[XFS] kill bhv_vnode_t
[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_open(
58         xfs_inode_t     *ip)
59 {
60         int             mode;
61
62         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
63                 return XFS_ERROR(EIO);
64
65         /*
66          * If it's a directory with any blocks, read-ahead block 0
67          * as we're almost certain to have the next operation be a read there.
68          */
69         if (S_ISDIR(ip->i_d.di_mode) && ip->i_d.di_nextents > 0) {
70                 mode = xfs_ilock_map_shared(ip);
71                 if (ip->i_d.di_nextents > 0)
72                         (void)xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
73                 xfs_iunlock(ip, mode);
74         }
75         return 0;
76 }
77
78 int
79 xfs_setattr(
80         struct xfs_inode        *ip,
81         struct iattr            *iattr,
82         int                     flags,
83         cred_t                  *credp)
84 {
85         xfs_mount_t             *mp = ip->i_mount;
86         struct inode            *inode = VFS_I(ip);
87         int                     mask = iattr->ia_valid;
88         xfs_trans_t             *tp;
89         int                     code;
90         uint                    lock_flags;
91         uint                    commit_flags=0;
92         uid_t                   uid=0, iuid=0;
93         gid_t                   gid=0, igid=0;
94         int                     timeflags = 0;
95         struct xfs_dquot        *udqp, *gdqp, *olddquot1, *olddquot2;
96         int                     file_owner;
97         int                     need_iolock = 1;
98
99         xfs_itrace_entry(ip);
100
101         if (mp->m_flags & XFS_MOUNT_RDONLY)
102                 return XFS_ERROR(EROFS);
103
104         if (XFS_FORCED_SHUTDOWN(mp))
105                 return XFS_ERROR(EIO);
106
107         olddquot1 = olddquot2 = NULL;
108         udqp = gdqp = NULL;
109
110         /*
111          * If disk quotas is on, we make sure that the dquots do exist on disk,
112          * before we start any other transactions. Trying to do this later
113          * is messy. We don't care to take a readlock to look at the ids
114          * in inode here, because we can't hold it across the trans_reserve.
115          * If the IDs do change before we take the ilock, we're covered
116          * because the i_*dquot fields will get updated anyway.
117          */
118         if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
119                 uint    qflags = 0;
120
121                 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
122                         uid = iattr->ia_uid;
123                         qflags |= XFS_QMOPT_UQUOTA;
124                 } else {
125                         uid = ip->i_d.di_uid;
126                 }
127                 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
128                         gid = iattr->ia_gid;
129                         qflags |= XFS_QMOPT_GQUOTA;
130                 }  else {
131                         gid = ip->i_d.di_gid;
132                 }
133
134                 /*
135                  * We take a reference when we initialize udqp and gdqp,
136                  * so it is important that we never blindly double trip on
137                  * the same variable. See xfs_create() for an example.
138                  */
139                 ASSERT(udqp == NULL);
140                 ASSERT(gdqp == NULL);
141                 code = XFS_QM_DQVOPALLOC(mp, ip, uid, gid, ip->i_d.di_projid,
142                                          qflags, &udqp, &gdqp);
143                 if (code)
144                         return code;
145         }
146
147         /*
148          * For the other attributes, we acquire the inode lock and
149          * first do an error checking pass.
150          */
151         tp = NULL;
152         lock_flags = XFS_ILOCK_EXCL;
153         if (flags & XFS_ATTR_NOLOCK)
154                 need_iolock = 0;
155         if (!(mask & ATTR_SIZE)) {
156                 if ((mask != (ATTR_CTIME|ATTR_ATIME|ATTR_MTIME)) ||
157                     (mp->m_flags & XFS_MOUNT_WSYNC)) {
158                         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
159                         commit_flags = 0;
160                         if ((code = xfs_trans_reserve(tp, 0,
161                                                      XFS_ICHANGE_LOG_RES(mp), 0,
162                                                      0, 0))) {
163                                 lock_flags = 0;
164                                 goto error_return;
165                         }
166                 }
167         } else {
168                 if (DM_EVENT_ENABLED(ip, DM_EVENT_TRUNCATE) &&
169                     !(flags & XFS_ATTR_DMI)) {
170                         int dmflags = AT_DELAY_FLAG(flags) | DM_SEM_FLAG_WR;
171                         code = XFS_SEND_DATA(mp, DM_EVENT_TRUNCATE, ip,
172                                 iattr->ia_size, 0, dmflags, NULL);
173                         if (code) {
174                                 lock_flags = 0;
175                                 goto error_return;
176                         }
177                 }
178                 if (need_iolock)
179                         lock_flags |= XFS_IOLOCK_EXCL;
180         }
181
182         xfs_ilock(ip, lock_flags);
183
184         /* boolean: are we the file owner? */
185         file_owner = (current_fsuid(credp) == ip->i_d.di_uid);
186
187         /*
188          * Change various properties of a file.
189          * Only the owner or users with CAP_FOWNER
190          * capability may do these things.
191          */
192         if (mask & (ATTR_MODE|ATTR_UID|ATTR_GID)) {
193                 /*
194                  * CAP_FOWNER overrides the following restrictions:
195                  *
196                  * The user ID of the calling process must be equal
197                  * to the file owner ID, except in cases where the
198                  * CAP_FSETID capability is applicable.
199                  */
200                 if (!file_owner && !capable(CAP_FOWNER)) {
201                         code = XFS_ERROR(EPERM);
202                         goto error_return;
203                 }
204
205                 /*
206                  * CAP_FSETID overrides the following restrictions:
207                  *
208                  * The effective user ID of the calling process shall match
209                  * the file owner when setting the set-user-ID and
210                  * set-group-ID bits on that file.
211                  *
212                  * The effective group ID or one of the supplementary group
213                  * IDs of the calling process shall match the group owner of
214                  * the file when setting the set-group-ID bit on that file
215                  */
216                 if (mask & ATTR_MODE) {
217                         mode_t m = 0;
218
219                         if ((iattr->ia_mode & S_ISUID) && !file_owner)
220                                 m |= S_ISUID;
221                         if ((iattr->ia_mode & S_ISGID) &&
222                             !in_group_p((gid_t)ip->i_d.di_gid))
223                                 m |= S_ISGID;
224 #if 0
225                         /* Linux allows this, Irix doesn't. */
226                         if ((iattr->ia_mode & S_ISVTX) && !S_ISDIR(ip->i_d.di_mode))
227                                 m |= S_ISVTX;
228 #endif
229                         if (m && !capable(CAP_FSETID))
230                                 iattr->ia_mode &= ~m;
231                 }
232         }
233
234         /*
235          * Change file ownership.  Must be the owner or privileged.
236          * If the system was configured with the "restricted_chown"
237          * option, the owner is not permitted to give away the file,
238          * and can change the group id only to a group of which he
239          * or she is a member.
240          */
241         if (mask & (ATTR_UID|ATTR_GID)) {
242                 /*
243                  * These IDs could have changed since we last looked at them.
244                  * But, we're assured that if the ownership did change
245                  * while we didn't have the inode locked, inode's dquot(s)
246                  * would have changed also.
247                  */
248                 iuid = ip->i_d.di_uid;
249                 igid = ip->i_d.di_gid;
250                 gid = (mask & ATTR_GID) ? iattr->ia_gid : igid;
251                 uid = (mask & ATTR_UID) ? iattr->ia_uid : iuid;
252
253                 /*
254                  * CAP_CHOWN overrides the following restrictions:
255                  *
256                  * If _POSIX_CHOWN_RESTRICTED is defined, this capability
257                  * shall override the restriction that a process cannot
258                  * change the user ID of a file it owns and the restriction
259                  * that the group ID supplied to the chown() function
260                  * shall be equal to either the group ID or one of the
261                  * supplementary group IDs of the calling process.
262                  */
263                 if (restricted_chown &&
264                     (iuid != uid || (igid != gid &&
265                                      !in_group_p((gid_t)gid))) &&
266                     !capable(CAP_CHOWN)) {
267                         code = XFS_ERROR(EPERM);
268                         goto error_return;
269                 }
270                 /*
271                  * Do a quota reservation only if uid/gid is actually
272                  * going to change.
273                  */
274                 if ((XFS_IS_UQUOTA_ON(mp) && iuid != uid) ||
275                     (XFS_IS_GQUOTA_ON(mp) && igid != gid)) {
276                         ASSERT(tp);
277                         code = XFS_QM_DQVOPCHOWNRESV(mp, tp, ip, udqp, gdqp,
278                                                 capable(CAP_FOWNER) ?
279                                                 XFS_QMOPT_FORCE_RES : 0);
280                         if (code)       /* out of quota */
281                                 goto error_return;
282                 }
283         }
284
285         /*
286          * Truncate file.  Must have write permission and not be a directory.
287          */
288         if (mask & ATTR_SIZE) {
289                 /* Short circuit the truncate case for zero length files */
290                 if (iattr->ia_size == 0 &&
291                     ip->i_size == 0 && ip->i_d.di_nextents == 0) {
292                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
293                         lock_flags &= ~XFS_ILOCK_EXCL;
294                         if (mask & ATTR_CTIME)
295                                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
296                         code = 0;
297                         goto error_return;
298                 }
299
300                 if (S_ISDIR(ip->i_d.di_mode)) {
301                         code = XFS_ERROR(EISDIR);
302                         goto error_return;
303                 } else if (!S_ISREG(ip->i_d.di_mode)) {
304                         code = XFS_ERROR(EINVAL);
305                         goto error_return;
306                 }
307                 /*
308                  * Make sure that the dquots are attached to the inode.
309                  */
310                 if ((code = XFS_QM_DQATTACH(mp, ip, XFS_QMOPT_ILOCKED)))
311                         goto error_return;
312         }
313
314         /*
315          * Change file access or modified times.
316          */
317         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
318                 if (!file_owner) {
319                         if ((mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)) &&
320                             !capable(CAP_FOWNER)) {
321                                 code = XFS_ERROR(EPERM);
322                                 goto error_return;
323                         }
324                 }
325         }
326
327         /*
328          * Now we can make the changes.  Before we join the inode
329          * to the transaction, if ATTR_SIZE is set then take care of
330          * the part of the truncation that must be done without the
331          * inode lock.  This needs to be done before joining the inode
332          * to the transaction, because the inode cannot be unlocked
333          * once it is a part of the transaction.
334          */
335         if (mask & ATTR_SIZE) {
336                 code = 0;
337                 if (iattr->ia_size > ip->i_size) {
338                         /*
339                          * Do the first part of growing a file: zero any data
340                          * in the last block that is beyond the old EOF.  We
341                          * need to do this before the inode is joined to the
342                          * transaction to modify the i_size.
343                          */
344                         code = xfs_zero_eof(ip, iattr->ia_size, ip->i_size);
345                 }
346                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
347
348                 /*
349                  * We are going to log the inode size change in this
350                  * transaction so any previous writes that are beyond the on
351                  * disk EOF and the new EOF that have not been written out need
352                  * to be written here. If we do not write the data out, we
353                  * expose ourselves to the null files problem.
354                  *
355                  * Only flush from the on disk size to the smaller of the in
356                  * memory file size or the new size as that's the range we
357                  * really care about here and prevents waiting for other data
358                  * not within the range we care about here.
359                  */
360                 if (!code &&
361                     ip->i_size != ip->i_d.di_size &&
362                     iattr->ia_size > ip->i_d.di_size) {
363                         code = xfs_flush_pages(ip,
364                                         ip->i_d.di_size, iattr->ia_size,
365                                         XFS_B_ASYNC, FI_NONE);
366                 }
367
368                 /* wait for all I/O to complete */
369                 vn_iowait(ip);
370
371                 if (!code)
372                         code = xfs_itruncate_data(ip, iattr->ia_size);
373                 if (code) {
374                         ASSERT(tp == NULL);
375                         lock_flags &= ~XFS_ILOCK_EXCL;
376                         ASSERT(lock_flags == XFS_IOLOCK_EXCL);
377                         goto error_return;
378                 }
379                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
380                 if ((code = xfs_trans_reserve(tp, 0,
381                                              XFS_ITRUNCATE_LOG_RES(mp), 0,
382                                              XFS_TRANS_PERM_LOG_RES,
383                                              XFS_ITRUNCATE_LOG_COUNT))) {
384                         xfs_trans_cancel(tp, 0);
385                         if (need_iolock)
386                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
387                         return code;
388                 }
389                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
390                 xfs_ilock(ip, XFS_ILOCK_EXCL);
391         }
392
393         if (tp) {
394                 xfs_trans_ijoin(tp, ip, lock_flags);
395                 xfs_trans_ihold(tp, ip);
396         }
397
398         /*
399          * Truncate file.  Must have write permission and not be a directory.
400          */
401         if (mask & ATTR_SIZE) {
402                 /*
403                  * Only change the c/mtime if we are changing the size
404                  * or we are explicitly asked to change it. This handles
405                  * the semantic difference between truncate() and ftruncate()
406                  * as implemented in the VFS.
407                  */
408                 if (iattr->ia_size != ip->i_size || (mask & ATTR_CTIME))
409                         timeflags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
410
411                 if (iattr->ia_size > ip->i_size) {
412                         ip->i_d.di_size = iattr->ia_size;
413                         ip->i_size = iattr->ia_size;
414                         if (!(flags & XFS_ATTR_DMI))
415                                 xfs_ichgtime(ip, XFS_ICHGTIME_CHG);
416                         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
417                 } else if (iattr->ia_size <= ip->i_size ||
418                            (iattr->ia_size == 0 && ip->i_d.di_nextents)) {
419                         /*
420                          * signal a sync transaction unless
421                          * we're truncating an already unlinked
422                          * file on a wsync filesystem
423                          */
424                         code = xfs_itruncate_finish(&tp, ip, iattr->ia_size,
425                                             XFS_DATA_FORK,
426                                             ((ip->i_d.di_nlink != 0 ||
427                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
428                                              ? 1 : 0));
429                         if (code)
430                                 goto abort_return;
431                         /*
432                          * Truncated "down", so we're removing references
433                          * to old data here - if we now delay flushing for
434                          * a long time, we expose ourselves unduly to the
435                          * notorious NULL files problem.  So, we mark this
436                          * vnode and flush it when the file is closed, and
437                          * do not wait the usual (long) time for writeout.
