans_cancel(tp, cancel_flags);
1456  out_release_inode:
1457         /*
1458          * Wait until after the current transaction is aborted to
1459          * release the inode.  This prevents recursive transactions
1460          * and deadlocks from xfs_inactive.
1461          */
1462         if (ip)
1463                 IRELE(ip);
1464
1465         xfs_qm_dqrele(udqp);
1466         xfs_qm_dqrele(gdqp);
1467
1468         if (unlock_dp_on_error)
1469                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1470         return error;
1471 }
1472
1473 #ifdef DEBUG
1474 int xfs_locked_n;
1475 int xfs_small_retries;
1476 int xfs_middle_retries;
1477 int xfs_lots_retries;
1478 int xfs_lock_delays;
1479 #endif
1480
1481 /*
1482  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1483  * a different value
1484  */
1485 static inline int
1486 xfs_lock_inumorder(int lock_mode, int subclass)
1487 {
1488         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1489                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1490         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1491                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1492
1493         return lock_mode;
1494 }
1495
1496 /*
1497  * The following routine will lock n inodes in exclusive mode.
1498  * We assume the caller calls us with the inodes in i_ino order.
1499  *
1500  * We need to detect deadlock where an inode that we lock
1501  * is in the AIL and we start waiting for another inode that is locked
1502  * by a thread in a long running transaction (such as truncate). This can
1503  * result in deadlock since the long running trans might need to wait
1504  * for the inode we just locked in order to push the tail and free space
1505  * in the log.
1506  */
1507 void
1508 xfs_lock_inodes(
1509         xfs_inode_t     **ips,
1510         int             inodes,
1511         uint            lock_mode)
1512 {
1513         int             attempts = 0, i, j, try_lock;
1514         xfs_log_item_t  *lp;
1515
1516         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1517
1518         try_lock = 0;
1519         i = 0;
1520
1521 again:
1522         for (; i < inodes; i++) {
1523                 ASSERT(ips[i]);
1524
1525                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1526                         continue;
1527
1528                 /*
1529                  * If try_lock is not set yet, make sure all locked inodes
1530                  * are not in the AIL.
1531                  * If any are, set try_lock to be used later.
1532                  */
1533
1534                 if (!try_lock) {
1535                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1536                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1537                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1538                                         try_lock++;
1539                                 }
1540                         }
1541                 }
1542
1543                 /*
1544                  * If any of the previous locks we have locked is in the AIL,
1545                  * we must TRY to get the second and subsequent locks. If
1546                  * we can't get any, we must release all we have
1547                  * and try again.
1548                  */
1549
1550                 if (try_lock) {
1551                         /* try_lock must be 0 if i is 0. */
1552                         /*
1553                          * try_lock means we have an inode locked
1554                          * that is in the AIL.
1555                          */
1556                         ASSERT(i != 0);
1557                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1558                                 attempts++;
1559
1560                                 /*
1561                                  * Unlock all previous guys and try again.
1562                                  * xfs_iunlock will try to push the tail
1563                                  * if the inode is in the AIL.
1564                                  */
1565
1566                                 for(j = i - 1; j >= 0; j--) {
1567
1568                                         /*
1569                                          * Check to see if we've already
1570                                          * unlocked this one.
1571                                          * Not the first one going back,
1572                                          * and the inode ptr is the same.
1573                                          */
1574                                         if ((j != (i - 1)) && ips[j] ==
1575                                                                 ips[j+1])
1576                                                 continue;
1577
1578                                         xfs_iunlock(ips[j], lock_mode);
1579                                 }
1580
1581                                 if ((attempts % 5) == 0) {
1582                                         delay(1); /* Don't just spin the CPU */
1583 #ifdef DEBUG
1584                                         xfs_lock_delays++;
1585 #endif
1586                                 }
1587                                 i = 0;
1588                                 try_lock = 0;
1589                                 goto again;
1590                         }
1591                 } else {
1592                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1593                 }
1594         }
1595
1596 #ifdef DEBUG
1597         if (attempts) {
1598                 if (attempts < 5) xfs_small_retries++;
1599                 else if (attempts < 100) xfs_middle_retries++;
1600                 else xfs_lots_retries++;
1601         } else {
1602                 xfs_locked_n++;
1603         }
1604 #endif
1605 }
1606
1607 /*
1608  * xfs_lock_two_inodes() can only be used to lock one type of lock
1609  * at a time - the iolock or the ilock, but not both at once. If
1610  * we lock both at once, lockdep will report false positives saying
1611  * we have violated locking orders.
1612  */
1613 void
1614 xfs_lock_two_inodes(
1615         xfs_inode_t             *ip0,
1616         xfs_inode_t             *ip1,
1617         uint                    lock_mode)
1618 {
1619         xfs_inode_t             *temp;
1620         int                     attempts = 0;
1621         xfs_log_item_t          *lp;
1622
1623         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1624                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1625         ASSERT(ip0->i_ino != ip1->i_ino);
1626
1627         if (ip0->i_ino > ip1->i_ino) {
1628                 temp = ip0;
1629                 ip0 = ip1;
1630                 ip1 = temp;
1631         }
1632
1633  again:
1634         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1635
1636         /*
1637          * If the first lock we have locked is in the AIL, we must TRY to get
1638          * the second lock. If we can't get it, we must release the first one
1639          * and try again.
