xfs: remove xfs_buf wrappers
[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                                 inode->i_mtime = inode->i_ctime =
189                                                 current_fs_time(inode->i_sb);
190                                 xfs_mark_inode_dirty_sync(ip);
191                         }
192                         code = 0;
193                         goto error_return;
194                 }
195
196                 if (S_ISDIR(ip->i_d.di_mode)) {
197                         code = XFS_ERROR(EISDIR);
198                         goto error_return;
199                 } else if (!S_ISREG(ip->i_d.di_mode)) {
200                         code = XFS_ERROR(EINVAL);
201                         goto error_return;
202                 }
203
204                 /*
205                  * Make sure that the dquots are attached to the inode.
206                  */
207                 code = xfs_qm_dqattach_locked(ip, 0);
208                 if (code)
209                         goto error_return;
210
211                 /*
212                  * Now we can make the changes.  Before we join the inode
213                  * to the transaction, if ATTR_SIZE is set then take care of
214                  * the part of the truncation that must be done without the
215                  * inode lock.  This needs to be done before joining the inode
216                  * to the transaction, because the inode cannot be unlocked
217                  * once it is a part of the transaction.
218                  */
219                 if (iattr->ia_size > ip->i_size) {
220                         /*
221                          * Do the first part of growing a file: zero any data
222                          * in the last block that is beyond the old EOF.  We
223                          * need to do this before the inode is joined to the
224                          * transaction to modify the i_size.
225                          */
226                         code = xfs_zero_eof(ip, iattr->ia_size, ip->i_size);
227                         if (code)
228                                 goto error_return;
229                 }
230                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
231                 lock_flags &= ~XFS_ILOCK_EXCL;
232
233                 /*
234                  * We are going to log the inode size change in this
235                  * transaction so any previous writes that are beyond the on
236                  * disk EOF and the new EOF that have not been written out need
237                  * to be written here. If we do not write the data out, we
238                  * expose ourselves to the null files problem.
239                  *
240                  * Only flush from the on disk size to the smaller of the in
241                  * memory file size or the new size as that's the range we
242                  * really care about here and prevents waiting for other data
243                  * not within the range we care about here.
244                  */
245                 if (ip->i_size != ip->i_d.di_size &&
246                     iattr->ia_size > ip->i_d.di_size) {
247                         code = xfs_flush_pages(ip,
248                                         ip->i_d.di_size, iattr->ia_size,
249                                         XBF_ASYNC, FI_NONE);
250                         if (code)
251                                 goto error_return;
252                 }
253
254                 /* wait for all I/O to complete */
255                 xfs_ioend_wait(ip);
256
257                 code = -block_truncate_page(inode->i_mapping, iattr->ia_size,
258                                             xfs_get_blocks);
259                 if (code)
260                         goto error_return;
261
262                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
263                 code = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
264                                          XFS_TRANS_PERM_LOG_RES,
265                                          XFS_ITRUNCATE_LOG_COUNT);
266                 if (code)
267                         goto error_return;
268
269                 truncate_setsize(inode, iattr->ia_size);
270
271                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
272                 lock_flags |= XFS_ILOCK_EXCL;
273
274                 xfs_ilock(ip, XFS_ILOCK_EXCL);
275
276                 xfs_trans_ijoin(tp, ip);
277
278                 /*
279                  * Only change the c/mtime if we are changing the size
280                  * or we are explicitly asked to change it. This handles
281                  * the semantic difference between truncate() and ftruncate()
282                  * as implemented in the VFS.
283                  *
284                  * The regular truncate() case without ATTR_CTIME and ATTR_MTIME
285                  * is a special case where we need to update the times despite
286                  * not having these flags set.  For all other operations the
287                  * VFS set these flags explicitly if it wants a timestamp
288                  * update.
289                  */
290                 if (iattr->ia_size != ip->i_size &&
291                     (!(mask & (ATTR_CTIME | ATTR_MTIME)))) {
292                         iattr->ia_ctime = iattr->ia_mtime =
293                                 current_fs_time(inode->i_sb);
294                         mask |= ATTR_CTIME | ATTR_MTIME;
295                 }
296
297                 if (iattr->ia_size > ip->i_size) {
298                         ip->i_d.di_size = iattr->ia_size;
299                         ip->i_size = iattr->ia_size;
300                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
301                 } else if (iattr->ia_size <= ip->i_size ||
302                            (iattr->ia_size == 0 && ip->i_d.di_nextents)) {
303                         /*
304                          * signal a sync transaction unless
305                          * we're truncating an already unlinked
306                          * file on a wsync filesystem
307                          */
308                         code = xfs_itruncate_finish(&tp, ip, iattr->ia_size,
309                                             XFS_DATA_FORK,
310                                             ((ip->i_d.di_nlink != 0 ||
311                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
312                                              ? 1 : 0));
313                         if (code)
314                                 goto abort_return;
315                         /*
316                          * Truncated "down", so we're removing references
317                          * to old data here - if we now delay flushing for
318                          * a long time, we expose ourselves unduly to the
319                          * notorious NULL files problem.  So, we mark this
320                          * vnode and flush it when the file is closed, and
321                          * do not wait the usual (long) time for writeout.
322                          */
323                         xfs_iflags_set(ip, XFS_ITRUNCATED);
324                 }
325         } else if (tp) {
326                 xfs_trans_ijoin(tp, ip);
327         }
328
329         /*
330          * Change file ownership.  Must be the owner or privileged.
331          */
332         if (mask & (ATTR_UID|ATTR_GID)) {
333                 /*
334                  * CAP_FSETID overrides the following restrictions:
335                  *
336                  * The set-user-ID and set-group-ID bits of a file will be
337                  * cleared upon successful return from chown()
338                  */
339                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
340                     !capable(CAP_FSETID)) {
341                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
342                 }
343
344                 /*
345                  * Change the ownerships and register quota modifications
346                  * in the transaction.
347                  */
348                 if (iuid != uid) {
349                         if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_UQUOTA_ON(mp)) {
350                                 ASSERT(mask & ATTR_UID);
351                                 ASSERT(udqp);
352                                 olddquot1 = xfs_qm_vop_chown(tp, ip,
353                                                         &ip->i_udquot, udqp);
354                         }
355                         ip->i_d.di_uid = uid;
356                         inode->i_uid = uid;
357                 }
358                 if (igid != gid) {
359                         if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_GQUOTA_ON(mp)) {
360                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
361                                 ASSERT(mask & ATTR_GID);
362                                 ASSERT(gdqp);
363                                 olddquot2 = xfs_qm_vop_chown(tp, ip,
364                                                         &ip->i_gdquot, gdqp);
365                         }
366                         ip->i_d.di_gid = gid;
367                         inode->i_gid = gid;
368                 }
369         }
370
371         /*
372          * Change file access modes.
373          */
374         if (mask & ATTR_MODE) {
375                 umode_t mode = iattr->ia_mode;
376
377                 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
378                         mode &= ~S_ISGID;
379
380                 ip->i_d.di_mode &= S_IFMT;
381                 ip->i_d.di_mode |= mode & ~S_IFMT;
382
383                 inode->i_mode &= S_IFMT;
384                 inode->i_mode |= mode & ~S_IFMT;
385         }
386
387         /*
388          * Change file access or modified times.
389          */
390         if (mask & ATTR_ATIME) {
391                 inode->i_atime = iattr->ia_atime;
392                 ip->i_d.di_atime.t_sec = iattr->ia_atime.tv_sec;
393                 ip->i_d.di_atime.t_nsec = iattr->ia_atime.tv_nsec;
394                 ip->i_update_core = 1;
395         }
396         if (mask & ATTR_CTIME) {
397                 inode->i_ctime = iattr->ia_ctime;
398                 ip->i_d.di_ctime.t_sec = iattr->ia_ctime.tv_sec;
399                 ip->i_d.di_ctime.t_nsec = iattr->ia_ctime.tv_nsec;
400                 ip->i_update_core = 1;
401         }
402         if (mask & ATTR_MTIME) {
403                 inode->i_mtime = iattr->ia_mtime;
404                 ip->i_d.di_mtime.t_sec = iattr->ia_mtime.tv_sec;
405                 ip->i_d.di_mtime.t_nsec = iattr->ia_mtime.tv_nsec;
406                 ip->i_update_core = 1;
407         }
408
409         /*
410          * And finally, log the inode core if any attribute in it
411          * has been changed.
412          */
413         if (mask & (ATTR_UID|ATTR_GID|ATTR_MODE|
414                     ATTR_ATIME|ATTR_CTIME|ATTR_MTIME))
415                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
416
417         XFS_STATS_INC(xs_ig_attrchg);
418
419         /*
420          * If this is a synchronous mount, make sure that the
421          * transaction goes to disk before returning to the user.
422          * This is slightly sub-optimal in that truncates require
423          * two sync transactions instead of one for wsync filesystems.
424          * One for the truncate and one for the timestamps since we
425          * don't want to change the timestamps unless we're sure the
426          * truncate worked.  Truncates are less than 1% of the laddis
427          * mix so this probably isn't worth the trouble to optimize.
428          */
429         code = 0;
430         if (mp->m_flags & XFS_MOUNT_WSYNC)
431                 xfs_trans_set_sync(tp);
432
433         code = xfs_trans_commit(tp, commit_flags);
434
435         xfs_iunlock(ip, lock_flags);
436
437         /*
438          * Release any dquot(s) the inode had kept before chown.
439          */
440         xfs_qm_dqrele(olddquot1);
441         xfs_qm_dqrele(olddquot2);
442         xfs_qm_dqrele(udqp);
443         xfs_qm_dqrele(gdqp);
444
445         if (code)
446                 return code;
447
448         /*
449          * XXX(hch): Updating the ACL entries is not atomic vs the i_mode
450          *           update.  We could avoid this with linked transactions
451          *           and passing down the transaction pointer all the way
452          *           to attr_set.  No previous user of the generic
453          *           Posix ACL code seems to care about this issue either.
