[XFS] Check return value of xfs_buf_get_noaddr()
[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_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_da_btree.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_ialloc_btree.h"
35 #include "xfs_dir2_sf.h"
36 #include "xfs_attr_sf.h"
37 #include "xfs_dinode.h"
38 #include "xfs_inode.h"
39 #include "xfs_inode_item.h"
40 #include "xfs_itable.h"
41 #include "xfs_btree.h"
42 #include "xfs_ialloc.h"
43 #include "xfs_alloc.h"
44 #include "xfs_bmap.h"
45 #include "xfs_attr.h"
46 #include "xfs_rw.h"
47 #include "xfs_error.h"
48 #include "xfs_quota.h"
49 #include "xfs_utils.h"
50 #include "xfs_rtalloc.h"
51 #include "xfs_trans_space.h"
52 #include "xfs_log_priv.h"
53 #include "xfs_filestream.h"
54 #include "xfs_vnodeops.h"
55
56 int
57 xfs_setattr(
58         struct xfs_inode        *ip,
59         struct iattr            *iattr,
60         int                     flags)
61 {
62         xfs_mount_t             *mp = ip->i_mount;
63         struct inode            *inode = VFS_I(ip);
64         int                     mask = iattr->ia_valid;
65         xfs_trans_t             *tp;
66         int                     code;
67         uint                    lock_flags;
68         uint                    commit_flags=0;
69         uid_t                   uid=0, iuid=0;
70         gid_t                   gid=0, igid=0;
71         int                     timeflags = 0;
72         struct xfs_dquot        *udqp, *gdqp, *olddquot1, *olddquot2;
73         int                     file_owner;
74         int                     need_iolock = 1;
75
76         xfs_itrace_entry(ip);
77
78         if (mp->m_flags & XFS_MOUNT_RDONLY)
79                 return XFS_ERROR(EROFS);
80
81         if (XFS_FORCED_SHUTDOWN(mp))
82                 return XFS_ERROR(EIO);
83
84         olddquot1 = olddquot2 = NULL;
85         udqp = gdqp = NULL;
86
87         /*
88          * If disk quotas is on, we make sure that the dquots do exist on disk,
89          * before we start any other transactions. Trying to do this later
90          * is messy. We don't care to take a readlock to look at the ids
91          * in inode here, because we can't hold it across the trans_reserve.
92          * If the IDs do change before we take the ilock, we're covered
93          * because the i_*dquot fields will get updated anyway.
94          */
95         if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
96                 uint    qflags = 0;
97
98                 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
99                         uid = iattr->ia_uid;
100                         qflags |= XFS_QMOPT_UQUOTA;
101                 } else {
102                         uid = ip->i_d.di_uid;
103                 }
104                 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
105                         gid = iattr->ia_gid;
106                         qflags |= XFS_QMOPT_GQUOTA;
107                 }  else {
108                         gid = ip->i_d.di_gid;
109                 }
110
111                 /*
112                  * We take a reference when we initialize udqp and gdqp,
113                  * so it is important that we never blindly double trip on
114                  * the same variable. See xfs_create() for an example.
115                  */
116                 ASSERT(udqp == NULL);
117                 ASSERT(gdqp == NULL);
118                 code = XFS_QM_DQVOPALLOC(mp, ip, uid, gid, ip->i_d.di_projid,
119                                          qflags, &udqp, &gdqp);
120                 if (code)
121                         return code;
122         }
123
124         /*
125          * For the other attributes, we acquire the inode lock and
126          * first do an error checking pass.
127          */
128         tp = NULL;
129         lock_flags = XFS_ILOCK_EXCL;
130         if (flags & XFS_ATTR_NOLOCK)
131                 need_iolock = 0;
132         if (!(mask & ATTR_SIZE)) {
133                 if ((mask != (ATTR_CTIME|ATTR_ATIME|ATTR_MTIME)) ||
134                     (mp->m_flags & XFS_MOUNT_WSYNC)) {
135                         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
136                         commit_flags = 0;
137                         if ((code = xfs_trans_reserve(tp, 0,
138                                                      XFS_ICHANGE_LOG_RES(mp), 0,
139                                                      0, 0))) {
140                                 lock_flags = 0;
141                                 goto error_return;
142                         }
143                 }
144         } else {
145                 if (DM_EVENT_ENABLED(ip, DM_EVENT_TRUNCATE) &&
146                     !(flags & XFS_ATTR_DMI)) {
147                         int dmflags = AT_DELAY_FLAG(flags) | DM_SEM_FLAG_WR;
148                         code = XFS_SEND_DATA(mp, DM_EVENT_TRUNCATE, ip,
149                                 iattr->ia_size, 0, dmflags, NULL);
150                         if (code) {
151                                 lock_flags = 0;
152                                 goto error_return;
153                         }
154                 }
155                 if (need_iolock)
156                         lock_flags |= XFS_IOLOCK_EXCL;
157         }
158
159         xfs_ilock(ip, lock_flags);
160
161         /* boolean: are we the file owner? */
162         file_owner = (current_fsuid() == ip->i_d.di_uid);
163
164         /*
165          * Change various properties of a file.
166          * Only the owner or users with CAP_FOWNER
167          * capability may do these things.
168          */
169         if (mask & (ATTR_MODE|ATTR_UID|ATTR_GID)) {
170                 /*
171                  * CAP_FOWNER overrides the following restrictions:
172                  *
173                  * The user ID of the calling process must be equal
174                  * to the file owner ID, except in cases where the
175                  * CAP_FSETID capability is applicable.
176                  */
177                 if (!file_owner && !capable(CAP_FOWNER)) {
178                         code = XFS_ERROR(EPERM);
179                         goto error_return;
180                 }
181
182                 /*
183                  * CAP_FSETID overrides the following restrictions:
184                  *
185                  * The effective user ID of the calling process shall match
186                  * the file owner when setting the set-user-ID and
187                  * set-group-ID bits on that file.
188                  *
189                  * The effective group ID or one of the supplementary group
190                  * IDs of the calling process shall match the group owner of
191                  * the file when setting the set-group-ID bit on that file
192                  */
193                 if (mask & ATTR_MODE) {
194                         mode_t m = 0;
195
196                         if ((iattr->ia_mode & S_ISUID) && !file_owner)
197                                 m |= S_ISUID;
198                         if ((iattr->ia_mode & S_ISGID) &&
199                             !in_group_p((gid_t)ip->i_d.di_gid))
200                                 m |= S_ISGID;
201 #if 0
202                         /* Linux allows this, Irix doesn't. */
203                         if ((iattr->ia_mode & S_ISVTX) && !S_ISDIR(ip->i_d.di_mode))
204                                 m |= S_ISVTX;
205 #endif
206                         if (m && !capable(CAP_FSETID))
207                                 iattr->ia_mode &= ~m;
208                 }
209         }
210
211         /*
212          * Change file ownership.  Must be the owner or privileged.
213          */
214         if (mask & (ATTR_UID|ATTR_GID)) {
215                 /*
216                  * These IDs could have changed since we last looked at them.
217                  * But, we're assured that if the ownership did change
218                  * while we didn't have the inode locked, inode's dquot(s)
219                  * would have changed also.
220                  */
221                 iuid = ip->i_d.di_uid;
222                 igid = ip->i_d.di_gid;
223                 gid = (mask & ATTR_GID) ? iattr->ia_gid : igid;
224                 uid = (mask & ATTR_UID) ? iattr->ia_uid : iuid;
225
226                 /*
227                  * CAP_CHOWN overrides the following restrictions:
228                  *
229                  * If _POSIX_CHOWN_RESTRICTED is defined, this capability
230                  * shall override the restriction that a process cannot
231                  * change the user ID of a file it owns and the restriction
232                  * that the group ID supplied to the chown() function
233                  * shall be equal to either the group ID or one of the
234                  * supplementary group IDs of the calling process.
235                  */
236                 if ((iuid != uid ||
237                      (igid != gid && !in_group_p((gid_t)gid))) &&
238                     !capable(CAP_CHOWN)) {
239                         code = XFS_ERROR(EPERM);
240                         goto error_return;
241                 }
242                 /*
243                  * Do a quota reservation only if uid/gid is actually
244                  * going to change.
245                  */
246                 if ((XFS_IS_UQUOTA_ON(mp) && iuid != uid) ||
247                     (XFS_IS_GQUOTA_ON(mp) && igid != gid)) {
248                         ASSERT(tp);
249                         code = XFS_QM_DQVOPCHOWNRESV(mp, tp, ip, udqp, gdqp,
250                                                 capable(CAP_FOWNER) ?
251                                                 XFS_QMOPT_FORCE_RES : 0);
252                         if (code)       /* out of quota */
253                                 goto error_return;
254                 }
255         }
256
257         /*
258          * Truncate file.  Must have write permission and not be a directory.
259          */
260         if (mask & ATTR_SIZE) {
261                 /* Short circuit the truncate case for zero length files */
262                 if (iattr->ia_size == 0 &&
263                     ip->i_size == 0 && ip->i_d.di_nextents == 0) {
264                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
265                         lock_flags &= ~XFS_ILOCK_EXCL;
266                         if (mask & ATTR_CTIME)
267                                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
268                         code = 0;
269                         goto error_return;
270                 }
271
272                 if (S_ISDIR(ip->i_d.di_mode)) {
273                         code = XFS_ERROR(EISDIR);
274                         goto error_return;
275                 } else if (!S_ISREG(ip->i_d.di_mode)) {
276                         code = XFS_ERROR(EINVAL);
277                         goto error_return;
278                 }
279                 /*
280                  * Make sure that the dquots are attached to the inode.
281                  */
282                 if ((code = XFS_QM_DQATTACH(mp, ip, XFS_QMOPT_ILOCKED)))
283                         goto error_return;
284         }
285
286         /*
287          * Change file access or modified times.
288          */
289         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
290                 if (!file_owner) {
291                         if ((mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)) &&
292                             !capable(CAP_FOWNER)) {
293                                 code = XFS_ERROR(EPERM);
294                                 goto error_return;
295                         }
296                 }
297         }
298
299         /*
300          * Now we can make the changes.  Before we join the inode
301          * to the transaction, if ATTR_SIZE is set then take care of
302          * the part of the truncation that must be done without the
303          * inode lock.  This needs to be done before joining the inode
304          * to the transaction, because the inode cannot be unlocked
305          * once it is a part of the transaction.
306          */
307         if (mask & ATTR_SIZE) {
308                 code = 0;
309                 if (iattr->ia_size > ip->i_size) {
310                         /*
311                          * Do the first part of growing a file: zero any data
312                          * in the last block that is beyond the old EOF.  We
313                          * need to do this before the inode is joined to the
314                          * transaction to modify the i_size.
315                          */
316                         code = xfs_zero_eof(ip, iattr->ia_size, ip->i_size);
317                 }
318                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
319
320                 /*
321                  * We are going to log the inode size change in this
322                  * transaction so any previous writes that are beyond the on
323                  * disk EOF and the new EOF that have not been written out need
324                  * to be written here. If we do not write the data out, we
325                  * expose ourselves to the null files problem.
326                  *
327                  * Only flush from the on disk size to the smaller of the in
328                  * memory file size or the new size as that's the range we
329                  * really care about here and prevents waiting for other data
330                  * not within the range we care about here.
331                  */
332                 if (!code &&
333                     ip->i_size != ip->i_d.di_size &&
334                     iattr->ia_size > ip->i_d.di_size) {
335                         code = xfs_flush_pages(ip,
336                                         ip->i_d.di_size, iattr->ia_size,
337                                         XFS_B_ASYNC, FI_NONE);
338                 }
339
340                 /* wait for all I/O to complete */
341                 xfs_ioend_wait(ip);
342
343                 if (!code)
344                         code = xfs_itruncate_data(ip, iattr->ia_size);
345                 if (code) {
346                         ASSERT(tp == NULL);
347                         lock_flags &= ~XFS_ILOCK_EXCL;
348                         ASSERT(lock_flags == XFS_IOLOCK_EXCL);
349                         goto error_return;
350                 }
351                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
352                 if ((code = xfs_trans_reserve(tp, 0,
353                                              XFS_ITRUNCATE_LOG_RES(mp), 0,
354                                              XFS_TRANS_PERM_LOG_RES,
355                                              XFS_ITRUNCATE_LOG_COUNT))) {
356                         xfs_trans_cancel(tp, 0);
357                         if (need_iolock)
358                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
359                         return code;
360                 }
361                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
362                 xfs_ilock(ip, XFS_ILOCK_EXCL);
363         }
364
365         if (tp) {
366                 xfs_trans_ijoin(tp, ip, lock_flags);
367                 xfs_trans_ihold(tp, ip);
368         }
369
370         /*
371          * Truncate file.  Must have write permission and not be a directory.
372          */
373         if (mask & ATTR_SIZE) {
374                 /*
375                  * Only change the c/mtime if we are changing the size
376                  * or we are explicitly asked to change it. This handles
377                  * the semantic difference between truncate() and ftruncate()
378                  * as implemented in the VFS.
379                  */
380                 if (iattr->ia_size != ip->i_size || (mask & ATTR_CTIME))
381                         timeflags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
382
383                 if (iattr->ia_size > ip->i_size) {
384                         ip->i_d.di_size = iattr->ia_size;
385                         ip->i_size = iattr->ia_size;
386                         if (!(flags & XFS_ATTR_DMI))
387                                 xfs_ichgtime(ip, XFS_ICHGTIME_CHG);
388                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
389                 } else if (iattr->ia_size <= ip->i_size ||
390                            (iattr->ia_size == 0 && ip->i_d.di_nextents)) {
391                         /*
392                          * signal a sync transaction unless
393                          * we're truncating an already unlinked
394                          * file on a wsync filesystem
395                          */
396                         code = xfs_itruncate_finish(&tp, ip, iattr->ia_size,
397                                             XFS_DATA_FORK,
398                                             ((ip->i_d.di_nlink != 0 ||
399                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
400                                              ? 1 : 0));
401                         if (code)
402                                 goto abort_return;
403                         /*
404                          * Truncated "down", so we're removing references
405                          * to old data here - if we now delay flushing for
406                          * a long time, we expose ourselves unduly to the
407                          * notorious NULL files problem.  So, we mark this
408                          * vnode and flush it when the file is closed, and
409                          * do not wait the usual (long) time for writeout.