438                          */
439                         xfs_iflags_set(ip, XFS_ITRUNCATED);
440                 }
441         }
442
443         /*
444          * Change file access modes.
445          */
446         if (mask & ATTR_MODE) {
447                 ip->i_d.di_mode &= S_IFMT;
448                 ip->i_d.di_mode |= iattr->ia_mode & ~S_IFMT;
449
450                 inode->i_mode &= S_IFMT;
451                 inode->i_mode |= iattr->ia_mode & ~S_IFMT;
452
453                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
454                 timeflags |= XFS_ICHGTIME_CHG;
455         }
456
457         /*
458          * Change file ownership.  Must be the owner or privileged.
459          * If the system was configured with the "restricted_chown"
460          * option, the owner is not permitted to give away the file,
461          * and can change the group id only to a group of which he
462          * or she is a member.
463          */
464         if (mask & (ATTR_UID|ATTR_GID)) {
465                 /*
466                  * CAP_FSETID overrides the following restrictions:
467                  *
468                  * The set-user-ID and set-group-ID bits of a file will be
469                  * cleared upon successful return from chown()
470                  */
471                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
472                     !capable(CAP_FSETID)) {
473                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
474                 }
475
476                 /*
477                  * Change the ownerships and register quota modifications
478                  * in the transaction.
479                  */
480                 if (iuid != uid) {
481                         if (XFS_IS_UQUOTA_ON(mp)) {
482                                 ASSERT(mask & ATTR_UID);
483                                 ASSERT(udqp);
484                                 olddquot1 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
485                                                         &ip->i_udquot, udqp);
486                         }
487                         ip->i_d.di_uid = uid;
488                         inode->i_uid = uid;
489                 }
490                 if (igid != gid) {
491                         if (XFS_IS_GQUOTA_ON(mp)) {
492                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
493                                 ASSERT(mask & ATTR_GID);
494                                 ASSERT(gdqp);
495                                 olddquot2 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
496                                                         &ip->i_gdquot, gdqp);
497                         }
498                         ip->i_d.di_gid = gid;
499                         inode->i_gid = gid;
500                 }
501
502                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
503                 timeflags |= XFS_ICHGTIME_CHG;
504         }
505
506
507         /*
508          * Change file access or modified times.
509          */
510         if (mask & (ATTR_ATIME|ATTR_MTIME)) {
511                 if (mask & ATTR_ATIME) {
512                         inode->i_atime = iattr->ia_atime;
513                         ip->i_d.di_atime.t_sec = iattr->ia_atime.tv_sec;
514                         ip->i_d.di_atime.t_nsec = iattr->ia_atime.tv_nsec;
515                         ip->i_update_core = 1;
516                         timeflags &= ~XFS_ICHGTIME_ACC;
517                 }
518                 if (mask & ATTR_MTIME) {
519                         inode->i_mtime = iattr->ia_mtime;
520                         ip->i_d.di_mtime.t_sec = iattr->ia_mtime.tv_sec;
521                         ip->i_d.di_mtime.t_nsec = iattr->ia_mtime.tv_nsec;
522                         timeflags &= ~XFS_ICHGTIME_MOD;
523                         timeflags |= XFS_ICHGTIME_CHG;
524                 }
525                 if (tp && (mask & (ATTR_MTIME_SET|ATTR_ATIME_SET)))
526                         xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
527         }
528
529         /*
530          * Change file inode change time only if ATTR_CTIME set
531          * AND we have been called by a DMI function.
532          */
533
534         if ((flags & XFS_ATTR_DMI) && (mask & ATTR_CTIME)) {
535                 inode->i_ctime = iattr->ia_ctime;
536                 ip->i_d.di_ctime.t_sec = iattr->ia_ctime.tv_sec;
537                 ip->i_d.di_ctime.t_nsec = iattr->ia_ctime.tv_nsec;
538                 ip->i_update_core = 1;
539                 timeflags &= ~XFS_ICHGTIME_CHG;
540         }
541
542         /*
543          * Send out timestamp changes that need to be set to the
544          * current time.  Not done when called by a DMI function.
545          */
546         if (timeflags && !(flags & XFS_ATTR_DMI))
547                 xfs_ichgtime(ip, timeflags);
548
549         XFS_STATS_INC(xs_ig_attrchg);
550
551         /*
552          * If this is a synchronous mount, make sure that the
553          * transaction goes to disk before returning to the user.
554          * This is slightly sub-optimal in that truncates require
555          * two sync transactions instead of one for wsync filesystems.
556          * One for the truncate and one for the timestamps since we
557          * don't want to change the timestamps unless we're sure the
558          * truncate worked.  Truncates are less than 1% of the laddis
559          * mix so this probably isn't worth the trouble to optimize.
560          */
561         code = 0;
562         if (tp) {
563                 if (mp->m_flags & XFS_MOUNT_WSYNC)
564                         xfs_trans_set_sync(tp);
565
566                 code = xfs_trans_commit(tp, commit_flags);
567         }
568
569         xfs_iunlock(ip, lock_flags);
570
571         /*
572          * Release any dquot(s) the inode had kept before chown.
573          */
574         XFS_QM_DQRELE(mp, olddquot1);
575         XFS_QM_DQRELE(mp, olddquot2);
576         XFS_QM_DQRELE(mp, udqp);
577         XFS_QM_DQRELE(mp, gdqp);
578
579         if (code) {
580                 return code;
581         }
582
583         if (DM_EVENT_ENABLED(ip, DM_EVENT_ATTRIBUTE) &&
584             !(flags & XFS_ATTR_DMI)) {
585                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_ATTRIBUTE, ip, DM_RIGHT_NULL,
586                                         NULL, DM_RIGHT_NULL, NULL, NULL,
587                                         0, 0, AT_DELAY_FLAG(flags));
588         }
589         return 0;
590
591  abort_return:
592         commit_flags |= XFS_TRANS_ABORT;
593         /* FALLTHROUGH */
594  error_return:
595         XFS_QM_DQRELE(mp, udqp);
596         XFS_QM_DQRELE(mp, gdqp);
597         if (tp) {
598                 xfs_trans_cancel(tp, commit_flags);
599         }
600         if (lock_flags != 0) {
601                 xfs_iunlock(ip, lock_flags);
602         }
603         return code;
604 }
605
606 /*
607  * The maximum pathlen is 1024 bytes. Since the minimum file system
608  * blocksize is 512 bytes, we can get a max of 2 extents back from
609  * bmapi.
610  */
611 #define SYMLINK_MAPS 2
612
613 STATIC int
614 xfs_readlink_bmap(
615         xfs_inode_t     *ip,
616         char            *link)
617 {
618         xfs_mount_t     *mp = ip->i_mount;
619         int             pathlen = ip->i_d.di_size;
620         int             nmaps = SYMLINK_MAPS;
621         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
622         xfs_daddr_t     d;
623         int             byte_cnt;
624         int             n;
625         xfs_buf_t       *bp;
626         int             error = 0;
627
628         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
629                         mval, &nmaps, NULL, NULL);
630         if (error)
631                 goto out;
632
633         for (n = 0; n < nmaps; n++) {
634                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
635                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
636
637                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
638                 error = XFS_BUF_GETERROR(bp);
639                 if (error) {
640                         xfs_ioerror_alert("xfs_readlink",
641                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
642                         xfs_buf_relse(bp);
643                         goto out;
644                 }
645                 if (pathlen < byte_cnt)
646                         byte_cnt = pathlen;
647                 pathlen -= byte_cnt;
648
649                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
650                 xfs_buf_relse(bp);
651         }
652
653         link[ip->i_d.di_size] = '\0';
654         error = 0;
655
656  out:
657         return error;
658 }
659
660 int
661 xfs_readlink(
662         xfs_inode_t     *ip,
663         char            *link)
664 {
665         xfs_mount_t     *mp = ip->i_mount;
666         int             pathlen;
667         int             error = 0;
668
669         xfs_itrace_entry(ip);
670
671         if (XFS_FORCED_SHUTDOWN(mp))
672                 return XFS_ERROR(EIO);
673
674         xfs_ilock(ip, XFS_ILOCK_SHARED);
675
676         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
677         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
678
679         pathlen = ip->i_d.di_size;
680         if (!pathlen)
681                 goto out;
682
683         if (ip->i_df.if_flags & XFS_IFINLINE) {
684                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
685                 link[pathlen] = '\0';
686         } else {
687                 error = xfs_readlink_bmap(ip, link);
688         }
689
690  out:
691         xfs_iunlock(ip, XFS_ILOCK_SHARED);
692         return error;
693 }
694
695 /*
696  * xfs_fsync
697  *
698  * This is called to sync the inode and its data out to disk.  We need to hold
699  * the I/O lock while flushing the data, and the inode lock while flushing the
700  * inode.  The inode lock CANNOT be held while flushing the data, so acquire
701  * after we're done with that.
702  */
703 int
704 xfs_fsync(
705         xfs_inode_t     *ip)
706 {
707         xfs_trans_t     *tp;
708         int             error;
709         int             log_flushed = 0, changed = 1;
710
711         xfs_itrace_entry(ip);
712
713         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
714                 return XFS_ERROR(EIO);
715
716         /* capture size updates in I/O completion before writing the inode. */
717         error = filemap_fdatawait(VFS_I(ip)->i_mapping);
718         if (error)
719                 return XFS_ERROR(error);
720
721         /*
722          * We always need to make sure that the required inode state is safe on
723          * disk.  The vnode might be clean but we still might need to force the
724          * log because of committed transactions that haven't hit the disk yet.
725          * Likewise, there could be unflushed non-transactional changes to the
726          * inode core that have to go to disk and this requires us to issue
727          * a synchronous transaction to capture these changes correctly.
728          *
729          * This code relies on the assumption that if the update_* fields
730          * of the inode are clear and the inode is unpinned then it is clean
731          * and no action is required.
732          */
733         xfs_ilock(ip, XFS_ILOCK_SHARED);
734
735         if (!(ip->i_update_size || ip->i_update_core)) {
736                 /*
737                  * Timestamps/size haven't changed since last inode flush or
738                  * inode transaction commit.  That means either nothing got
739                  * written or a transaction committed which caught the updates.
740                  * If the latter happened and the transaction hasn't hit the
741                  * disk yet, the inode will be still be pinned.  If it is,
742                  * force the log.
743                  */
744
745                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
746
747                 if (xfs_ipincount(ip)) {
748                         error = _xfs_log_force(ip->i_mount, (xfs_lsn_t)0,
749                                       XFS_LOG_FORCE | XFS_LOG_SYNC,
750                                       &log_flushed);
751                 } else {
752                         /*
753                          * If the inode is not pinned and nothing has changed
754                          * we don't need to flush the cache.
755                          */
756                         changed = 0;
757                 }
758         } else  {
759                 /*
760                  * Kick off a transaction to log the inode core to get the
761                  * updates.  The sync transaction will also force the log.
762                  */
763                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
764                 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
765                 error = xfs_trans_reserve(tp, 0,
766                                 XFS_FSYNC_TS_LOG_RES(ip->i_mount), 0, 0, 0);
767                 if (error) {
768                         xfs_trans_cancel(tp, 0);
769                         return error;
770                 }
771                 xfs_ilock(ip, XFS_ILOCK_EXCL);
772
773                 /*
774                  * Note - it's possible that we might have pushed ourselves out
775                  * of the way during trans_reserve which would flush the inode.
776                  * But there's no guarantee that the inode buffer has actually
777                  * gone out yet (it's delwri).  Plus the buffer could be pinned
778                  * anyway if it's part of an inode in another recent
779                  * transaction.  So we play it safe and fire off the
780                  * transaction anyway.
781                  */
782                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
783                 xfs_trans_ihold(tp, ip);
784                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
785                 xfs_trans_set_sync(tp);
786                 error = _xfs_trans_commit(tp, 0, &log_flushed);
787
788                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
789         }
790
791         if ((ip->i_mount->m_flags & XFS_MOUNT_BARRIER) && changed) {
792                 /*
793                  * If the log write didn't issue an ordered tag we need
794                  * to flush the disk cache for the data device now.
795                  */
796                 if (!log_flushed)
797                         xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
798
799                 /*
800                  * If this inode is on the RT dev we need to flush that
801                  * cache as well.
802                  */
803                 if (XFS_IS_REALTIME_INODE(ip))
804                         xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
805         }
806
807         return error;
808 }
809
810 /*
811  * This is called by xfs_inactive to free any blocks beyond eof
812  * when the link count isn't zero and by xfs_dm_punch_hole() when
813  * punching a hole to EOF.
814  */
815 int
816 xfs_free_eofblocks(
817         xfs_mount_t     *mp,
818         xfs_inode_t     *ip,
819         int             flags)
820 {
821         xfs_trans_t     *tp;
822         int             error;
823         xfs_fileoff_t   end_fsb;
824         xfs_fileoff_t   last_fsb;
825         xfs_filblks_t   map_len;
826         int             nimaps;
827         xfs_bmbt_irec_t imap;
828         int             use_iolock = (flags & XFS_FREE_EOF_LOCK);
829
830         /*
831          * Figure out if there are any blocks beyond the end
832          * of the file.  If not, then there is nothing to do.
833          */
834         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
835         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
836         map_len = last_fsb - end_fsb;
837         if (map_len <= 0)
838                 return 0;
839
840         nimaps = 1;
841         xfs_ilock(ip, XFS_ILOCK_SHARED);
842         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
843                           NULL, 0, &imap, &nimaps, NULL, NULL);
844         xfs_iunlock(ip, XFS_ILOCK_SHARED);
845
846         if (!error && (nimaps != 0) &&
847             (imap.br_startblock != HOLESTARTBLOCK ||
848              ip->i_delayed_blks)) {
849                 /*
850                  * Attach the dquots to the inode up front.
851                  */
852                 if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
853                         return error;
854
855                 /*
856                  * There are blocks after the end of file.
857                  * Free them up now by truncating the file to
858                  * its current size.
859                  */
860                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
861
862                 /*
863                  * Do the xfs_itruncate_start() call before
864                  * reserving any log space because
865                  * itruncate_start will call into the buffer
866                  * cache and we can't
867                  * do that within a transaction.