454          */
455         if ((mask & ATTR_MODE) && !(flags & XFS_ATTR_NOACL)) {
456                 code = -xfs_acl_chmod(inode);
457                 if (code)
458                         return XFS_ERROR(code);
459         }
460
461         return 0;
462
463  abort_return:
464         commit_flags |= XFS_TRANS_ABORT;
465  error_return:
466         xfs_qm_dqrele(udqp);
467         xfs_qm_dqrele(gdqp);
468         if (tp) {
469                 xfs_trans_cancel(tp, commit_flags);
470         }
471         if (lock_flags != 0) {
472                 xfs_iunlock(ip, lock_flags);
473         }
474         return code;
475 }
476
477 /*
478  * The maximum pathlen is 1024 bytes. Since the minimum file system
479  * blocksize is 512 bytes, we can get a max of 2 extents back from
480  * bmapi.
481  */
482 #define SYMLINK_MAPS 2
483
484 STATIC int
485 xfs_readlink_bmap(
486         xfs_inode_t     *ip,
487         char            *link)
488 {
489         xfs_mount_t     *mp = ip->i_mount;
490         int             pathlen = ip->i_d.di_size;
491         int             nmaps = SYMLINK_MAPS;
492         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
493         xfs_daddr_t     d;
494         int             byte_cnt;
495         int             n;
496         xfs_buf_t       *bp;
497         int             error = 0;
498
499         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
500                         mval, &nmaps, NULL);
501         if (error)
502                 goto out;
503
504         for (n = 0; n < nmaps; n++) {
505                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
506                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
507
508                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt),
509                                   XBF_LOCK | XBF_MAPPED | XBF_DONT_BLOCK);
510                 error = XFS_BUF_GETERROR(bp);
511                 if (error) {
512                         xfs_ioerror_alert("xfs_readlink",
513                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
514                         xfs_buf_relse(bp);
515                         goto out;
516                 }
517                 if (pathlen < byte_cnt)
518                         byte_cnt = pathlen;
519                 pathlen -= byte_cnt;
520
521                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
522                 xfs_buf_relse(bp);
523         }
524
525         link[ip->i_d.di_size] = '\0';
526         error = 0;
527
528  out:
529         return error;
530 }
531
532 int
533 xfs_readlink(
534         xfs_inode_t     *ip,
535         char            *link)
536 {
537         xfs_mount_t     *mp = ip->i_mount;
538         int             pathlen;
539         int             error = 0;
540
541         trace_xfs_readlink(ip);
542
543         if (XFS_FORCED_SHUTDOWN(mp))
544                 return XFS_ERROR(EIO);
545
546         xfs_ilock(ip, XFS_ILOCK_SHARED);
547
548         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
549         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
550
551         pathlen = ip->i_d.di_size;
552         if (!pathlen)
553                 goto out;
554
555         if (ip->i_df.if_flags & XFS_IFINLINE) {
556                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
557                 link[pathlen] = '\0';
558         } else {
559                 error = xfs_readlink_bmap(ip, link);
560         }
561
562  out:
563         xfs_iunlock(ip, XFS_ILOCK_SHARED);
564         return error;
565 }
566
567 /*
568  * Flags for xfs_free_eofblocks
569  */
570 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
571
572 /*
573  * This is called by xfs_inactive to free any blocks beyond eof
574  * when the link count isn't zero and by xfs_dm_punch_hole() when
575  * punching a hole to EOF.
576  */
577 STATIC int
578 xfs_free_eofblocks(
579         xfs_mount_t     *mp,
580         xfs_inode_t     *ip,
581         int             flags)
582 {
583         xfs_trans_t     *tp;
584         int             error;
585         xfs_fileoff_t   end_fsb;
586         xfs_fileoff_t   last_fsb;
587         xfs_filblks_t   map_len;
588         int             nimaps;
589         xfs_bmbt_irec_t imap;
590
591         /*
592          * Figure out if there are any blocks beyond the end
593          * of the file.  If not, then there is nothing to do.
594          */
595         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
596         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
597         if (last_fsb <= end_fsb)
598                 return 0;
599         map_len = last_fsb - end_fsb;
600
601         nimaps = 1;
602         xfs_ilock(ip, XFS_ILOCK_SHARED);
603         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
604                           NULL, 0, &imap, &nimaps, NULL);
605         xfs_iunlock(ip, XFS_ILOCK_SHARED);
606
607         if (!error && (nimaps != 0) &&
608             (imap.br_startblock != HOLESTARTBLOCK ||
609              ip->i_delayed_blks)) {
610                 /*
611                  * Attach the dquots to the inode up front.
612                  */
613                 error = xfs_qm_dqattach(ip, 0);
614                 if (error)
615                         return error;
616
617                 /*
618                  * There are blocks after the end of file.
619                  * Free them up now by truncating the file to
620                  * its current size.
621                  */
622                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
623
624                 /*
625                  * Do the xfs_itruncate_start() call before
626                  * reserving any log space because
627                  * itruncate_start will call into the buffer
628                  * cache and we can't
629                  * do that within a transaction.
630                  */
631                 if (flags & XFS_FREE_EOF_TRYLOCK) {
632                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
633                                 xfs_trans_cancel(tp, 0);
634                                 return 0;
635                         }
636                 } else {
637                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
638                 }
639                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
640                                     ip->i_size);
641                 if (error) {
642                         xfs_trans_cancel(tp, 0);
643                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
644                         return error;
645                 }
646
647                 error = xfs_trans_reserve(tp, 0,
648                                           XFS_ITRUNCATE_LOG_RES(mp),
649                                           0, XFS_TRANS_PERM_LOG_RES,
650                                           XFS_ITRUNCATE_LOG_COUNT);
651                 if (error) {
652                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
653                         xfs_trans_cancel(tp, 0);
654                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
655                         return error;
656                 }
657
658                 xfs_ilock(ip, XFS_ILOCK_EXCL);
659                 xfs_trans_ijoin(tp, ip);
660
661                 error = xfs_itruncate_finish(&tp, ip,
662                                              ip->i_size,
663                                              XFS_DATA_FORK,
664                                              0);
665                 /*
666                  * If we get an error at this point we
667                  * simply don't bother truncating the file.
668                  */
669                 if (error) {
670                         xfs_trans_cancel(tp,
671                                          (XFS_TRANS_RELEASE_LOG_RES |
672                                           XFS_TRANS_ABORT));
673                 } else {
674                         error = xfs_trans_commit(tp,
675                                                 XFS_TRANS_RELEASE_LOG_RES);
676                 }
677                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
678         }
679         return error;
680 }
681
682 /*
683  * Free a symlink that has blocks associated with it.
684  */
685 STATIC int
686 xfs_inactive_symlink_rmt(
687         xfs_inode_t     *ip,
688         xfs_trans_t     **tpp)
689 {
690         xfs_buf_t       *bp;
691         int             committed;
692         int             done;
693         int             error;
694         xfs_fsblock_t   first_block;
695         xfs_bmap_free_t free_list;
696         int             i;
697         xfs_mount_t     *mp;
698         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
699         int             nmaps;
700         xfs_trans_t     *ntp;
701         int             size;
702         xfs_trans_t     *tp;
703
704         tp = *tpp;
705         mp = ip->i_mount;
706         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
707         /*
708          * We're freeing a symlink that has some
709          * blocks allocated to it.  Free the
710          * blocks here.  We know that we've got
711          * either 1 or 2 extents and that we can
712          * free them all in one bunmapi call.
713          */
714         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
715         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
716                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
717                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
718                 xfs_trans_cancel(tp, 0);
719                 *tpp = NULL;
720                 return error;
721         }
722         /*
723          * Lock the inode, fix the size, and join it to the transaction.
724          * Hold it so in the normal path, we still have it locked for
725          * the second transaction.  In the error paths we need it
726          * held so the cancel won't rele it, see below.
727          */
728         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
729         size = (int)ip->i_d.di_size;
730         ip->i_d.di_size = 0;
731         xfs_trans_ijoin(tp, ip);
732         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
733         /*
734          * Find the block(s) so we can inval and unmap them.
735          */
736         done = 0;
737         xfs_bmap_init(&free_list, &first_block);
738         nmaps = ARRAY_SIZE(mval);
739         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
740                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
741                         &free_list)))
742                 goto error0;
743         /*
744          * Invalidate the block(s).
745          */
746         for (i = 0; i < nmaps; i++) {
747                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
748                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
749                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
750                 xfs_trans_binval(tp, bp);
751         }
752         /*
753          * Unmap the dead block(s) to the free_list.
754          */
755         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
756                         &first_block, &free_list, &done)))
757                 goto error1;
758         ASSERT(done);
759         /*
760          * Commit the first transaction.  This logs the EFI and the inode.
761          */
762         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
763                 goto error1;
764         /*
765          * The transaction must have been committed, since there were
766          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
767          * The new tp has the extent freeing and EFDs.
768          */
769         ASSERT(committed);
770         /*
771          * The first xact was committed, so add the inode to the new one.
772          * Mark it dirty so it will be logged and moved forward in the log as
773          * part of every commit.
774          */
775         xfs_trans_ijoin(tp, ip);
776         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
777         /*
778          * Get a new, empty transaction to return to our caller.
779          */
780         ntp = xfs_trans_dup(tp);
781         /*
782          * Commit the transaction containing extent freeing and EFDs.
783          * If we get an error on the commit here or on the reserve below,
784          * we need to unlock the inode since the new transaction doesn't
785          * have the inode attached.
786          */
787         error = xfs_trans_commit(tp, 0);
788         tp = ntp;
789         if (error) {
790                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
791                 goto error0;
792         }
793         /*
794          * transaction commit worked ok so we can drop the extra ticket
795          * reference that we gained in xfs_trans_dup()
796          */
797         xfs_log_ticket_put(tp->t_ticket);
798
799         /*
800          * Remove the memory for extent descriptions (just bookkeeping).
801          */
802         if (ip->i_df.if_bytes)
803                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
804         ASSERT(ip->i_df.if_bytes == 0);
805         /*
806          * Put an itruncate log reservation in the new transaction
807          * for our caller.
808          */
809         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
810                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
811                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
812                 goto error0;
813         }
814         /*
815          * Return with the inode locked but not joined to the transaction.