1640          */
1641         lp = (xfs_log_item_t *)ip0->i_itemp;
1642         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1643                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1644                         xfs_iunlock(ip0, lock_mode);
1645                         if ((++attempts % 5) == 0)
1646                                 delay(1); /* Don't just spin the CPU */
1647                         goto again;
1648                 }
1649         } else {
1650                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1651         }
1652 }
1653
1654 int
1655 xfs_remove(
1656         xfs_inode_t             *dp,
1657         struct xfs_name         *name,
1658         xfs_inode_t             *ip)
1659 {
1660         xfs_mount_t             *mp = dp->i_mount;
1661         xfs_trans_t             *tp = NULL;
1662         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1663         int                     error = 0;
1664         xfs_bmap_free_t         free_list;
1665         xfs_fsblock_t           first_block;
1666         int                     cancel_flags;
1667         int                     committed;
1668         int                     link_zero;
1669         uint                    resblks;
1670         uint                    log_count;
1671
1672         trace_xfs_remove(dp, name);
1673
1674         if (XFS_FORCED_SHUTDOWN(mp))
1675                 return XFS_ERROR(EIO);
1676
1677         error = xfs_qm_dqattach(dp, 0);
1678         if (error)
1679                 goto std_return;
1680
1681         error = xfs_qm_dqattach(ip, 0);
1682         if (error)
1683                 goto std_return;
1684
1685         if (is_dir) {
1686                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1687                 log_count = XFS_DEFAULT_LOG_COUNT;
1688         } else {
1689                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1690                 log_count = XFS_REMOVE_LOG_COUNT;
1691         }
1692         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1693
1694         /*
1695          * We try to get the real space reservation first,
1696          * allowing for directory btree deletion(s) implying
1697          * possible bmap insert(s).  If we can't get the space
1698          * reservation then we use 0 instead, and avoid the bmap
1699          * btree insert(s) in the directory code by, if the bmap
1700          * insert tries to happen, instead trimming the LAST
1701          * block from the directory.
1702          */
1703         resblks = XFS_REMOVE_SPACE_RES(mp);
1704         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1705                                   XFS_TRANS_PERM_LOG_RES, log_count);
1706         if (error == ENOSPC) {
1707                 resblks = 0;
1708                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1709                                           XFS_TRANS_PERM_LOG_RES, log_count);
1710         }
1711         if (error) {
1712                 ASSERT(error != ENOSPC);
1713                 cancel_flags = 0;
1714                 goto out_trans_cancel;
1715         }
1716
1717         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1718
1719         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1720         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1721
1722         /*
1723          * If we're removing a directory perform some additional validation.
1724          */
1725         if (is_dir) {
1726                 ASSERT(ip->i_d.di_nlink >= 2);
1727                 if (ip->i_d.di_nlink != 2) {
1728                         error = XFS_ERROR(ENOTEMPTY);
1729                         goto out_trans_cancel;
1730                 }
1731                 if (!xfs_dir_isempty(ip)) {
1732                         error = XFS_ERROR(ENOTEMPTY);
1733                         goto out_trans_cancel;
1734                 }
1735         }
1736
1737         xfs_bmap_init(&free_list, &first_block);
1738         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1739                                         &first_block, &free_list, resblks);
1740         if (error) {
1741                 ASSERT(error != ENOENT);
1742                 goto out_bmap_cancel;
1743         }
1744         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1745
1746         if (is_dir) {
1747                 /*
1748                  * Drop the link from ip's "..".
1749                  */
1750                 error = xfs_droplink(tp, dp);
1751                 if (error)
1752                         goto out_bmap_cancel;
1753
1754                 /*
1755                  * Drop the "." link from ip to self.
1756                  */
1757                 error = xfs_droplink(tp, ip);
1758                 if (error)
1759                         goto out_bmap_cancel;
1760         } else {
1761                 /*
1762                  * When removing a non-directory we need to log the parent
1763                  * inode here.  For a directory this is done implicitly
1764                  * by the xfs_droplink call for the ".." entry.
1765                  */
1766                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1767         }
1768
1769         /*
1770          * Drop the link from dp to ip.
1771          */
1772         error = xfs_droplink(tp, ip);
1773         if (error)
1774                 goto out_bmap_cancel;
1775
1776         /*
1777          * Determine if this is the last link while
1778          * we are in the transaction.
1779          */
1780         link_zero = (ip->i_d.di_nlink == 0);
1781
1782         /*
1783          * If this is a synchronous mount, make sure that the
1784          * remove transaction goes to disk before returning to
1785          * the user.
1786          */
1787         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1788                 xfs_trans_set_sync(tp);
1789
1790         error = xfs_bmap_finish(&tp, &free_list, &committed);
1791         if (error)
1792                 goto out_bmap_cancel;
1793
1794         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1795         if (error)
1796                 goto std_return;
1797
1798         /*
1799          * If we are using filestreams, kill the stream association.
1800          * If the file is still open it may get a new one but that
1801          * will get killed on last close in xfs_close() so we don't
1802          * have to worry about that.