868                  */
869                 if (use_iolock)
870                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
871                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
872                                     ip->i_size);
873                 if (error) {
874                         xfs_trans_cancel(tp, 0);
875                         if (use_iolock)
876                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
877                         return error;
878                 }
879
880                 error = xfs_trans_reserve(tp, 0,
881                                           XFS_ITRUNCATE_LOG_RES(mp),
882                                           0, XFS_TRANS_PERM_LOG_RES,
883                                           XFS_ITRUNCATE_LOG_COUNT);
884                 if (error) {
885                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
886                         xfs_trans_cancel(tp, 0);
887                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
888                         return error;
889                 }
890
891                 xfs_ilock(ip, XFS_ILOCK_EXCL);
892                 xfs_trans_ijoin(tp, ip,
893                                 XFS_IOLOCK_EXCL |
894                                 XFS_ILOCK_EXCL);
895                 xfs_trans_ihold(tp, ip);
896
897                 error = xfs_itruncate_finish(&tp, ip,
898                                              ip->i_size,
899                                              XFS_DATA_FORK,
900                                              0);
901                 /*
902                  * If we get an error at this point we
903                  * simply don't bother truncating the file.
904                  */
905                 if (error) {
906                         xfs_trans_cancel(tp,
907                                          (XFS_TRANS_RELEASE_LOG_RES |
908                                           XFS_TRANS_ABORT));
909                 } else {
910                         error = xfs_trans_commit(tp,
911                                                 XFS_TRANS_RELEASE_LOG_RES);
912                 }
913                 xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
914                                             : XFS_ILOCK_EXCL));
915         }
916         return error;
917 }
918
919 /*
920  * Free a symlink that has blocks associated with it.
921  */
922 STATIC int
923 xfs_inactive_symlink_rmt(
924         xfs_inode_t     *ip,
925         xfs_trans_t     **tpp)
926 {
927         xfs_buf_t       *bp;
928         int             committed;
929         int             done;
930         int             error;
931         xfs_fsblock_t   first_block;
932         xfs_bmap_free_t free_list;
933         int             i;
934         xfs_mount_t     *mp;
935         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
936         int             nmaps;
937         xfs_trans_t     *ntp;
938         int             size;
939         xfs_trans_t     *tp;
940
941         tp = *tpp;
942         mp = ip->i_mount;
943         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
944         /*
945          * We're freeing a symlink that has some
946          * blocks allocated to it.  Free the
947          * blocks here.  We know that we've got
948          * either 1 or 2 extents and that we can
949          * free them all in one bunmapi call.
950          */
951         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
952         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
953                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
954                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
955                 xfs_trans_cancel(tp, 0);
956                 *tpp = NULL;
957                 return error;
958         }
959         /*
960          * Lock the inode, fix the size, and join it to the transaction.
961          * Hold it so in the normal path, we still have it locked for
962          * the second transaction.  In the error paths we need it
963          * held so the cancel won't rele it, see below.
964          */
965         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
966         size = (int)ip->i_d.di_size;
967         ip->i_d.di_size = 0;
968         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
969         xfs_trans_ihold(tp, ip);
970         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
971         /*
972          * Find the block(s) so we can inval and unmap them.
973          */
974         done = 0;
975         XFS_BMAP_INIT(&free_list, &first_block);
976         nmaps = ARRAY_SIZE(mval);
977         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
978                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
979                         &free_list, NULL)))
980                 goto error0;
981         /*
982          * Invalidate the block(s).
983          */
984         for (i = 0; i < nmaps; i++) {
985                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
986                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
987                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
988                 xfs_trans_binval(tp, bp);
989         }
990         /*
991          * Unmap the dead block(s) to the free_list.
992          */
993         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
994                         &first_block, &free_list, NULL, &done)))
995                 goto error1;
996         ASSERT(done);
997         /*
998          * Commit the first transaction.  This logs the EFI and the inode.
999          */
1000         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
1001                 goto error1;
1002         /*
1003          * The transaction must have been committed, since there were
1004          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
1005          * The new tp has the extent freeing and EFDs.
1006          */
1007         ASSERT(committed);
1008         /*
1009          * The first xact was committed, so add the inode to the new one.
1010          * Mark it dirty so it will be logged and moved forward in the log as
1011          * part of every commit.
1012          */
1013         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1014         xfs_trans_ihold(tp, ip);
1015         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1016         /*
1017          * Get a new, empty transaction to return to our caller.
1018          */
1019         ntp = xfs_trans_dup(tp);
1020         /*
1021          * Commit the transaction containing extent freeing and EFDs.
1022          * If we get an error on the commit here or on the reserve below,
1023          * we need to unlock the inode since the new transaction doesn't
1024          * have the inode attached.
1025          */
1026         error = xfs_trans_commit(tp, 0);
1027         tp = ntp;
1028         if (error) {
1029                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1030                 goto error0;
1031         }
1032         /*
1033          * Remove the memory for extent descriptions (just bookkeeping).
1034          */
1035         if (ip->i_df.if_bytes)
1036                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
1037         ASSERT(ip->i_df.if_bytes == 0);
1038         /*
1039          * Put an itruncate log reservation in the new transaction
1040          * for our caller.
1041          */
1042         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
1043                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
1044                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1045                 goto error0;
1046         }
1047         /*
1048          * Return with the inode locked but not joined to the transaction.
1049          */
1050         *tpp = tp;
1051         return 0;
1052
1053  error1:
1054         xfs_bmap_cancel(&free_list);
1055  error0:
1056         /*
1057          * Have to come here with the inode locked and either
1058          * (held and in the transaction) or (not in the transaction).
1059          * If the inode isn't held then cancel would iput it, but
1060          * that's wrong since this is inactive and the vnode ref
1061          * count is 0 already.
1062          * Cancel won't do anything to the inode if held, but it still
1063          * needs to be locked until the cancel is done, if it was
1064          * joined to the transaction.
1065          */
1066         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1067         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1068         *tpp = NULL;
1069         return error;
1070
1071 }
1072
1073 STATIC int
1074 xfs_inactive_symlink_local(
1075         xfs_inode_t     *ip,
1076         xfs_trans_t     **tpp)
1077 {
1078         int             error;
1079
1080         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
1081         /*
1082          * We're freeing a symlink which fit into
1083          * the inode.  Just free the memory used
1084          * to hold the old symlink.
1085          */
1086         error = xfs_trans_reserve(*tpp, 0,
1087                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
1088                                   0, XFS_TRANS_PERM_LOG_RES,
1089                                   XFS_ITRUNCATE_LOG_COUNT);
1090
1091         if (error) {
1092                 xfs_trans_cancel(*tpp, 0);
1093                 *tpp = NULL;
1094                 return error;
1095         }
1096         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1097
1098         /*
1099          * Zero length symlinks _can_ exist.
1100          */
1101         if (ip->i_df.if_bytes > 0) {
1102                 xfs_idata_realloc(ip,
1103                                   -(ip->i_df.if_bytes),
1104                                   XFS_DATA_FORK);
1105                 ASSERT(ip->i_df.if_bytes == 0);
1106         }
1107         return 0;
1108 }
1109
1110 STATIC int
1111 xfs_inactive_attrs(
1112         xfs_inode_t     *ip,
1113         xfs_trans_t     **tpp)
1114 {
1115         xfs_trans_t     *tp;
1116         int             error;
1117         xfs_mount_t     *mp;
1118
1119         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1120         tp = *tpp;
1121         mp = ip->i_mount;
1122         ASSERT(ip->i_d.di_forkoff != 0);
1123         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1124         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1125         if (error)
1126                 goto error_unlock;
1127
1128         error = xfs_attr_inactive(ip);
1129         if (error)
1130                 goto error_unlock;
1131
1132         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1133         error = xfs_trans_reserve(tp, 0,
1134                                   XFS_IFREE_LOG_RES(mp),
1135                                   0, XFS_TRANS_PERM_LOG_RES,
1136                                   XFS_INACTIVE_LOG_COUNT);
1137         if (error)
1138                 goto error_cancel;
1139
1140         xfs_ilock(ip, XFS_ILOCK_EXCL);
1141         xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1142         xfs_trans_ihold(tp, ip);
1143         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1144
1145         ASSERT(ip->i_d.di_anextents == 0);
1146
1147         *tpp = tp;
1148         return 0;
1149
1150 error_cancel:
1151         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1152         xfs_trans_cancel(tp, 0);
1153 error_unlock:
1154         *tpp = NULL;
1155         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1156         return error;
1157 }
1158
1159 int
1160 xfs_release(
1161         xfs_inode_t     *ip)
1162 {
1163         xfs_mount_t     *mp = ip->i_mount;
1164         int             error;
1165
1166         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
1167                 return 0;
1168
1169         /* If this is a read-only mount, don't do this (would generate I/O) */
1170         if (mp->m_flags & XFS_MOUNT_RDONLY)
1171                 return 0;
1172
1173         if (!XFS_FORCED_SHUTDOWN(mp)) {
1174                 int truncated;
1175
1176                 /*
1177                  * If we are using filestreams, and we have an unlinked
1178                  * file that we are processing the last close on, then nothing
1179                  * will be able to reopen and write to this file. Purge this
1180                  * inode from the filestreams cache so that it doesn't delay
1181                  * teardown of the inode.
1182                  */
1183                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
1184                         xfs_filestream_deassociate(ip);
1185
1186                 /*
1187                  * If we previously truncated this file and removed old data
1188                  * in the process, we want to initiate "early" writeout on
1189                  * the last close.  This is an attempt to combat the notorious
1190                  * NULL files problem which is particularly noticable from a
1191                  * truncate down, buffered (re-)write (delalloc), followed by
1192                  * a crash.  What we are effectively doing here is
1193                  * significantly reducing the time window where we'd otherwise
1194                  * be exposed to that problem.
1195                  */
1196                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1197                 if (truncated && VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
1198                         xfs_flush_pages(ip, 0, -1, XFS_B_ASYNC, FI_NONE);
1199         }
1200
1201         if (ip->i_d.di_nlink != 0) {
1202                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1203                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1204                        ip->i_delayed_blks > 0)) &&
1205                      (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
1206                     (!(ip->i_d.di_flags &
1207                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
1208                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1209                         if (error)
1210                                 return error;
1211                 }
1212         }
1213
1214         return 0;
1215 }
1216
1217 /*
1218  * xfs_inactive
1219  *
1220  * This is called when the vnode reference count for the vnode
1221  * goes to zero.  If the file has been unlinked, then it must
1222  * now be truncated.  Also, we clear all of the read-ahead state
1223  * kept for the inode here since the file is now closed.
1224  */
1225 int
1226 xfs_inactive(
1227         xfs_inode_t     *ip)
1228 {
1229         xfs_bmap_free_t free_list;
1230         xfs_fsblock_t   first_block;
1231         int             committed;
1232         xfs_trans_t     *tp;
1233         xfs_mount_t     *mp;
1234         int             error;
1235         int             truncate;
1236
1237         xfs_itrace_entry(ip);
1238
1239         /*
1240          * If the inode is already free, then there can be nothing
1241          * to clean up here.
1242          */
1243         if (ip->i_d.di_mode == 0 || VN_BAD(VFS_I(ip))) {
1244                 ASSERT(ip->i_df.if_real_bytes == 0);
1245                 ASSERT(ip->i_df.if_broot_bytes == 0);
1246                 return VN_INACTIVE_CACHE;
1247         }
1248
1249         /*
1250          * Only do a truncate if it's a regular file with
1251          * some actual space in it.  It's OK to look at the
1252          * inode's fields without the lock because we're the
1253          * only one with a reference to the inode.
1254          */
1255         truncate = ((ip->i_d.di_nlink == 0) &&
1256             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1257              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1258             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1259
1260         mp = ip->i_mount;
1261
1262         if (ip->i_d.di_nlink == 0 && DM_EVENT_ENABLED(ip, DM_EVENT_DESTROY))
1263                 XFS_SEND_DESTROY(mp, ip, DM_RIGHT_NULL);
1264
1265         error = 0;
1266
1267         /* If this is a read-only mount, don't do this (would generate I/O) */
1268         if (mp->m_flags & XFS_MOUNT_RDONLY)
1269                 goto out;
1270
1271         if (ip->i_d.di_nlink != 0) {
1272                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1273                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
1274                        ip->i_delayed_blks > 0)) &&
1275                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1276                      (!(ip->i_d.di_flags &
1277                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1278                       (ip->i_delayed_blks != 0)))) {
1279                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1280                         if (error)
1281                                 return VN_INACTIVE_CACHE;
1282                 }
1283                 goto out;
1284         }
1285
1286         ASSERT(ip->i_d.di_nlink == 0);
1287
1288         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
1289                 return VN_INACTIVE_CACHE;
1290
1291         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1292         if (truncate) {
1293                 /*
1294                  * Do the xfs_itruncate_start() call before
1295                  * reserving any log space because itruncate_start
1296                  * will call into the buffer cache and we can't
1297                  * do that within a transaction.
1298                  */
1299                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1300
1301                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1302                 if (error) {
1303                         xfs_trans_cancel(tp, 0);
1304                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1305                         return VN_INACTIVE_CACHE;
1306                 }
1307
1308                 error = xfs_trans_reserve(tp, 0,
1309                                           XFS_ITRUNCATE_LOG_RES(mp),
1310                                           0, XFS_TRANS_PERM_LOG_RES,
1311                                           XFS_ITRUNCATE_LOG_COUNT);
1312                 if (error) {
1313                         /* Don't call itruncate_cleanup */
1314                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1315                         xfs_trans_cancel(tp, 0);
1316                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1317                         return VN_INACTIVE_CACHE;
1318                 }
1319
1320                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1321                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1322                 xfs_trans_ihold(tp, ip);
1323
1324                 /*
1325                  * normally, we have to run xfs_itruncate_finish sync.
1326                  * But if filesystem is wsync and we're in the inactive
1327                  * path, then we know that nlink == 0, and that the
1328                  * xaction that made nlink == 0 is permanently committed
1329                  * since xfs_remove runs as a synchronous transaction.
1330                  */
1331                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1332                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1333
1334                 if (error) {
1335                         xfs_trans_cancel(tp,
1336                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1337                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1338                         return VN_INACTIVE_CACHE;
1339                 }
1340         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1341
1342                 /*
1343                  * If we get an error while cleaning up a
1344                  * symlink we bail out.
