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