816          */
817         *tpp = tp;
818         return 0;
819
820  error1:
821         xfs_bmap_cancel(&free_list);
822  error0:
823         /*
824          * Have to come here with the inode locked and either
825          * (held and in the transaction) or (not in the transaction).
826          * If the inode isn't held then cancel would iput it, but
827          * that's wrong since this is inactive and the vnode ref
828          * count is 0 already.
829          * Cancel won't do anything to the inode if held, but it still
830          * needs to be locked until the cancel is done, if it was
831          * joined to the transaction.
832          */
833         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
834         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
835         *tpp = NULL;
836         return error;
837
838 }
839
840 STATIC int
841 xfs_inactive_symlink_local(
842         xfs_inode_t     *ip,
843         xfs_trans_t     **tpp)
844 {
845         int             error;
846
847         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
848         /*
849          * We're freeing a symlink which fit into
850          * the inode.  Just free the memory used
851          * to hold the old symlink.
852          */
853         error = xfs_trans_reserve(*tpp, 0,
854                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
855                                   0, XFS_TRANS_PERM_LOG_RES,
856                                   XFS_ITRUNCATE_LOG_COUNT);
857
858         if (error) {
859                 xfs_trans_cancel(*tpp, 0);
860                 *tpp = NULL;
861                 return error;
862         }
863         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
864
865         /*
866          * Zero length symlinks _can_ exist.
867          */
868         if (ip->i_df.if_bytes > 0) {
869                 xfs_idata_realloc(ip,
870                                   -(ip->i_df.if_bytes),
871                                   XFS_DATA_FORK);
872                 ASSERT(ip->i_df.if_bytes == 0);
873         }
874         return 0;
875 }
876
877 STATIC int
878 xfs_inactive_attrs(
879         xfs_inode_t     *ip,
880         xfs_trans_t     **tpp)
881 {
882         xfs_trans_t     *tp;
883         int             error;
884         xfs_mount_t     *mp;
885
886         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
887         tp = *tpp;
888         mp = ip->i_mount;
889         ASSERT(ip->i_d.di_forkoff != 0);
890         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
891         xfs_iunlock(ip, XFS_ILOCK_EXCL);
892         if (error)
893                 goto error_unlock;
894
895         error = xfs_attr_inactive(ip);
896         if (error)
897                 goto error_unlock;
898
899         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
900         error = xfs_trans_reserve(tp, 0,
901                                   XFS_IFREE_LOG_RES(mp),
902                                   0, XFS_TRANS_PERM_LOG_RES,
903                                   XFS_INACTIVE_LOG_COUNT);
904         if (error)
905                 goto error_cancel;
906
907         xfs_ilock(ip, XFS_ILOCK_EXCL);
908         xfs_trans_ijoin(tp, ip);
909         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
910
911         ASSERT(ip->i_d.di_anextents == 0);
912
913         *tpp = tp;
914         return 0;
915
916 error_cancel:
917         ASSERT(XFS_FORCED_SHUTDOWN(mp));
918         xfs_trans_cancel(tp, 0);
919 error_unlock:
920         *tpp = NULL;
921         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
922         return error;
923 }
924
925 int
926 xfs_release(
927         xfs_inode_t     *ip)
928 {
929         xfs_mount_t     *mp = ip->i_mount;
930         int             error;
931
932         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
933                 return 0;
934
935         /* If this is a read-only mount, don't do this (would generate I/O) */
936         if (mp->m_flags & XFS_MOUNT_RDONLY)
937                 return 0;
938
939         if (!XFS_FORCED_SHUTDOWN(mp)) {
940                 int truncated;
941
942                 /*
943                  * If we are using filestreams, and we have an unlinked
944                  * file that we are processing the last close on, then nothing
945                  * will be able to reopen and write to this file. Purge this
946                  * inode from the filestreams cache so that it doesn't delay
947                  * teardown of the inode.
948                  */
949                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
950                         xfs_filestream_deassociate(ip);
951
952                 /*
953                  * If we previously truncated this file and removed old data
954                  * in the process, we want to initiate "early" writeout on
955                  * the last close.  This is an attempt to combat the notorious
956                  * NULL files problem which is particularly noticable from a
957                  * truncate down, buffered (re-)write (delalloc), followed by
958                  * a crash.  What we are effectively doing here is
959                  * significantly reducing the time window where we'd otherwise
960                  * be exposed to that problem.
961                  */
962                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
963                 if (truncated && VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
964                         xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
965         }
966
967         if (ip->i_d.di_nlink != 0) {
968                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
969                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
970                        ip->i_delayed_blks > 0)) &&
971                      (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
972                     (!(ip->i_d.di_flags &
973                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
974
975                         /*
976                          * If we can't get the iolock just skip truncating
977                          * the blocks past EOF because we could deadlock
978                          * with the mmap_sem otherwise.  We'll get another
979                          * chance to drop them once the last reference to
980                          * the inode is dropped, so we'll never leak blocks
981                          * permanently.
982                          */
983                         error = xfs_free_eofblocks(mp, ip,
984                                                    XFS_FREE_EOF_TRYLOCK);
985                         if (error)
986                                 return error;
987                 }
988         }
989
990         return 0;
991 }
992
993 /*
994  * xfs_inactive
995  *
996  * This is called when the vnode reference count for the vnode
997  * goes to zero.  If the file has been unlinked, then it must
998  * now be truncated.  Also, we clear all of the read-ahead state
999  * kept for the inode here since the file is now closed.
1000  */
1001 int
1002 xfs_inactive(
1003         xfs_inode_t     *ip)
1004 {
1005         xfs_bmap_free_t free_list;
1006         xfs_fsblock_t   first_block;
1007         int             committed;
1008         xfs_trans_t     *tp;
1009         xfs_mount_t     *mp;
1010         int             error;
1011         int             truncate;
1012
1013         /*
1014          * If the inode is already free, then there can be nothing
1015          * to clean up here.
1016          */
1017         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
1018                 ASSERT(ip->i_df.if_real_bytes == 0);
1019                 ASSERT(ip->i_df.if_broot_bytes == 0);
1020                 return VN_INACTIVE_CACHE;
1021         }
1022
1023         /*
1024          * Only do a truncate if it's a regular file with
1025          * some actual space in it.  It's OK to look at the
1026          * inode's fields without the lock because we're the
1027          * only one with a reference to the inode.
1028          */
1029         truncate = ((ip->i_d.di_nlink == 0) &&
1030             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1031              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1032             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1033
1034         mp = ip->i_mount;
1035
1036         error = 0;
1037
1038         /* If this is a read-only mount, don't do this (would generate I/O) */
1039         if (mp->m_flags & XFS_MOUNT_RDONLY)
1040                 goto out;
1041
1042         if (ip->i_d.di_nlink != 0) {
1043                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1044                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1045                        ip->i_delayed_blks > 0)) &&
1046                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1047                      (!(ip->i_d.di_flags &
1048                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1049                       (ip->i_delayed_blks != 0)))) {
1050                         error = xfs_free_eofblocks(mp, ip, 0);
1051                         if (error)
1052                                 return VN_INACTIVE_CACHE;
1053                 }
1054                 goto out;
1055         }
1056
1057         ASSERT(ip->i_d.di_nlink == 0);
1058
1059         error = xfs_qm_dqattach(ip, 0);
1060         if (error)
1061                 return VN_INACTIVE_CACHE;
1062
1063         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1064         if (truncate) {
1065                 /*
1066                  * Do the xfs_itruncate_start() call before
1067                  * reserving any log space because itruncate_start
1068                  * will call into the buffer cache and we can't
1069                  * do that within a transaction.
1070                  */
1071                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1072
1073                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1074                 if (error) {
1075                         xfs_trans_cancel(tp, 0);
1076                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1077                         return VN_INACTIVE_CACHE;
1078                 }
1079
1080                 error = xfs_trans_reserve(tp, 0,
1081                                           XFS_ITRUNCATE_LOG_RES(mp),
1082                                           0, XFS_TRANS_PERM_LOG_RES,
1083                                           XFS_ITRUNCATE_LOG_COUNT);
1084                 if (error) {
1085                         /* Don't call itruncate_cleanup */
1086                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1087                         xfs_trans_cancel(tp, 0);
1088                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1089                         return VN_INACTIVE_CACHE;
1090                 }
1091
1092                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1093                 xfs_trans_ijoin(tp, ip);
1094
1095                 /*
1096                  * normally, we have to run xfs_itruncate_finish sync.
1097                  * But if filesystem is wsync and we're in the inactive
1098                  * path, then we know that nlink == 0, and that the
1099                  * xaction that made nlink == 0 is permanently committed
1100                  * since xfs_remove runs as a synchronous transaction.
1101                  */
1102                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1103                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1104
1105                 if (error) {
1106                         xfs_trans_cancel(tp,
1107                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1108                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1109                         return VN_INACTIVE_CACHE;
1110                 }
1111         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1112
1113                 /*
1114                  * If we get an error while cleaning up a
1115                  * symlink we bail out.
1116                  */
1117                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1118                         xfs_inactive_symlink_rmt(ip, &tp) :
1119                         xfs_inactive_symlink_local(ip, &tp);
1120
1121                 if (error) {
1122                         ASSERT(tp == NULL);
1123                         return VN_INACTIVE_CACHE;
1124                 }
1125
1126                 xfs_trans_ijoin(tp, ip);
1127         } else {
1128                 error = xfs_trans_reserve(tp, 0,
1129                                           XFS_IFREE_LOG_RES(mp),
1130                                           0, XFS_TRANS_PERM_LOG_RES,
1131                                           XFS_INACTIVE_LOG_COUNT);
1132                 if (error) {
1133                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1134                         xfs_trans_cancel(tp, 0);
1135                         return VN_INACTIVE_CACHE;
1136                 }
1137
1138                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1139                 xfs_trans_ijoin(tp, ip);
1140         }
1141
1142         /*
1143          * If there are attributes associated with the file
1144          * then blow them away now.  The code calls a routine
1145          * that recursively deconstructs the attribute fork.
1146          * We need to just commit the current transaction
1147          * because we can't use it for xfs_attr_inactive().