1803          */
1804         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1805                 xfs_filestream_deassociate(ip);
1806
1807         return 0;
1808
1809  out_bmap_cancel:
1810         xfs_bmap_cancel(&free_list);
1811         cancel_flags |= XFS_TRANS_ABORT;
1812  out_trans_cancel:
1813         xfs_trans_cancel(tp, cancel_flags);
1814  std_return:
1815         return error;
1816 }
1817
1818 int
1819 xfs_link(
1820         xfs_inode_t             *tdp,
1821         xfs_inode_t             *sip,
1822         struct xfs_name         *target_name)
1823 {
1824         xfs_mount_t             *mp = tdp->i_mount;
1825         xfs_trans_t             *tp;
1826         int                     error;
1827         xfs_bmap_free_t         free_list;
1828         xfs_fsblock_t           first_block;
1829         int                     cancel_flags;
1830         int                     committed;
1831         int                     resblks;
1832
1833         trace_xfs_link(tdp, target_name);
1834
1835         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1836
1837         if (XFS_FORCED_SHUTDOWN(mp))
1838                 return XFS_ERROR(EIO);
1839
1840         error = xfs_qm_dqattach(sip, 0);
1841         if (error)
1842                 goto std_return;
1843
1844         error = xfs_qm_dqattach(tdp, 0);
1845         if (error)
1846                 goto std_return;
1847
1848         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1849         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1850         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1851         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1852                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1853         if (error == ENOSPC) {
1854                 resblks = 0;
1855                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1856                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1857         }
1858         if (error) {
1859                 cancel_flags = 0;
1860                 goto error_return;
1861         }
1862
1863         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1864
1865         xfs_trans_ijoin_ref(tp, sip, XFS_ILOCK_EXCL);
1866         xfs_trans_ijoin_ref(tp, tdp, XFS_ILOCK_EXCL);
1867
1868         /*
1869          * If the source has too many links, we can't make any more to it.
1870          */
1871         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1872                 error = XFS_ERROR(EMLINK);
1873                 goto error_return;
1874         }
1875
1876         /*
1877          * If we are using project inheritance, we only allow hard link
1878          * creation in our tree when the project IDs are the same; else
1879          * the tree quota mechanism could be circumvented.
1880          */
1881         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1882                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1883                 error = XFS_ERROR(EXDEV);
1884                 goto error_return;
1885         }
1886
1887         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1888         if (error)
1889                 goto error_return;
1890
1891         xfs_bmap_init(&free_list, &first_block);
1892
1893         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1894                                         &first_block, &free_list, resblks);
1895         if (error)
1896                 goto abort_return;
1897         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1898         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1899
1900         error = xfs_bumplink(tp, sip);
1901         if (error)
1902                 goto abort_return;
1903
1904         /*
1905          * If this is a synchronous mount, make sure that the
1906          * link transaction goes to disk before returning to
1907          * the user.
1908          */
1909         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1910                 xfs_trans_set_sync(tp);
1911         }
1912
1913         error = xfs_bmap_finish (&tp, &free_list, &committed);
1914         if (error) {
1915                 xfs_bmap_cancel(&free_list);
1916                 goto abort_return;
1917         }
1918
1919         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1920
1921  abort_return:
1922         cancel_flags |= XFS_TRANS_ABORT;
1923  error_return:
1924         xfs_trans_cancel(tp, cancel_flags);
1925  std_return:
1926         return error;
1927 }
1928
1929 int
1930 xfs_symlink(
1931         xfs_inode_t             *dp,
1932         struct xfs_name         *link_name,
1933         const char              *target_path,
1934         mode_t                  mode,
1935         xfs_inode_t             **ipp)
1936 {
1937         xfs_mount_t             *mp = dp->i_mount;
1938         xfs_trans_t             *tp;
1939         xfs_inode_t             *ip;
1940         int                     error;
1941         int                     pathlen;
1942         xfs_bmap_free_t         free_list;
1943         xfs_fsblock_t           first_block;
1944         boolean_t               unlock_dp_on_error = B_FALSE;
1945         uint                    cancel_flags;
1946         int                     committed;
1947         xfs_fileoff_t           first_fsb;
1948         xfs_filblks_t           fs_blocks;
1949         int                     nmaps;
1950         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1951         xfs_daddr_t             d;
1952         const char              *cur_chunk;
1953         int                     byte_cnt;
1954         int                     n;
1955         xfs_buf_t               *bp;
1956         prid_t                  prid;
1957         struct xfs_dquot        *udqp, *gdqp;
1958         uint                    resblks;
1959
1960         *ipp = NULL;
1961         error = 0;
1962         ip = NULL;
1963         tp = NULL;
1964
1965         trace_xfs_symlink(dp, link_name);
1966
1967         if (XFS_FORCED_SHUTDOWN(mp))
1968                 return XFS_ERROR(EIO);
1969
1970         /*
1971          * Check component lengths of the target path name.
1972          */
1973         pathlen = strlen(target_path);
1974         if (pathlen >= MAXPATHLEN)      /* total string too long */
1975                 return XFS_ERROR(ENAMETOOLONG);
1976
1977         udqp = gdqp = NULL;
1978         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1979                 prid = xfs_get_projid(dp);
1980         else
1981                 prid = XFS_PROJID_DEFAULT;
1982
1983         /*
1984          * Make sure that we have allocated dquot(s) on disk.