410                          */
411                         xfs_iflags_set(ip, XFS_ITRUNCATED);
412                 }
413         }
414
415         /*
416          * Change file access modes.
417          */
418         if (mask & ATTR_MODE) {
419                 ip->i_d.di_mode &= S_IFMT;
420                 ip->i_d.di_mode |= iattr->ia_mode & ~S_IFMT;
421
422                 inode->i_mode &= S_IFMT;
423                 inode->i_mode |= iattr->ia_mode & ~S_IFMT;
424
425                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
426                 timeflags |= XFS_ICHGTIME_CHG;
427         }
428
429         /*
430          * Change file ownership.  Must be the owner or privileged.
431          */
432         if (mask & (ATTR_UID|ATTR_GID)) {
433                 /*
434                  * CAP_FSETID overrides the following restrictions:
435                  *
436                  * The set-user-ID and set-group-ID bits of a file will be
437                  * cleared upon successful return from chown()
438                  */
439                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
440                     !capable(CAP_FSETID)) {
441                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
442                 }
443
444                 /*
445                  * Change the ownerships and register quota modifications
446                  * in the transaction.
447                  */
448                 if (iuid != uid) {
449                         if (XFS_IS_UQUOTA_ON(mp)) {
450                                 ASSERT(mask & ATTR_UID);
451                                 ASSERT(udqp);
452                                 olddquot1 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
453                                                         &ip->i_udquot, udqp);
454                         }
455                         ip->i_d.di_uid = uid;
456                         inode->i_uid = uid;
457                 }
458                 if (igid != gid) {
459                         if (XFS_IS_GQUOTA_ON(mp)) {
460                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
461                                 ASSERT(mask & ATTR_GID);
462                                 ASSERT(gdqp);
463                                 olddquot2 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
464                                                         &ip->i_gdquot, gdqp);
465                         }
466                         ip->i_d.di_gid = gid;
467                         inode->i_gid = gid;
468                 }
469
470                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
471                 timeflags |= XFS_ICHGTIME_CHG;
472         }
473
474
475         /*
476          * Change file access or modified times.
477          */
478         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
479                 if (mask & ATTR_ATIME) {
480                         inode->i_atime = iattr->ia_atime;
481                         ip->i_d.di_atime.t_sec = iattr->ia_atime.tv_sec;
482                         ip->i_d.di_atime.t_nsec = iattr->ia_atime.tv_nsec;
483                         ip->i_update_core = 1;
484                 }
485                 if (mask & ATTR_MTIME) {
486                         inode->i_mtime = iattr->ia_mtime;
487                         ip->i_d.di_mtime.t_sec = iattr->ia_mtime.tv_sec;
488                         ip->i_d.di_mtime.t_nsec = iattr->ia_mtime.tv_nsec;
489                         timeflags &= ~XFS_ICHGTIME_MOD;
490                         timeflags |= XFS_ICHGTIME_CHG;
491                 }
492                 if (tp && (mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)))
493                         xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
494         }
495
496         /*
497          * Change file inode change time only if ATTR_CTIME set
498          * AND we have been called by a DMI function.
499          */
500
501         if ((flags & XFS_ATTR_DMI) && (mask & ATTR_CTIME)) {
502                 inode->i_ctime = iattr->ia_ctime;
503                 ip->i_d.di_ctime.t_sec = iattr->ia_ctime.tv_sec;
504                 ip->i_d.di_ctime.t_nsec = iattr->ia_ctime.tv_nsec;
505                 ip->i_update_core = 1;
506                 timeflags &= ~XFS_ICHGTIME_CHG;
507         }
508
509         /*
510          * Send out timestamp changes that need to be set to the
511          * current time.  Not done when called by a DMI function.
512          */
513         if (timeflags && !(flags & XFS_ATTR_DMI))
514                 xfs_ichgtime(ip, timeflags);
515
516         XFS_STATS_INC(xs_ig_attrchg);
517
518         /*
519          * If this is a synchronous mount, make sure that the
520          * transaction goes to disk before returning to the user.
521          * This is slightly sub-optimal in that truncates require
522          * two sync transactions instead of one for wsync filesystems.
523          * One for the truncate and one for the timestamps since we
524          * don't want to change the timestamps unless we're sure the
525          * truncate worked.  Truncates are less than 1% of the laddis
526          * mix so this probably isn't worth the trouble to optimize.
527          */
528         code = 0;
529         if (tp) {
530                 if (mp->m_flags & XFS_MOUNT_WSYNC)
531                         xfs_trans_set_sync(tp);
532
533                 code = xfs_trans_commit(tp, commit_flags);
534         }
535
536         xfs_iunlock(ip, lock_flags);
537
538         /*
539          * Release any dquot(s) the inode had kept before chown.
540          */
541         XFS_QM_DQRELE(mp, olddquot1);
542         XFS_QM_DQRELE(mp, olddquot2);
543         XFS_QM_DQRELE(mp, udqp);
544         XFS_QM_DQRELE(mp, gdqp);
545
546         if (code) {
547                 return code;
548         }
549
550         if (DM_EVENT_ENABLED(ip, DM_EVENT_ATTRIBUTE) &&
551             !(flags & XFS_ATTR_DMI)) {
552                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_ATTRIBUTE, ip, DM_RIGHT_NULL,
553                                         NULL, DM_RIGHT_NULL, NULL, NULL,
554                                         0, 0, AT_DELAY_FLAG(flags));
555         }
556         return 0;
557
558  abort_return:
559         commit_flags |= XFS_TRANS_ABORT;
560         /* FALLTHROUGH */
561  error_return:
562         XFS_QM_DQRELE(mp, udqp);
563         XFS_QM_DQRELE(mp, gdqp);
564         if (tp) {
565                 xfs_trans_cancel(tp, commit_flags);
566         }
567         if (lock_flags != 0) {
568                 xfs_iunlock(ip, lock_flags);
569         }
570         return code;
571 }
572
573 /*
574  * The maximum pathlen is 1024 bytes. Since the minimum file system
575  * blocksize is 512 bytes, we can get a max of 2 extents back from
576  * bmapi.
577  */
578 #define SYMLINK_MAPS 2
579
580 STATIC int
581 xfs_readlink_bmap(
582         xfs_inode_t     *ip,
583         char            *link)
584 {
585         xfs_mount_t     *mp = ip->i_mount;
586         int             pathlen = ip->i_d.di_size;
587         int             nmaps = SYMLINK_MAPS;
588         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
589         xfs_daddr_t     d;
590         int             byte_cnt;
591         int             n;
592         xfs_buf_t       *bp;
593         int             error = 0;
594
595         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
596                         mval, &nmaps, NULL, NULL);
597         if (error)
598                 goto out;
599
600         for (n = 0; n < nmaps; n++) {
601                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
602                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
603
604                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
605                 error = XFS_BUF_GETERROR(bp);
606                 if (error) {
607                         xfs_ioerror_alert("xfs_readlink",
608                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
609                         xfs_buf_relse(bp);
610                         goto out;
611                 }
612                 if (pathlen < byte_cnt)
613                         byte_cnt = pathlen;
614                 pathlen -= byte_cnt;
615
616                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
617                 xfs_buf_relse(bp);
618         }
619
620         link[ip->i_d.di_size] = '\0';
621         error = 0;
622
623  out:
624         return error;
625 }
626
627 int
628 xfs_readlink(
629         xfs_inode_t     *ip,
630         char            *link)
631 {
632         xfs_mount_t     *mp = ip->i_mount;
633         int             pathlen;
634         int             error = 0;
635
636         xfs_itrace_entry(ip);
637
638         if (XFS_FORCED_SHUTDOWN(mp))
639                 return XFS_ERROR(EIO);
640
641         xfs_ilock(ip, XFS_ILOCK_SHARED);
642
643         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
644         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
645
646         pathlen = ip->i_d.di_size;
647         if (!pathlen)
648                 goto out;
649
650         if (ip->i_df.if_flags & XFS_IFINLINE) {
651                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
652                 link[pathlen] = '\0';
653         } else {
654                 error = xfs_readlink_bmap(ip, link);
655         }
656
657  out:
658         xfs_iunlock(ip, XFS_ILOCK_SHARED);
659         return error;
660 }
661
662 /*
663  * xfs_fsync
664  *
665  * This is called to sync the inode and its data out to disk.  We need to hold
666  * the I/O lock while flushing the data, and the inode lock while flushing the
667  * inode.  The inode lock CANNOT be held while flushing the data, so acquire
668  * after we're done with that.
669  */
670 int
671 xfs_fsync(
672         xfs_inode_t     *ip)
673 {
674         xfs_trans_t     *tp;
675         int             error;
676         int             log_flushed = 0, changed = 1;
677
678         xfs_itrace_entry(ip);
679
680         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
681                 return XFS_ERROR(EIO);
682
683         /* capture size updates in I/O completion before writing the inode. */
684         error = xfs_wait_on_pages(ip, 0, -1);
685         if (error)
686                 return XFS_ERROR(error);
687
688         /*
689          * We always need to make sure that the required inode state is safe on
690          * disk.  The vnode might be clean but we still might need to force the
691          * log because of committed transactions that haven't hit the disk yet.
692          * Likewise, there could be unflushed non-transactional changes to the
693          * inode core that have to go to disk and this requires us to issue
694          * a synchronous transaction to capture these changes correctly.
695          *
696          * This code relies on the assumption that if the update_* fields
697          * of the inode are clear and the inode is unpinned then it is clean
698          * and no action is required.
699          */
700         xfs_ilock(ip, XFS_ILOCK_SHARED);
701
702         if (!(ip->i_update_size || ip->i_update_core)) {
703                 /*
704                  * Timestamps/size haven't changed since last inode flush or
705                  * inode transaction commit.  That means either nothing got
706                  * written or a transaction committed which caught the updates.
707                  * If the latter happened and the transaction hasn't hit the
708                  * disk yet, the inode will be still be pinned.  If it is,
709                  * force the log.
710                  */
711
712                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
713
714                 if (xfs_ipincount(ip)) {
715                         error = _xfs_log_force(ip->i_mount, (xfs_lsn_t)0,
716                                       XFS_LOG_FORCE | XFS_LOG_SYNC,
717                                       &log_flushed);
718                 } else {
719                         /*
720                          * If the inode is not pinned and nothing has changed
721                          * we don't need to flush the cache.
722                          */
723                         changed = 0;
724                 }
725         } else  {
726                 /*
727                  * Kick off a transaction to log the inode core to get the
728                  * updates.  The sync transaction will also force the log.
729                  */
730                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
731                 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
732                 error = xfs_trans_reserve(tp, 0,
733                                 XFS_FSYNC_TS_LOG_RES(ip->i_mount), 0, 0, 0);
734                 if (error) {
735                         xfs_trans_cancel(tp, 0);
736                         return error;
737                 }
738                 xfs_ilock(ip, XFS_ILOCK_EXCL);
739
740                 /*
741                  * Note - it's possible that we might have pushed ourselves out
742                  * of the way during trans_reserve which would flush the inode.
743                  * But there's no guarantee that the inode buffer has actually
744                  * gone out yet (it's delwri).  Plus the buffer could be pinned
745                  * anyway if it's part of an inode in another recent
746                  * transaction.  So we play it safe and fire off the
747                  * transaction anyway.
748                  */
749                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
750                 xfs_trans_ihold(tp, ip);
751                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
752                 xfs_trans_set_sync(tp);
753                 error = _xfs_trans_commit(tp, 0, &log_flushed);
754
755                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
756         }
757
758         if ((ip->i_mount->m_flags & XFS_MOUNT_BARRIER) && changed) {
759                 /*
760                  * If the log write didn't issue an ordered tag we need
761                  * to flush the disk cache for the data device now.
762                  */
763                 if (!log_flushed)
764                         xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
765
766                 /*
767                  * If this inode is on the RT dev we need to flush that
768                  * cache as well.
769                  */
770                 if (XFS_IS_REALTIME_INODE(ip))
771                         xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
772         }
773
774         return error;
775 }
776
777 /*
778  * This is called by xfs_inactive to free any blocks beyond eof
779  * when the link count isn't zero and by xfs_dm_punch_hole() when
780  * punching a hole to EOF.
781  */
782 int
783 xfs_free_eofblocks(
784         xfs_mount_t     *mp,
785         xfs_inode_t     *ip,
786         int             flags)
787 {
788         xfs_trans_t     *tp;
789         int             error;
790         xfs_fileoff_t   end_fsb;
791         xfs_fileoff_t   last_fsb;
792         xfs_filblks_t   map_len;
793         int             nimaps;
794         xfs_bmbt_irec_t imap;
795         int             use_iolock = (flags & XFS_FREE_EOF_LOCK);
796
797         /*
798          * Figure out if there are any blocks beyond the end
799          * of the file.  If not, then there is nothing to do.