1148          */
1149         if (ip->i_d.di_anextents > 0) {
1150                 error = xfs_inactive_attrs(ip, &tp);
1151                 /*
1152                  * If we got an error, the transaction is already
1153                  * cancelled, and the inode is unlocked. Just get out.
1154                  */
1155                  if (error)
1156                          return VN_INACTIVE_CACHE;
1157         } else if (ip->i_afp) {
1158                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1159         }
1160
1161         /*
1162          * Free the inode.
1163          */
1164         xfs_bmap_init(&free_list, &first_block);
1165         error = xfs_ifree(tp, ip, &free_list);
1166         if (error) {
1167                 /*
1168                  * If we fail to free the inode, shut down.  The cancel
1169                  * might do that, we need to make sure.  Otherwise the
1170                  * inode might be lost for a long time or forever.
1171                  */
1172                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1173                         cmn_err(CE_NOTE,
1174                 "xfs_inactive:  xfs_ifree() returned an error = %d on %s",
1175                                 error, mp->m_fsname);
1176                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1177                 }
1178                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1179         } else {
1180                 /*
1181                  * Credit the quota account(s). The inode is gone.
1182                  */
1183                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1184
1185                 /*
1186                  * Just ignore errors at this point.  There is nothing we can
1187                  * do except to try to keep going. Make sure it's not a silent
1188                  * error.
1189                  */
1190                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1191                 if (error)
1192                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1193                                 "xfs_bmap_finish() returned error %d", error);
1194                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1195                 if (error)
1196                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1197                                 "xfs_trans_commit() returned error %d", error);
1198         }
1199
1200         /*
1201          * Release the dquots held by inode, if any.
1202          */
1203         xfs_qm_dqdetach(ip);
1204         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1205
1206  out:
1207         return VN_INACTIVE_CACHE;
1208 }
1209
1210 /*
1211  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1212  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1213  * ci_name->name will point to a the actual name (caller must free) or
1214  * will be set to NULL if an exact match is found.
1215  */
1216 int
1217 xfs_lookup(
1218         xfs_inode_t             *dp,
1219         struct xfs_name         *name,
1220         xfs_inode_t             **ipp,
1221         struct xfs_name         *ci_name)
1222 {
1223         xfs_ino_t               inum;
1224         int                     error;
1225         uint                    lock_mode;
1226
1227         trace_xfs_lookup(dp, name);
1228
1229         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1230                 return XFS_ERROR(EIO);
1231
1232         lock_mode = xfs_ilock_map_shared(dp);
1233         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1234         xfs_iunlock_map_shared(dp, lock_mode);
1235
1236         if (error)
1237                 goto out;
1238
1239         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
1240         if (error)
1241                 goto out_free_name;
1242
1243         return 0;
1244
1245 out_free_name:
1246         if (ci_name)
1247                 kmem_free(ci_name->name);
1248 out:
1249         *ipp = NULL;
1250         return error;
1251 }
1252
1253 int
1254 xfs_create(
1255         xfs_inode_t             *dp,
1256         struct xfs_name         *name,
1257         mode_t                  mode,
1258         xfs_dev_t               rdev,
1259         xfs_inode_t             **ipp)
1260 {
1261         int                     is_dir = S_ISDIR(mode);
1262         struct xfs_mount        *mp = dp->i_mount;
1263         struct xfs_inode        *ip = NULL;
1264         struct xfs_trans        *tp = NULL;
1265         int                     error;
1266         xfs_bmap_free_t         free_list;
1267         xfs_fsblock_t           first_block;
1268         boolean_t               unlock_dp_on_error = B_FALSE;
1269         uint                    cancel_flags;
1270         int                     committed;
1271         xfs_prid_t              prid;
1272         struct xfs_dquot        *udqp = NULL;
1273         struct xfs_dquot        *gdqp = NULL;
1274         uint                    resblks;
1275         uint                    log_res;
1276         uint                    log_count;
1277
1278         trace_xfs_create(dp, name);
1279
1280         if (XFS_FORCED_SHUTDOWN(mp))
1281                 return XFS_ERROR(EIO);
1282
1283         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1284                 prid = dp->i_d.di_projid;
1285         else
1286                 prid = dfltprid;
1287
1288         /*
1289          * Make sure that we have allocated dquot(s) on disk.
1290          */
1291         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1292                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1293         if (error)
1294                 goto std_return;
1295
1296         if (is_dir) {
1297                 rdev = 0;
1298                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1299                 log_res = XFS_MKDIR_LOG_RES(mp);
1300                 log_count = XFS_MKDIR_LOG_COUNT;
1301                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
1302         } else {
1303                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1304                 log_res = XFS_CREATE_LOG_RES(mp);
1305                 log_count = XFS_CREATE_LOG_COUNT;
1306                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1307         }
1308
1309         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1310
1311         /*
1312          * Initially assume that the file does not exist and
1313          * reserve the resources for that case.  If that is not
1314          * the case we'll drop the one we have and get a more
1315          * appropriate transaction later.
1316          */
1317         error = xfs_trans_reserve(tp, resblks, log_res, 0,
1318                         XFS_TRANS_PERM_LOG_RES, log_count);
1319         if (error == ENOSPC) {
1320                 /* flush outstanding delalloc blocks and retry */
1321                 xfs_flush_inodes(dp);
1322                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
1323                                 XFS_TRANS_PERM_LOG_RES, log_count);
1324         }
1325         if (error == ENOSPC) {
1326                 /* No space at all so try a "no-allocation" reservation */
1327                 resblks = 0;
1328                 error = xfs_trans_reserve(tp, 0, log_res, 0,
1329                                 XFS_TRANS_PERM_LOG_RES, log_count);
1330         }
1331         if (error) {
1332                 cancel_flags = 0;
1333                 goto out_trans_cancel;
1334         }
1335
1336         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1337         unlock_dp_on_error = B_TRUE;
1338
1339         /*
1340          * Check for directory link count overflow.
1341          */
1342         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
1343                 error = XFS_ERROR(EMLINK);
1344                 goto out_trans_cancel;
1345         }
1346
1347         xfs_bmap_init(&free_list, &first_block);
1348
1349         /*
1350          * Reserve disk quota and the inode.
1351          */
1352         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1353         if (error)
1354                 goto out_trans_cancel;
1355
1356         error = xfs_dir_canenter(tp, dp, name, resblks);
1357         if (error)
1358                 goto out_trans_cancel;
1359
1360         /*
1361          * A newly created regular or special file just has one directory
1362          * entry pointing to them, but a directory also the "." entry
1363          * pointing to itself.
1364          */
1365         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1366                                prid, resblks > 0, &ip, &committed);
1367         if (error) {
1368                 if (error == ENOSPC)
1369                         goto out_trans_cancel;
1370                 goto out_trans_abort;
1371         }
1372
1373         /*
1374          * At this point, we've gotten a newly allocated inode.
1375          * It is locked (and joined to the transaction).
1376          */
1377         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1378
1379         /*
1380          * Now we join the directory inode to the transaction.  We do not do it
1381          * earlier because xfs_dir_ialloc might commit the previous transaction
1382          * (and release all the locks).  An error from here on will result in
1383          * the transaction cancel unlocking dp so don't do it explicitly in the
1384          * error path.
1385          */
1386         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1387         unlock_dp_on_error = B_FALSE;
1388
1389         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1390                                         &first_block, &free_list, resblks ?
1391                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1392         if (error) {
1393                 ASSERT(error != ENOSPC);
1394                 goto out_trans_abort;
1395         }
1396         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1397         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1398
1399         if (is_dir) {
1400                 error = xfs_dir_init(tp, ip, dp);
1401                 if (error)
1402                         goto out_bmap_cancel;
1403
1404                 error = xfs_bumplink(tp, dp);
1405                 if (error)
1406                         goto out_bmap_cancel;
1407         }
1408
1409         /*
1410          * If this is a synchronous mount, make sure that the
1411          * create transaction goes to disk before returning to
1412          * the user.
1413          */
1414         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1415                 xfs_trans_set_sync(tp);
1416
1417         /*
1418          * Attach the dquot(s) to the inodes and modify them incore.
1419          * These ids of the inode couldn't have changed since the new
1420          * inode has been locked ever since it was created.
1421          */
1422         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1423
1424         /*
1425          * xfs_trans_commit normally decrements the vnode ref count
1426          * when it unlocks the inode. Since we want to return the
1427          * vnode to the caller, we bump the vnode ref count now.
1428          */
1429         IHOLD(ip);
1430
1431         error = xfs_bmap_finish(&tp, &free_list, &committed);
1432         if (error)
1433                 goto out_abort_rele;
1434
1435         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1436         if (error) {
1437                 IRELE(ip);
1438                 goto out_dqrele;
1439         }
1440
1441         xfs_qm_dqrele(udqp);
1442         xfs_qm_dqrele(gdqp);
1443
1444         *ipp = ip;
1445         return 0;
1446
1447  out_bmap_cancel:
1448         xfs_bmap_cancel(&free_list);
1449  out_trans_abort:
1450         cancel_flags |= XFS_TRANS_ABORT;
1451  out_trans_cancel:
1452         xfs_trans_cancel(tp, cancel_flags);
1453  out_dqrele:
1454         xfs_qm_dqrele(udqp);
1455         xfs_qm_dqrele(gdqp);
1456
1457         if (unlock_dp_on_error)
1458                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1459  std_return:
1460         return error;
1461
1462  out_abort_rele:
1463         /*
1464          * Wait until after the current transaction is aborted to
1465          * release the inode.  This prevents recursive transactions
1466          * and deadlocks from xfs_inactive.
1467          */
1468         xfs_bmap_cancel(&free_list);
1469         cancel_flags |= XFS_TRANS_ABORT;
1470         xfs_trans_cancel(tp, cancel_flags);
1471         IRELE(ip);
1472         unlock_dp_on_error = B_FALSE;
1473         goto out_dqrele;
1474 }
1475
1476 #ifdef DEBUG
1477 int xfs_locked_n;
1478 int xfs_small_retries;
1479 int xfs_middle_retries;
1480 int xfs_lots_retries;
1481 int xfs_lock_delays;
1482 #endif
1483
1484 /*
1485  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1486  * a different value
1487  */
1488 static inline int
1489 xfs_lock_inumorder(int lock_mode, int subclass)
1490 {
1491         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1492                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1493         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1494                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1495
1496         return lock_mode;
1497 }
1498
1499 /*
1500  * The following routine will lock n inodes in exclusive mode.