1985          */
1986         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1987                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1988         if (error)
1989                 goto std_return;
1990
1991         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1992         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1993         /*
1994          * The symlink will fit into the inode data fork?
1995          * There can't be any attributes so we get the whole variable part.
1996          */
1997         if (pathlen <= XFS_LITINO(mp))
1998                 fs_blocks = 0;
1999         else
2000                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2001         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2002         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2003                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2004         if (error == ENOSPC && fs_blocks == 0) {
2005                 resblks = 0;
2006                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2007                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2008         }
2009         if (error) {
2010                 cancel_flags = 0;
2011                 goto error_return;
2012         }
2013
2014         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2015         unlock_dp_on_error = B_TRUE;
2016
2017         /*
2018          * Check whether the directory allows new symlinks or not.
2019          */
2020         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2021                 error = XFS_ERROR(EPERM);
2022                 goto error_return;
2023         }
2024
2025         /*
2026          * Reserve disk quota : blocks and inode.
2027          */
2028         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
2029         if (error)
2030                 goto error_return;
2031
2032         /*
2033          * Check for ability to enter directory entry, if no space reserved.
2034          */
2035         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2036         if (error)
2037                 goto error_return;
2038         /*
2039          * Initialize the bmap freelist prior to calling either
2040          * bmapi or the directory create code.
2041          */
2042         xfs_bmap_init(&free_list, &first_block);
2043
2044         /*
2045          * Allocate an inode for the symlink.
2046          */
2047         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
2048                                prid, resblks > 0, &ip, NULL);
2049         if (error) {
2050                 if (error == ENOSPC)
2051                         goto error_return;
2052                 goto error1;
2053         }
2054
2055         /*
2056          * An error after we've joined dp to the transaction will result in the
2057          * transaction cancel unlocking dp so don't do it explicitly in the
2058          * error path.
2059          */
2060         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
2061         unlock_dp_on_error = B_FALSE;
2062
2063         /*
2064          * Also attach the dquot(s) to it, if applicable.
2065          */
2066         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
2067
2068         if (resblks)
2069                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2070         /*
2071          * If the symlink will fit into the inode, write it inline.
2072          */
2073         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2074                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2075                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2076                 ip->i_d.di_size = pathlen;
2077
2078                 /*
2079                  * The inode was initially created in extent format.
2080                  */
2081                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2082                 ip->i_df.if_flags |= XFS_IFINLINE;
2083
2084                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2085                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2086
2087         } else {
2088                 first_fsb = 0;
2089                 nmaps = SYMLINK_MAPS;
2090
2091                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2092                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2093                                   &first_block, resblks, mval, &nmaps,
2094                                   &free_list);
2095                 if (error)
2096                         goto error2;
2097
2098                 if (resblks)
2099                         resblks -= fs_blocks;
2100                 ip->i_d.di_size = pathlen;
2101                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2102
2103                 cur_chunk = target_path;
2104                 for (n = 0; n < nmaps; n++) {
2105                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2106                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2107                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2108                                                BTOBB(byte_cnt), 0);
2109                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2110                         if (pathlen < byte_cnt) {
2111                                 byte_cnt = pathlen;
2112                         }
2113                         pathlen -= byte_cnt;
2114
2115                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2116                         cur_chunk += byte_cnt;
2117
2118                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2119                 }
2120         }
2121
2122         /*
2123          * Create the directory entry for the symlink.
2124          */
2125         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2126                                         &first_block, &free_list, resblks);
2127         if (error)
2128                 goto error2;
2129         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2130         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2131
2132         /*
2133          * If this is a synchronous mount, make sure that the
2134          * symlink transaction goes to disk before returning to
2135          * the user.
2136          */
2137         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2138                 xfs_trans_set_sync(tp);
2139         }
2140
2141         error = xfs_bmap_finish(&tp, &free_list, &committed);
2142         if (error) {
2143                 goto error2;
2144         }
2145         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2146         xfs_qm_dqrele(udqp);
2147         xfs_qm_dqrele(gdqp);
2148
2149         *ipp = ip;
2150         return 0;
2151
2152  error2:
2153         IRELE(ip);
2154  error1:
2155         xfs_bmap_cancel(&free_list);
2156         cancel_flags |= XFS_TRANS_ABORT;
2157  error_return:
2158         xfs_trans_cancel(tp, cancel_flags);
2159         xfs_qm_dqrele(udqp);
2160         xfs_qm_dqrele(gdqp);
2161
2162         if (unlock_dp_on_error)
2163                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2164  std_return:
2165         return error;
2166 }
2167
2168 int
2169 xfs_set_dmattrs(
2170         xfs_inode_t     *ip,
2171         u_int           evmask,
2172         u_int16_t       state)
2173 {
2174         xfs_mount_t     *mp = ip->i_mount;
2175         xfs_trans_t     *tp;
2176         int             error;
2177
2178         if (!capable(CAP_SYS_ADMIN))
2179                 return XFS_ERROR(EPERM);
2180
2181         if (XFS_FORCED_SHUTDOWN(mp))
2182                 return XFS_ERROR(EIO);
2183
2184         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2185         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2186         if (error) {
2187                 xfs_trans_cancel(tp, 0);
2188                 return error;
2189         }
2190         xfs_ilock(ip, XFS_ILOCK_EXCL);
2191         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
2192
2193         ip->i_d.di_dmevmask = evmask;
2194         ip->i_d.di_dmstate  = state;
2195
2196         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2197         error = xfs_trans_commit(tp, 0);
2198
2199         return error;
2200 }
2201
2202 /*
2203  * xfs_alloc_file_space()
2204  *      This routine allocates disk space for the given file.