1345                  */
1346                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1347                         xfs_inactive_symlink_rmt(ip, &tp) :
1348                         xfs_inactive_symlink_local(ip, &tp);
1349
1350                 if (error) {
1351                         ASSERT(tp == NULL);
1352                         return VN_INACTIVE_CACHE;
1353                 }
1354
1355                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1356                 xfs_trans_ihold(tp, ip);
1357         } else {
1358                 error = xfs_trans_reserve(tp, 0,
1359                                           XFS_IFREE_LOG_RES(mp),
1360                                           0, XFS_TRANS_PERM_LOG_RES,
1361                                           XFS_INACTIVE_LOG_COUNT);
1362                 if (error) {
1363                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1364                         xfs_trans_cancel(tp, 0);
1365                         return VN_INACTIVE_CACHE;
1366                 }
1367
1368                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1369                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1370                 xfs_trans_ihold(tp, ip);
1371         }
1372
1373         /*
1374          * If there are attributes associated with the file
1375          * then blow them away now.  The code calls a routine
1376          * that recursively deconstructs the attribute fork.
1377          * We need to just commit the current transaction
1378          * because we can't use it for xfs_attr_inactive().
1379          */
1380         if (ip->i_d.di_anextents > 0) {
1381                 error = xfs_inactive_attrs(ip, &tp);
1382                 /*
1383                  * If we got an error, the transaction is already
1384                  * cancelled, and the inode is unlocked. Just get out.
1385                  */
1386                  if (error)
1387                          return VN_INACTIVE_CACHE;
1388         } else if (ip->i_afp) {
1389                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1390         }
1391
1392         /*
1393          * Free the inode.
1394          */
1395         XFS_BMAP_INIT(&free_list, &first_block);
1396         error = xfs_ifree(tp, ip, &free_list);
1397         if (error) {
1398                 /*
1399                  * If we fail to free the inode, shut down.  The cancel
1400                  * might do that, we need to make sure.  Otherwise the
1401                  * inode might be lost for a long time or forever.
1402                  */
1403                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1404                         cmn_err(CE_NOTE,
1405                 "xfs_inactive:  xfs_ifree() returned an error = %d on %s",
1406                                 error, mp->m_fsname);
1407                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1408                 }
1409                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1410         } else {
1411                 /*
1412                  * Credit the quota account(s). The inode is gone.
1413                  */
1414                 XFS_TRANS_MOD_DQUOT_BYINO(mp, tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1415
1416                 /*
1417                  * Just ignore errors at this point.  There is nothing we can
1418                  * do except to try to keep going. Make sure it's not a silent
1419                  * error.
1420                  */
1421                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
1422                 if (error)
1423                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1424                                 "xfs_bmap_finish() returned error %d", error);
1425                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1426                 if (error)
1427                         xfs_fs_cmn_err(CE_NOTE, mp, "xfs_inactive: "
1428                                 "xfs_trans_commit() returned error %d", error);
1429         }
1430         /*
1431          * Release the dquots held by inode, if any.
1432          */
1433         XFS_QM_DQDETACH(mp, ip);
1434
1435         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1436
1437  out:
1438         return VN_INACTIVE_CACHE;
1439 }
1440
1441 /*
1442  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
1443  * is allowed, otherwise it has to be an exact match. If a CI match is found,
1444  * ci_name->name will point to a the actual name (caller must free) or
1445  * will be set to NULL if an exact match is found.
1446  */
1447 int
1448 xfs_lookup(
1449         xfs_inode_t             *dp,
1450         struct xfs_name         *name,
1451         xfs_inode_t             **ipp,
1452         struct xfs_name         *ci_name)
1453 {
1454         xfs_ino_t               inum;
1455         int                     error;
1456         uint                    lock_mode;
1457
1458         xfs_itrace_entry(dp);
1459
1460         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1461                 return XFS_ERROR(EIO);
1462
1463         lock_mode = xfs_ilock_map_shared(dp);
1464         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
1465         xfs_iunlock_map_shared(dp, lock_mode);
1466
1467         if (error)
1468                 goto out;
1469
1470         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp, 0);
1471         if (error)
1472                 goto out_free_name;
1473
1474         xfs_itrace_ref(*ipp);
1475         return 0;
1476
1477 out_free_name:
1478         if (ci_name)
1479                 kmem_free(ci_name->name);
1480 out:
1481         *ipp = NULL;
1482         return error;
1483 }
1484
1485 int
1486 xfs_create(
1487         xfs_inode_t             *dp,
1488         struct xfs_name         *name,
1489         mode_t                  mode,
1490         xfs_dev_t               rdev,
1491         xfs_inode_t             **ipp,
1492         cred_t                  *credp)
1493 {
1494         xfs_mount_t             *mp = dp->i_mount;
1495         xfs_inode_t             *ip;
1496         xfs_trans_t             *tp;
1497         int                     error;
1498         xfs_bmap_free_t         free_list;
1499         xfs_fsblock_t           first_block;
1500         boolean_t               unlock_dp_on_error = B_FALSE;
1501         int                     dm_event_sent = 0;
1502         uint                    cancel_flags;
1503         int                     committed;
1504         xfs_prid_t              prid;
1505         struct xfs_dquot        *udqp, *gdqp;
1506         uint                    resblks;
1507
1508         ASSERT(!*ipp);
1509         xfs_itrace_entry(dp);
1510
1511         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
1512                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
1513                                 dp, DM_RIGHT_NULL, NULL,
1514                                 DM_RIGHT_NULL, name->name, NULL,
1515                                 mode, 0, 0);
1516
1517                 if (error)
1518                         return error;
1519                 dm_event_sent = 1;
1520         }
1521
1522         if (XFS_FORCED_SHUTDOWN(mp))
1523                 return XFS_ERROR(EIO);
1524
1525         /* Return through std_return after this point. */
1526
1527         udqp = gdqp = NULL;
1528         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1529                 prid = dp->i_d.di_projid;
1530         else
1531                 prid = (xfs_prid_t)dfltprid;
1532
1533         /*
1534          * Make sure that we have allocated dquot(s) on disk.
1535          */
1536         error = XFS_QM_DQVOPALLOC(mp, dp,
1537                         current_fsuid(credp), current_fsgid(credp), prid,
1538                         XFS_QMOPT_QUOTALL|XFS_QMOPT_INHERIT, &udqp, &gdqp);
1539         if (error)
1540                 goto std_return;
1541
1542         ip = NULL;
1543
1544         tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1545         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1546         resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1547         /*
1548          * Initially assume that the file does not exist and
1549          * reserve the resources for that case.  If that is not
1550          * the case we'll drop the one we have and get a more
1551          * appropriate transaction later.
1552          */
1553         error = xfs_trans_reserve(tp, resblks, XFS_CREATE_LOG_RES(mp), 0,
1554                         XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1555         if (error == ENOSPC) {
1556                 resblks = 0;
1557                 error = xfs_trans_reserve(tp, 0, XFS_CREATE_LOG_RES(mp), 0,
1558                                 XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1559         }
1560         if (error) {
1561                 cancel_flags = 0;
1562                 goto error_return;
1563         }
1564
1565         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1566         unlock_dp_on_error = B_TRUE;
1567
1568         XFS_BMAP_INIT(&free_list, &first_block);
1569
1570         ASSERT(ip == NULL);
1571
1572         /*
1573          * Reserve disk quota and the inode.
1574          */
1575         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
1576         if (error)
1577                 goto error_return;
1578
1579         error = xfs_dir_canenter(tp, dp, name, resblks);
1580         if (error)
1581                 goto error_return;
1582         error = xfs_dir_ialloc(&tp, dp, mode, 1,
1583                         rdev, credp, prid, resblks > 0,
1584                         &ip, &committed);
1585         if (error) {
1586                 if (error == ENOSPC)
1587                         goto error_return;
1588                 goto abort_return;
1589         }
1590         xfs_itrace_ref(ip);
1591
1592         /*
1593          * At this point, we've gotten a newly allocated inode.
1594          * It is locked (and joined to the transaction).
1595          */
1596
1597         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1598
1599         /*
1600          * Now we join the directory inode to the transaction.  We do not do it
1601          * earlier because xfs_dir_ialloc might commit the previous transaction
1602          * (and release all the locks).  An error from here on will result in
1603          * the transaction cancel unlocking dp so don't do it explicitly in the
1604          * error path.
1605          */
1606         IHOLD(dp);
1607         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1608         unlock_dp_on_error = B_FALSE;
1609
1610         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1611                                         &first_block, &free_list, resblks ?
1612                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1613         if (error) {
1614                 ASSERT(error != ENOSPC);
1615                 goto abort_return;
1616         }
1617         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1618         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1619
1620         /*
1621          * If this is a synchronous mount, make sure that the
1622          * create transaction goes to disk before returning to
1623          * the user.
1624          */
1625         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1626                 xfs_trans_set_sync(tp);
1627         }
1628
1629         dp->i_gen++;
1630
1631         /*
1632          * Attach the dquot(s) to the inodes and modify them incore.
1633          * These ids of the inode couldn't have changed since the new
1634          * inode has been locked ever since it was created.
1635          */
1636         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
1637
1638         /*
1639          * xfs_trans_commit normally decrements the vnode ref count
1640          * when it unlocks the inode. Since we want to return the
1641          * vnode to the caller, we bump the vnode ref count now.
1642          */
1643         IHOLD(ip);
1644
1645         error = xfs_bmap_finish(&tp, &free_list, &committed);
1646         if (error) {
1647                 xfs_bmap_cancel(&free_list);
1648                 goto abort_rele;
1649         }
1650
1651         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1652         if (error) {
1653                 IRELE(ip);
1654                 tp = NULL;
1655                 goto error_return;
1656         }
1657
1658         XFS_QM_DQRELE(mp, udqp);
1659         XFS_QM_DQRELE(mp, gdqp);
1660
1661         *ipp = ip;
1662
1663         /* Fallthrough to std_return with error = 0  */
1664
1665 std_return:
1666         if ((*ipp || (error != 0 && dm_event_sent != 0)) &&
1667             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
1668                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
1669                         dp, DM_RIGHT_NULL,
1670                         *ipp ? ip : NULL,
1671                         DM_RIGHT_NULL, name->name, NULL,
1672                         mode, error, 0);
1673         }
1674         return error;
1675
1676  abort_return:
1677         cancel_flags |= XFS_TRANS_ABORT;
1678         /* FALLTHROUGH */
1679
1680  error_return:
1681         if (tp != NULL)
1682                 xfs_trans_cancel(tp, cancel_flags);
1683
1684         XFS_QM_DQRELE(mp, udqp);
1685         XFS_QM_DQRELE(mp, gdqp);
1686
1687         if (unlock_dp_on_error)
1688                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1689
1690         goto std_return;
1691
1692  abort_rele:
1693         /*
1694          * Wait until after the current transaction is aborted to
1695          * release the inode.  This prevents recursive transactions
1696          * and deadlocks from xfs_inactive.
1697          */
1698         cancel_flags |= XFS_TRANS_ABORT;
1699         xfs_trans_cancel(tp, cancel_flags);
1700         IRELE(ip);
1701
1702         XFS_QM_DQRELE(mp, udqp);
1703         XFS_QM_DQRELE(mp, gdqp);
1704
1705         goto std_return;
1706 }
1707
1708 #ifdef DEBUG
1709 int xfs_locked_n;
1710 int xfs_small_retries;
1711 int xfs_middle_retries;
1712 int xfs_lots_retries;
1713 int xfs_lock_delays;
1714 #endif
1715
1716 /*
1717  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1718  * a different value
1719  */
1720 static inline int
1721 xfs_lock_inumorder(int lock_mode, int subclass)
1722 {
1723         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1724                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1725         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1726                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1727
1728         return lock_mode;
1729 }
1730
1731 /*
1732  * The following routine will lock n inodes in exclusive mode.
1733  * We assume the caller calls us with the inodes in i_ino order.
1734  *
1735  * We need to detect deadlock where an inode that we lock
1736  * is in the AIL and we start waiting for another inode that is locked
1737  * by a thread in a long running transaction (such as truncate). This can
1738  * result in deadlock since the long running trans might need to wait
1739  * for the inode we just locked in order to push the tail and free space
1740  * in the log.
1741  */
1742 void
1743 xfs_lock_inodes(
1744         xfs_inode_t     **ips,
1745         int             inodes,
1746         uint            lock_mode)
1747 {
1748         int             attempts = 0, i, j, try_lock;
1749         xfs_log_item_t  *lp;
1750
1751         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1752
1753         try_lock = 0;
1754         i = 0;
1755
1756 again:
1757         for (; i < inodes; i++) {
1758                 ASSERT(ips[i]);
1759
1760                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1761                         continue;
1762
1763                 /*
1764                  * If try_lock is not set yet, make sure all locked inodes
1765                  * are not in the AIL.
1766                  * If any are, set try_lock to be used later.
1767                  */
1768
1769                 if (!try_lock) {
1770                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1771                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1772                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1773                                         try_lock++;
1774                                 }
1775                         }
1776                 }
1777
1778                 /*
1779                  * If any of the previous locks we have locked is in the AIL,
1780                  * we must TRY to get the second and subsequent locks. If
1781                  * we can't get any, we must release all we have
1782                  * and try again.
1783                  */
1784
1785                 if (try_lock) {
1786                         /* try_lock must be 0 if i is 0. */
1787                         /*
1788                          * try_lock means we have an inode locked
1789                          * that is in the AIL.
1790                          */
1791                         ASSERT(i != 0);
1792                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1793                                 attempts++;
1794
1795                                 /*
1796                                  * Unlock all previous guys and try again.
1797                                  * xfs_iunlock will try to push the tail
1798                                  * if the inode is in the AIL.
1799                                  */
1800
1801                                 for(j = i - 1; j >= 0; j--) {
1802
1803                                         /*
1804                                          * Check to see if we've already
1805                                          * unlocked this one.
1806                                          * Not the first one going back,
1807                                          * and the inode ptr is the same.