800          */
801         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
802         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
803         map_len = last_fsb - end_fsb;
804         if (map_len <= 0)
805                 return 0;
806
807         nimaps = 1;
808         xfs_ilock(ip, XFS_ILOCK_SHARED);
809         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
810                           NULL, 0, &imap, &nimaps, NULL, NULL);
811         xfs_iunlock(ip, XFS_ILOCK_SHARED);
812
813         if (!error && (nimaps != 0) &&
814             (imap.br_startblock != HOLESTARTBLOCK ||
815              ip->i_delayed_blks)) {
816                 /*
817                  * Attach the dquots to the inode up front.
818                  */
819                 if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
820                         return error;
821
822                 /*
823                  * There are blocks after the end of file.
824                  * Free them up now by truncating the file to
825                  * its current size.
826                  */
827                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
828
829                 /*
830                  * Do the xfs_itruncate_start() call before
831                  * reserving any log space because
832                  * itruncate_start will call into the buffer
833                  * cache and we can't
834                  * do that within a transaction.
835                  */
836                 if (use_iolock)
837                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
838                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
839                                     ip->i_size);
840                 if (error) {
841                         xfs_trans_cancel(tp, 0);
842                         if (use_iolock)
843                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
844                         return error;
845                 }
846
847                 error = xfs_trans_reserve(tp, 0,
848                                           XFS_ITRUNCATE_LOG_RES(mp),
849                                           0, XFS_TRANS_PERM_LOG_RES,
850                                           XFS_ITRUNCATE_LOG_COUNT);
851                 if (error) {
852                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
853                         xfs_trans_cancel(tp, 0);
854                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
855                         return error;
856                 }
857
858                 xfs_ilock(ip, XFS_ILOCK_EXCL);
859                 xfs_trans_ijoin(tp, ip,
860                                 XFS_IOLOCK_EXCL |
861                                 XFS_ILOCK_EXCL);
862                 xfs_trans_ihold(tp, ip);
863
864                 error = xfs_itruncate_finish(&tp, ip,
865                                              ip->i_size,
866                                              XFS_DATA_FORK,
867                                              0);
868                 /*
869                  * If we get an error at this point we
870                  * simply don't bother truncating the file.
871                  */
872                 if (error) {
873                         xfs_trans_cancel(tp,
874                                          (XFS_TRANS_RELEASE_LOG_RES |
875                                           XFS_TRANS_ABORT));
876                 } else {
877                         error = xfs_trans_commit(tp,
878                                                 XFS_TRANS_RELEASE_LOG_RES);
879                 }
880                 xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
881                                             : XFS_ILOCK_EXCL));
882         }
883         return error;
884 }
885
886 /*
887  * Free a symlink that has blocks associated with it.
888  */
889 STATIC int
890 xfs_inactive_symlink_rmt(
891         xfs_inode_t     *ip,
892         xfs_trans_t     **tpp)
893 {
894         xfs_buf_t       *bp;
895         int             committed;
896         int             done;
897         int             error;
898         xfs_fsblock_t   first_block;
899         xfs_bmap_free_t free_list;
900         int             i;
901         xfs_mount_t     *mp;
902         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
903         int             nmaps;
904         xfs_trans_t     *ntp;
905         int             size;
906         xfs_trans_t     *tp;
907
908         tp = *tpp;
909         mp = ip->i_mount;
910         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
911         /*
912          * We're freeing a symlink that has some
913          * blocks allocated to it.  Free the
914          * blocks here.  We know that we've got
915          * either 1 or 2 extents and that we can
916          * free them all in one bunmapi call.
917          */
918         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
919         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
920                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
921                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
922                 xfs_trans_cancel(tp, 0);
923                 *tpp = NULL;
924                 return error;
925         }
926         /*
927          * Lock the inode, fix the size, and join it to the transaction.
928          * Hold it so in the normal path, we still have it locked for
929          * the second transaction.  In the error paths we need it
930          * held so the cancel won't rele it, see below.
931          */
932         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
933         size = (int)ip->i_d.di_size;
934         ip->i_d.di_size = 0;
935         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
936         xfs_trans_ihold(tp, ip);
937         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
938         /*
939          * Find the block(s) so we can inval and unmap them.
940          */
941         done = 0;
942         XFS_BMAP_INIT(&free_list, &first_block);
943         nmaps = ARRAY_SIZE(mval);
944         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
945                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
946                         &free_list, NULL)))
947                 goto error0;
948         /*
949          * Invalidate the block(s).
950          */
951         for (i = 0; i < nmaps; i++) {
952                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
953                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
954                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
955                 xfs_trans_binval(tp, bp);
956         }
957         /*
958          * Unmap the dead block(s) to the free_list.
959          */
960         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
961                         &first_block, &free_list, NULL, &done)))
962                 goto error1;
963         ASSERT(done);
964         /*
965          * Commit the first transaction.  This logs the EFI and the inode.
966          */
967         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
968                 goto error1;
969         /*
970          * The transaction must have been committed, since there were
971          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
972          * The new tp has the extent freeing and EFDs.
973          */
974         ASSERT(committed);
975         /*
976          * The first xact was committed, so add the inode to the new one.
977          * Mark it dirty so it will be logged and moved forward in the log as
978          * part of every commit.
979          */
980         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
981         xfs_trans_ihold(tp, ip);
982         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
983         /*
984          * Get a new, empty transaction to return to our caller.
985          */
986         ntp = xfs_trans_dup(tp);
987         /*
988          * Commit the transaction containing extent freeing and EFDs.
989          * If we get an error on the commit here or on the reserve below,
990          * we need to unlock the inode since the new transaction doesn't
991          * have the inode attached.
992          */
993         error = xfs_trans_commit(tp, 0);
994         tp = ntp;
995         if (error) {
996                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
997                 goto error0;
998         }
999         /*
1000          * transaction commit worked ok so we can drop the extra ticket
1001          * reference that we gained in xfs_trans_dup()
1002          */
1003         xfs_log_ticket_put(tp->t_ticket);
1004
1005         /*
1006          * Remove the memory for extent descriptions (just bookkeeping).
1007          */
1008         if (ip->i_df.if_bytes)
1009                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
1010         ASSERT(ip->i_df.if_bytes == 0);
1011         /*
1012          * Put an itruncate log reservation in the new transaction
1013          * for our caller.
1014          */
1015         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
1016                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
1017                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1018                 goto error0;
1019         }
1020         /*
1021          * Return with the inode locked but not joined to the transaction.
1022          */
1023         *tpp = tp;
1024         return 0;
1025
1026  error1:
1027         xfs_bmap_cancel(&free_list);
1028  error0:
1029         /*
1030          * Have to come here with the inode locked and either
1031          * (held and in the transaction) or (not in the transaction).
1032          * If the inode isn't held then cancel would iput it, but
1033          * that's wrong since this is inactive and the vnode ref
1034          * count is 0 already.
1035          * Cancel won't do anything to the inode if held, but it still
1036          * needs to be locked until the cancel is done, if it was
1037          * joined to the transaction.
1038          */
1039         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1040         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1041         *tpp = NULL;
1042         return error;
1043
1044 }
1045
1046 STATIC int
1047 xfs_inactive_symlink_local(
1048         xfs_inode_t     *ip,
1049         xfs_trans_t     **tpp)
1050 {
1051         int             error;
1052
1053         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
1054         /*
1055          * We're freeing a symlink which fit into
1056          * the inode.  Just free the memory used
1057          * to hold the old symlink.
1058          */
1059         error = xfs_trans_reserve(*tpp, 0,
1060                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
1061                                   0, XFS_TRANS_PERM_LOG_RES,
1062                                   XFS_ITRUNCATE_LOG_COUNT);
1063
1064         if (error) {
1065                 xfs_trans_cancel(*tpp, 0);
1066                 *tpp = NULL;
1067                 return error;
1068         }
1069         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1070
1071         /*
1072          * Zero length symlinks _can_ exist.
1073          */
1074         if (ip->i_df.if_bytes > 0) {
1075                 xfs_idata_realloc(ip,
1076                                   -(ip->i_df.if_bytes),
1077                                   XFS_DATA_FORK);
1078                 ASSERT(ip->i_df.if_bytes == 0);
1079         }
1080         return 0;
1081 }
1082
1083 STATIC int
1084 xfs_inactive_attrs(
1085         xfs_inode_t     *ip,
1086         xfs_trans_t     **tpp)
1087 {
1088         xfs_trans_t     *tp;
1089         int             error;
1090         xfs_mount_t     *mp;
1091
1092         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1093         tp = *tpp;
1094         mp = ip->i_mount;
1095         ASSERT(ip->i_d.di_forkoff != 0);
1096         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1097         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1098         if (error)
1099                 goto error_unlock;
1100
1101         error = xfs_attr_inactive(ip);
1102         if (error)
1103                 goto error_unlock;
1104
1105         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1106         error = xfs_trans_reserve(tp, 0,
1107                                   XFS_IFREE_LOG_RES(mp),
1108                                   0, XFS_TRANS_PERM_LOG_RES,
1109                                   XFS_INACTIVE_LOG_COUNT);
1110         if (error)
1111                 goto error_cancel;
1112
1113         xfs_ilock(ip, XFS_ILOCK_EXCL);
1114         xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1115         xfs_trans_ihold(tp, ip);
1116         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1117
1118         ASSERT(ip->i_d.di_anextents == 0);
1119
1120         *tpp = tp;
1121         return 0;
1122
1123 error_cancel:
1124         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1125         xfs_trans_cancel(tp, 0);
1126 error_unlock:
1127         *tpp = NULL;
1128         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1129         return error;
1130 }
1131
1132 int
1133 xfs_release(
1134         xfs_inode_t     *ip)
1135 {
1136         xfs_mount_t     *mp = ip->i_mount;
1137         int             error;
1138
1139         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
1140                 return 0;
1141
1142         /* If this is a read-only mount, don't do this (would generate I/O) */
1143         if (mp->m_flags & XFS_MOUNT_RDONLY)
1144                 return 0;
1145
1146         if (!XFS_FORCED_SHUTDOWN(mp)) {
1147                 int truncated;
1148
1149                 /*
1150                  * If we are using filestreams, and we have an unlinked
1151                  * file that we are processing the last close on, then nothing
1152                  * will be able to reopen and write to this file. Purge this
1153                  * inode from the filestreams cache so that it doesn't delay
1154                  * teardown of the inode.
1155                  */
1156                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
1157                         xfs_filestream_deassociate(ip);
1158
1159                 /*
1160                  * If we previously truncated this file and removed old data
1161                  * in the process, we want to initiate "early" writeout on
1162                  * the last close.  This is an attempt to combat the notorious
1163                  * NULL files problem which is particularly noticable from a
1164                  * truncate down, buffered (re-)write (delalloc), followed by
1165                  * a crash.  What we are effectively doing here is
1166                  * significantly reducing the time window where we'd otherwise
1167                  * be exposed to that problem.
1168                  */
1169                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1170                 if (truncated && VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
1171                         xfs_flush_pages(ip, 0, -1, XFS_B_ASYNC, FI_NONE);
1172         }
1173
1174         if (ip->i_d.di_nlink != 0) {
1175                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1176                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1177                        ip->i_delayed_blks > 0)) &&
1178                      (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
1179                     (!(ip->i_d.di_flags &
1180                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
1181                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1182                         if (error)
1183                                 return error;
1184                 }
1185         }
1186
1187         return 0;
1188 }
1189
1190 /*
1191  * xfs_inactive
1192  *
1193  * This is called when the vnode reference count for the vnode
1194  * goes to zero.  If the file has been unlinked, then it must
1195  * now be truncated.  Also, we clear all of the read-ahead state
1196  * kept for the inode here since the file is now closed.
1197  */
1198 int
1199 xfs_inactive(
1200         xfs_inode_t     *ip)
1201 {
1202         xfs_bmap_free_t free_list;
1203         xfs_fsblock_t   first_block;
1204         int             committed;
1205         xfs_trans_t     *tp;
1206         xfs_mount_t     *mp;
1207         int             error;
1208         int             truncate;
1209
1210         xfs_itrace_entry(ip);
1211
1212         /*
1213          * If the inode is already free, then there can be nothing
1214          * to clean up here.
1215          */
1216         if (ip->i_d.di_mode == 0 || VN_BAD(VFS_I(ip))) {
1217                 ASSERT(ip->i_df.if_real_bytes == 0);
1218                 ASSERT(ip->i_df.if_broot_bytes == 0);
1219                 return VN_INACTIVE_CACHE;
1220         }
1221
1222         /*
1223          * Only do a truncate if it's a regular file with
1224          * some actual space in it.  It's OK to look at the
1225          * inode's fields without the lock because we're the
1226          * only one with a reference to the inode.
1227          */
1228         truncate = ((ip->i_d.di_nlink == 0) &&
1229             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1230              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1231             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1232
1233         mp = ip->i_mount;
1234
1235         if (ip->i_d.di_nlink == 0 && DM_EVENT_ENABLED(ip, DM_EVENT_DESTROY))
1236                 XFS_SEND_DESTROY(mp, ip, DM_RIGHT_NULL);
1237
1238         error = 0;
1239
1240         /* If this is a read-only mount, don't do this (would generate I/O) */
1241         if (mp->m_flags & XFS_MOUNT_RDONLY)
1242                 goto out;
1243
1244         if (ip->i_d.di_nlink != 0) {
1245                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1246                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1247                        ip->i_delayed_blks > 0)) &&
1248                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1249                      (!(ip->i_d.di_flags &
1250                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1251                       (ip->i_delayed_blks != 0)))) {
1252                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1253                         if (error)
1254                                 return VN_INACTIVE_CACHE;
1255                 }
1256                 goto out;
1257         }
1258
1259         ASSERT(ip->i_d.di_nlink == 0);
1260
1261         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
1262                 return VN_INACTIVE_CACHE;
1263
1264         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1265         if (truncate) {
1266                 /*
1267                  * Do the xfs_itruncate_start() call before
1268                  * reserving any log space because itruncate_start
1269                  * will call into the buffer cache and we can't
1270                  * do that within a transaction.