2205  *
2206  *      If alloc_type == 0, this request is for an ALLOCSP type
2207  *      request which will change the file size.  In this case, no
2208  *      DMAPI event will be generated by the call.  A TRUNCATE event
2209  *      will be generated later by xfs_setattr.
2210  *
2211  *      If alloc_type != 0, this request is for a RESVSP type
2212  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2213  *      lower block boundary byte address is less than the file's
2214  *      length.
2215  *
2216  * RETURNS:
2217  *       0 on success
2218  *      errno on error
2219  *
2220  */
2221 STATIC int
2222 xfs_alloc_file_space(
2223         xfs_inode_t             *ip,
2224         xfs_off_t               offset,
2225         xfs_off_t               len,
2226         int                     alloc_type,
2227         int                     attr_flags)
2228 {
2229         xfs_mount_t             *mp = ip->i_mount;
2230         xfs_off_t               count;
2231         xfs_filblks_t           allocated_fsb;
2232         xfs_filblks_t           allocatesize_fsb;
2233         xfs_extlen_t            extsz, temp;
2234         xfs_fileoff_t           startoffset_fsb;
2235         xfs_fsblock_t           firstfsb;
2236         int                     nimaps;
2237         int                     bmapi_flag;
2238         int                     quota_flag;
2239         int                     rt;
2240         xfs_trans_t             *tp;
2241         xfs_bmbt_irec_t         imaps[1], *imapp;
2242         xfs_bmap_free_t         free_list;
2243         uint                    qblocks, resblks, resrtextents;
2244         int                     committed;
2245         int                     error;
2246
2247         trace_xfs_alloc_file_space(ip);
2248
2249         if (XFS_FORCED_SHUTDOWN(mp))
2250                 return XFS_ERROR(EIO);
2251
2252         error = xfs_qm_dqattach(ip, 0);
2253         if (error)
2254                 return error;
2255
2256         if (len <= 0)
2257                 return XFS_ERROR(EINVAL);
2258
2259         rt = XFS_IS_REALTIME_INODE(ip);
2260         extsz = xfs_get_extsz_hint(ip);
2261
2262         count = len;
2263         imapp = &imaps[0];
2264         nimaps = 1;
2265         bmapi_flag = XFS_BMAPI_WRITE | alloc_type;
2266         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
2267         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
2268
2269         /*
2270          * Allocate file space until done or until there is an error
2271          */
2272         while (allocatesize_fsb && !error) {
2273                 xfs_fileoff_t   s, e;
2274
2275                 /*
2276                  * Determine space reservations for data/realtime.
2277                  */
2278                 if (unlikely(extsz)) {
2279                         s = startoffset_fsb;
2280                         do_div(s, extsz);
2281                         s *= extsz;
2282                         e = startoffset_fsb + allocatesize_fsb;
2283                         if ((temp = do_mod(startoffset_fsb, extsz)))
2284                                 e += temp;
2285                         if ((temp = do_mod(e, extsz)))
2286                                 e += extsz - temp;
2287                 } else {
2288                         s = 0;
2289                         e = allocatesize_fsb;
2290                 }
2291
2292                 /*
2293                  * The transaction reservation is limited to a 32-bit block
2294                  * count, hence we need to limit the number of blocks we are
2295                  * trying to reserve to avoid an overflow. We can't allocate
2296                  * more than @nimaps extents, and an extent is limited on disk
2297                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
2298                  */
2299                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
2300                 if (unlikely(rt)) {
2301                         resrtextents = qblocks = resblks;
2302                         resrtextents /= mp->m_sb.sb_rextsize;
2303                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2304                         quota_flag = XFS_QMOPT_RES_RTBLKS;
2305                 } else {
2306                         resrtextents = 0;
2307                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
2308                         quota_flag = XFS_QMOPT_RES_REGBLKS;
2309                 }
2310
2311                 /*
2312                  * Allocate and setup the transaction.
2313                  */
2314                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2315                 error = xfs_trans_reserve(tp, resblks,
2316                                           XFS_WRITE_LOG_RES(mp), resrtextents,
2317                                           XFS_TRANS_PERM_LOG_RES,
2318                                           XFS_WRITE_LOG_COUNT);
2319                 /*
2320                  * Check for running out of space
2321                  */
2322                 if (error) {
2323                         /*
2324                          * Free the transaction structure.