1808                                          */
1809                                         if ((j != (i - 1)) && ips[j] ==
1810                                                                 ips[j+1])
1811                                                 continue;
1812
1813                                         xfs_iunlock(ips[j], lock_mode);
1814                                 }
1815
1816                                 if ((attempts % 5) == 0) {
1817                                         delay(1); /* Don't just spin the CPU */
1818 #ifdef DEBUG
1819                                         xfs_lock_delays++;
1820 #endif
1821                                 }
1822                                 i = 0;
1823                                 try_lock = 0;
1824                                 goto again;
1825                         }
1826                 } else {
1827                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1828                 }
1829         }
1830
1831 #ifdef DEBUG
1832         if (attempts) {
1833                 if (attempts < 5) xfs_small_retries++;
1834                 else if (attempts < 100) xfs_middle_retries++;
1835                 else xfs_lots_retries++;
1836         } else {
1837                 xfs_locked_n++;
1838         }
1839 #endif
1840 }
1841
1842 void
1843 xfs_lock_two_inodes(
1844         xfs_inode_t             *ip0,
1845         xfs_inode_t             *ip1,
1846         uint                    lock_mode)
1847 {
1848         xfs_inode_t             *temp;
1849         int                     attempts = 0;
1850         xfs_log_item_t          *lp;
1851
1852         ASSERT(ip0->i_ino != ip1->i_ino);
1853
1854         if (ip0->i_ino > ip1->i_ino) {
1855                 temp = ip0;
1856                 ip0 = ip1;
1857                 ip1 = temp;
1858         }
1859
1860  again:
1861         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1862
1863         /*
1864          * If the first lock we have locked is in the AIL, we must TRY to get
1865          * the second lock. If we can't get it, we must release the first one
1866          * and try again.
1867          */
1868         lp = (xfs_log_item_t *)ip0->i_itemp;
1869         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1870                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1871                         xfs_iunlock(ip0, lock_mode);
1872                         if ((++attempts % 5) == 0)
1873                                 delay(1); /* Don't just spin the CPU */
1874                         goto again;
1875                 }
1876         } else {
1877                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1878         }
1879 }
1880
1881 int
1882 xfs_remove(
1883         xfs_inode_t             *dp,
1884         struct xfs_name         *name,
1885         xfs_inode_t             *ip)
1886 {
1887         xfs_mount_t             *mp = dp->i_mount;
1888         xfs_trans_t             *tp = NULL;
1889         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1890         int                     error = 0;
1891         xfs_bmap_free_t         free_list;
1892         xfs_fsblock_t           first_block;
1893         int                     cancel_flags;
1894         int                     committed;
1895         int                     link_zero;
1896         uint                    resblks;
1897         uint                    log_count;
1898
1899         xfs_itrace_entry(dp);
1900         xfs_itrace_entry(ip);
1901
1902         if (XFS_FORCED_SHUTDOWN(mp))
1903                 return XFS_ERROR(EIO);
1904
1905         if (DM_EVENT_ENABLED(dp, DM_EVENT_REMOVE)) {
1906                 error = XFS_SEND_NAMESP(mp, DM_EVENT_REMOVE, dp, DM_RIGHT_NULL,
1907                                         NULL, DM_RIGHT_NULL, name->name, NULL,
1908                                         ip->i_d.di_mode, 0, 0);
1909                 if (error)
1910                         return error;
1911         }
1912
1913         error = XFS_QM_DQATTACH(mp, dp, 0);
1914         if (error)
1915                 goto std_return;
1916
1917         error = XFS_QM_DQATTACH(mp, ip, 0);
1918         if (error)
1919                 goto std_return;
1920
1921         if (is_dir) {
1922                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1923                 log_count = XFS_DEFAULT_LOG_COUNT;
1924         } else {
1925                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1926                 log_count = XFS_REMOVE_LOG_COUNT;
1927         }
1928         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1929
1930         /*
1931          * We try to get the real space reservation first,
1932          * allowing for directory btree deletion(s) implying
1933          * possible bmap insert(s).  If we can't get the space
1934          * reservation then we use 0 instead, and avoid the bmap
1935          * btree insert(s) in the directory code by, if the bmap
1936          * insert tries to happen, instead trimming the LAST
1937          * block from the directory.
1938          */
1939         resblks = XFS_REMOVE_SPACE_RES(mp);
1940         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1941                                   XFS_TRANS_PERM_LOG_RES, log_count);
1942         if (error == ENOSPC) {
1943                 resblks = 0;
1944                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1945                                           XFS_TRANS_PERM_LOG_RES, log_count);
1946         }
1947         if (error) {
1948                 ASSERT(error != ENOSPC);
1949                 cancel_flags = 0;
1950                 goto out_trans_cancel;
1951         }
1952
1953         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1954
1955         /*
1956          * At this point, we've gotten both the directory and the entry
1957          * inodes locked.
1958          */
1959         IHOLD(ip);
1960         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1961
1962         IHOLD(dp);
1963         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1964
1965         /*
1966          * If we're removing a directory perform some additional validation.
1967          */
1968         if (is_dir) {
1969                 ASSERT(ip->i_d.di_nlink >= 2);
1970                 if (ip->i_d.di_nlink != 2) {
1971                         error = XFS_ERROR(ENOTEMPTY);
1972                         goto out_trans_cancel;
1973                 }
1974                 if (!xfs_dir_isempty(ip)) {
1975                         error = XFS_ERROR(ENOTEMPTY);
1976                         goto out_trans_cancel;
1977                 }
1978         }
1979
1980         XFS_BMAP_INIT(&free_list, &first_block);
1981         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1982                                         &first_block, &free_list, resblks);
1983         if (error) {
1984                 ASSERT(error != ENOENT);
1985                 goto out_bmap_cancel;
1986         }
1987         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1988
1989         /*
1990          * Bump the in memory generation count on the parent
1991          * directory so that other can know that it has changed.
1992          */
1993         dp->i_gen++;
1994         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1995
1996         if (is_dir) {
1997                 /*
1998                  * Drop the link from ip's "..".
1999                  */
2000                 error = xfs_droplink(tp, dp);
2001                 if (error)
2002                         goto out_bmap_cancel;
2003
2004                 /*
2005                  * Drop the link from dp to ip.
2006                  */
2007                 error = xfs_droplink(tp, ip);
2008                 if (error)
2009                         goto out_bmap_cancel;
2010         } else {
2011                 /*
2012                  * When removing a non-directory we need to log the parent
2013                  * inode here for the i_gen update.  For a directory this is
2014                  * done implicitly by the xfs_droplink call for the ".." entry.
2015                  */
2016                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2017         }
2018
2019         /*
2020          * Drop the "." link from ip to self.
2021          */
2022         error = xfs_droplink(tp, ip);
2023         if (error)
2024                 goto out_bmap_cancel;
2025
2026         /*
2027          * Determine if this is the last link while
2028          * we are in the transaction.
2029          */
2030         link_zero = (ip->i_d.di_nlink == 0);
2031
2032         /*
2033          * If this is a synchronous mount, make sure that the
2034          * remove transaction goes to disk before returning to
2035          * the user.
2036          */
2037         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2038                 xfs_trans_set_sync(tp);
2039
2040         error = xfs_bmap_finish(&tp, &free_list, &committed);
2041         if (error)
2042                 goto out_bmap_cancel;
2043
2044         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2045         if (error)
2046                 goto std_return;
2047
2048         /*
2049          * If we are using filestreams, kill the stream association.
2050          * If the file is still open it may get a new one but that
2051          * will get killed on last close in xfs_close() so we don't
2052          * have to worry about that.
2053          */
2054         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
2055                 xfs_filestream_deassociate(ip);
2056
2057         xfs_itrace_exit(ip);
2058         xfs_itrace_exit(dp);
2059
2060  std_return:
2061         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTREMOVE)) {
2062                 XFS_SEND_NAMESP(mp, DM_EVENT_POSTREMOVE, dp, DM_RIGHT_NULL,
2063                                 NULL, DM_RIGHT_NULL, name->name, NULL,
2064                                 ip->i_d.di_mode, error, 0);
2065         }
2066
2067         return error;
2068
2069  out_bmap_cancel:
2070         xfs_bmap_cancel(&free_list);
2071         cancel_flags |= XFS_TRANS_ABORT;
2072  out_trans_cancel:
2073         xfs_trans_cancel(tp, cancel_flags);
2074         goto std_return;
2075 }
2076
2077 int
2078 xfs_link(
2079         xfs_inode_t             *tdp,
2080         xfs_inode_t             *sip,
2081         struct xfs_name         *target_name)
2082 {
2083         xfs_mount_t             *mp = tdp->i_mount;
2084         xfs_trans_t             *tp;
2085         int                     error;
2086         xfs_bmap_free_t         free_list;
2087         xfs_fsblock_t           first_block;
2088         int                     cancel_flags;
2089         int                     committed;
2090         int                     resblks;
2091
2092         xfs_itrace_entry(tdp);
2093         xfs_itrace_entry(sip);
2094
2095         ASSERT(!S_ISDIR(sip->i_d.di_mode));
2096
2097         if (XFS_FORCED_SHUTDOWN(mp))
2098                 return XFS_ERROR(EIO);
2099
2100         if (DM_EVENT_ENABLED(tdp, DM_EVENT_LINK)) {
2101                 error = XFS_SEND_NAMESP(mp, DM_EVENT_LINK,
2102                                         tdp, DM_RIGHT_NULL,
2103                                         sip, DM_RIGHT_NULL,
2104                                         target_name->name, NULL, 0, 0, 0);
2105                 if (error)
2106                         return error;
2107         }
2108
2109         /* Return through std_return after this point. */
2110
2111         error = XFS_QM_DQATTACH(mp, sip, 0);
2112         if (!error && sip != tdp)
2113                 error = XFS_QM_DQATTACH(mp, tdp, 0);
2114         if (error)
2115                 goto std_return;
2116
2117         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
2118         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2119         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
2120         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
2121                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2122         if (error == ENOSPC) {
2123                 resblks = 0;
2124                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
2125                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2126         }
2127         if (error) {
2128                 cancel_flags = 0;
2129                 goto error_return;
2130         }
2131
2132         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
2133
2134         /*
2135          * Increment vnode ref counts since xfs_trans_commit &
2136          * xfs_trans_cancel will both unlock the inodes and
2137          * decrement the associated ref counts.
2138          */
2139         IHOLD(sip);
2140         IHOLD(tdp);
2141         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
2142         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
2143
2144         /*
2145          * If the source has too many links, we can't make any more to it.
2146          */
2147         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
2148                 error = XFS_ERROR(EMLINK);
2149                 goto error_return;
2150         }
2151
2152         /*
2153          * If we are using project inheritance, we only allow hard link
2154          * creation in our tree when the project IDs are the same; else
2155          * the tree quota mechanism could be circumvented.
2156          */
2157         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2158                      (tdp->i_d.di_projid != sip->i_d.di_projid))) {
2159                 error = XFS_ERROR(EXDEV);
2160                 goto error_return;
2161         }
2162
2163         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
2164         if (error)
2165                 goto error_return;
2166
2167         XFS_BMAP_INIT(&free_list, &first_block);
2168
2169         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
2170                                         &first_block, &free_list, resblks);
2171         if (error)
2172                 goto abort_return;
2173         xfs_ichgtime(tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2174         tdp->i_gen++;
2175         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
2176
2177         error = xfs_bumplink(tp, sip);
2178         if (error)
2179                 goto abort_return;
2180
2181         /*
2182          * If this is a synchronous mount, make sure that the
2183          * link transaction goes to disk before returning to
2184          * the user.
2185          */
2186         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2187                 xfs_trans_set_sync(tp);
2188         }
2189
2190         error = xfs_bmap_finish (&tp, &free_list, &committed);
2191         if (error) {
2192                 xfs_bmap_cancel(&free_list);
2193                 goto abort_return;
2194         }
2195
2196         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2197         if (error)
2198                 goto std_return;
2199
2200         /* Fall through to std_return with error = 0. */
2201 std_return:
2202         if (DM_EVENT_ENABLED(sip, DM_EVENT_POSTLINK)) {
2203                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTLINK,
2204                                 tdp, DM_RIGHT_NULL,
2205                                 sip, DM_RIGHT_NULL,
2206                                 target_name->name, NULL, 0, error, 0);
2207         }
2208         return error;
2209
2210  abort_return:
2211         cancel_flags |= XFS_TRANS_ABORT;
2212         /* FALLTHROUGH */
2213
2214  error_return:
2215         xfs_trans_cancel(tp, cancel_flags);
2216         goto std_return;
2217 }
2218
2219
2220 int
2221 xfs_mkdir(
2222         xfs_inode_t             *dp,
2223         struct xfs_name         *dir_name,
2224         mode_t                  mode,
2225         xfs_inode_t             **ipp,
2226         cred_t                  *credp)
2227 {
2228         xfs_mount_t             *mp = dp->i_mount;
2229         xfs_inode_t             *cdp;   /* inode of created dir */
2230         xfs_trans_t             *tp;
2231         int                     cancel_flags;
2232         int                     error;
2233         int                     committed;
2234         xfs_bmap_free_t         free_list;
2235         xfs_fsblock_t           first_block;
2236         boolean_t               unlock_dp_on_error = B_FALSE;
2237         boolean_t               created = B_FALSE;
2238         int                     dm_event_sent = 0;
2239         xfs_prid_t              prid;
2240         struct xfs_dquot        *udqp, *gdqp;
2241         uint                    resblks;
2242
2243         if (XFS_FORCED_SHUTDOWN(mp))
2244                 return XFS_ERROR(EIO);
2245
2246         tp = NULL;
2247
2248         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
2249                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
2250                                         dp, DM_RIGHT_NULL, NULL,
2251                                         DM_RIGHT_NULL, dir_name->name, NULL,
2252                                         mode, 0, 0);
2253                 if (error)
2254                         return error;
2255                 dm_event_sent = 1;
2256         }
2257
2258         /* Return through std_return after this point. */
2259
2260         xfs_itrace_entry(dp);
2261
2262         mp = dp->i_mount;
2263         udqp = gdqp = NULL;
2264         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2265                 prid = dp->i_d.di_projid;
2266         else
2267                 prid = (xfs_prid_t)dfltprid;
2268
2269         /*
2270          * Make sure that we have allocated dquot(s) on disk.