1271                  */
1272                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1273
1274                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1275                 if (error) {
1276                         xfs_trans_cancel(tp, 0);
1277                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1278                         return VN_INACTIVE_CACHE;
1279                 }
1280
1281                 error = xfs_trans_reserve(tp, 0,
1282                                           XFS_ITRUNCATE_LOG_RES(mp),
1283                                           0, XFS_TRANS_PERM_LOG_RES,
1284                                           XFS_ITRUNCATE_LOG_COUNT);
1285                 if (error) {
1286                         /* Don't call itruncate_cleanup */
1287                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1288                         xfs_trans_cancel(tp, 0);
1289                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1290                         return VN_INACTIVE_CACHE;
1291                 }
1292
1293                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1294                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1295                 xfs_trans_ihold(tp, ip);
1296
1297                 /*
1298                  * normally, we have to run xfs_itruncate_finish sync.
1299                  * But if filesystem is wsync and we're in the inactive
1300                  * path, then we know that nlink == 0, and that the
1301                  * xaction that made nlink == 0 is permanently committed
1302                  * since xfs_remove runs as a synchronous transaction.
1303                  */
1304                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1305                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1306
1307                 if (error) {
1308                         xfs_trans_cancel(tp,
1309                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1310                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1311                         return VN_INACTIVE_CACHE;
1312                 }
1313         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1314
1315                 /*
1316                  * If we get an error while cleaning up a
1317                  * symlink we bail out.
1318                  */
1319                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1320                         xfs_inactive_symlink_rmt(ip, &tp) :
1321                         xfs_inactive_symlink_local(ip, &tp);
1322
1323                 if (error) {
1324                         ASSERT(tp == NULL);
1325                         return VN_INACTIVE_CACHE;
1326                 }
1327
1328                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1329                 xfs_trans_ihold(tp, ip);
1330         } else {
1331                 error = xfs_trans_reserve(tp, 0,
1332                                           XFS_IFREE_LOG_RES(mp),
1333                                           0, XFS_TRANS_PERM_LOG_RES,
1334                                           XFS_INACTIVE_LOG_COUNT);
1335                 if (error) {
1336                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1337                         xfs_trans_cancel(tp, 0);
1338                         return VN_INACTIVE_CACHE;
1339                 }
1340
1341                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1342                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1343                 xfs_trans_ihold(tp, ip);
1344         }
1345
1346         /*
1347          * If there are attributes associated with the file
1348          * then blow them away now.  The code calls a routine
1349          * that recursively deconstructs the attribute fork.
1350          * We need to just commit the current transaction
1351          * because we can't use it for xfs_attr_inactive().
1352          */
1353         if (ip->i_d.di_anextents > 0) {
1354                 error = xfs_inactive_attrs(ip, &tp);
1355                 /*
1356                  * If we got an error, the transaction is already
1357                  * cancelled, and the inode is unlocked. Just get out.
1358                  */
1359                  if (error)
1360                          return VN_INACTIVE_CACHE;
1361         } else if (ip->i_afp) {
1362                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1363         }
1364
1365         /*
1366          * Free the inode.
1367          */
1368         XFS_BMAP_INIT(&free_list, &first_block);
1369         error = xfs_ifree(tp, ip, &free_list);
1370         if (error) {
1371                 /*
1372                  * If we fail to free the inode, shut down.  The cancel
1373                  * might do that, we need to make sure.  Otherwise the
1374                  * inode might be lost for a long time or forever.
1375                  */
1376                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1377                         cmn_err(CE_NOTE,
1378                 "xfs_inactive:  xfs_ifree() returned an error = %d on %s",
1379                                 error, mp->m_fsname);
1380                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1381                 }
1382                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1383         } else {
1384                 /*
1385                  * Credit the quota account(s). The inode is gone.
1386                  */
1387                 XFS_TRANS_MOD_DQUOT_BYINO(mp, tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1388
1389                 /*
1390                  * Just ignore errors at this point.  There is nothing we can
1391                  * do except to try to keep going. Make sure it's not a silent
1392                  * error.
1393                  */
1394                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1395                 if (error)
1396                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1397                                 "xfs_bmap_finish() returned error %d", error);
1398                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1399                 if (error)
1400                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1401                                 "xfs_trans_commit() returned error %d", error);
1402         }
1403         /*
1404          * Release the dquots held by inode, if any.
1405          */
1406         XFS_QM_DQDETACH(mp, ip);
1407
1408         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1409
1410  out:
1411         return VN_INACTIVE_CACHE;
1412 }
1413
1414 /*
1415  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1416  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1417  * ci_name->name will point to a the actual name (caller must free) or
1418  * will be set to NULL if an exact match is found.
1419  */
1420 int
1421 xfs_lookup(
1422         xfs_inode_t             *dp,
1423         struct xfs_name         *name,
1424         xfs_inode_t             **ipp,
1425         struct xfs_name         *ci_name)
1426 {
1427         xfs_ino_t               inum;
1428         int                     error;
1429         uint                    lock_mode;
1430
1431         xfs_itrace_entry(dp);
1432
1433         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1434                 return XFS_ERROR(EIO);
1435
1436         lock_mode = xfs_ilock_map_shared(dp);
1437         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1438         xfs_iunlock_map_shared(dp, lock_mode);
1439
1440         if (error)
1441                 goto out;
1442
1443         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp, 0);
1444         if (error)
1445                 goto out_free_name;
1446
1447         xfs_itrace_ref(*ipp);
1448         return 0;
1449
1450 out_free_name:
1451         if (ci_name)
1452                 kmem_free(ci_name->name);
1453 out:
1454         *ipp = NULL;
1455         return error;
1456 }
1457
1458 int
1459 xfs_create(
1460         xfs_inode_t             *dp,
1461         struct xfs_name         *name,
1462         mode_t                  mode,
1463         xfs_dev_t               rdev,
1464         xfs_inode_t             **ipp,
1465         cred_t                  *credp)
1466 {
1467         xfs_mount_t             *mp = dp->i_mount;
1468         xfs_inode_t             *ip;
1469         xfs_trans_t             *tp;
1470         int                     error;
1471         xfs_bmap_free_t         free_list;
1472         xfs_fsblock_t           first_block;
1473         boolean_t               unlock_dp_on_error = B_FALSE;
1474         int                     dm_event_sent = 0;
1475         uint                    cancel_flags;
1476         int                     committed;
1477         xfs_prid_t              prid;
1478         struct xfs_dquot        *udqp, *gdqp;
1479         uint                    resblks;
1480
1481         ASSERT(!*ipp);
1482         xfs_itrace_entry(dp);
1483
1484         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
1485                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
1486                                 dp, DM_RIGHT_NULL, NULL,
1487                                 DM_RIGHT_NULL, name->name, NULL,
1488                                 mode, 0, 0);
1489
1490                 if (error)
1491                         return error;
1492                 dm_event_sent = 1;
1493         }
1494
1495         if (XFS_FORCED_SHUTDOWN(mp))
1496                 return XFS_ERROR(EIO);
1497
1498         /* Return through std_return after this point. */
1499
1500         udqp = gdqp = NULL;
1501         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1502                 prid = dp->i_d.di_projid;
1503         else
1504                 prid = (xfs_prid_t)dfltprid;
1505
1506         /*
1507          * Make sure that we have allocated dquot(s) on disk.
1508          */
1509         error = XFS_QM_DQVOPALLOC(mp, dp,
1510                         current_fsuid(), current_fsgid(), prid,
1511                         XFS_QMOPT_QUOTALL|XFS_QMOPT_INHERIT, &udqp, &gdqp);
1512         if (error)
1513                 goto std_return;
1514
1515         ip = NULL;
1516
1517         tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1518         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1519         resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1520         /*
1521          * Initially assume that the file does not exist and
1522          * reserve the resources for that case.  If that is not
1523          * the case we'll drop the one we have and get a more
1524          * appropriate transaction later.
1525          */
1526         error = xfs_trans_reserve(tp, resblks, XFS_CREATE_LOG_RES(mp), 0,
1527                         XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1528         if (error == ENOSPC) {
1529                 resblks = 0;
1530                 error = xfs_trans_reserve(tp, 0, XFS_CREATE_LOG_RES(mp), 0,
1531                                 XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1532         }
1533         if (error) {
1534                 cancel_flags = 0;
1535                 goto error_return;
1536         }
1537
1538         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1539         unlock_dp_on_error = B_TRUE;
1540
1541         XFS_BMAP_INIT(&free_list, &first_block);
1542
1543         ASSERT(ip == NULL);
1544
1545         /*
1546          * Reserve disk quota and the inode.
1547          */
1548         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
1549         if (error)
1550                 goto error_return;
1551
1552         error = xfs_dir_canenter(tp, dp, name, resblks);
1553         if (error)
1554                 goto error_return;
1555         error = xfs_dir_ialloc(&tp, dp, mode, 1,
1556                         rdev, credp, prid, resblks > 0,
1557                         &ip, &committed);
1558         if (error) {
1559                 if (error == ENOSPC)
1560                         goto error_return;
1561                 goto abort_return;
1562         }
1563         xfs_itrace_ref(ip);
1564
1565         /*
1566          * At this point, we've gotten a newly allocated inode.
1567          * It is locked (and joined to the transaction).
1568          */
1569
1570         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1571
1572         /*
1573          * Now we join the directory inode to the transaction.  We do not do it
1574          * earlier because xfs_dir_ialloc might commit the previous transaction
1575          * (and release all the locks).  An error from here on will result in
1576          * the transaction cancel unlocking dp so don't do it explicitly in the
1577          * error path.
1578          */
1579         IHOLD(dp);
1580         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1581         unlock_dp_on_error = B_FALSE;
1582
1583         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1584                                         &first_block, &free_list, resblks ?
1585                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1586         if (error) {
1587                 ASSERT(error != ENOSPC);
1588                 goto abort_return;
1589         }
1590         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1591         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1592
1593         /*
1594          * If this is a synchronous mount, make sure that the
1595          * create transaction goes to disk before returning to
1596          * the user.
1597          */
1598         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1599                 xfs_trans_set_sync(tp);
1600         }
1601
1602         /*
1603          * Attach the dquot(s) to the inodes and modify them incore.
1604          * These ids of the inode couldn't have changed since the new
1605          * inode has been locked ever since it was created.
1606          */
1607         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
1608
1609         /*
1610          * xfs_trans_commit normally decrements the vnode ref count
1611          * when it unlocks the inode. Since we want to return the
1612          * vnode to the caller, we bump the vnode ref count now.
1613          */
1614         IHOLD(ip);
1615
1616         error = xfs_bmap_finish(&tp, &free_list, &committed);
1617         if (error) {
1618                 xfs_bmap_cancel(&free_list);
1619                 goto abort_rele;
1620         }
1621
1622         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1623         if (error) {
1624                 IRELE(ip);
1625                 tp = NULL;
1626                 goto error_return;
1627         }
1628
1629         XFS_QM_DQRELE(mp, udqp);
1630         XFS_QM_DQRELE(mp, gdqp);
1631
1632         *ipp = ip;
1633
1634         /* Fallthrough to std_return with error = 0  */
1635
1636 std_return:
1637         if ((*ipp || (error != 0 && dm_event_sent != 0)) &&
1638             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
1639                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
1640                         dp, DM_RIGHT_NULL,
1641                         *ipp ? ip : NULL,
1642                         DM_RIGHT_NULL, name->name, NULL,
1643                         mode, error, 0);
1644         }
1645         return error;
1646
1647  abort_return:
1648         cancel_flags |= XFS_TRANS_ABORT;
1649         /* FALLTHROUGH */
1650
1651  error_return:
1652         if (tp != NULL)
1653                 xfs_trans_cancel(tp, cancel_flags);
1654
1655         XFS_QM_DQRELE(mp, udqp);
1656         XFS_QM_DQRELE(mp, gdqp);
1657
1658         if (unlock_dp_on_error)
1659                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1660
1661         goto std_return;
1662
1663  abort_rele:
1664         /*
1665          * Wait until after the current transaction is aborted to
1666          * release the inode.  This prevents recursive transactions
1667          * and deadlocks from xfs_inactive.
1668          */
1669         cancel_flags |= XFS_TRANS_ABORT;
1670         xfs_trans_cancel(tp, cancel_flags);
1671         IRELE(ip);
1672
1673         XFS_QM_DQRELE(mp, udqp);
1674         XFS_QM_DQRELE(mp, gdqp);
1675
1676         goto std_return;
1677 }
1678
1679 #ifdef DEBUG
1680 int xfs_locked_n;
1681 int xfs_small_retries;
1682 int xfs_middle_retries;
1683 int xfs_lots_retries;
1684 int xfs_lock_delays;
1685 #endif
1686
1687 /*
1688  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1689  * a different value
1690  */
1691 static inline int
1692 xfs_lock_inumorder(int lock_mode, int subclass)
1693 {
1694         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1695                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1696         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1697                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1698
1699         return lock_mode;
1700 }
1701
1702 /*
1703  * The following routine will lock n inodes in exclusive mode.
1704  * We assume the caller calls us with the inodes in i_ino order.
1705  *
1706  * We need to detect deadlock where an inode that we lock
1707  * is in the AIL and we start waiting for another inode that is locked
1708  * by a thread in a long running transaction (such as truncate). This can
1709  * result in deadlock since the long running trans might need to wait
1710  * for the inode we just locked in order to push the tail and free space
1711  * in the log.