2325                          */
2326                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2327                         xfs_trans_cancel(tp, 0);
2328                         break;
2329                 }
2330                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2331                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
2332                                                       0, quota_flag);
2333                 if (error)
2334                         goto error1;
2335
2336                 xfs_trans_ijoin(tp, ip);
2337
2338                 /*
2339                  * Issue the xfs_bmapi() call to allocate the blocks
2340                  */
2341                 xfs_bmap_init(&free_list, &firstfsb);
2342                 error = xfs_bmapi(tp, ip, startoffset_fsb,
2343                                   allocatesize_fsb, bmapi_flag,
2344                                   &firstfsb, 0, imapp, &nimaps,
2345                                   &free_list);
2346                 if (error) {
2347                         goto error0;
2348                 }
2349
2350                 /*
2351                  * Complete the transaction
2352                  */
2353                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2354                 if (error) {
2355                         goto error0;
2356                 }
2357
2358                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2359                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2360                 if (error) {
2361                         break;
2362                 }
2363
2364                 allocated_fsb = imapp->br_blockcount;
2365
2366                 if (nimaps == 0) {
2367                         error = XFS_ERROR(ENOSPC);
2368                         break;
2369                 }
2370
2371                 startoffset_fsb += allocated_fsb;
2372                 allocatesize_fsb -= allocated_fsb;
2373         }
2374
2375         return error;
2376
2377 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
2378         xfs_bmap_cancel(&free_list);
2379         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
2380
2381 error1: /* Just cancel transaction */
2382         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2383         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2384         return error;
2385 }
2386
2387 /*
2388  * Zero file bytes between startoff and endoff inclusive.
2389  * The iolock is held exclusive and no blocks are buffered.
2390  *
2391  * This function is used by xfs_free_file_space() to zero
2392  * partial blocks when the range to free is not block aligned.
2393  * When unreserving space with boundaries that are not block
2394  * aligned we round up the start and round down the end
2395  * boundaries and then use this function to zero the parts of
2396  * the blocks that got dropped during the rounding.
2397  */
2398 STATIC int
2399 xfs_zero_remaining_bytes(
2400         xfs_inode_t             *ip,
2401         xfs_off_t               startoff,
2402         xfs_off_t               endoff)
2403 {
2404         xfs_bmbt_irec_t         imap;
2405         xfs_fileoff_t           offset_fsb;
2406         xfs_off_t               lastoffset;
2407         xfs_off_t               offset;
2408         xfs_buf_t               *bp;
2409         xfs_mount_t             *mp = ip->i_mount;
2410         int                     nimap;
2411         int                     error = 0;
2412
2413         /*
2414          * Avoid doing I/O beyond eof - it's not necessary
2415          * since nothing can read beyond eof.  The space will
2416          * be zeroed when the file is extended anyway.
2417          */
2418         if (startoff >= ip->i_size)
2419                 return 0;
2420
2421         if (endoff > ip->i_size)
2422                 endoff = ip->i_size;
2423
2424         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
2425                                         mp->m_rtdev_targp : mp->m_ddev_targp,
2426                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
2427         if (!bp)
2428                 return XFS_ERROR(ENOMEM);
2429
2430         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
2431                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
2432                 nimap = 1;
2433                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
2434                         NULL, 0, &imap, &nimap, NULL);
2435                 if (error || nimap < 1)
2436                         break;
2437                 ASSERT(imap.br_blockcount >= 1);
2438                 ASSERT(imap.br_startoff == offset_fsb);
2439                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
2440                 if (lastoffset > endoff)
2441                         lastoffset = endoff;
2442                 if (imap.br_startblock == HOLESTARTBLOCK)
2443                         continue;
2444                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2445                 if (imap.br_state == XFS_EXT_UNWRITTEN)
2446                         continue;
2447                 XFS_BUF_UNDONE(bp);
2448                 XFS_BUF_UNWRITE(bp);
2449                 XFS_BUF_READ(bp);
2450                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
2451                 xfsbdstrat(mp, bp);
2452                 error = xfs_buf_iowait(bp);
2453                 if (error) {
2454                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
2455                                           mp, bp, XFS_BUF_ADDR(bp));
2456                         break;
2457                 }
2458                 memset(XFS_BUF_PTR(bp) +
2459                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2460                       0, lastoffset - offset + 1);
2461                 XFS_BUF_UNDONE(bp);
2462                 XFS_BUF_UNREAD(bp);
2463                 XFS_BUF_WRITE(bp);
2464                 xfsbdstrat(mp, bp);
2465                 error = xfs_buf_iowait(bp);
2466                 if (error) {
2467                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
2468                                           mp, bp, XFS_BUF_ADDR(bp));
2469                         break;
2470                 }
2471         }
2472         xfs_buf_free(bp);
2473         return error;
2474 }
2475
2476 /*
2477  * xfs_free_file_space()
2478  *      This routine frees disk space for the given file.
2479  *
2480  *      This routine is only called by xfs_change_file_space
2481  *      for an UNRESVSP type call.