2271          */
2272         error = XFS_QM_DQVOPALLOC(mp, dp,
2273                         current_fsuid(credp), current_fsgid(credp), prid,
2274                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2275         if (error)
2276                 goto std_return;
2277
2278         tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
2279         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2280         resblks = XFS_MKDIR_SPACE_RES(mp, dir_name->len);
2281         error = xfs_trans_reserve(tp, resblks, XFS_MKDIR_LOG_RES(mp), 0,
2282                                   XFS_TRANS_PERM_LOG_RES, XFS_MKDIR_LOG_COUNT);
2283         if (error == ENOSPC) {
2284                 resblks = 0;
2285                 error = xfs_trans_reserve(tp, 0, XFS_MKDIR_LOG_RES(mp), 0,
2286                                           XFS_TRANS_PERM_LOG_RES,
2287                                           XFS_MKDIR_LOG_COUNT);
2288         }
2289         if (error) {
2290                 cancel_flags = 0;
2291                 goto error_return;
2292         }
2293
2294         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2295         unlock_dp_on_error = B_TRUE;
2296
2297         /*
2298          * Check for directory link count overflow.
2299          */
2300         if (dp->i_d.di_nlink >= XFS_MAXLINK) {
2301                 error = XFS_ERROR(EMLINK);
2302                 goto error_return;
2303         }
2304
2305         /*
2306          * Reserve disk quota and the inode.
2307          */
2308         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2309         if (error)
2310                 goto error_return;
2311
2312         error = xfs_dir_canenter(tp, dp, dir_name, resblks);
2313         if (error)
2314                 goto error_return;
2315         /*
2316          * create the directory inode.
2317          */
2318         error = xfs_dir_ialloc(&tp, dp, mode, 2,
2319                         0, credp, prid, resblks > 0,
2320                 &cdp, NULL);
2321         if (error) {
2322                 if (error == ENOSPC)
2323                         goto error_return;
2324                 goto abort_return;
2325         }
2326         xfs_itrace_ref(cdp);
2327
2328         /*
2329          * Now we add the directory inode to the transaction.
2330          * We waited until now since xfs_dir_ialloc might start
2331          * a new transaction.  Had we joined the transaction
2332          * earlier, the locks might have gotten released. An error
2333          * from here on will result in the transaction cancel
2334          * unlocking dp so don't do it explicitly in the error path.
2335          */
2336         IHOLD(dp);
2337         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2338         unlock_dp_on_error = B_FALSE;
2339
2340         XFS_BMAP_INIT(&free_list, &first_block);
2341
2342         error = xfs_dir_createname(tp, dp, dir_name, cdp->i_ino,
2343                                         &first_block, &free_list, resblks ?
2344                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
2345         if (error) {
2346                 ASSERT(error != ENOSPC);
2347                 goto error1;
2348         }
2349         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2350
2351         /*
2352          * Bump the in memory version number of the parent directory
2353          * so that other processes accessing it will recognize that
2354          * the directory has changed.
2355          */
2356         dp->i_gen++;
2357
2358         error = xfs_dir_init(tp, cdp, dp);
2359         if (error)
2360                 goto error2;
2361
2362         cdp->i_gen = 1;
2363         error = xfs_bumplink(tp, dp);
2364         if (error)
2365                 goto error2;
2366
2367         created = B_TRUE;
2368
2369         *ipp = cdp;
2370         IHOLD(cdp);
2371
2372         /*
2373          * Attach the dquots to the new inode and modify the icount incore.
2374          */
2375         XFS_QM_DQVOPCREATE(mp, tp, cdp, udqp, gdqp);
2376
2377         /*
2378          * If this is a synchronous mount, make sure that the
2379          * mkdir transaction goes to disk before returning to
2380          * the user.
2381          */
2382         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2383                 xfs_trans_set_sync(tp);
2384         }
2385
2386         error = xfs_bmap_finish(&tp, &free_list, &committed);
2387         if (error) {
2388                 IRELE(cdp);
2389                 goto error2;
2390         }
2391
2392         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2393         XFS_QM_DQRELE(mp, udqp);
2394         XFS_QM_DQRELE(mp, gdqp);
2395         if (error) {
2396                 IRELE(cdp);
2397         }
2398
2399         /* Fall through to std_return with error = 0 or errno from
2400          * xfs_trans_commit. */
2401
2402 std_return:
2403         if ((created || (error != 0 && dm_event_sent != 0)) &&
2404             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
2405                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
2406                                         dp, DM_RIGHT_NULL,
2407                                         created ? cdp : NULL,
2408                                         DM_RIGHT_NULL,
2409                                         dir_name->name, NULL,
2410                                         mode, error, 0);
2411         }
2412         return error;
2413
2414  error2:
2415  error1:
2416         xfs_bmap_cancel(&free_list);
2417  abort_return:
2418         cancel_flags |= XFS_TRANS_ABORT;
2419  error_return:
2420         xfs_trans_cancel(tp, cancel_flags);
2421         XFS_QM_DQRELE(mp, udqp);
2422         XFS_QM_DQRELE(mp, gdqp);
2423
2424         if (unlock_dp_on_error)
2425                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2426
2427         goto std_return;
2428 }
2429
2430 int
2431 xfs_symlink(
2432         xfs_inode_t             *dp,
2433         struct xfs_name         *link_name,
2434         const char              *target_path,
2435         mode_t                  mode,
2436         xfs_inode_t             **ipp,
2437         cred_t                  *credp)
2438 {
2439         xfs_mount_t             *mp = dp->i_mount;
2440         xfs_trans_t             *tp;
2441         xfs_inode_t             *ip;
2442         int                     error;
2443         int                     pathlen;
2444         xfs_bmap_free_t         free_list;
2445         xfs_fsblock_t           first_block;
2446         boolean_t               unlock_dp_on_error = B_FALSE;
2447         uint                    cancel_flags;
2448         int                     committed;
2449         xfs_fileoff_t           first_fsb;
2450         xfs_filblks_t           fs_blocks;
2451         int                     nmaps;
2452         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
2453         xfs_daddr_t             d;
2454         const char              *cur_chunk;
2455         int                     byte_cnt;
2456         int                     n;
2457         xfs_buf_t               *bp;
2458         xfs_prid_t              prid;
2459         struct xfs_dquot        *udqp, *gdqp;
2460         uint                    resblks;
2461
2462         *ipp = NULL;
2463         error = 0;
2464         ip = NULL;
2465         tp = NULL;
2466
2467         xfs_itrace_entry(dp);
2468
2469         if (XFS_FORCED_SHUTDOWN(mp))
2470                 return XFS_ERROR(EIO);
2471
2472         /*
2473          * Check component lengths of the target path name.
2474          */
2475         pathlen = strlen(target_path);
2476         if (pathlen >= MAXPATHLEN)      /* total string too long */
2477                 return XFS_ERROR(ENAMETOOLONG);
2478
2479         if (DM_EVENT_ENABLED(dp, DM_EVENT_SYMLINK)) {
2480                 error = XFS_SEND_NAMESP(mp, DM_EVENT_SYMLINK, dp,
2481                                         DM_RIGHT_NULL, NULL, DM_RIGHT_NULL,
2482                                         link_name->name, target_path, 0, 0, 0);
2483                 if (error)
2484                         return error;
2485         }
2486
2487         /* Return through std_return after this point. */
2488
2489         udqp = gdqp = NULL;
2490         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2491                 prid = dp->i_d.di_projid;
2492         else
2493                 prid = (xfs_prid_t)dfltprid;
2494
2495         /*
2496          * Make sure that we have allocated dquot(s) on disk.
2497          */
2498         error = XFS_QM_DQVOPALLOC(mp, dp,
2499                         current_fsuid(credp), current_fsgid(credp), prid,
2500                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2501         if (error)
2502                 goto std_return;
2503
2504         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
2505         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2506         /*
2507          * The symlink will fit into the inode data fork?
2508          * There can't be any attributes so we get the whole variable part.
2509          */
2510         if (pathlen <= XFS_LITINO(mp))
2511                 fs_blocks = 0;
2512         else
2513                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
2514         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
2515         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
2516                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2517         if (error == ENOSPC && fs_blocks == 0) {
2518                 resblks = 0;
2519                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
2520                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
2521         }
2522         if (error) {
2523                 cancel_flags = 0;
2524                 goto error_return;
2525         }
2526
2527         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2528         unlock_dp_on_error = B_TRUE;
2529
2530         /*
2531          * Check whether the directory allows new symlinks or not.
2532          */
2533         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
2534                 error = XFS_ERROR(EPERM);
2535                 goto error_return;
2536         }
2537
2538         /*
2539          * Reserve disk quota : blocks and inode.
2540          */
2541         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2542         if (error)
2543                 goto error_return;
2544
2545         /*
2546          * Check for ability to enter directory entry, if no space reserved.
2547          */
2548         error = xfs_dir_canenter(tp, dp, link_name, resblks);
2549         if (error)
2550                 goto error_return;
2551         /*
2552          * Initialize the bmap freelist prior to calling either
2553          * bmapi or the directory create code.
2554          */
2555         XFS_BMAP_INIT(&free_list, &first_block);
2556
2557         /*
2558          * Allocate an inode for the symlink.
2559          */
2560         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT),
2561                                1, 0, credp, prid, resblks > 0, &ip, NULL);
2562         if (error) {
2563                 if (error == ENOSPC)
2564                         goto error_return;
2565                 goto error1;
2566         }
2567         xfs_itrace_ref(ip);
2568
2569         /*
2570          * An error after we've joined dp to the transaction will result in the
2571          * transaction cancel unlocking dp so don't do it explicitly in the
2572          * error path.
2573          */
2574         IHOLD(dp);
2575         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2576         unlock_dp_on_error = B_FALSE;
2577
2578         /*
2579          * Also attach the dquot(s) to it, if applicable.
2580          */
2581         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
2582
2583         if (resblks)
2584                 resblks -= XFS_IALLOC_SPACE_RES(mp);
2585         /*
2586          * If the symlink will fit into the inode, write it inline.
2587          */
2588         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
2589                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
2590                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
2591                 ip->i_d.di_size = pathlen;
2592
2593                 /*
2594                  * The inode was initially created in extent format.
2595                  */
2596                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
2597                 ip->i_df.if_flags |= XFS_IFINLINE;
2598
2599                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
2600                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
2601
2602         } else {
2603                 first_fsb = 0;
2604                 nmaps = SYMLINK_MAPS;
2605
2606                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
2607                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
2608                                   &first_block, resblks, mval, &nmaps,
2609                                   &free_list, NULL);
2610                 if (error) {
2611                         goto error1;
2612                 }
2613
2614                 if (resblks)
2615                         resblks -= fs_blocks;
2616                 ip->i_d.di_size = pathlen;
2617                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2618
2619                 cur_chunk = target_path;
2620                 for (n = 0; n < nmaps; n++) {
2621                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
2622                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
2623                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
2624                                                BTOBB(byte_cnt), 0);
2625                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
2626                         if (pathlen < byte_cnt) {
2627                                 byte_cnt = pathlen;
2628                         }
2629                         pathlen -= byte_cnt;
2630
2631                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
2632                         cur_chunk += byte_cnt;
2633
2634                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
2635                 }
2636         }
2637
2638         /*
2639          * Create the directory entry for the symlink.
2640          */
2641         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
2642                                         &first_block, &free_list, resblks);
2643         if (error)
2644                 goto error1;
2645         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2646         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2647
2648         /*
2649          * Bump the in memory version number of the parent directory
2650          * so that other processes accessing it will recognize that
2651          * the directory has changed.
2652          */
2653         dp->i_gen++;
2654
2655         /*
2656          * If this is a synchronous mount, make sure that the
2657          * symlink transaction goes to disk before returning to
2658          * the user.
2659          */
2660         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2661                 xfs_trans_set_sync(tp);
2662         }
2663
2664         /*
2665          * xfs_trans_commit normally decrements the vnode ref count
2666          * when it unlocks the inode. Since we want to return the
2667          * vnode to the caller, we bump the vnode ref count now.
2668          */
2669         IHOLD(ip);
2670
2671         error = xfs_bmap_finish(&tp, &free_list, &committed);
2672         if (error) {
2673                 goto error2;
2674         }
2675         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2676         XFS_QM_DQRELE(mp, udqp);
2677         XFS_QM_DQRELE(mp, gdqp);
2678
2679         /* Fall through to std_return with error = 0 or errno from
2680          * xfs_trans_commit     */
2681 std_return:
2682         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTSYMLINK)) {
2683                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTSYMLINK,
2684                                         dp, DM_RIGHT_NULL,
2685                                         error ? NULL : ip,
2686                                         DM_RIGHT_NULL, link_name->name,
2687                                         target_path, 0, error, 0);
2688         }
2689
2690         if (!error)
2691                 *ipp = ip;
2692         return error;
2693
2694  error2:
2695         IRELE(ip);
2696  error1:
2697         xfs_bmap_cancel(&free_list);
2698         cancel_flags |= XFS_TRANS_ABORT;
2699  error_return:
2700         xfs_trans_cancel(tp, cancel_flags);
2701         XFS_QM_DQRELE(mp, udqp);
2702         XFS_QM_DQRELE(mp, gdqp);
2703
2704         if (unlock_dp_on_error)
2705                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2706
2707         goto std_return;
2708 }
2709
2710 int
2711 xfs_inode_flush(
2712         xfs_inode_t     *ip,
2713         int             flags)
2714 {
2715         xfs_mount_t     *mp = ip->i_mount;
2716         int             error = 0;
2717
2718         if (XFS_FORCED_SHUTDOWN(mp))
2719                 return XFS_ERROR(EIO);
2720
2721         /*
2722          * Bypass inodes which have already been cleaned by
2723          * the inode flush clustering code inside xfs_iflush
2724          */
2725         if (xfs_inode_clean(ip))
2726                 return 0;
2727
2728         /*
2729          * We make this non-blocking if the inode is contended,
2730          * return EAGAIN to indicate to the caller that they
2731          * did not succeed. This prevents the flush path from
2732          * blocking on inodes inside another operation right
2733          * now, they get caught later by xfs_sync.