1712  */
1713 void
1714 xfs_lock_inodes(
1715         xfs_inode_t     **ips,
1716         int             inodes,
1717         uint            lock_mode)
1718 {
1719         int             attempts = 0, i, j, try_lock;
1720         xfs_log_item_t  *lp;
1721
1722         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1723
1724         try_lock = 0;
1725         i = 0;
1726
1727 again:
1728         for (; i < inodes; i++) {
1729                 ASSERT(ips[i]);
1730
1731                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1732                         continue;
1733
1734                 /*
1735                  * If try_lock is not set yet, make sure all locked inodes
1736                  * are not in the AIL.
1737                  * If any are, set try_lock to be used later.
1738                  */
1739
1740                 if (!try_lock) {
1741                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1742                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1743                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1744                                         try_lock++;
1745                                 }
1746                         }
1747                 }
1748
1749                 /*
1750                  * If any of the previous locks we have locked is in the AIL,
1751                  * we must TRY to get the second and subsequent locks. If
1752                  * we can't get any, we must release all we have
1753                  * and try again.
1754                  */
1755
1756                 if (try_lock) {
1757                         /* try_lock must be 0 if i is 0. */
1758                         /*
1759                          * try_lock means we have an inode locked
1760                          * that is in the AIL.
1761                          */
1762                         ASSERT(i != 0);
1763                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1764                                 attempts++;
1765
1766                                 /*
1767                                  * Unlock all previous guys and try again.
1768                                  * xfs_iunlock will try to push the tail
1769                                  * if the inode is in the AIL.
1770                                  */
1771
1772                                 for(j = i - 1; j >= 0; j--) {
1773
1774                                         /*
1775                                          * Check to see if we've already
1776                                          * unlocked this one.
1777                                          * Not the first one going back,
1778                                          * and the inode ptr is the same.
1779                                          */
1780                                         if ((j != (i - 1)) && ips[j] ==
1781                                                                 ips[j+1])
1782                                                 continue;
1783
1784                                         xfs_iunlock(ips[j], lock_mode);
1785                                 }
1786
1787                                 if ((attempts % 5) == 0) {
1788                                         delay(1); /* Don't just spin the CPU */
1789 #ifdef DEBUG
1790                                         xfs_lock_delays++;
1791 #endif
1792                                 }
1793                                 i = 0;
1794                                 try_lock = 0;
1795                                 goto again;
1796                         }
1797                 } else {
1798                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1799                 }
1800         }
1801
1802 #ifdef DEBUG
1803         if (attempts) {
1804                 if (attempts < 5) xfs_small_retries++;
1805                 else if (attempts < 100) xfs_middle_retries++;
1806                 else xfs_lots_retries++;
1807         } else {
1808                 xfs_locked_n++;
1809         }
1810 #endif
1811 }
1812
1813 /*
1814  * xfs_lock_two_inodes() can only be used to lock one type of lock
1815  * at a time - the iolock or the ilock, but not both at once. If
1816  * we lock both at once, lockdep will report false positives saying
1817  * we have violated locking orders.
1818  */
1819 void
1820 xfs_lock_two_inodes(
1821         xfs_inode_t             *ip0,
1822         xfs_inode_t             *ip1,
1823         uint                    lock_mode)
1824 {
1825         xfs_inode_t             *temp;
1826         int                     attempts = 0;
1827         xfs_log_item_t          *lp;
1828
1829         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1830                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1831         ASSERT(ip0->i_ino != ip1->i_ino);
1832
1833         if (ip0->i_ino > ip1->i_ino) {
1834                 temp = ip0;
1835                 ip0 = ip1;
1836                 ip1 = temp;
1837         }
1838
1839  again:
1840         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1841
1842         /*
1843          * If the first lock we have locked is in the AIL, we must TRY to get
1844          * the second lock. If we can't get it, we must release the first one
1845          * and try again.
1846          */
1847         lp = (xfs_log_item_t *)ip0->i_itemp;
1848         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1849                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1850                         xfs_iunlock(ip0, lock_mode);
1851                         if ((++attempts % 5) == 0)
1852                                 delay(1); /* Don't just spin the CPU */
1853                         goto again;
1854                 }
1855         } else {
1856                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1857         }
1858 }
1859
1860 int
1861 xfs_remove(
1862         xfs_inode_t             *dp,
1863         struct xfs_name         *name,
1864         xfs_inode_t             *ip)
1865 {
1866         xfs_mount_t             *mp = dp->i_mount;
1867         xfs_trans_t             *tp = NULL;
1868         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1869         int                     error = 0;
1870         xfs_bmap_free_t         free_list;
1871         xfs_fsblock_t           first_block;
1872         int                     cancel_flags;
1873         int                     committed;
1874         int                     link_zero;
1875         uint                    resblks;
1876         uint                    log_count;
1877
1878         xfs_itrace_entry(dp);
1879         xfs_itrace_entry(ip);
1880
1881         if (XFS_FORCED_SHUTDOWN(mp))
1882                 return XFS_ERROR(EIO);
1883
1884         if (DM_EVENT_ENABLED(dp, DM_EVENT_REMOVE)) {
1885                 error = XFS_SEND_NAMESP(mp, DM_EVENT_REMOVE, dp, DM_RIGHT_NULL,
1886                                         NULL, DM_RIGHT_NULL, name->name, NULL,
1887                                         ip->i_d.di_mode, 0, 0);
1888                 if (error)
1889                         return error;
1890         }
1891
1892         error = XFS_QM_DQATTACH(mp, dp, 0);
1893         if (error)
1894                 goto std_return;
1895
1896         error = XFS_QM_DQATTACH(mp, ip, 0);
1897         if (error)
1898                 goto std_return;
1899
1900         if (is_dir) {
1901                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1902                 log_count = XFS_DEFAULT_LOG_COUNT;
1903         } else {
1904                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1905                 log_count = XFS_REMOVE_LOG_COUNT;
1906         }
1907         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1908
1909         /*
1910          * We try to get the real space reservation first,
1911          * allowing for directory btree deletion(s) implying
1912          * possible bmap insert(s).  If we can't get the space
1913          * reservation then we use 0 instead, and avoid the bmap
1914          * btree insert(s) in the directory code by, if the bmap
1915          * insert tries to happen, instead trimming the LAST
1916          * block from the directory.
1917          */
1918         resblks = XFS_REMOVE_SPACE_RES(mp);
1919         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1920                                   XFS_TRANS_PERM_LOG_RES, log_count);
1921         if (error == ENOSPC) {
1922                 resblks = 0;
1923                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1924                                           XFS_TRANS_PERM_LOG_RES, log_count);
1925         }
1926         if (error) {
1927                 ASSERT(error != ENOSPC);
1928                 cancel_flags = 0;
1929                 goto out_trans_cancel;
1930         }
1931
1932         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1933
1934         /*
1935          * At this point, we've gotten both the directory and the entry
1936          * inodes locked.
1937          */
1938         IHOLD(ip);
1939         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1940
1941         IHOLD(dp);
1942         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1943
1944         /*
1945          * If we're removing a directory perform some additional validation.
1946          */
1947         if (is_dir) {
1948                 ASSERT(ip->i_d.di_nlink >= 2);
1949                 if (ip->i_d.di_nlink != 2) {
1950                         error = XFS_ERROR(ENOTEMPTY);
1951                         goto out_trans_cancel;
1952                 }
1953                 if (!xfs_dir_isempty(ip)) {
1954                         error = XFS_ERROR(ENOTEMPTY);
1955                         goto out_trans_cancel;
1956                 }
1957         }
1958
1959         XFS_BMAP_INIT(&free_list, &first_block);
1960         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1961                                         &first_block, &free_list, resblks);
1962         if (error) {
1963                 ASSERT(error != ENOENT);
1964                 goto out_bmap_cancel;
1965         }
1966         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1967
1968         if (is_dir) {
1969                 /*
1970                  * Drop the link from ip's "..".
1971                  */
1972                 error = xfs_droplink(tp, dp);
1973                 if (error)
1974                         goto out_bmap_cancel;
1975
1976                 /*
1977                  * Drop the "." link from ip to self.
1978                  */
1979                 error = xfs_droplink(tp, ip);
1980                 if (error)
1981                         goto out_bmap_cancel;
1982         } else {
1983                 /*
1984                  * When removing a non-directory we need to log the parent
1985                  * inode here.  For a directory this is done implicitly
1986                  * by the xfs_droplink call for the ".." entry.
1987                  */
1988                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1989         }
1990
1991         /*
1992          * Drop the link from dp to ip.
1993          */
1994         error = xfs_droplink(tp, ip);
1995         if (error)
1996                 goto out_bmap_cancel;
1997
1998         /*
1999          * Determine if this is the last link while
2000          * we are in the transaction.
2001          */
2002         link_zero = (ip->i_d.di_nlink == 0);
2003
2004         /*
2005          * If this is a synchronous mount, make sure that the
2006          * remove transaction goes to disk before returning to
2007          * the user.
2008          */
2009         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2010                 xfs_trans_set_sync(tp);
2011
2012         error = xfs_bmap_finish(&tp, &free_list, &committed);
2013         if (error)
2014                 goto out_bmap_cancel;
2015
2016         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2017         if (error)
2018                 goto std_return;
2019
2020         /*
2021          * If we are using filestreams, kill the stream association.
2022          * If the file is still open it may get a new one but that
2023          * will get killed on last close in xfs_close() so we don't
2024          * have to worry about that.
2025          */
2026         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
2027                 xfs_filestream_deassociate(ip);
2028
2029         xfs_itrace_exit(ip);
2030         xfs_itrace_exit(dp);
2031
2032  std_return:
2033         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTREMOVE)) {
2034                 XFS_SEND_NAMESP(mp, DM_EVENT_POSTREMOVE, dp, DM_RIGHT_NULL,
2035                                 NULL, DM_RIGHT_NULL, name->name, NULL,
2036                                 ip->i_d.di_mode, error, 0);
2037         }
2038
2039         return error;
2040
2041  out_bmap_cancel:
2042         xfs_bmap_cancel(&free_list);
2043         cancel_flags |= XFS_TRANS_ABORT;
2044  out_trans_cancel:
2045         xfs_trans_cancel(tp, cancel_flags);
2046         goto std_return;
2047 }
2048
2049 int
2050 xfs_link(
2051         xfs_inode_t             *tdp,
2052         xfs_inode_t             *sip,
2053         struct xfs_name         *target_name)
2054 {
2055         xfs_mount_t             *mp = tdp->i_mount;
2056         xfs_trans_t             *tp;
2057         int                     error;
2058         xfs_bmap_free_t         free_list;
2059         xfs_fsblock_t           first_block;
2060         int                     cancel_flags;
2061         int                     committed;
2062         int                     resblks;
2063
2064         xfs_itrace_entry(tdp);
2065         xfs_itrace_entry(sip);
2066
2067         ASSERT(!S_ISDIR(sip->i_d.di_mode));
2068
2069         if (XFS_FORCED_SHUTDOWN(mp))
2070                 return XFS_ERROR(EIO);
2071
2072         if (DM_EVENT_ENABLED(tdp, DM_EVENT_LINK)) {
2073                 error = XFS_SEND_NAMESP(mp, DM_EVENT_LINK,
2074                                         tdp, DM_RIGHT_NULL,
2075                                         sip, DM_RIGHT_NULL,
2076                                         target_name->name, NULL, 0, 0, 0);
2077                 if (error)
2078                         return error;
2079         }
2080
2081         /* Return through std_return after this point. */
2082
2083         error = XFS_QM_DQATTACH(mp, sip, 0);
2084         if (!error && sip != tdp)
2085                 error = XFS_QM_DQATTACH(mp, tdp, 0);
2086         if (error)
2087                 goto std_return;
2088
2089         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
2090         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2091         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
2092         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
2093                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2094         if (error == ENOSPC) {
2095                 resblks = 0;
2096                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
2097                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2098         }
2099         if (error) {
2100                 cancel_flags = 0;
2101                 goto error_return;
2102         }
2103
2104         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
2105
2106         /*
2107          * Increment vnode ref counts since xfs_trans_commit &
2108          * xfs_trans_cancel will both unlock the inodes and
2109          * decrement the associated ref counts.
2110          */
2111         IHOLD(sip);
2112         IHOLD(tdp);
2113         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
2114         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
2115
2116         /*
2117          * If the source has too many links, we can't make any more to it.
2118          */
2119         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
2120                 error = XFS_ERROR(EMLINK);
2121                 goto error_return;
2122         }
2123
2124         /*
2125          * If we are using project inheritance, we only allow hard link
2126          * creation in our tree when the project IDs are the same; else
2127          * the tree quota mechanism could be circumvented.
2128          */
2129         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2130                      (tdp->i_d.di_projid != sip->i_d.di_projid))) {
2131                 error = XFS_ERROR(EXDEV);
2132                 goto error_return;
2133         }
2134
2135         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
2136         if (error)
2137                 goto error_return;
2138
2139         XFS_BMAP_INIT(&free_list, &first_block);
2140
2141         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
2142                                         &first_block, &free_list, resblks);
2143         if (error)
2144                 goto abort_return;
2145         xfs_ichgtime(tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2146         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
2147
2148         error = xfs_bumplink(tp, sip);
2149         if (error)
2150                 goto abort_return;
2151
2152         /*
2153          * If this is a synchronous mount, make sure that the
2154          * link transaction goes to disk before returning to
2155          * the user.