1501  * We assume the caller calls us with the inodes in i_ino order.
1502  *
1503  * We need to detect deadlock where an inode that we lock
1504  * is in the AIL and we start waiting for another inode that is locked
1505  * by a thread in a long running transaction (such as truncate). This can
1506  * result in deadlock since the long running trans might need to wait
1507  * for the inode we just locked in order to push the tail and free space
1508  * in the log.
1509  */
1510 void
1511 xfs_lock_inodes(
1512         xfs_inode_t     **ips,
1513         int             inodes,
1514         uint            lock_mode)
1515 {
1516         int             attempts = 0, i, j, try_lock;
1517         xfs_log_item_t  *lp;
1518
1519         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1520
1521         try_lock = 0;
1522         i = 0;
1523
1524 again:
1525         for (; i < inodes; i++) {
1526                 ASSERT(ips[i]);
1527
1528                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1529                         continue;
1530
1531                 /*
1532                  * If try_lock is not set yet, make sure all locked inodes
1533                  * are not in the AIL.
1534                  * If any are, set try_lock to be used later.
1535                  */
1536
1537                 if (!try_lock) {
1538                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1539                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1540                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1541                                         try_lock++;
1542                                 }
1543                         }
1544                 }
1545
1546                 /*
1547                  * If any of the previous locks we have locked is in the AIL,
1548                  * we must TRY to get the second and subsequent locks. If
1549                  * we can't get any, we must release all we have
1550                  * and try again.
1551                  */
1552
1553                 if (try_lock) {
1554                         /* try_lock must be 0 if i is 0. */
1555                         /*
1556                          * try_lock means we have an inode locked
1557                          * that is in the AIL.
1558                          */
1559                         ASSERT(i != 0);
1560                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1561                                 attempts++;
1562
1563                                 /*
1564                                  * Unlock all previous guys and try again.
1565                                  * xfs_iunlock will try to push the tail
1566                                  * if the inode is in the AIL.
1567                                  */
1568
1569                                 for(j = i - 1; j >= 0; j--) {
1570
1571                                         /*
1572                                          * Check to see if we've already
1573                                          * unlocked this one.
1574                                          * Not the first one going back,
1575                                          * and the inode ptr is the same.
1576                                          */
1577                                         if ((j != (i - 1)) && ips[j] ==
1578                                                                 ips[j+1])
1579                                                 continue;
1580
1581                                         xfs_iunlock(ips[j], lock_mode);
1582                                 }
1583
1584                                 if ((attempts % 5) == 0) {
1585                                         delay(1); /* Don't just spin the CPU */
1586 #ifdef DEBUG
1587                                         xfs_lock_delays++;
1588 #endif
1589                                 }
1590                                 i = 0;
1591                                 try_lock = 0;
1592                                 goto again;
1593                         }
1594                 } else {
1595                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1596                 }
1597         }
1598
1599 #ifdef DEBUG
1600         if (attempts) {
1601                 if (attempts < 5) xfs_small_retries++;
1602                 else if (attempts < 100) xfs_middle_retries++;
1603                 else xfs_lots_retries++;
1604         } else {
1605                 xfs_locked_n++;
1606         }
1607 #endif
1608 }
1609
1610 /*
1611  * xfs_lock_two_inodes() can only be used to lock one type of lock
1612  * at a time - the iolock or the ilock, but not both at once. If
1613  * we lock both at once, lockdep will report false positives saying
1614  * we have violated locking orders.
1615  */
1616 void
1617 xfs_lock_two_inodes(
1618         xfs_inode_t             *ip0,
1619         xfs_inode_t             *ip1,
1620         uint                    lock_mode)
1621 {
1622         xfs_inode_t             *temp;
1623         int                     attempts = 0;
1624         xfs_log_item_t          *lp;
1625
1626         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1627                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1628         ASSERT(ip0->i_ino != ip1->i_ino);
1629
1630         if (ip0->i_ino > ip1->i_ino) {
1631                 temp = ip0;
1632                 ip0 = ip1;
1633                 ip1 = temp;
1634         }
1635
1636  again:
1637         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1638
1639         /*
1640          * If the first lock we have locked is in the AIL, we must TRY to get
1641          * the second lock. If we can't get it, we must release the first one
1642          * and try again.
1643          */
1644         lp = (xfs_log_item_t *)ip0->i_itemp;
1645         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1646                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1647                         xfs_iunlock(ip0, lock_mode);
1648                         if ((++attempts % 5) == 0)
1649                                 delay(1); /* Don't just spin the CPU */
1650                         goto again;
1651                 }
1652         } else {
1653                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1654         }
1655 }
1656
1657 int
1658 xfs_remove(
1659         xfs_inode_t             *dp,
1660         struct xfs_name         *name,
1661         xfs_inode_t             *ip)
1662 {
1663         xfs_mount_t             *mp = dp->i_mount;
1664         xfs_trans_t             *tp = NULL;
1665         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1666         int                     error = 0;
1667         xfs_bmap_free_t         free_list;
1668         xfs_fsblock_t           first_block;
1669         int                     cancel_flags;
1670         int                     committed;
1671         int                     link_zero;
1672         uint                    resblks;
1673         uint                    log_count;
1674
1675         trace_xfs_remove(dp, name);
1676
1677         if (XFS_FORCED_SHUTDOWN(mp))
1678                 return XFS_ERROR(EIO);
1679
1680         error = xfs_qm_dqattach(dp, 0);
1681         if (error)
1682                 goto std_return;
1683
1684         error = xfs_qm_dqattach(ip, 0);
1685         if (error)
1686                 goto std_return;
1687
1688         if (is_dir) {
1689                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1690                 log_count = XFS_DEFAULT_LOG_COUNT;
1691         } else {
1692                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1693                 log_count = XFS_REMOVE_LOG_COUNT;
1694         }
1695         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1696
1697         /*
1698          * We try to get the real space reservation first,
1699          * allowing for directory btree deletion(s) implying
1700          * possible bmap insert(s).  If we can't get the space
1701          * reservation then we use 0 instead, and avoid the bmap
1702          * btree insert(s) in the directory code by, if the bmap
1703          * insert tries to happen, instead trimming the LAST
1704          * block from the directory.
1705          */
1706         resblks = XFS_REMOVE_SPACE_RES(mp);
1707         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1708                                   XFS_TRANS_PERM_LOG_RES, log_count);
1709         if (error == ENOSPC) {
1710                 resblks = 0;
1711                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1712                                           XFS_TRANS_PERM_LOG_RES, log_count);
1713         }
1714         if (error) {
1715                 ASSERT(error != ENOSPC);
1716                 cancel_flags = 0;
1717                 goto out_trans_cancel;
1718         }
1719
1720         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1721
1722         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1723         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1724
1725         /*
1726          * If we're removing a directory perform some additional validation.
1727          */
1728         if (is_dir) {
1729                 ASSERT(ip->i_d.di_nlink >= 2);
1730                 if (ip->i_d.di_nlink != 2) {
1731                         error = XFS_ERROR(ENOTEMPTY);
1732                         goto out_trans_cancel;
1733                 }
1734                 if (!xfs_dir_isempty(ip)) {
1735                         error = XFS_ERROR(ENOTEMPTY);
1736                         goto out_trans_cancel;
1737                 }
1738         }
1739
1740         xfs_bmap_init(&free_list, &first_block);
1741         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1742                                         &first_block, &free_list, resblks);
1743         if (error) {
1744                 ASSERT(error != ENOENT);
1745                 goto out_bmap_cancel;
1746         }
1747         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1748
1749         if (is_dir) {
1750                 /*
1751                  * Drop the link from ip's "..".
1752                  */
1753                 error = xfs_droplink(tp, dp);
1754                 if (error)
1755                         goto out_bmap_cancel;
1756
1757                 /*
1758                  * Drop the "." link from ip to self.
1759                  */
1760                 error = xfs_droplink(tp, ip);
1761                 if (error)
1762                         goto out_bmap_cancel;
1763         } else {
1764                 /*
1765                  * When removing a non-directory we need to log the parent
1766                  * inode here.  For a directory this is done implicitly
1767                  * by the xfs_droplink call for the ".." entry.
1768                  */
1769                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1770         }
1771
1772         /*
1773          * Drop the link from dp to ip.
1774          */
1775         error = xfs_droplink(tp, ip);
1776         if (error)
1777                 goto out_bmap_cancel;
1778
1779         /*
1780          * Determine if this is the last link while
1781          * we are in the transaction.
1782          */
1783         link_zero = (ip->i_d.di_nlink == 0);
1784
1785         /*
1786          * If this is a synchronous mount, make sure that the
1787          * remove transaction goes to disk before returning to
1788          * the user.
1789          */
1790         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1791                 xfs_trans_set_sync(tp);
1792
1793         error = xfs_bmap_finish(&tp, &free_list, &committed);
1794         if (error)
1795                 goto out_bmap_cancel;
1796
1797         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1798         if (error)
1799                 goto std_return;
1800
1801         /*
1802          * If we are using filestreams, kill the stream association.
1803          * If the file is still open it may get a new one but that
1804          * will get killed on last close in xfs_close() so we don't
1805          * have to worry about that.