2734          */
2735         if (flags & FLUSH_SYNC) {
2736                 xfs_ilock(ip, XFS_ILOCK_SHARED);
2737                 xfs_iflock(ip);
2738         } else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
2739                 if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
2740                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2741                         return EAGAIN;
2742                 }
2743         } else {
2744                 return EAGAIN;
2745         }
2746
2747         error = xfs_iflush(ip, (flags & FLUSH_SYNC) ? XFS_IFLUSH_SYNC
2748                                                     : XFS_IFLUSH_ASYNC_NOBLOCK);
2749         xfs_iunlock(ip, XFS_ILOCK_SHARED);
2750
2751         return error;
2752 }
2753
2754
2755 int
2756 xfs_set_dmattrs(
2757         xfs_inode_t     *ip,
2758         u_int           evmask,
2759         u_int16_t       state)
2760 {
2761         xfs_mount_t     *mp = ip->i_mount;
2762         xfs_trans_t     *tp;
2763         int             error;
2764
2765         if (!capable(CAP_SYS_ADMIN))
2766                 return XFS_ERROR(EPERM);
2767
2768         if (XFS_FORCED_SHUTDOWN(mp))
2769                 return XFS_ERROR(EIO);
2770
2771         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
2772         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
2773         if (error) {
2774                 xfs_trans_cancel(tp, 0);
2775                 return error;
2776         }
2777         xfs_ilock(ip, XFS_ILOCK_EXCL);
2778         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2779
2780         ip->i_d.di_dmevmask = evmask;
2781         ip->i_d.di_dmstate  = state;
2782
2783         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2784         IHOLD(ip);
2785         error = xfs_trans_commit(tp, 0);
2786
2787         return error;
2788 }
2789
2790 int
2791 xfs_reclaim(
2792         xfs_inode_t     *ip)
2793 {
2794
2795         xfs_itrace_entry(ip);
2796
2797         ASSERT(!VN_MAPPED(VFS_I(ip)));
2798
2799         /* bad inode, get out here ASAP */
2800         if (VN_BAD(VFS_I(ip))) {
2801                 xfs_ireclaim(ip);
2802                 return 0;
2803         }
2804
2805         vn_iowait(ip);
2806
2807         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0);
2808
2809         /*
2810          * Make sure the atime in the XFS inode is correct before freeing the
2811          * Linux inode.
2812          */
2813         xfs_synchronize_atime(ip);
2814
2815         /*
2816          * If we have nothing to flush with this inode then complete the
2817          * teardown now, otherwise break the link between the xfs inode and the
2818          * linux inode and clean up the xfs inode later. This avoids flushing
2819          * the inode to disk during the delete operation itself.
2820          *
2821          * When breaking the link, we need to set the XFS_IRECLAIMABLE flag
2822          * first to ensure that xfs_iunpin() will never see an xfs inode
2823          * that has a linux inode being reclaimed. Synchronisation is provided
2824          * by the i_flags_lock.
2825          */
2826         if (!ip->i_update_core && (ip->i_itemp == NULL)) {
2827                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2828                 xfs_iflock(ip);
2829                 return xfs_finish_reclaim(ip, 1, XFS_IFLUSH_DELWRI_ELSE_SYNC);
2830         } else {
2831                 xfs_mount_t     *mp = ip->i_mount;
2832
2833                 /* Protect sync and unpin from us */
2834                 XFS_MOUNT_ILOCK(mp);
2835                 spin_lock(&ip->i_flags_lock);
2836                 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
2837                 VFS_I(ip)->i_private = NULL;
2838                 ip->i_vnode = NULL;
2839                 spin_unlock(&ip->i_flags_lock);
2840                 list_add_tail(&ip->i_reclaim, &mp->m_del_inodes);
2841                 XFS_MOUNT_IUNLOCK(mp);
2842         }
2843         return 0;
2844 }
2845
2846 int
2847 xfs_finish_reclaim(
2848         xfs_inode_t     *ip,
2849         int             locked,
2850         int             sync_mode)
2851 {
2852         xfs_perag_t     *pag = xfs_get_perag(ip->i_mount, ip->i_ino);
2853         struct inode    *vp = VFS_I(ip);
2854
2855         if (vp && VN_BAD(vp))
2856                 goto reclaim;
2857
2858         /* The hash lock here protects a thread in xfs_iget_core from
2859          * racing with us on linking the inode back with a vnode.
2860          * Once we have the XFS_IRECLAIM flag set it will not touch
2861          * us.
2862          */
2863         write_lock(&pag->pag_ici_lock);
2864         spin_lock(&ip->i_flags_lock);
2865         if (__xfs_iflags_test(ip, XFS_IRECLAIM) ||
2866             (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) && vp == NULL)) {
2867                 spin_unlock(&ip->i_flags_lock);
2868                 write_unlock(&pag->pag_ici_lock);
2869                 if (locked) {
2870                         xfs_ifunlock(ip);
2871                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2872                 }
2873                 return 1;
2874         }
2875         __xfs_iflags_set(ip, XFS_IRECLAIM);
2876         spin_unlock(&ip->i_flags_lock);
2877         write_unlock(&pag->pag_ici_lock);
2878         xfs_put_perag(ip->i_mount, pag);
2879
2880         /*
2881          * If the inode is still dirty, then flush it out.  If the inode
2882          * is not in the AIL, then it will be OK to flush it delwri as
2883          * long as xfs_iflush() does not keep any references to the inode.
2884          * We leave that decision up to xfs_iflush() since it has the
2885          * knowledge of whether it's OK to simply do a delwri flush of
2886          * the inode or whether we need to wait until the inode is
2887          * pulled from the AIL.
2888          * We get the flush lock regardless, though, just to make sure
2889          * we don't free it while it is being flushed.
2890          */
2891         if (!locked) {
2892                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2893                 xfs_iflock(ip);
2894         }
2895
2896         /*
2897          * In the case of a forced shutdown we rely on xfs_iflush() to
2898          * wait for the inode to be unpinned before returning an error.
2899          */
2900         if (xfs_iflush(ip, sync_mode) == 0) {
2901                 /* synchronize with xfs_iflush_done */
2902                 xfs_iflock(ip);
2903                 xfs_ifunlock(ip);
2904         }
2905
2906         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2907
2908  reclaim:
2909         xfs_ireclaim(ip);
2910         return 0;
2911 }
2912
2913 int
2914 xfs_finish_reclaim_all(xfs_mount_t *mp, int noblock)
2915 {
2916         int             purged;
2917         xfs_inode_t     *ip, *n;
2918         int             done = 0;
2919
2920         while (!done) {
2921                 purged = 0;
2922                 XFS_MOUNT_ILOCK(mp);
2923                 list_for_each_entry_safe(ip, n, &mp->m_del_inodes, i_reclaim) {
2924                         if (noblock) {
2925                                 if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0)
2926                                         continue;
2927                                 if (xfs_ipincount(ip) ||
2928                                     !xfs_iflock_nowait(ip)) {
2929                                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2930                                         continue;
2931                                 }
2932                         }
2933                         XFS_MOUNT_IUNLOCK(mp);
2934                         if (xfs_finish_reclaim(ip, noblock,
2935                                         XFS_IFLUSH_DELWRI_ELSE_ASYNC))
2936                                 delay(1);
2937                         purged = 1;
2938                         break;
2939                 }
2940
2941                 done = !purged;
2942         }
2943
2944         XFS_MOUNT_IUNLOCK(mp);
2945         return 0;
2946 }
2947
2948 /*
2949  * xfs_alloc_file_space()
2950  *      This routine allocates disk space for the given file.
2951  *
2952  *      If alloc_type == 0, this request is for an ALLOCSP type
2953  *      request which will change the file size.  In this case, no
2954  *      DMAPI event will be generated by the call.  A TRUNCATE event
2955  *      will be generated later by xfs_setattr.
2956  *
2957  *      If alloc_type != 0, this request is for a RESVSP type
2958  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
2959  *      lower block boundary byte address is less than the file's
2960  *      length.
2961  *
2962  * RETURNS:
2963  *       0 on success
2964  *      errno on error
2965  *
2966  */
2967 STATIC int
2968 xfs_alloc_file_space(
2969         xfs_inode_t             *ip,
2970         xfs_off_t               offset,
2971         xfs_off_t               len,
2972         int                     alloc_type,
2973         int                     attr_flags)
2974 {
2975         xfs_mount_t             *mp = ip->i_mount;
2976         xfs_off_t               count;
2977         xfs_filblks_t           allocated_fsb;
2978         xfs_filblks_t           allocatesize_fsb;
2979         xfs_extlen_t            extsz, temp;
2980         xfs_fileoff_t           startoffset_fsb;
2981         xfs_fsblock_t           firstfsb;
2982         int                     nimaps;
2983         int                     bmapi_flag;
2984         int                     quota_flag;
2985         int                     rt;
2986         xfs_trans_t             *tp;
2987         xfs_bmbt_irec_t         imaps[1], *imapp;
2988         xfs_bmap_free_t         free_list;
2989         uint                    qblocks, resblks, resrtextents;
2990         int                     committed;
2991         int                     error;
2992
2993         xfs_itrace_entry(ip);
2994
2995         if (XFS_FORCED_SHUTDOWN(mp))
2996                 return XFS_ERROR(EIO);
2997
2998         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
2999                 return error;
3000
3001         if (len <= 0)
3002                 return XFS_ERROR(EINVAL);
3003
3004         rt = XFS_IS_REALTIME_INODE(ip);
3005         extsz = xfs_get_extsz_hint(ip);
3006
3007         count = len;
3008         imapp = &imaps[0];
3009         nimaps = 1;
3010         bmapi_flag = XFS_BMAPI_WRITE | (alloc_type ? XFS_BMAPI_PREALLOC : 0);
3011         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
3012         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
3013
3014         /*      Generate a DMAPI event if needed.       */
3015         if (alloc_type != 0 && offset < ip->i_size &&
3016                         (attr_flags & XFS_ATTR_DMI) == 0  &&
3017                         DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3018                 xfs_off_t           end_dmi_offset;
3019
3020                 end_dmi_offset = offset+len;
3021                 if (end_dmi_offset > ip->i_size)
3022                         end_dmi_offset = ip->i_size;
3023                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, offset,
3024                                       end_dmi_offset - offset, 0, NULL);
3025                 if (error)
3026                         return error;
3027         }
3028
3029         /*
3030          * Allocate file space until done or until there is an error
3031          */
3032 retry:
3033         while (allocatesize_fsb && !error) {
3034                 xfs_fileoff_t   s, e;
3035
3036                 /*
3037                  * Determine space reservations for data/realtime.
3038                  */
3039                 if (unlikely(extsz)) {
3040                         s = startoffset_fsb;
3041                         do_div(s, extsz);
3042                         s *= extsz;
3043                         e = startoffset_fsb + allocatesize_fsb;
3044                         if ((temp = do_mod(startoffset_fsb, extsz)))
3045                                 e += temp;
3046                         if ((temp = do_mod(e, extsz)))
3047                                 e += extsz - temp;
3048                 } else {
3049                         s = 0;
3050                         e = allocatesize_fsb;
3051                 }
3052
3053                 if (unlikely(rt)) {
3054                         resrtextents = qblocks = (uint)(e - s);
3055                         resrtextents /= mp->m_sb.sb_rextsize;
3056                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3057                         quota_flag = XFS_QMOPT_RES_RTBLKS;
3058                 } else {
3059                         resrtextents = 0;
3060                         resblks = qblocks = \
3061                                 XFS_DIOSTRAT_SPACE_RES(mp, (uint)(e - s));
3062                         quota_flag = XFS_QMOPT_RES_REGBLKS;
3063                 }
3064
3065                 /*
3066                  * Allocate and setup the transaction.
3067                  */
3068                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3069                 error = xfs_trans_reserve(tp, resblks,
3070                                           XFS_WRITE_LOG_RES(mp), resrtextents,
3071                                           XFS_TRANS_PERM_LOG_RES,
3072                                           XFS_WRITE_LOG_COUNT);
3073                 /*
3074                  * Check for running out of space
3075                  */
3076                 if (error) {
3077                         /*
3078                          * Free the transaction structure.
3079                          */
3080                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3081                         xfs_trans_cancel(tp, 0);
3082                         break;
3083                 }
3084                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3085                 error = XFS_TRANS_RESERVE_QUOTA_NBLKS(mp, tp, ip,
3086                                                       qblocks, 0, quota_flag);
3087                 if (error)
3088                         goto error1;
3089
3090                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3091                 xfs_trans_ihold(tp, ip);
3092
3093                 /*
3094                  * Issue the xfs_bmapi() call to allocate the blocks
3095                  */
3096                 XFS_BMAP_INIT(&free_list, &firstfsb);
3097                 error = xfs_bmapi(tp, ip, startoffset_fsb,
3098                                   allocatesize_fsb, bmapi_flag,
3099                                   &firstfsb, 0, imapp, &nimaps,
3100                                   &free_list, NULL);
3101                 if (error) {
3102                         goto error0;
3103                 }
3104
3105                 /*
3106                  * Complete the transaction
3107                  */
3108                 error = xfs_bmap_finish(&tp, &free_list, &committed);
3109                 if (error) {
3110                         goto error0;
3111                 }
3112
3113                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3114                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3115                 if (error) {
3116                         break;
3117                 }
3118
3119                 allocated_fsb = imapp->br_blockcount;
3120
3121                 if (nimaps == 0) {
3122                         error = XFS_ERROR(ENOSPC);
3123                         break;
3124                 }
3125
3126                 startoffset_fsb += allocated_fsb;
3127                 allocatesize_fsb -= allocated_fsb;
3128         }
3129 dmapi_enospc_check:
3130         if (error == ENOSPC && (attr_flags & XFS_ATTR_DMI) == 0 &&
3131             DM_EVENT_ENABLED(ip, DM_EVENT_NOSPACE)) {
3132                 error = XFS_SEND_NAMESP(mp, DM_EVENT_NOSPACE,
3133                                 ip, DM_RIGHT_NULL,
3134                                 ip, DM_RIGHT_NULL,
3135                                 NULL, NULL, 0, 0, 0); /* Delay flag intentionally unused */
3136                 if (error == 0)
3137                         goto retry;     /* Maybe DMAPI app. has made space */
3138                 /* else fall through with error from XFS_SEND_DATA */
3139         }
3140
3141         return error;
3142
3143 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
3144         xfs_bmap_cancel(&free_list);
3145         XFS_TRANS_UNRESERVE_QUOTA_NBLKS(mp, tp, ip, qblocks, 0, quota_flag);
3146
3147 error1: /* Just cancel transaction */
3148         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
3149         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3150         goto dmapi_enospc_check;
3151 }
3152
3153 /*
3154  * Zero file bytes between startoff and endoff inclusive.