2156          */
2157         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2158                 xfs_trans_set_sync(tp);
2159         }
2160
2161         error = xfs_bmap_finish (&tp, &free_list, &committed);
2162         if (error) {
2163                 xfs_bmap_cancel(&free_list);
2164                 goto abort_return;
2165         }
2166
2167         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2168         if (error)
2169                 goto std_return;
2170
2171         /* Fall through to std_return with error = 0. */
2172 std_return:
2173         if (DM_EVENT_ENABLED(sip, DM_EVENT_POSTLINK)) {
2174                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTLINK,
2175                                 tdp, DM_RIGHT_NULL,
2176                                 sip, DM_RIGHT_NULL,
2177                                 target_name->name, NULL, 0, error, 0);
2178         }
2179         return error;
2180
2181  abort_return:
2182         cancel_flags |= XFS_TRANS_ABORT;
2183         /* FALLTHROUGH */
2184
2185  error_return:
2186         xfs_trans_cancel(tp, cancel_flags);
2187         goto std_return;
2188 }
2189
2190
2191 int
2192 xfs_mkdir(
2193         xfs_inode_t             *dp,
2194         struct xfs_name         *dir_name,
2195         mode_t                  mode,
2196         xfs_inode_t             **ipp,
2197         cred_t                  *credp)
2198 {
2199         xfs_mount_t             *mp = dp->i_mount;
2200         xfs_inode_t             *cdp;   /* inode of created dir */
2201         xfs_trans_t             *tp;
2202         int                     cancel_flags;
2203         int                     error;
2204         int                     committed;
2205         xfs_bmap_free_t         free_list;
2206         xfs_fsblock_t           first_block;
2207         boolean_t               unlock_dp_on_error = B_FALSE;
2208         boolean_t               created = B_FALSE;
2209         int                     dm_event_sent = 0;
2210         xfs_prid_t              prid;
2211         struct xfs_dquot        *udqp, *gdqp;
2212         uint                    resblks;
2213
2214         if (XFS_FORCED_SHUTDOWN(mp))
2215                 return XFS_ERROR(EIO);
2216
2217         tp = NULL;
2218
2219         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
2220                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
2221                                         dp, DM_RIGHT_NULL, NULL,
2222                                         DM_RIGHT_NULL, dir_name->name, NULL,
2223                                         mode, 0, 0);
2224                 if (error)
2225                         return error;
2226                 dm_event_sent = 1;
2227         }
2228
2229         /* Return through std_return after this point. */
2230
2231         xfs_itrace_entry(dp);
2232
2233         mp = dp->i_mount;
2234         udqp = gdqp = NULL;
2235         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2236                 prid = dp->i_d.di_projid;
2237         else
2238                 prid = (xfs_prid_t)dfltprid;
2239
2240         /*
2241          * Make sure that we have allocated dquot(s) on disk.
2242          */
2243         error = XFS_QM_DQVOPALLOC(mp, dp,
2244                         current_fsuid(), current_fsgid(), prid,
2245                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2246         if (error)
2247                 goto std_return;
2248
2249         tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
2250         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2251         resblks = XFS_MKDIR_SPACE_RES(mp, dir_name->len);
2252         error = xfs_trans_reserve(tp, resblks, XFS_MKDIR_LOG_RES(mp), 0,
2253                                   XFS_TRANS_PERM_LOG_RES, XFS_MKDIR_LOG_COUNT);
2254         if (error == ENOSPC) {
2255                 resblks = 0;
2256                 error = xfs_trans_reserve(tp, 0, XFS_MKDIR_LOG_RES(mp), 0,
2257                                           XFS_TRANS_PERM_LOG_RES,
2258                                           XFS_MKDIR_LOG_COUNT);
2259         }
2260         if (error) {
2261                 cancel_flags = 0;
2262                 goto error_return;
2263         }
2264
2265         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2266         unlock_dp_on_error = B_TRUE;
2267
2268         /*
2269          * Check for directory link count overflow.
2270          */
2271         if (dp->i_d.di_nlink >= XFS_MAXLINK) {
2272                 error = XFS_ERROR(EMLINK);
2273                 goto error_return;
2274         }
2275
2276         /*
2277          * Reserve disk quota and the inode.
2278          */
2279         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2280         if (error)
2281                 goto error_return;
2282
2283         error = xfs_dir_canenter(tp, dp, dir_name, resblks);
2284         if (error)
2285                 goto error_return;
2286         /*
2287          * create the directory inode.
2288          */
2289         error = xfs_dir_ialloc(&tp, dp, mode, 2,
2290                         0, credp, prid, resblks > 0,
2291                 &cdp, NULL);
2292         if (error) {
2293                 if (error == ENOSPC)
2294                         goto error_return;
2295                 goto abort_return;
2296         }
2297         xfs_itrace_ref(cdp);
2298
2299         /*
2300          * Now we add the directory inode to the transaction.
2301          * We waited until now since xfs_dir_ialloc might start
2302          * a new transaction.  Had we joined the transaction
2303          * earlier, the locks might have gotten released. An error
2304          * from here on will result in the transaction cancel
2305          * unlocking dp so don't do it explicitly in the error path.
2306          */
2307         IHOLD(dp);
2308         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2309         unlock_dp_on_error = B_FALSE;
2310
2311         XFS_BMAP_INIT(&free_list, &first_block);
2312
2313         error = xfs_dir_createname(tp, dp, dir_name, cdp->i_ino,
2314                                         &first_block, &free_list, resblks ?
2315                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
2316         if (error) {
2317                 ASSERT(error != ENOSPC);
2318                 goto error1;
2319         }
2320         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2321
2322         error = xfs_dir_init(tp, cdp, dp);
2323         if (error)
2324                 goto error2;
2325
2326         error = xfs_bumplink(tp, dp);
2327         if (error)
2328                 goto error2;
2329
2330         created = B_TRUE;
2331
2332         *ipp = cdp;
2333         IHOLD(cdp);
2334
2335         /*
2336          * Attach the dquots to the new inode and modify the icount incore.
2337          */
2338         XFS_QM_DQVOPCREATE(mp, tp, cdp, udqp, gdqp);
2339
2340         /*
2341          * If this is a synchronous mount, make sure that the
2342          * mkdir transaction goes to disk before returning to
2343          * the user.
2344          */
2345         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2346                 xfs_trans_set_sync(tp);
2347         }
2348
2349         error = xfs_bmap_finish(&tp, &free_list, &committed);
2350         if (error) {
2351                 IRELE(cdp);
2352                 goto error2;
2353         }
2354
2355         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2356         XFS_QM_DQRELE(mp, udqp);
2357         XFS_QM_DQRELE(mp, gdqp);
2358         if (error) {
2359                 IRELE(cdp);
2360         }
2361
2362         /* Fall through to std_return with error = 0 or errno from
2363          * xfs_trans_commit. */
2364
2365 std_return:
2366         if ((created || (error != 0 && dm_event_sent != 0)) &&
2367             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
2368                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
2369                                         dp, DM_RIGHT_NULL,
2370                                         created ? cdp : NULL,
2371                                         DM_RIGHT_NULL,
2372                                         dir_name->name, NULL,
2373                                         mode, error, 0);
2374         }
2375         return error;
2376
2377  error2:
2378  error1:
2379         xfs_bmap_cancel(&free_list);
2380  abort_return:
2381         cancel_flags |= XFS_TRANS_ABORT;
2382  error_return:
2383         xfs_trans_cancel(tp, cancel_flags);
2384         XFS_QM_DQRELE(mp, udqp);
2385         XFS_QM_DQRELE(mp, gdqp);
2386
2387         if (unlock_dp_on_error)
2388                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2389
2390         goto std_return;
2391 }
2392
2393 int
2394 xfs_symlink(
2395         xfs_inode_t             *dp,
2396         struct xfs_name         *link_name,
2397         const char              *target_path,
2398         mode_t                  mode,
2399         xfs_inode_t             **ipp,
2400         cred_t                  *credp)
2401 {
2402         xfs_mount_t             *mp = dp->i_mount;
2403         xfs_trans_t             *tp;
2404         xfs_inode_t             *ip;
2405         int                     error;
2406         int                     pathlen;
2407         xfs_bmap_free_t         free_list;
2408         xfs_fsblock_t           first_block;
2409         boolean_t               unlock_dp_on_error = B_FALSE;
2410         uint                    cancel_flags;
2411         int                     committed;
2412         xfs_fileoff_t           first_fsb;
2413         xfs_filblks_t           fs_blocks;
2414         int                     nmaps;
2415         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
2416         xfs_daddr_t             d;
2417         const char              *cur_chunk;
2418         int                     byte_cnt;
2419         int                     n;
2420         xfs_buf_t               *bp;
2421         xfs_prid_t              prid;
2422         struct xfs_dquot        *udqp, *gdqp;
2423         uint                    resblks;
2424
2425         *ipp = NULL;
2426         error = 0;
2427         ip = NULL;
2428         tp = NULL;
2429
2430         xfs_itrace_entry(dp);
2431
2432         if (XFS_FORCED_SHUTDOWN(mp))
2433                 return XFS_ERROR(EIO);
2434
2435         /*
2436          * Check component lengths of the target path name.
2437          */
2438         pathlen = strlen(target_path);
2439         if (pathlen >= MAXPATHLEN)      /* total string too long */
2440                 return XFS_ERROR(ENAMETOOLONG);
2441
2442         if (DM_EVENT_ENABLED(dp, DM_EVENT_SYMLINK)) {
2443                 error = XFS_SEND_NAMESP(mp, DM_EVENT_SYMLINK, dp,
2444                                         DM_RIGHT_NULL, NULL, DM_RIGHT_NULL,
2445                                         link_name->name, target_path, 0, 0, 0);
2446                 if (error)
2447                         return error;
2448         }
2449
2450         /* Return through std_return after this point. */
2451
2452         udqp = gdqp = NULL;
2453         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2454                 prid = dp->i_d.di_projid;
2455         else
2456                 prid = (xfs_prid_t)dfltprid;
2457
2458         /*
2459          * Make sure that we have allocated dquot(s) on disk.
2460          */
2461         error = XFS_QM_DQVOPALLOC(mp, dp,
2462                         current_fsuid(), current_fsgid(), prid,
2463                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2464         if (error)
2465                 goto std_return;
2466
2467         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
2468         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2469         /*
2470          * The symlink will fit into the inode data fork?
2471          * There can't be any attributes so we get the whole variable part.
2472          */
2473         if (pathlen <= XFS_LITINO(mp))
2474                 fs_blocks = 0;
2475         else
2476                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2477         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2478         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2479                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2480         if (error == ENOSPC && fs_blocks == 0) {
2481                 resblks = 0;
2482                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2483                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2484         }
2485         if (error) {
2486                 cancel_flags = 0;
2487                 goto error_return;
2488         }
2489
2490         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2491         unlock_dp_on_error = B_TRUE;
2492
2493         /*
2494          * Check whether the directory allows new symlinks or not.
2495          */
2496         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2497                 error = XFS_ERROR(EPERM);
2498                 goto error_return;
2499         }
2500
2501         /*
2502          * Reserve disk quota : blocks and inode.
2503          */
2504         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2505         if (error)
2506                 goto error_return;
2507
2508         /*
2509          * Check for ability to enter directory entry, if no space reserved.
2510          */
2511         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2512         if (error)
2513                 goto error_return;
2514         /*
2515          * Initialize the bmap freelist prior to calling either
2516          * bmapi or the directory create code.
2517          */
2518         XFS_BMAP_INIT(&free_list, &first_block);
2519
2520         /*
2521          * Allocate an inode for the symlink.
2522          */
2523         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT),
2524                                1, 0, credp, prid, resblks > 0, &ip, NULL);
2525         if (error) {
2526                 if (error == ENOSPC)
2527                         goto error_return;
2528                 goto error1;
2529         }
2530         xfs_itrace_ref(ip);
2531
2532         /*
2533          * An error after we've joined dp to the transaction will result in the
2534          * transaction cancel unlocking dp so don't do it explicitly in the
2535          * error path.
2536          */
2537         IHOLD(dp);
2538         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2539         unlock_dp_on_error = B_FALSE;
2540
2541         /*
2542          * Also attach the dquot(s) to it, if applicable.
2543          */
2544         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
2545
2546         if (resblks)
2547                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2548         /*
2549          * If the symlink will fit into the inode, write it inline.
2550          */
2551         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2552                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2553                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2554                 ip->i_d.di_size = pathlen;
2555
2556                 /*
2557                  * The inode was initially created in extent format.
2558                  */
2559                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2560                 ip->i_df.if_flags |= XFS_IFINLINE;
2561
2562                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2563                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2564
2565         } else {
2566                 first_fsb = 0;
2567                 nmaps = SYMLINK_MAPS;
2568
2569                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2570                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2571                                   &first_block, resblks, mval, &nmaps,
2572                                   &free_list, NULL);
2573                 if (error) {
2574                         goto error1;
2575                 }
2576
2577                 if (resblks)
2578                         resblks -= fs_blocks;
2579                 ip->i_d.di_size = pathlen;
2580                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2581
2582                 cur_chunk = target_path;
2583                 for (n = 0; n < nmaps; n++) {
2584                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2585                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2586                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2587                                                BTOBB(byte_cnt), 0);
2588                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2589                         if (pathlen < byte_cnt) {
2590                                 byte_cnt = pathlen;
2591                         }
2592                         pathlen -= byte_cnt;
2593
2594                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2595                         cur_chunk += byte_cnt;
2596
2597                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2598                 }
2599         }
2600
2601         /*
2602          * Create the directory entry for the symlink.
2603          */
2604         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2605                                         &first_block, &free_list, resblks);
2606         if (error)
2607                 goto error1;
2608         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2609         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2610
2611         /*
2612          * If this is a synchronous mount, make sure that the
2613          * symlink transaction goes to disk before returning to
2614          * the user.
2615          */
2616         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2617                 xfs_trans_set_sync(tp);
2618         }
2619
2620         /*
2621          * xfs_trans_commit normally decrements the vnode ref count
2622          * when it unlocks the inode. Since we want to return the
2623          * vnode to the caller, we bump the vnode ref count now.