1806          */
1807         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1808                 xfs_filestream_deassociate(ip);
1809
1810         return 0;
1811
1812  out_bmap_cancel:
1813         xfs_bmap_cancel(&free_list);
1814         cancel_flags |= XFS_TRANS_ABORT;
1815  out_trans_cancel:
1816         xfs_trans_cancel(tp, cancel_flags);
1817  std_return:
1818         return error;
1819 }
1820
1821 int
1822 xfs_link(
1823         xfs_inode_t             *tdp,
1824         xfs_inode_t             *sip,
1825         struct xfs_name         *target_name)
1826 {
1827         xfs_mount_t             *mp = tdp->i_mount;
1828         xfs_trans_t             *tp;
1829         int                     error;
1830         xfs_bmap_free_t         free_list;
1831         xfs_fsblock_t           first_block;
1832         int                     cancel_flags;
1833         int                     committed;
1834         int                     resblks;
1835
1836         trace_xfs_link(tdp, target_name);
1837
1838         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1839
1840         if (XFS_FORCED_SHUTDOWN(mp))
1841                 return XFS_ERROR(EIO);
1842
1843         error = xfs_qm_dqattach(sip, 0);
1844         if (error)
1845                 goto std_return;
1846
1847         error = xfs_qm_dqattach(tdp, 0);
1848         if (error)
1849                 goto std_return;
1850
1851         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1852         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1853         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1854         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1855                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1856         if (error == ENOSPC) {
1857                 resblks = 0;
1858                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1859                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1860         }
1861         if (error) {
1862                 cancel_flags = 0;
1863                 goto error_return;
1864         }
1865
1866         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1867
1868         xfs_trans_ijoin_ref(tp, sip, XFS_ILOCK_EXCL);
1869         xfs_trans_ijoin_ref(tp, tdp, XFS_ILOCK_EXCL);
1870
1871         /*
1872          * If the source has too many links, we can't make any more to it.
1873          */
1874         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1875                 error = XFS_ERROR(EMLINK);
1876                 goto error_return;
1877         }
1878
1879         /*
1880          * If we are using project inheritance, we only allow hard link
1881          * creation in our tree when the project IDs are the same; else
1882          * the tree quota mechanism could be circumvented.
1883          */
1884         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1885                      (tdp->i_d.di_projid != sip->i_d.di_projid))) {
1886                 error = XFS_ERROR(EXDEV);
1887                 goto error_return;
1888         }
1889
1890         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1891         if (error)
1892                 goto error_return;
1893
1894         xfs_bmap_init(&free_list, &first_block);
1895
1896         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1897                                         &first_block, &free_list, resblks);
1898         if (error)
1899                 goto abort_return;
1900         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1901         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1902
1903         error = xfs_bumplink(tp, sip);
1904         if (error)
1905                 goto abort_return;
1906
1907         /*
1908          * If this is a synchronous mount, make sure that the
1909          * link transaction goes to disk before returning to
1910          * the user.
1911          */
1912         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1913                 xfs_trans_set_sync(tp);
1914         }
1915
1916         error = xfs_bmap_finish (&tp, &free_list, &committed);
1917         if (error) {
1918                 xfs_bmap_cancel(&free_list);
1919                 goto abort_return;
1920         }
1921
1922         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1923
1924  abort_return:
1925         cancel_flags |= XFS_TRANS_ABORT;
1926  error_return:
1927         xfs_trans_cancel(tp, cancel_flags);
1928  std_return:
1929         return error;
1930 }
1931
1932 int
1933 xfs_symlink(
1934         xfs_inode_t             *dp,
1935         struct xfs_name         *link_name,
1936         const char              *target_path,
1937         mode_t                  mode,
1938         xfs_inode_t             **ipp)
1939 {
1940         xfs_mount_t             *mp = dp->i_mount;
1941         xfs_trans_t             *tp;
1942         xfs_inode_t             *ip;
1943         int                     error;
1944         int                     pathlen;
1945         xfs_bmap_free_t         free_list;
1946         xfs_fsblock_t           first_block;
1947         boolean_t               unlock_dp_on_error = B_FALSE;
1948         uint                    cancel_flags;
1949         int                     committed;
1950         xfs_fileoff_t           first_fsb;
1951         xfs_filblks_t           fs_blocks;
1952         int                     nmaps;
1953         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1954         xfs_daddr_t             d;
1955         const char              *cur_chunk;
1956         int                     byte_cnt;
1957         int                     n;
1958         xfs_buf_t               *bp;
1959         xfs_prid_t              prid;
1960         struct xfs_dquot        *udqp, *gdqp;
1961         uint                    resblks;
1962
1963         *ipp = NULL;
1964         error = 0;
1965         ip = NULL;
1966         tp = NULL;
1967
1968         trace_xfs_symlink(dp, link_name);
1969
1970         if (XFS_FORCED_SHUTDOWN(mp))
1971                 return XFS_ERROR(EIO);
1972
1973         /*
1974          * Check component lengths of the target path name.
1975          */
1976         pathlen = strlen(target_path);
1977         if (pathlen >= MAXPATHLEN)      /* total string too long */
1978                 return XFS_ERROR(ENAMETOOLONG);
1979
1980         udqp = gdqp = NULL;
1981         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1982                 prid = dp->i_d.di_projid;
1983         else
1984                 prid = (xfs_prid_t)dfltprid;
1985
1986         /*
1987          * Make sure that we have allocated dquot(s) on disk.
1988          */
1989         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1990                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1991         if (error)
1992                 goto std_return;
1993
1994         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1995         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1996         /*
1997          * The symlink will fit into the inode data fork?
1998          * There can't be any attributes so we get the whole variable part.
1999          */
2000         if (pathlen <= XFS_LITINO(mp))
2001                 fs_blocks = 0;
2002         else
2003                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2004         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2005         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2006                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2007         if (error == ENOSPC && fs_blocks == 0) {
2008                 resblks = 0;
2009                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2010                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2011         }
2012         if (error) {
2013                 cancel_flags = 0;
2014                 goto error_return;
2015         }
2016
2017         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2018         unlock_dp_on_error = B_TRUE;
2019
2020         /*
2021          * Check whether the directory allows new symlinks or not.
2022          */
2023         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2024                 error = XFS_ERROR(EPERM);
2025                 goto error_return;
2026         }
2027
2028         /*
2029          * Reserve disk quota : blocks and inode.
2030          */
2031         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
2032         if (error)
2033                 goto error_return;
2034
2035         /*
2036          * Check for ability to enter directory entry, if no space reserved.
2037          */
2038         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2039         if (error)
2040                 goto error_return;
2041         /*
2042          * Initialize the bmap freelist prior to calling either
2043          * bmapi or the directory create code.
2044          */
2045         xfs_bmap_init(&free_list, &first_block);
2046
2047         /*
2048          * Allocate an inode for the symlink.
2049          */
2050         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
2051                                prid, resblks > 0, &ip, NULL);
2052         if (error) {
2053                 if (error == ENOSPC)
2054                         goto error_return;
2055                 goto error1;
2056         }
2057
2058         /*
2059          * An error after we've joined dp to the transaction will result in the
2060          * transaction cancel unlocking dp so don't do it explicitly in the
2061          * error path.
2062          */
2063         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
2064         unlock_dp_on_error = B_FALSE;
2065
2066         /*
2067          * Also attach the dquot(s) to it, if applicable.
2068          */
2069         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
2070
2071         if (resblks)
2072                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2073         /*
2074          * If the symlink will fit into the inode, write it inline.
2075          */
2076         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2077                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2078                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2079                 ip->i_d.di_size = pathlen;
2080
2081                 /*
2082                  * The inode was initially created in extent format.
2083                  */
2084                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2085                 ip->i_df.if_flags |= XFS_IFINLINE;
2086
2087                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2088                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2089
2090         } else {
2091                 first_fsb = 0;
2092                 nmaps = SYMLINK_MAPS;
2093
2094                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2095                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2096                                   &first_block, resblks, mval, &nmaps,
2097                                   &free_list);
2098                 if (error) {
2099                         goto error1;
2100                 }
2101
2102                 if (resblks)
2103                         resblks -= fs_blocks;
2104                 ip->i_d.di_size = pathlen;
2105                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2106
2107                 cur_chunk = target_path;
2108                 for (n = 0; n < nmaps; n++) {
2109                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2110                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2111                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2112                                                BTOBB(byte_cnt), 0);
2113                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2114                         if (pathlen < byte_cnt) {
2115                                 byte_cnt = pathlen;
2116                         }
2117                         pathlen -= byte_cnt;
2118
2119                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2120                         cur_chunk += byte_cnt;
2121
2122                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2123                 }
2124         }
2125
2126         /*
2127          * Create the directory entry for the symlink.
2128          */
2129         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2130                                         &first_block, &free_list, resblks);
2131         if (error)
2132                 goto error1;
2133         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2134         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2135
2136         /*
2137          * If this is a synchronous mount, make sure that the
2138          * symlink transaction goes to disk before returning to
2139          * the user.
2140          */
2141         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2142                 xfs_trans_set_sync(tp);
2143         }
2144
2145         /*
2146          * xfs_trans_commit normally decrements the vnode ref count
2147          * when it unlocks the inode. Since we want to return the
2148          * vnode to the caller, we bump the vnode ref count now.
2149          */
2150         IHOLD(ip);
2151
2152         error = xfs_bmap_finish(&tp, &free_list, &committed);
2153         if (error) {
2154                 goto error2;
2155         }
2156         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2157         xfs_qm_dqrele(udqp);
2158         xfs_qm_dqrele(gdqp);
2159
2160         *ipp = ip;
2161         return 0;
2162
2163  error2:
2164         IRELE(ip);
2165  error1:
2166         xfs_bmap_cancel(&free_list);
2167         cancel_flags |= XFS_TRANS_ABORT;
2168  error_return:
2169         xfs_trans_cancel(tp, cancel_flags);
2170         xfs_qm_dqrele(udqp);
2171         xfs_qm_dqrele(gdqp);
2172
2173         if (unlock_dp_on_error)
2174                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2175  std_return:
2176         return error;
2177 }
2178
2179 int
2180 xfs_set_dmattrs(
2181         xfs_inode_t     *ip,
2182         u_int           evmask,
2183         u_int16_t       state)
2184 {
2185         xfs_mount_t     *mp = ip->i_mount;
2186         xfs_trans_t     *tp;
2187         int             error;
2188
2189         if (!capable(CAP_SYS_ADMIN))
2190                 return XFS_ERROR(EPERM);
2191
2192         if (XFS_FORCED_SHUTDOWN(mp))
2193                 return XFS_ERROR(EIO);
2194
2195         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2196         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2197         if (error) {
2198                 xfs_trans_cancel(tp, 0);
2199                 return error;
2200         }
2201         xfs_ilock(ip, XFS_ILOCK_EXCL);
2202         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
2203
2204         ip->i_d.di_dmevmask = evmask;
2205         ip->i_d.di_dmstate  = state;
2206
2207         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2208         error = xfs_trans_commit(tp, 0);
2209
2210         return error;
2211 }
2212
2213 /*
2214  * xfs_alloc_file_space()
2215  *      This routine allocates disk space for the given file.