2482  *
2483  * RETURNS:
2484  *       0 on success
2485  *      errno on error
2486  *
2487  */
2488 STATIC int
2489 xfs_free_file_space(
2490         xfs_inode_t             *ip,
2491         xfs_off_t               offset,
2492         xfs_off_t               len,
2493         int                     attr_flags)
2494 {
2495         int                     committed;
2496         int                     done;
2497         xfs_fileoff_t           endoffset_fsb;
2498         int                     error;
2499         xfs_fsblock_t           firstfsb;
2500         xfs_bmap_free_t         free_list;
2501         xfs_bmbt_irec_t         imap;
2502         xfs_off_t               ioffset;
2503         xfs_extlen_t            mod=0;
2504         xfs_mount_t             *mp;
2505         int                     nimap;
2506         uint                    resblks;
2507         uint                    rounding;
2508         int                     rt;
2509         xfs_fileoff_t           startoffset_fsb;
2510         xfs_trans_t             *tp;
2511         int                     need_iolock = 1;
2512
2513         mp = ip->i_mount;
2514
2515         trace_xfs_free_file_space(ip);
2516
2517         error = xfs_qm_dqattach(ip, 0);
2518         if (error)
2519                 return error;
2520
2521         error = 0;
2522         if (len <= 0)   /* if nothing being freed */
2523                 return error;
2524         rt = XFS_IS_REALTIME_INODE(ip);
2525         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2526         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2527
2528         if (attr_flags & XFS_ATTR_NOLOCK)
2529                 need_iolock = 0;
2530         if (need_iolock) {
2531                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2532                 /* wait for the completion of any pending DIOs */
2533                 xfs_ioend_wait(ip);
2534         }
2535
2536         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2537         ioffset = offset & ~(rounding - 1);
2538
2539         if (VN_CACHED(VFS_I(ip)) != 0) {
2540                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2541                 if (error)
2542                         goto out_unlock_iolock;
2543         }
2544
2545         /*
2546          * Need to zero the stuff we're not freeing, on disk.
2547          * If it's a realtime file & can't use unwritten extents then we
2548          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2549          * will take care of it for us.
2550          */
2551         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2552                 nimap = 1;
2553                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
2554                         1, 0, NULL, 0, &imap, &nimap, NULL);
2555                 if (error)
2556                         goto out_unlock_iolock;
2557                 ASSERT(nimap == 0 || nimap == 1);
2558                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2559                         xfs_daddr_t     block;
2560
2561                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2562                         block = imap.br_startblock;
2563                         mod = do_div(block, mp->m_sb.sb_rextsize);
2564                         if (mod)
2565                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2566                 }
2567                 nimap = 1;
2568                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
2569                         1, 0, NULL, 0, &imap, &nimap, NULL);
2570                 if (error)
2571                         goto out_unlock_iolock;
2572                 ASSERT(nimap == 0 || nimap == 1);
2573                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2574                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2575                         mod++;
2576                         if (mod && (mod != mp->m_sb.sb_rextsize))
2577                                 endoffset_fsb -= mod;
2578                 }
2579         }
2580         if ((done = (endoffset_fsb <= startoffset_fsb)))
2581                 /*
2582                  * One contiguous piece to clear
2583                  */
2584                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2585         else {
2586                 /*
2587                  * Some full blocks, possibly two pieces to clear
2588                  */
2589                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2590                         error = xfs_zero_remaining_bytes(ip, offset,
2591                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2592                 if (!error &&
2593                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2594                         error = xfs_zero_remaining_bytes(ip,
2595                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2596                                 offset + len - 1);
2597         }
2598
2599         /*
2600          * free file space until done or until there is an error
2601          */
2602         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2603         while (!error && !done) {
2604
2605                 /*
2606                  * allocate and setup the transaction. Allow this
2607                  * transaction to dip into the reserve blocks to ensure
2608                  * the freeing of the space succeeds at ENOSPC.
2609                  */
2610                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2611                 tp->t_flags |= XFS_TRANS_RESERVE;
2612                 error = xfs_trans_reserve(tp,
2613                                           resblks,
2614                                           XFS_WRITE_LOG_RES(mp),
2615                                           0,
2616                                           XFS_TRANS_PERM_LOG_RES,
2617                                           XFS_WRITE_LOG_COUNT);
2618
2619                 /*
2620                  * check for running out of space
2621                  */
2622                 if (error) {
2623                         /*
2624                          * Free the transaction structure.
2625                          */
2626                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2627                         xfs_trans_cancel(tp, 0);
2628                         break;
2629                 }
2630                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2631                 error = xfs_trans_reserve_quota(tp, mp,
2632                                 ip->i_udquot, ip->i_gdquot,
2633                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2634                 if (error)
2635                         goto error1;
2636
2637                 xfs_trans_ijoin(tp, ip);
2638
2639                 /*
2640                  * issue the bunmapi() call to free the blocks
2641                  */
2642                 xfs_bmap_init(&free_list, &firstfsb);
2643                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2644                                   endoffset_fsb - startoffset_fsb,
2645                                   0, 2, &firstfsb, &free_list, &done);
2646                 if (error) {
2647                         goto error0;
2648                 }
2649
2650                 /*
2651                  * complete the transaction
2652                  */
2653                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2654                 if (error) {
2655                         goto error0;
2656                 }
2657
2658                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2659                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2660         }
2661
2662  out_unlock_iolock:
2663         if (need_iolock)
2664                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2665         return error;
2666
2667  error0:
2668         xfs_bmap_cancel(&free_list);
2669  error1:
2670         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2671         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2672                     XFS_ILOCK_EXCL);
2673         return error;
2674 }
2675
2676 /*
2677  * xfs_change_file_space()
2678  *      This routine allocates or frees disk space for the given file.