2624          */
2625         IHOLD(ip);
2626
2627         error = xfs_bmap_finish(&tp, &free_list, &committed);
2628         if (error) {
2629                 goto error2;
2630         }
2631         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2632         XFS_QM_DQRELE(mp, udqp);
2633         XFS_QM_DQRELE(mp, gdqp);
2634
2635         /* Fall through to std_return with error = 0 or errno from
2636          * xfs_trans_commit     */
2637 std_return:
2638         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTSYMLINK)) {
2639                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTSYMLINK,
2640                                         dp, DM_RIGHT_NULL,
2641                                         error ? NULL : ip,
2642                                         DM_RIGHT_NULL, link_name->name,
2643                                         target_path, 0, error, 0);
2644         }
2645
2646         if (!error)
2647                 *ipp = ip;
2648         return error;
2649
2650  error2:
2651         IRELE(ip);
2652  error1:
2653         xfs_bmap_cancel(&free_list);
2654         cancel_flags |= XFS_TRANS_ABORT;
2655  error_return:
2656         xfs_trans_cancel(tp, cancel_flags);
2657         XFS_QM_DQRELE(mp, udqp);
2658         XFS_QM_DQRELE(mp, gdqp);
2659
2660         if (unlock_dp_on_error)
2661                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2662
2663         goto std_return;
2664 }
2665
2666 int
2667 xfs_inode_flush(
2668         xfs_inode_t     *ip,
2669         int             flags)
2670 {
2671         xfs_mount_t     *mp = ip->i_mount;
2672         int             error = 0;
2673
2674         if (XFS_FORCED_SHUTDOWN(mp))
2675                 return XFS_ERROR(EIO);
2676
2677         /*
2678          * Bypass inodes which have already been cleaned by
2679          * the inode flush clustering code inside xfs_iflush
2680          */
2681         if (xfs_inode_clean(ip))
2682                 return 0;
2683
2684         /*
2685          * We make this non-blocking if the inode is contended,
2686          * return EAGAIN to indicate to the caller that they
2687          * did not succeed. This prevents the flush path from
2688          * blocking on inodes inside another operation right
2689          * now, they get caught later by xfs_sync.
2690          */
2691         if (flags & FLUSH_SYNC) {
2692                 xfs_ilock(ip, XFS_ILOCK_SHARED);
2693                 xfs_iflock(ip);
2694         } else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2695                 if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
2696                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2697                         return EAGAIN;
2698                 }
2699         } else {
2700                 return EAGAIN;
2701         }
2702
2703         error = xfs_iflush(ip, (flags & FLUSH_SYNC) ? XFS_IFLUSH_SYNC
2704                                                     : XFS_IFLUSH_ASYNC_NOBLOCK);
2705         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2706
2707         return error;
2708 }
2709
2710
2711 int
2712 xfs_set_dmattrs(
2713         xfs_inode_t     *ip,
2714         u_int           evmask,
2715         u_int16_t       state)
2716 {
2717         xfs_mount_t     *mp = ip->i_mount;
2718         xfs_trans_t     *tp;
2719         int             error;
2720
2721         if (!capable(CAP_SYS_ADMIN))
2722                 return XFS_ERROR(EPERM);
2723
2724         if (XFS_FORCED_SHUTDOWN(mp))
2725                 return XFS_ERROR(EIO);
2726
2727         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2728         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2729         if (error) {
2730                 xfs_trans_cancel(tp, 0);
2731                 return error;
2732         }
2733         xfs_ilock(ip, XFS_ILOCK_EXCL);
2734         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2735
2736         ip->i_d.di_dmevmask = evmask;
2737         ip->i_d.di_dmstate  = state;
2738
2739         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2740         IHOLD(ip);
2741         error = xfs_trans_commit(tp, 0);
2742
2743         return error;
2744 }
2745
2746 int
2747 xfs_reclaim(
2748         xfs_inode_t     *ip)
2749 {
2750
2751         xfs_itrace_entry(ip);
2752
2753         ASSERT(!VN_MAPPED(VFS_I(ip)));
2754
2755         /* bad inode, get out here ASAP */
2756         if (VN_BAD(VFS_I(ip))) {
2757                 xfs_ireclaim(ip);
2758                 return 0;
2759         }
2760
2761         xfs_ioend_wait(ip);
2762
2763         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0);
2764
2765         /*
2766          * Make sure the atime in the XFS inode is correct before freeing the
2767          * Linux inode.
2768          */
2769         xfs_synchronize_atime(ip);
2770
2771         /*
2772          * If we have nothing to flush with this inode then complete the
2773          * teardown now, otherwise break the link between the xfs inode and the
2774          * linux inode and clean up the xfs inode later. This avoids flushing
2775          * the inode to disk during the delete operation itself.
2776          *
2777          * When breaking the link, we need to set the XFS_IRECLAIMABLE flag
2778          * first to ensure that xfs_iunpin() will never see an xfs inode
2779          * that has a linux inode being reclaimed. Synchronisation is provided
2780          * by the i_flags_lock.
2781          */
2782         if (!ip->i_update_core && (ip->i_itemp == NULL)) {
2783                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2784                 xfs_iflock(ip);
2785                 xfs_iflags_set(ip, XFS_IRECLAIMABLE);
2786                 return xfs_reclaim_inode(ip, 1, XFS_IFLUSH_DELWRI_ELSE_SYNC);
2787         }
2788         xfs_inode_set_reclaim_tag(ip);
2789         return 0;
2790 }
2791
2792 /*
2793  * xfs_alloc_file_space()
2794  *      This routine allocates disk space for the given file.
2795  *
2796  *      If alloc_type == 0, this request is for an ALLOCSP type
2797  *      request which will change the file size.  In this case, no
2798  *      DMAPI event will be generated by the call.  A TRUNCATE event
2799  *      will be generated later by xfs_setattr.
2800  *
2801  *      If alloc_type != 0, this request is for a RESVSP type
2802  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2803  *      lower block boundary byte address is less than the file's
2804  *      length.
2805  *
2806  * RETURNS:
2807  *       0 on success
2808  *      errno on error
2809  *
2810  */
2811 STATIC int
2812 xfs_alloc_file_space(
2813         xfs_inode_t             *ip,
2814         xfs_off_t               offset,
2815         xfs_off_t               len,
2816         int                     alloc_type,
2817         int                     attr_flags)
2818 {
2819         xfs_mount_t             *mp = ip->i_mount;
2820         xfs_off_t               count;
2821         xfs_filblks_t           allocated_fsb;
2822         xfs_filblks_t           allocatesize_fsb;
2823         xfs_extlen_t            extsz, temp;
2824         xfs_fileoff_t           startoffset_fsb;
2825         xfs_fsblock_t           firstfsb;
2826         int                     nimaps;
2827         int                     bmapi_flag;
2828         int                     quota_flag;
2829         int                     rt;
2830         xfs_trans_t             *tp;
2831         xfs_bmbt_irec_t         imaps[1], *imapp;
2832         xfs_bmap_free_t         free_list;
2833         uint                    qblocks, resblks, resrtextents;
2834         int                     committed;
2835         int                     error;
2836
2837         xfs_itrace_entry(ip);
2838
2839         if (XFS_FORCED_SHUTDOWN(mp))
2840                 return XFS_ERROR(EIO);
2841
2842         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
2843                 return error;
2844
2845         if (len <= 0)
2846                 return XFS_ERROR(EINVAL);
2847
2848         rt = XFS_IS_REALTIME_INODE(ip);
2849         extsz = xfs_get_extsz_hint(ip);
2850
2851         count = len;
2852         imapp = &imaps[0];
2853         nimaps = 1;
2854         bmapi_flag = XFS_BMAPI_WRITE | (alloc_type ? XFS_BMAPI_PREALLOC : 0);
2855         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
2856         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
2857
2858         /*      Generate a DMAPI event if needed.       */
2859         if (alloc_type != 0 && offset < ip->i_size &&
2860                         (attr_flags & XFS_ATTR_DMI) == 0  &&
2861                         DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
2862                 xfs_off_t           end_dmi_offset;
2863
2864                 end_dmi_offset = offset+len;
2865                 if (end_dmi_offset > ip->i_size)
2866                         end_dmi_offset = ip->i_size;
2867                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, offset,
2868                                       end_dmi_offset - offset, 0, NULL);
2869                 if (error)
2870                         return error;
2871         }
2872
2873         /*
2874          * Allocate file space until done or until there is an error
2875          */
2876 retry:
2877         while (allocatesize_fsb && !error) {
2878                 xfs_fileoff_t   s, e;
2879
2880                 /*
2881                  * Determine space reservations for data/realtime.
2882                  */
2883                 if (unlikely(extsz)) {
2884                         s = startoffset_fsb;
2885                         do_div(s, extsz);
2886                         s *= extsz;
2887                         e = startoffset_fsb + allocatesize_fsb;
2888                         if ((temp = do_mod(startoffset_fsb, extsz)))
2889                                 e += temp;
2890                         if ((temp = do_mod(e, extsz)))
2891                                 e += extsz - temp;
2892                 } else {
2893                         s = 0;
2894                         e = allocatesize_fsb;
2895                 }
2896
2897                 if (unlikely(rt)) {
2898                         resrtextents = qblocks = (uint)(e - s);
2899                         resrtextents /= mp->m_sb.sb_rextsize;
2900                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2901                         quota_flag = XFS_QMOPT_RES_RTBLKS;
2902                 } else {
2903                         resrtextents = 0;
2904                         resblks = qblocks = \
2905                                 XFS_DIOSTRAT_SPACE_RES(mp, (uint)(e - s));
2906                         quota_flag = XFS_QMOPT_RES_REGBLKS;
2907                 }
2908
2909                 /*
2910                  * Allocate and setup the transaction.
2911                  */
2912                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2913                 error = xfs_trans_reserve(tp, resblks,
2914                                           XFS_WRITE_LOG_RES(mp), resrtextents,
2915                                           XFS_TRANS_PERM_LOG_RES,
2916                                           XFS_WRITE_LOG_COUNT);
2917                 /*
2918                  * Check for running out of space
2919                  */
2920                 if (error) {
2921                         /*
2922                          * Free the transaction structure.
2923                          */
2924                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2925                         xfs_trans_cancel(tp, 0);
2926                         break;
2927                 }
2928                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2929                 error = XFS_TRANS_RESERVE_QUOTA_NBLKS(mp, tp, ip,
2930                                                       qblocks, 0, quota_flag);
2931                 if (error)
2932                         goto error1;
2933
2934                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2935                 xfs_trans_ihold(tp, ip);
2936
2937                 /*
2938                  * Issue the xfs_bmapi() call to allocate the blocks
2939                  */
2940                 XFS_BMAP_INIT(&free_list, &firstfsb);
2941                 error = xfs_bmapi(tp, ip, startoffset_fsb,
2942                                   allocatesize_fsb, bmapi_flag,
2943                                   &firstfsb, 0, imapp, &nimaps,
2944                                   &free_list, NULL);
2945                 if (error) {
2946                         goto error0;
2947                 }
2948
2949                 /*
2950                  * Complete the transaction
2951                  */
2952                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2953                 if (error) {
2954                         goto error0;
2955                 }
2956
2957                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2958                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2959                 if (error) {
2960                         break;
2961                 }
2962
2963                 allocated_fsb = imapp->br_blockcount;
2964
2965                 if (nimaps == 0) {
2966                         error = XFS_ERROR(ENOSPC);
2967                         break;
2968                 }
2969
2970                 startoffset_fsb += allocated_fsb;
2971                 allocatesize_fsb -= allocated_fsb;
2972         }
2973 dmapi_enospc_check:
2974         if (error == ENOSPC && (attr_flags & XFS_ATTR_DMI) == 0 &&
2975             DM_EVENT_ENABLED(ip, DM_EVENT_NOSPACE)) {
2976                 error = XFS_SEND_NAMESP(mp, DM_EVENT_NOSPACE,
2977                                 ip, DM_RIGHT_NULL,
2978                                 ip, DM_RIGHT_NULL,
2979                                 NULL, NULL, 0, 0, 0); /* Delay flag intentionally unused */
2980                 if (error == 0)
2981                         goto retry;     /* Maybe DMAPI app. has made space */
2982                 /* else fall through with error from XFS_SEND_DATA */
2983         }
2984
2985         return error;
2986
2987 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
2988         xfs_bmap_cancel(&free_list);
2989         XFS_TRANS_UNRESERVE_QUOTA_NBLKS(mp, tp, ip, qblocks, 0, quota_flag);
2990
2991 error1: /* Just cancel transaction */
2992         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2993         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2994         goto dmapi_enospc_check;
2995 }
2996
2997 /*
2998  * Zero file bytes between startoff and endoff inclusive.
2999  * The iolock is held exclusive and no blocks are buffered.
3000  *
3001  * This function is used by xfs_free_file_space() to zero
3002  * partial blocks when the range to free is not block aligned.
3003  * When unreserving space with boundaries that are not block
3004  * aligned we round up the start and round down the end
3005  * boundaries and then use this function to zero the parts of
3006  * the blocks that got dropped during the rounding.
3007  */
3008 STATIC int
3009 xfs_zero_remaining_bytes(
3010         xfs_inode_t             *ip,
3011         xfs_off_t               startoff,
3012         xfs_off_t               endoff)
3013 {
3014         xfs_bmbt_irec_t         imap;
3015         xfs_fileoff_t           offset_fsb;
3016         xfs_off_t               lastoffset;
3017         xfs_off_t               offset;
3018         xfs_buf_t               *bp;
3019         xfs_mount_t             *mp = ip->i_mount;
3020         int                     nimap;
3021         int                     error = 0;
3022
3023         /*
3024          * Avoid doing I/O beyond eof - it's not necessary
3025          * since nothing can read beyond eof.  The space will
3026          * be zeroed when the file is extended anyway.