2216  *
2217  *      If alloc_type == 0, this request is for an ALLOCSP type
2218  *      request which will change the file size.  In this case, no
2219  *      DMAPI event will be generated by the call.  A TRUNCATE event
2220  *      will be generated later by xfs_setattr.
2221  *
2222  *      If alloc_type != 0, this request is for a RESVSP type
2223  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2224  *      lower block boundary byte address is less than the file's
2225  *      length.
2226  *
2227  * RETURNS:
2228  *       0 on success
2229  *      errno on error
2230  *
2231  */
2232 STATIC int
2233 xfs_alloc_file_space(
2234         xfs_inode_t             *ip,
2235         xfs_off_t               offset,
2236         xfs_off_t               len,
2237         int                     alloc_type,
2238         int                     attr_flags)
2239 {
2240         xfs_mount_t             *mp = ip->i_mount;
2241         xfs_off_t               count;
2242         xfs_filblks_t           allocated_fsb;
2243         xfs_filblks_t           allocatesize_fsb;
2244         xfs_extlen_t            extsz, temp;
2245         xfs_fileoff_t           startoffset_fsb;
2246         xfs_fsblock_t           firstfsb;
2247         int                     nimaps;
2248         int                     bmapi_flag;
2249         int                     quota_flag;
2250         int                     rt;
2251         xfs_trans_t             *tp;
2252         xfs_bmbt_irec_t         imaps[1], *imapp;
2253         xfs_bmap_free_t         free_list;
2254         uint                    qblocks, resblks, resrtextents;
2255         int                     committed;
2256         int                     error;
2257
2258         trace_xfs_alloc_file_space(ip);
2259
2260         if (XFS_FORCED_SHUTDOWN(mp))
2261                 return XFS_ERROR(EIO);
2262
2263         error = xfs_qm_dqattach(ip, 0);
2264         if (error)
2265                 return error;
2266
2267         if (len <= 0)
2268                 return XFS_ERROR(EINVAL);
2269
2270         rt = XFS_IS_REALTIME_INODE(ip);
2271         extsz = xfs_get_extsz_hint(ip);
2272
2273         count = len;
2274         imapp = &imaps[0];
2275         nimaps = 1;
2276         bmapi_flag = XFS_BMAPI_WRITE | alloc_type;
2277         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
2278         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
2279
2280         /*
2281          * Allocate file space until done or until there is an error
2282          */
2283         while (allocatesize_fsb && !error) {
2284                 xfs_fileoff_t   s, e;
2285
2286                 /*
2287                  * Determine space reservations for data/realtime.
2288                  */
2289                 if (unlikely(extsz)) {
2290                         s = startoffset_fsb;
2291                         do_div(s, extsz);
2292                         s *= extsz;
2293                         e = startoffset_fsb + allocatesize_fsb;
2294                         if ((temp = do_mod(startoffset_fsb, extsz)))
2295                                 e += temp;
2296                         if ((temp = do_mod(e, extsz)))
2297                                 e += extsz - temp;
2298                 } else {
2299                         s = 0;
2300                         e = allocatesize_fsb;
2301                 }
2302
2303                 /*
2304                  * The transaction reservation is limited to a 32-bit block
2305                  * count, hence we need to limit the number of blocks we are
2306                  * trying to reserve to avoid an overflow. We can't allocate
2307                  * more than @nimaps extents, and an extent is limited on disk
2308                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
2309                  */
2310                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
2311                 if (unlikely(rt)) {
2312                         resrtextents = qblocks = resblks;
2313                         resrtextents /= mp->m_sb.sb_rextsize;
2314                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2315                         quota_flag = XFS_QMOPT_RES_RTBLKS;
2316                 } else {
2317                         resrtextents = 0;
2318                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
2319                         quota_flag = XFS_QMOPT_RES_REGBLKS;
2320                 }
2321
2322                 /*
2323                  * Allocate and setup the transaction.
2324                  */
2325                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2326                 error = xfs_trans_reserve(tp, resblks,
2327                                           XFS_WRITE_LOG_RES(mp), resrtextents,
2328                                           XFS_TRANS_PERM_LOG_RES,
2329                                           XFS_WRITE_LOG_COUNT);
2330                 /*
2331                  * Check for running out of space
2332                  */
2333                 if (error) {
2334                         /*
2335                          * Free the transaction structure.
2336                          */
2337                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2338                         xfs_trans_cancel(tp, 0);
2339                         break;
2340                 }
2341                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2342                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
2343                                                       0, quota_flag);
2344                 if (error)
2345                         goto error1;
2346
2347                 xfs_trans_ijoin(tp, ip);
2348
2349                 /*
2350                  * Issue the xfs_bmapi() call to allocate the blocks
2351                  */
2352                 xfs_bmap_init(&free_list, &firstfsb);
2353                 error = xfs_bmapi(tp, ip, startoffset_fsb,
2354                                   allocatesize_fsb, bmapi_flag,
2355                                   &firstfsb, 0, imapp, &nimaps,
2356                                   &free_list);
2357                 if (error) {
2358                         goto error0;
2359                 }
2360
2361                 /*
2362                  * Complete the transaction
2363                  */
2364                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2365                 if (error) {
2366                         goto error0;
2367                 }
2368
2369                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2370                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2371                 if (error) {
2372                         break;
2373                 }
2374
2375                 allocated_fsb = imapp->br_blockcount;
2376
2377                 if (nimaps == 0) {
2378                         error = XFS_ERROR(ENOSPC);
2379                         break;
2380                 }
2381
2382                 startoffset_fsb += allocated_fsb;
2383                 allocatesize_fsb -= allocated_fsb;
2384         }
2385
2386         return error;
2387
2388 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
2389         xfs_bmap_cancel(&free_list);
2390         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
2391
2392 error1: /* Just cancel transaction */
2393         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2394         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2395         return error;
2396 }
2397
2398 /*
2399  * Zero file bytes between startoff and endoff inclusive.
2400  * The iolock is held exclusive and no blocks are buffered.
2401  *
2402  * This function is used by xfs_free_file_space() to zero
2403  * partial blocks when the range to free is not block aligned.
2404  * When unreserving space with boundaries that are not block
2405  * aligned we round up the start and round down the end
2406  * boundaries and then use this function to zero the parts of
2407  * the blocks that got dropped during the rounding.
2408  */
2409 STATIC int
2410 xfs_zero_remaining_bytes(
2411         xfs_inode_t             *ip,
2412         xfs_off_t               startoff,
2413         xfs_off_t               endoff)
2414 {
2415         xfs_bmbt_irec_t         imap;
2416         xfs_fileoff_t           offset_fsb;
2417         xfs_off_t               lastoffset;
2418         xfs_off_t               offset;
2419         xfs_buf_t               *bp;
2420         xfs_mount_t             *mp = ip->i_mount;
2421         int                     nimap;
2422         int                     error = 0;
2423
2424         /*
2425          * Avoid doing I/O beyond eof - it's not necessary
2426          * since nothing can read beyond eof.  The space will
2427          * be zeroed when the file is extended anyway.
2428          */
2429         if (startoff >= ip->i_size)
2430                 return 0;
2431
2432         if (endoff > ip->i_size)
2433                 endoff = ip->i_size;
2434
2435         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
2436                                         mp->m_rtdev_targp : mp->m_ddev_targp,
2437                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
2438         if (!bp)
2439                 return XFS_ERROR(ENOMEM);
2440
2441         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
2442                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
2443                 nimap = 1;
2444                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
2445                         NULL, 0, &imap, &nimap, NULL);
2446                 if (error || nimap < 1)
2447                         break;
2448                 ASSERT(imap.br_blockcount >= 1);
2449                 ASSERT(imap.br_startoff == offset_fsb);
2450                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
2451                 if (lastoffset > endoff)
2452                         lastoffset = endoff;
2453                 if (imap.br_startblock == HOLESTARTBLOCK)
2454                         continue;
2455                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2456                 if (imap.br_state == XFS_EXT_UNWRITTEN)
2457                         continue;
2458                 XFS_BUF_UNDONE(bp);
2459                 XFS_BUF_UNWRITE(bp);
2460                 XFS_BUF_READ(bp);
2461                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
2462                 xfsbdstrat(mp, bp);
2463                 error = xfs_buf_iowait(bp);
2464                 if (error) {
2465                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
2466                                           mp, bp, XFS_BUF_ADDR(bp));
2467                         break;
2468                 }
2469                 memset(XFS_BUF_PTR(bp) +
2470                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2471                       0, lastoffset - offset + 1);
2472                 XFS_BUF_UNDONE(bp);
2473                 XFS_BUF_UNREAD(bp);
2474                 XFS_BUF_WRITE(bp);
2475                 xfsbdstrat(mp, bp);
2476                 error = xfs_buf_iowait(bp);
2477                 if (error) {
2478                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
2479                                           mp, bp, XFS_BUF_ADDR(bp));
2480                         break;
2481                 }
2482         }
2483         xfs_buf_free(bp);
2484         return error;
2485 }
2486
2487 /*
2488  * xfs_free_file_space()
2489  *      This routine frees disk space for the given file.
2490  *
2491  *      This routine is only called by xfs_change_file_space
2492  *      for an UNRESVSP type call.