2679  *      The user specified parameters are checked for alignment and size
2680  *      limitations.
2681  *
2682  * RETURNS:
2683  *       0 on success
2684  *      errno on error
2685  *
2686  */
2687 int
2688 xfs_change_file_space(
2689         xfs_inode_t     *ip,
2690         int             cmd,
2691         xfs_flock64_t   *bf,
2692         xfs_off_t       offset,
2693         int             attr_flags)
2694 {
2695         xfs_mount_t     *mp = ip->i_mount;
2696         int             clrprealloc;
2697         int             error;
2698         xfs_fsize_t     fsize;
2699         int             setprealloc;
2700         xfs_off_t       startoffset;
2701         xfs_off_t       llen;
2702         xfs_trans_t     *tp;
2703         struct iattr    iattr;
2704         int             prealloc_type;
2705
2706         if (!S_ISREG(ip->i_d.di_mode))
2707                 return XFS_ERROR(EINVAL);
2708
2709         switch (bf->l_whence) {
2710         case 0: /*SEEK_SET*/
2711                 break;
2712         case 1: /*SEEK_CUR*/
2713                 bf->l_start += offset;
2714                 break;
2715         case 2: /*SEEK_END*/
2716                 bf->l_start += ip->i_size;
2717                 break;
2718         default:
2719                 return XFS_ERROR(EINVAL);
2720         }
2721
2722         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2723
2724         if (   (bf->l_start < 0)
2725             || (bf->l_start > XFS_MAXIOFFSET(mp))
2726             || (bf->l_start + llen < 0)
2727             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2728                 return XFS_ERROR(EINVAL);
2729
2730         bf->l_whence = 0;
2731
2732         startoffset = bf->l_start;
2733         fsize = ip->i_size;
2734
2735         /*
2736          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2737          * file space.
2738          * These calls do NOT zero the data space allocated to the file,
2739          * nor do they change the file size.
2740          *
2741          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2742          * space.
2743          * These calls cause the new file data to be zeroed and the file
2744          * size to be changed.
2745          */
2746         setprealloc = clrprealloc = 0;
2747         prealloc_type = XFS_BMAPI_PREALLOC;
2748
2749         switch (cmd) {
2750         case XFS_IOC_ZERO_RANGE:
2751                 prealloc_type |= XFS_BMAPI_CONVERT;
2752                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2753                 /* FALLTHRU */
2754         case XFS_IOC_RESVSP:
2755         case XFS_IOC_RESVSP64:
2756                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2757                                                 prealloc_type, attr_flags);
2758                 if (error)
2759                         return error;
2760                 setprealloc = 1;
2761                 break;
2762
2763         case XFS_IOC_UNRESVSP:
2764         case XFS_IOC_UNRESVSP64:
2765                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2766                                                                 attr_flags)))
2767                         return error;
2768                 break;
2769
2770         case XFS_IOC_ALLOCSP:
2771         case XFS_IOC_ALLOCSP64:
2772         case XFS_IOC_FREESP:
2773         case XFS_IOC_FREESP64:
2774                 if (startoffset > fsize) {
2775                         error = xfs_alloc_file_space(ip, fsize,
2776                                         startoffset - fsize, 0, attr_flags);
2777                         if (error)
2778                                 break;
2779                 }
2780
2781                 iattr.ia_valid = ATTR_SIZE;
2782                 iattr.ia_size = startoffset;
2783
2784                 error = xfs_setattr(ip, &iattr, attr_flags);
2785
2786                 if (error)
2787                         return error;
2788
2789                 clrprealloc = 1;
2790                 break;
2791
2792         default:
2793                 ASSERT(0);
2794                 return XFS_ERROR(EINVAL);
2795         }
2796
2797         /*
2798          * update the inode timestamp, mode, and prealloc flag bits
2799          */
2800         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2801
2802         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2803                                       0, 0, 0))) {
2804                 /* ASSERT(0); */
2805                 xfs_trans_cancel(tp, 0);
2806                 return error;
2807         }
2808
2809         xfs_ilock(ip, XFS_ILOCK_EXCL);
2810
2811         xfs_trans_ijoin(tp, ip);
2812
2813         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2814                 ip->i_d.di_mode &= ~S_ISUID;
2815
2816                 /*
2817                  * Note that we don't have to worry about mandatory
2818                  * file locking being disabled here because we only
2819                  * clear the S_ISGID bit if the Group execute bit is
2820                  * on, but if it was on then mandatory locking wouldn't
2821                  * have been enabled.
2822                  */
2823                 if (ip->i_d.di_mode & S_IXGRP)
2824                         ip->i_d.di_mode &= ~S_ISGID;
2825
2826                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2827         }
2828         if (setprealloc)
2829                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2830         else if (clrprealloc)
2831                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2832
2833         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2834         xfs_trans_set_sync(tp);
2835
2836         error = xfs_trans_commit(tp, 0);
2837
2838         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2839
2840         return error;
2841 }