3027          */
3028         if (startoff >= ip->i_size)
3029                 return 0;
3030
3031         if (endoff > ip->i_size)
3032                 endoff = ip->i_size;
3033
3034         bp = xfs_buf_get_noaddr(mp->m_sb.sb_blocksize,
3035                                 XFS_IS_REALTIME_INODE(ip) ?
3036                                 mp->m_rtdev_targp : mp->m_ddev_targp);
3037         if (!bp)
3038                 return XFS_ERROR(ENOMEM);
3039
3040         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
3041                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
3042                 nimap = 1;
3043                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
3044                         NULL, 0, &imap, &nimap, NULL, NULL);
3045                 if (error || nimap < 1)
3046                         break;
3047                 ASSERT(imap.br_blockcount >= 1);
3048                 ASSERT(imap.br_startoff == offset_fsb);
3049                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
3050                 if (lastoffset > endoff)
3051                         lastoffset = endoff;
3052                 if (imap.br_startblock == HOLESTARTBLOCK)
3053                         continue;
3054                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3055                 if (imap.br_state == XFS_EXT_UNWRITTEN)
3056                         continue;
3057                 XFS_BUF_UNDONE(bp);
3058                 XFS_BUF_UNWRITE(bp);
3059                 XFS_BUF_READ(bp);
3060                 XFS_BUF_SET_ADDR(bp, XFS_FSB_TO_DB(ip, imap.br_startblock));
3061                 xfsbdstrat(mp, bp);
3062                 error = xfs_iowait(bp);
3063                 if (error) {
3064                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
3065                                           mp, bp, XFS_BUF_ADDR(bp));
3066                         break;
3067                 }
3068                 memset(XFS_BUF_PTR(bp) +
3069                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
3070                       0, lastoffset - offset + 1);
3071                 XFS_BUF_UNDONE(bp);
3072                 XFS_BUF_UNREAD(bp);
3073                 XFS_BUF_WRITE(bp);
3074                 xfsbdstrat(mp, bp);
3075                 error = xfs_iowait(bp);
3076                 if (error) {
3077                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
3078                                           mp, bp, XFS_BUF_ADDR(bp));
3079                         break;
3080                 }
3081         }
3082         xfs_buf_free(bp);
3083         return error;
3084 }
3085
3086 /*
3087  * xfs_free_file_space()
3088  *      This routine frees disk space for the given file.
3089  *
3090  *      This routine is only called by xfs_change_file_space
3091  *      for an UNRESVSP type call.
3092  *
3093  * RETURNS:
3094  *       0 on success
3095  *      errno on error
3096  *
3097  */
3098 STATIC int
3099 xfs_free_file_space(
3100         xfs_inode_t             *ip,
3101         xfs_off_t               offset,
3102         xfs_off_t               len,
3103         int                     attr_flags)
3104 {
3105         int                     committed;
3106         int                     done;
3107         xfs_off_t               end_dmi_offset;
3108         xfs_fileoff_t           endoffset_fsb;
3109         int                     error;
3110         xfs_fsblock_t           firstfsb;
3111         xfs_bmap_free_t         free_list;
3112         xfs_bmbt_irec_t         imap;
3113         xfs_off_t               ioffset;
3114         xfs_extlen_t            mod=0;
3115         xfs_mount_t             *mp;
3116         int                     nimap;
3117         uint                    resblks;
3118         uint                    rounding;
3119         int                     rt;
3120         xfs_fileoff_t           startoffset_fsb;
3121         xfs_trans_t             *tp;
3122         int                     need_iolock = 1;
3123
3124         mp = ip->i_mount;
3125
3126         xfs_itrace_entry(ip);
3127
3128         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
3129                 return error;
3130
3131         error = 0;
3132         if (len <= 0)   /* if nothing being freed */
3133                 return error;
3134         rt = XFS_IS_REALTIME_INODE(ip);
3135         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
3136         end_dmi_offset = offset + len;
3137         endoffset_fsb = XFS_B_TO_FSBT(mp, end_dmi_offset);
3138
3139         if (offset < ip->i_size && (attr_flags & XFS_ATTR_DMI) == 0 &&
3140             DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3141                 if (end_dmi_offset > ip->i_size)
3142                         end_dmi_offset = ip->i_size;
3143                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip,
3144                                 offset, end_dmi_offset - offset,
3145                                 AT_DELAY_FLAG(attr_flags), NULL);
3146                 if (error)
3147                         return error;
3148         }
3149
3150         if (attr_flags & XFS_ATTR_NOLOCK)
3151                 need_iolock = 0;
3152         if (need_iolock) {
3153                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
3154                 /* wait for the completion of any pending DIOs */
3155                 xfs_ioend_wait(ip);
3156         }
3157
3158         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
3159         ioffset = offset & ~(rounding - 1);
3160
3161         if (VN_CACHED(VFS_I(ip)) != 0) {
3162                 xfs_inval_cached_trace(ip, ioffset, -1, ioffset, -1);
3163                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
3164                 if (error)
3165                         goto out_unlock_iolock;
3166         }
3167
3168         /*
3169          * Need to zero the stuff we're not freeing, on disk.
3170          * If its a realtime file & can't use unwritten extents then we
3171          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
3172          * will take care of it for us.
3173          */
3174         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
3175                 nimap = 1;
3176                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
3177                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3178                 if (error)
3179                         goto out_unlock_iolock;
3180                 ASSERT(nimap == 0 || nimap == 1);
3181                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3182                         xfs_daddr_t     block;
3183
3184                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3185                         block = imap.br_startblock;
3186                         mod = do_div(block, mp->m_sb.sb_rextsize);
3187                         if (mod)
3188                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
3189                 }
3190                 nimap = 1;
3191                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
3192                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3193                 if (error)
3194                         goto out_unlock_iolock;
3195                 ASSERT(nimap == 0 || nimap == 1);
3196                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3197                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3198                         mod++;
3199                         if (mod && (mod != mp->m_sb.sb_rextsize))
3200                                 endoffset_fsb -= mod;
3201                 }
3202         }
3203         if ((done = (endoffset_fsb <= startoffset_fsb)))
3204                 /*
3205                  * One contiguous piece to clear
3206                  */
3207                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
3208         else {
3209                 /*
3210                  * Some full blocks, possibly two pieces to clear
3211                  */
3212                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
3213                         error = xfs_zero_remaining_bytes(ip, offset,
3214                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
3215                 if (!error &&
3216                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
3217                         error = xfs_zero_remaining_bytes(ip,
3218                                 XFS_FSB_TO_B(mp, endoffset_fsb),
3219                                 offset + len - 1);
3220         }
3221
3222         /*
3223          * free file space until done or until there is an error
3224          */
3225         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3226         while (!error && !done) {
3227
3228                 /*
3229                  * allocate and setup the transaction. Allow this
3230                  * transaction to dip into the reserve blocks to ensure
3231                  * the freeing of the space succeeds at ENOSPC.
3232                  */
3233                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3234                 tp->t_flags |= XFS_TRANS_RESERVE;
3235                 error = xfs_trans_reserve(tp,
3236                                           resblks,
3237                                           XFS_WRITE_LOG_RES(mp),
3238                                           0,
3239                                           XFS_TRANS_PERM_LOG_RES,
3240                                           XFS_WRITE_LOG_COUNT);
3241
3242                 /*
3243                  * check for running out of space
3244                  */
3245                 if (error) {
3246                         /*
3247                          * Free the transaction structure.
3248                          */
3249                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3250                         xfs_trans_cancel(tp, 0);
3251                         break;
3252                 }
3253                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3254                 error = XFS_TRANS_RESERVE_QUOTA(mp, tp,
3255                                 ip->i_udquot, ip->i_gdquot, resblks, 0,
3256                                 XFS_QMOPT_RES_REGBLKS);
3257                 if (error)
3258                         goto error1;
3259
3260                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3261                 xfs_trans_ihold(tp, ip);
3262
3263                 /*
3264                  * issue the bunmapi() call to free the blocks
3265                  */
3266                 XFS_BMAP_INIT(&free_list, &firstfsb);
3267                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
3268                                   endoffset_fsb - startoffset_fsb,
3269                                   0, 2, &firstfsb, &free_list, NULL, &done);
3270                 if (error) {
3271                         goto error0;
3272                 }
3273
3274                 /*
3275                  * complete the transaction
3276                  */
3277                 error = xfs_bmap_finish(&tp, &free_list, &committed);
3278                 if (error) {
3279                         goto error0;
3280                 }
3281
3282                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3283                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3284         }
3285
3286  out_unlock_iolock:
3287         if (need_iolock)
3288                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
3289         return error;
3290
3291  error0:
3292         xfs_bmap_cancel(&free_list);
3293  error1:
3294         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
3295         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
3296                     XFS_ILOCK_EXCL);
3297         return error;
3298 }
3299
3300 /*
3301  * xfs_change_file_space()
3302  *      This routine allocates or frees disk space for the given file.
3303  *      The user specified parameters are checked for alignment and size
3304  *      limitations.
3305  *
3306  * RETURNS:
3307  *       0 on success
3308  *      errno on error
3309  *
3310  */
3311 int
3312 xfs_change_file_space(
3313         xfs_inode_t     *ip,
3314         int             cmd,
3315         xfs_flock64_t   *bf,
3316         xfs_off_t       offset,
3317         int             attr_flags)
3318 {
3319         xfs_mount_t     *mp = ip->i_mount;
3320         int             clrprealloc;
3321         int             error;
3322         xfs_fsize_t     fsize;
3323         int             setprealloc;
3324         xfs_off_t       startoffset;
3325         xfs_off_t       llen;
3326         xfs_trans_t     *tp;
3327         struct iattr    iattr;
3328
3329         xfs_itrace_entry(ip);
3330
3331         if (!S_ISREG(ip->i_d.di_mode))
3332                 return XFS_ERROR(EINVAL);
3333
3334         switch (bf->l_whence) {
3335         case 0: /*SEEK_SET*/
3336                 break;
3337         case 1: /*SEEK_CUR*/
3338                 bf->l_start += offset;
3339                 break;
3340         case 2: /*SEEK_END*/
3341                 bf->l_start += ip->i_size;
3342                 break;
3343         default:
3344                 return XFS_ERROR(EINVAL);
3345         }
3346
3347         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
3348
3349         if (   (bf->l_start < 0)
3350             || (bf->l_start > XFS_MAXIOFFSET(mp))
3351             || (bf->l_start + llen < 0)
3352             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
3353                 return XFS_ERROR(EINVAL);
3354
3355         bf->l_whence = 0;
3356
3357         startoffset = bf->l_start;
3358         fsize = ip->i_size;
3359
3360         /*
3361          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
3362          * file space.
3363          * These calls do NOT zero the data space allocated to the file,
3364          * nor do they change the file size.
3365          *
3366          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
3367          * space.
3368          * These calls cause the new file data to be zeroed and the file
3369          * size to be changed.
3370          */
3371         setprealloc = clrprealloc = 0;
3372
3373         switch (cmd) {
3374         case XFS_IOC_RESVSP:
3375         case XFS_IOC_RESVSP64:
3376                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
3377                                                                 1, attr_flags);
3378                 if (error)
3379                         return error;
3380                 setprealloc = 1;
3381                 break;
3382
3383         case XFS_IOC_UNRESVSP:
3384         case XFS_IOC_UNRESVSP64:
3385                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
3386                                                                 attr_flags)))
3387                         return error;
3388                 break;
3389
3390         case XFS_IOC_ALLOCSP:
3391         case XFS_IOC_ALLOCSP64:
3392         case XFS_IOC_FREESP:
3393         case XFS_IOC_FREESP64:
3394                 if (startoffset > fsize) {
3395                         error = xfs_alloc_file_space(ip, fsize,
3396                                         startoffset - fsize, 0, attr_flags);
3397                         if (error)
3398                                 break;
3399                 }
3400
3401                 iattr.ia_valid = ATTR_SIZE;
3402                 iattr.ia_size = startoffset;
3403
3404                 error = xfs_setattr(ip, &iattr, attr_flags);
3405
3406                 if (error)
3407                         return error;
3408
3409                 clrprealloc = 1;
3410                 break;
3411
3412         default:
3413                 ASSERT(0);
3414                 return XFS_ERROR(EINVAL);
3415         }
3416
3417         /*
3418          * update the inode timestamp, mode, and prealloc flag bits
3419          */
3420         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
3421
3422         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
3423                                       0, 0, 0))) {
3424                 /* ASSERT(0); */
3425                 xfs_trans_cancel(tp, 0);
3426                 return error;
3427         }
3428
3429         xfs_ilock(ip, XFS_ILOCK_EXCL);
3430
3431         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3432         xfs_trans_ihold(tp, ip);
3433
3434         if ((attr_flags & XFS_ATTR_DMI) == 0) {
3435                 ip->i_d.di_mode &= ~S_ISUID;
3436
3437                 /*
3438                  * Note that we don't have to worry about mandatory
3439                  * file locking being disabled here because we only
3440                  * clear the S_ISGID bit if the Group execute bit is
3441                  * on, but if it was on then mandatory locking wouldn't
3442                  * have been enabled.
3443                  */
3444                 if (ip->i_d.di_mode & S_IXGRP)
3445                         ip->i_d.di_mode &= ~S_ISGID;
3446
3447                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3448         }
3449         if (setprealloc)
3450                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
3451         else if (clrprealloc)
3452                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
3453
3454         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
3455         xfs_trans_set_sync(tp);
3456
3457         error = xfs_trans_commit(tp, 0);
3458
3459         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3460
3461         return error;
3462 }