2493  *
2494  * RETURNS:
2495  *       0 on success
2496  *      errno on error
2497  *
2498  */
2499 STATIC int
2500 xfs_free_file_space(
2501         xfs_inode_t             *ip,
2502         xfs_off_t               offset,
2503         xfs_off_t               len,
2504         int                     attr_flags)
2505 {
2506         int                     committed;
2507         int                     done;
2508         xfs_fileoff_t           endoffset_fsb;
2509         int                     error;
2510         xfs_fsblock_t           firstfsb;
2511         xfs_bmap_free_t         free_list;
2512         xfs_bmbt_irec_t         imap;
2513         xfs_off_t               ioffset;
2514         xfs_extlen_t            mod=0;
2515         xfs_mount_t             *mp;
2516         int                     nimap;
2517         uint                    resblks;
2518         uint                    rounding;
2519         int                     rt;
2520         xfs_fileoff_t           startoffset_fsb;
2521         xfs_trans_t             *tp;
2522         int                     need_iolock = 1;
2523
2524         mp = ip->i_mount;
2525
2526         trace_xfs_free_file_space(ip);
2527
2528         error = xfs_qm_dqattach(ip, 0);
2529         if (error)
2530                 return error;
2531
2532         error = 0;
2533         if (len <= 0)   /* if nothing being freed */
2534                 return error;
2535         rt = XFS_IS_REALTIME_INODE(ip);
2536         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2537         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2538
2539         if (attr_flags & XFS_ATTR_NOLOCK)
2540                 need_iolock = 0;
2541         if (need_iolock) {
2542                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2543                 /* wait for the completion of any pending DIOs */
2544                 xfs_ioend_wait(ip);
2545         }
2546
2547         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2548         ioffset = offset & ~(rounding - 1);
2549
2550         if (VN_CACHED(VFS_I(ip)) != 0) {
2551                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2552                 if (error)
2553                         goto out_unlock_iolock;
2554         }
2555
2556         /*
2557          * Need to zero the stuff we're not freeing, on disk.
2558          * If it's a realtime file & can't use unwritten extents then we
2559          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2560          * will take care of it for us.
2561          */
2562         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2563                 nimap = 1;
2564                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
2565                         1, 0, NULL, 0, &imap, &nimap, NULL);
2566                 if (error)
2567                         goto out_unlock_iolock;
2568                 ASSERT(nimap == 0 || nimap == 1);
2569                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2570                         xfs_daddr_t     block;
2571
2572                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2573                         block = imap.br_startblock;
2574                         mod = do_div(block, mp->m_sb.sb_rextsize);
2575                         if (mod)
2576                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2577                 }
2578                 nimap = 1;
2579                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
2580                         1, 0, NULL, 0, &imap, &nimap, NULL);
2581                 if (error)
2582                         goto out_unlock_iolock;
2583                 ASSERT(nimap == 0 || nimap == 1);
2584                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2585                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2586                         mod++;
2587                         if (mod && (mod != mp->m_sb.sb_rextsize))
2588                                 endoffset_fsb -= mod;
2589                 }
2590         }
2591         if ((done = (endoffset_fsb <= startoffset_fsb)))
2592                 /*
2593                  * One contiguous piece to clear
2594                  */
2595                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2596         else {
2597                 /*
2598                  * Some full blocks, possibly two pieces to clear
2599                  */
2600                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2601                         error = xfs_zero_remaining_bytes(ip, offset,
2602                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2603                 if (!error &&
2604                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2605                         error = xfs_zero_remaining_bytes(ip,
2606                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2607                                 offset + len - 1);
2608         }
2609
2610         /*
2611          * free file space until done or until there is an error
2612          */
2613         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2614         while (!error && !done) {
2615
2616                 /*
2617                  * allocate and setup the transaction. Allow this
2618                  * transaction to dip into the reserve blocks to ensure
2619                  * the freeing of the space succeeds at ENOSPC.
2620                  */
2621                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2622                 tp->t_flags |= XFS_TRANS_RESERVE;
2623                 error = xfs_trans_reserve(tp,
2624                                           resblks,
2625                                           XFS_WRITE_LOG_RES(mp),
2626                                           0,
2627                                           XFS_TRANS_PERM_LOG_RES,
2628                                           XFS_WRITE_LOG_COUNT);
2629
2630                 /*
2631                  * check for running out of space
2632                  */
2633                 if (error) {
2634                         /*
2635                          * Free the transaction structure.
2636                          */
2637                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2638                         xfs_trans_cancel(tp, 0);
2639                         break;
2640                 }
2641                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2642                 error = xfs_trans_reserve_quota(tp, mp,
2643                                 ip->i_udquot, ip->i_gdquot,
2644                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2645                 if (error)
2646                         goto error1;
2647
2648                 xfs_trans_ijoin(tp, ip);
2649
2650                 /*
2651                  * issue the bunmapi() call to free the blocks
2652                  */
2653                 xfs_bmap_init(&free_list, &firstfsb);
2654                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2655                                   endoffset_fsb - startoffset_fsb,
2656                                   0, 2, &firstfsb, &free_list, &done);
2657                 if (error) {
2658                         goto error0;
2659                 }
2660
2661                 /*
2662                  * complete the transaction
2663                  */
2664                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2665                 if (error) {
2666                         goto error0;
2667                 }
2668
2669                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2670                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2671         }
2672
2673  out_unlock_iolock:
2674         if (need_iolock)
2675                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2676         return error;
2677
2678  error0:
2679         xfs_bmap_cancel(&free_list);
2680  error1:
2681         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2682         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2683                     XFS_ILOCK_EXCL);
2684         return error;
2685 }
2686
2687 /*
2688  * xfs_change_file_space()
2689  *      This routine allocates or frees disk space for the given file.
2690  *      The user specified parameters are checked for alignment and size
2691  *      limitations.
2692  *
2693  * RETURNS:
2694  *       0 on success
2695  *      errno on error
2696  *
2697  */
2698 int
2699 xfs_change_file_space(
2700         xfs_inode_t     *ip,
2701         int             cmd,
2702         xfs_flock64_t   *bf,
2703         xfs_off_t       offset,
2704         int             attr_flags)
2705 {
2706         xfs_mount_t     *mp = ip->i_mount;
2707         int             clrprealloc;
2708         int             error;
2709         xfs_fsize_t     fsize;
2710         int             setprealloc;
2711         xfs_off_t       startoffset;
2712         xfs_off_t       llen;
2713         xfs_trans_t     *tp;
2714         struct iattr    iattr;
2715         int             prealloc_type;
2716
2717         if (!S_ISREG(ip->i_d.di_mode))
2718                 return XFS_ERROR(EINVAL);
2719
2720         switch (bf->l_whence) {
2721         case 0: /*SEEK_SET*/
2722                 break;
2723         case 1: /*SEEK_CUR*/
2724                 bf->l_start += offset;
2725                 break;
2726         case 2: /*SEEK_END*/
2727                 bf->l_start += ip->i_size;
2728                 break;
2729         default:
2730                 return XFS_ERROR(EINVAL);
2731         }
2732
2733         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2734
2735         if (   (bf->l_start < 0)
2736             || (bf->l_start > XFS_MAXIOFFSET(mp))
2737             || (bf->l_start + llen < 0)
2738             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2739                 return XFS_ERROR(EINVAL);
2740
2741         bf->l_whence = 0;
2742
2743         startoffset = bf->l_start;
2744         fsize = ip->i_size;
2745
2746         /*
2747          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2748          * file space.
2749          * These calls do NOT zero the data space allocated to the file,
2750          * nor do they change the file size.
2751          *
2752          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2753          * space.
2754          * These calls cause the new file data to be zeroed and the file
2755          * size to be changed.
2756          */
2757         setprealloc = clrprealloc = 0;
2758         prealloc_type = XFS_BMAPI_PREALLOC;
2759
2760         switch (cmd) {
2761         case XFS_IOC_ZERO_RANGE:
2762                 prealloc_type |= XFS_BMAPI_CONVERT;
2763                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2764                 /* FALLTHRU */
2765         case XFS_IOC_RESVSP:
2766         case XFS_IOC_RESVSP64:
2767                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2768                                                 prealloc_type, attr_flags);
2769                 if (error)
2770                         return error;
2771                 setprealloc = 1;
2772                 break;
2773
2774         case XFS_IOC_UNRESVSP:
2775         case XFS_IOC_UNRESVSP64:
2776                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2777                                                                 attr_flags)))
2778                         return error;
2779                 break;
2780
2781         case XFS_IOC_ALLOCSP:
2782         case XFS_IOC_ALLOCSP64:
2783         case XFS_IOC_FREESP:
2784         case XFS_IOC_FREESP64:
2785                 if (startoffset > fsize) {
2786                         error = xfs_alloc_file_space(ip, fsize,
2787                                         startoffset - fsize, 0, attr_flags);
2788                         if (error)
2789                                 break;
2790                 }
2791
2792                 iattr.ia_valid = ATTR_SIZE;
2793                 iattr.ia_size = startoffset;
2794
2795                 error = xfs_setattr(ip, &iattr, attr_flags);
2796
2797                 if (error)
2798                         return error;
2799
2800                 clrprealloc = 1;
2801                 break;
2802
2803         default:
2804                 ASSERT(0);
2805                 return XFS_ERROR(EINVAL);
2806         }
2807
2808         /*
2809          * update the inode timestamp, mode, and prealloc flag bits
2810          */
2811         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2812
2813         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2814                                       0, 0, 0))) {
2815                 /* ASSERT(0); */
2816                 xfs_trans_cancel(tp, 0);
2817                 return error;
2818         }
2819
2820         xfs_ilock(ip, XFS_ILOCK_EXCL);
2821
2822         xfs_trans_ijoin(tp, ip);
2823
2824         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2825                 ip->i_d.di_mode &= ~S_ISUID;
2826
2827                 /*
2828                  * Note that we don't have to worry about mandatory
2829                  * file locking being disabled here because we only
2830                  * clear the S_ISGID bit if the Group execute bit is
2831                  * on, but if it was on then mandatory locking wouldn't
2832                  * have been enabled.
2833                  */
2834                 if (ip->i_d.di_mode & S_IXGRP)
2835                         ip->i_d.di_mode &= ~S_ISGID;
2836
2837                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2838         }
2839         if (setprealloc)
2840                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2841         else if (clrprealloc)
2842                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2843
2844         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2845         xfs_trans_set_sync(tp);
2846
2847         error = xfs_trans_commit(tp, 0);
2848
2849         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2850
2851         return error;
2852 }