3155  * The iolock is held exclusive and no blocks are buffered.
3156  */
3157 STATIC int
3158 xfs_zero_remaining_bytes(
3159         xfs_inode_t             *ip,
3160         xfs_off_t               startoff,
3161         xfs_off_t               endoff)
3162 {
3163         xfs_bmbt_irec_t         imap;
3164         xfs_fileoff_t           offset_fsb;
3165         xfs_off_t               lastoffset;
3166         xfs_off_t               offset;
3167         xfs_buf_t               *bp;
3168         xfs_mount_t             *mp = ip->i_mount;
3169         int                     nimap;
3170         int                     error = 0;
3171
3172         bp = xfs_buf_get_noaddr(mp->m_sb.sb_blocksize,
3173                                 XFS_IS_REALTIME_INODE(ip) ?
3174                                 mp->m_rtdev_targp : mp->m_ddev_targp);
3175
3176         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
3177                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
3178                 nimap = 1;
3179                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
3180                         NULL, 0, &imap, &nimap, NULL, NULL);
3181                 if (error || nimap < 1)
3182                         break;
3183                 ASSERT(imap.br_blockcount >= 1);
3184                 ASSERT(imap.br_startoff == offset_fsb);
3185                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
3186                 if (lastoffset > endoff)
3187                         lastoffset = endoff;
3188                 if (imap.br_startblock == HOLESTARTBLOCK)
3189                         continue;
3190                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3191                 if (imap.br_state == XFS_EXT_UNWRITTEN)
3192                         continue;
3193                 XFS_BUF_UNDONE(bp);
3194                 XFS_BUF_UNWRITE(bp);
3195                 XFS_BUF_READ(bp);
3196                 XFS_BUF_SET_ADDR(bp, XFS_FSB_TO_DB(ip, imap.br_startblock));
3197                 xfsbdstrat(mp, bp);
3198                 error = xfs_iowait(bp);
3199                 if (error) {
3200                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
3201                                           mp, bp, XFS_BUF_ADDR(bp));
3202                         break;
3203                 }
3204                 memset(XFS_BUF_PTR(bp) +
3205                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
3206                       0, lastoffset - offset + 1);
3207                 XFS_BUF_UNDONE(bp);
3208                 XFS_BUF_UNREAD(bp);
3209                 XFS_BUF_WRITE(bp);
3210                 xfsbdstrat(mp, bp);
3211                 error = xfs_iowait(bp);
3212                 if (error) {
3213                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
3214                                           mp, bp, XFS_BUF_ADDR(bp));
3215                         break;
3216                 }
3217         }
3218         xfs_buf_free(bp);
3219         return error;
3220 }
3221
3222 /*
3223  * xfs_free_file_space()
3224  *      This routine frees disk space for the given file.
3225  *
3226  *      This routine is only called by xfs_change_file_space
3227  *      for an UNRESVSP type call.
3228  *
3229  * RETURNS:
3230  *       0 on success
3231  *      errno on error
3232  *
3233  */
3234 STATIC int
3235 xfs_free_file_space(
3236         xfs_inode_t             *ip,
3237         xfs_off_t               offset,
3238         xfs_off_t               len,
3239         int                     attr_flags)
3240 {
3241         int                     committed;
3242         int                     done;
3243         xfs_off_t               end_dmi_offset;
3244         xfs_fileoff_t           endoffset_fsb;
3245         int                     error;
3246         xfs_fsblock_t           firstfsb;
3247         xfs_bmap_free_t         free_list;
3248         xfs_bmbt_irec_t         imap;
3249         xfs_off_t               ioffset;
3250         xfs_extlen_t            mod=0;
3251         xfs_mount_t             *mp;
3252         int                     nimap;
3253         uint                    resblks;
3254         uint                    rounding;
3255         int                     rt;
3256         xfs_fileoff_t           startoffset_fsb;
3257         xfs_trans_t             *tp;
3258         int                     need_iolock = 1;
3259
3260         mp = ip->i_mount;
3261
3262         xfs_itrace_entry(ip);
3263
3264         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
3265                 return error;
3266
3267         error = 0;
3268         if (len <= 0)   /* if nothing being freed */
3269                 return error;
3270         rt = XFS_IS_REALTIME_INODE(ip);
3271         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
3272         end_dmi_offset = offset + len;
3273         endoffset_fsb = XFS_B_TO_FSBT(mp, end_dmi_offset);
3274
3275         if (offset < ip->i_size && (attr_flags & XFS_ATTR_DMI) == 0 &&
3276             DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3277                 if (end_dmi_offset > ip->i_size)
3278                         end_dmi_offset = ip->i_size;
3279                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip,
3280                                 offset, end_dmi_offset - offset,
3281                                 AT_DELAY_FLAG(attr_flags), NULL);
3282                 if (error)
3283                         return error;
3284         }
3285
3286         if (attr_flags & XFS_ATTR_NOLOCK)
3287                 need_iolock = 0;
3288         if (need_iolock) {
3289                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
3290                 vn_iowait(ip);  /* wait for the completion of any pending DIOs */
3291         }
3292
3293         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
3294         ioffset = offset & ~(rounding - 1);
3295
3296         if (VN_CACHED(VFS_I(ip)) != 0) {
3297                 xfs_inval_cached_trace(ip, ioffset, -1, ioffset, -1);
3298                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
3299                 if (error)
3300                         goto out_unlock_iolock;
3301         }
3302
3303         /*
3304          * Need to zero the stuff we're not freeing, on disk.
3305          * If its a realtime file & can't use unwritten extents then we
3306          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
3307          * will take care of it for us.
3308          */
3309         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
3310                 nimap = 1;
3311                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
3312                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3313                 if (error)
3314                         goto out_unlock_iolock;
3315                 ASSERT(nimap == 0 || nimap == 1);
3316                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3317                         xfs_daddr_t     block;
3318
3319                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3320                         block = imap.br_startblock;
3321                         mod = do_div(block, mp->m_sb.sb_rextsize);
3322                         if (mod)
3323                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
3324                 }
3325                 nimap = 1;
3326                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
3327                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
3328                 if (error)
3329                         goto out_unlock_iolock;
3330                 ASSERT(nimap == 0 || nimap == 1);
3331                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
3332                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
3333                         mod++;
3334                         if (mod && (mod != mp->m_sb.sb_rextsize))
3335                                 endoffset_fsb -= mod;
3336                 }
3337         }
3338         if ((done = (endoffset_fsb <= startoffset_fsb)))
3339                 /*
3340                  * One contiguous piece to clear
3341                  */
3342                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
3343         else {
3344                 /*
3345                  * Some full blocks, possibly two pieces to clear
3346                  */
3347                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
3348                         error = xfs_zero_remaining_bytes(ip, offset,
3349                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
3350                 if (!error &&
3351                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
3352                         error = xfs_zero_remaining_bytes(ip,
3353                                 XFS_FSB_TO_B(mp, endoffset_fsb),
3354                                 offset + len - 1);
3355         }
3356
3357         /*
3358          * free file space until done or until there is an error
3359          */
3360         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3361         while (!error && !done) {
3362
3363                 /*
3364                  * allocate and setup the transaction. Allow this
3365                  * transaction to dip into the reserve blocks to ensure
3366                  * the freeing of the space succeeds at ENOSPC.
3367                  */
3368                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3369                 tp->t_flags |= XFS_TRANS_RESERVE;
3370                 error = xfs_trans_reserve(tp,
3371                                           resblks,
3372                                           XFS_WRITE_LOG_RES(mp),
3373                                           0,
3374                                           XFS_TRANS_PERM_LOG_RES,
3375                                           XFS_WRITE_LOG_COUNT);
3376
3377                 /*
3378                  * check for running out of space
3379                  */
3380                 if (error) {
3381                         /*
3382                          * Free the transaction structure.
3383                          */
3384                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3385                         xfs_trans_cancel(tp, 0);
3386                         break;
3387                 }
3388                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3389                 error = XFS_TRANS_RESERVE_QUOTA(mp, tp,
3390                                 ip->i_udquot, ip->i_gdquot, resblks, 0,
3391                                 XFS_QMOPT_RES_REGBLKS);
3392                 if (error)
3393                         goto error1;
3394
3395                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3396                 xfs_trans_ihold(tp, ip);
3397
3398                 /*
3399                  * issue the bunmapi() call to free the blocks
3400                  */
3401                 XFS_BMAP_INIT(&free_list, &firstfsb);
3402                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
3403                                   endoffset_fsb - startoffset_fsb,
3404                                   0, 2, &firstfsb, &free_list, NULL, &done);
3405                 if (error) {
3406                         goto error0;
3407                 }
3408
3409                 /*
3410                  * complete the transaction
3411                  */
3412                 error = xfs_bmap_finish(&tp, &free_list, &committed);
3413                 if (error) {
3414                         goto error0;
3415                 }
3416
3417                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3418                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3419         }
3420
3421  out_unlock_iolock:
3422         if (need_iolock)
3423                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
3424         return error;
3425
3426  error0:
3427         xfs_bmap_cancel(&free_list);
3428  error1:
3429         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
3430         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
3431                     XFS_ILOCK_EXCL);
3432         return error;
3433 }
3434
3435 /*
3436  * xfs_change_file_space()
3437  *      This routine allocates or frees disk space for the given file.
3438  *      The user specified parameters are checked for alignment and size
3439  *      limitations.
3440  *
3441  * RETURNS:
3442  *       0 on success
3443  *      errno on error
3444  *
3445  */
3446 int
3447 xfs_change_file_space(
3448         xfs_inode_t     *ip,
3449         int             cmd,
3450         xfs_flock64_t   *bf,
3451         xfs_off_t       offset,
3452         cred_t          *credp,
3453         int             attr_flags)
3454 {
3455         xfs_mount_t     *mp = ip->i_mount;
3456         int             clrprealloc;
3457         int             error;
3458         xfs_fsize_t     fsize;
3459         int             setprealloc;
3460         xfs_off_t       startoffset;
3461         xfs_off_t       llen;
3462         xfs_trans_t     *tp;
3463         struct iattr    iattr;
3464
3465         xfs_itrace_entry(ip);
3466
3467         if (!S_ISREG(ip->i_d.di_mode))
3468                 return XFS_ERROR(EINVAL);
3469
3470         switch (bf->l_whence) {
3471         case 0: /*SEEK_SET*/
3472                 break;
3473         case 1: /*SEEK_CUR*/
3474                 bf->l_start += offset;
3475                 break;
3476         case 2: /*SEEK_END*/
3477                 bf->l_start += ip->i_size;
3478                 break;
3479         default:
3480                 return XFS_ERROR(EINVAL);
3481         }
3482
3483         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
3484
3485         if (   (bf->l_start < 0)
3486             || (bf->l_start > XFS_MAXIOFFSET(mp))
3487             || (bf->l_start + llen < 0)
3488             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
3489                 return XFS_ERROR(EINVAL);
3490
3491         bf->l_whence = 0;
3492
3493         startoffset = bf->l_start;
3494         fsize = ip->i_size;
3495
3496         /*
3497          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
3498          * file space.
3499          * These calls do NOT zero the data space allocated to the file,
3500          * nor do they change the file size.
3501          *
3502          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
3503          * space.
3504          * These calls cause the new file data to be zeroed and the file
3505          * size to be changed.
3506          */
3507         setprealloc = clrprealloc = 0;
3508
3509         switch (cmd) {
3510         case XFS_IOC_RESVSP:
3511         case XFS_IOC_RESVSP64:
3512                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
3513                                                                 1, attr_flags);
3514                 if (error)
3515                         return error;
3516                 setprealloc = 1;
3517                 break;
3518
3519         case XFS_IOC_UNRESVSP:
3520         case XFS_IOC_UNRESVSP64:
3521                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
3522                                                                 attr_flags)))
3523                         return error;
3524                 break;
3525
3526         case XFS_IOC_ALLOCSP:
3527         case XFS_IOC_ALLOCSP64:
3528         case XFS_IOC_FREESP:
3529         case XFS_IOC_FREESP64:
3530                 if (startoffset > fsize) {
3531                         error = xfs_alloc_file_space(ip, fsize,
3532                                         startoffset - fsize, 0, attr_flags);
3533                         if (error)
3534                                 break;
3535                 }
3536
3537                 iattr.ia_valid = ATTR_SIZE;
3538                 iattr.ia_size = startoffset;
3539
3540                 error = xfs_setattr(ip, &iattr, attr_flags, credp);
3541
3542                 if (error)
3543                         return error;
3544
3545                 clrprealloc = 1;
3546                 break;
3547
3548         default:
3549                 ASSERT(0);
3550                 return XFS_ERROR(EINVAL);
3551         }
3552
3553         /*
3554          * update the inode timestamp, mode, and prealloc flag bits
3555          */
3556         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
3557
3558         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
3559                                       0, 0, 0))) {
3560                 /* ASSERT(0); */
3561                 xfs_trans_cancel(tp, 0);
3562                 return error;
3563         }
3564
3565         xfs_ilock(ip, XFS_ILOCK_EXCL);
3566
3567         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3568         xfs_trans_ihold(tp, ip);
3569
3570         if ((attr_flags & XFS_ATTR_DMI) == 0) {
3571                 ip->i_d.di_mode &= ~S_ISUID;
3572
3573                 /*
3574                  * Note that we don't have to worry about mandatory
3575                  * file locking being disabled here because we only
3576                  * clear the S_ISGID bit if the Group execute bit is
3577                  * on, but if it was on then mandatory locking wouldn't
3578                  * have been enabled.
3579                  */
3580                 if (ip->i_d.di_mode & S_IXGRP)
3581                         ip->i_d.di_mode &= ~S_ISGID;
3582
3583                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3584         }
3585         if (setprealloc)
3586                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
3587         else if (clrprealloc)
3588                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
3589
3590         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
3591         xfs_trans_set_sync(tp);
3592
3593         error = xfs_trans_commit(tp, 0);
3594
3595         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3596
3597         return error;
3598 }