Merge master.kernel.org:/pub/scm/linux/kernel/git/bart/ide-2.6
[linux-2.6] / 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_refcache.h"
52 #include "xfs_trans_space.h"
53 #include "xfs_log_priv.h"
54 #include "xfs_filestream.h"
55 #include "xfs_vnodeops.h"
56
57 int
58 xfs_open(
59         xfs_inode_t     *ip)
60 {
61         int             mode;
62
63         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
64                 return XFS_ERROR(EIO);
65
66         /*
67          * If it's a directory with any blocks, read-ahead block 0
68          * as we're almost certain to have the next operation be a read there.
69          */
70         if (S_ISDIR(ip->i_d.di_mode) && ip->i_d.di_nextents > 0) {
71                 mode = xfs_ilock_map_shared(ip);
72                 if (ip->i_d.di_nextents > 0)
73                         (void)xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
74                 xfs_iunlock(ip, mode);
75         }
76         return 0;
77 }
78
79 /*
80  * xfs_getattr
81  */
82 int
83 xfs_getattr(
84         xfs_inode_t     *ip,
85         bhv_vattr_t     *vap,
86         int             flags)
87 {
88         bhv_vnode_t     *vp = XFS_ITOV(ip);
89         xfs_mount_t     *mp = ip->i_mount;
90
91         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
92
93         if (XFS_FORCED_SHUTDOWN(mp))
94                 return XFS_ERROR(EIO);
95
96         if (!(flags & ATTR_LAZY))
97                 xfs_ilock(ip, XFS_ILOCK_SHARED);
98
99         vap->va_size = XFS_ISIZE(ip);
100         if (vap->va_mask == XFS_AT_SIZE)
101                 goto all_done;
102
103         vap->va_nblocks =
104                 XFS_FSB_TO_BB(mp, ip->i_d.di_nblocks + ip->i_delayed_blks);
105         vap->va_nodeid = ip->i_ino;
106 #if XFS_BIG_INUMS
107         vap->va_nodeid += mp->m_inoadd;
108 #endif
109         vap->va_nlink = ip->i_d.di_nlink;
110
111         /*
112          * Quick exit for non-stat callers
113          */
114         if ((vap->va_mask &
115             ~(XFS_AT_SIZE|XFS_AT_FSID|XFS_AT_NODEID|
116               XFS_AT_NLINK|XFS_AT_BLKSIZE)) == 0)
117                 goto all_done;
118
119         /*
120          * Copy from in-core inode.
121          */
122         vap->va_mode = ip->i_d.di_mode;
123         vap->va_uid = ip->i_d.di_uid;
124         vap->va_gid = ip->i_d.di_gid;
125         vap->va_projid = ip->i_d.di_projid;
126
127         /*
128          * Check vnode type block/char vs. everything else.
129          */
130         switch (ip->i_d.di_mode & S_IFMT) {
131         case S_IFBLK:
132         case S_IFCHR:
133                 vap->va_rdev = ip->i_df.if_u2.if_rdev;
134                 vap->va_blocksize = BLKDEV_IOSIZE;
135                 break;
136         default:
137                 vap->va_rdev = 0;
138
139                 if (!(ip->i_d.di_flags & XFS_DIFLAG_REALTIME)) {
140                         vap->va_blocksize = xfs_preferred_iosize(mp);
141                 } else {
142
143                         /*
144                          * If the file blocks are being allocated from a
145                          * realtime partition, then return the inode's
146                          * realtime extent size or the realtime volume's
147                          * extent size.
148                          */
149                         vap->va_blocksize =
150                                 xfs_get_extsz_hint(ip) << mp->m_sb.sb_blocklog;
151                 }
152                 break;
153         }
154
155         vn_atime_to_timespec(vp, &vap->va_atime);
156         vap->va_mtime.tv_sec = ip->i_d.di_mtime.t_sec;
157         vap->va_mtime.tv_nsec = ip->i_d.di_mtime.t_nsec;
158         vap->va_ctime.tv_sec = ip->i_d.di_ctime.t_sec;
159         vap->va_ctime.tv_nsec = ip->i_d.di_ctime.t_nsec;
160
161         /*
162          * Exit for stat callers.  See if any of the rest of the fields
163          * to be filled in are needed.
164          */
165         if ((vap->va_mask &
166              (XFS_AT_XFLAGS|XFS_AT_EXTSIZE|XFS_AT_NEXTENTS|XFS_AT_ANEXTENTS|
167               XFS_AT_GENCOUNT|XFS_AT_VCODE)) == 0)
168                 goto all_done;
169
170         /*
171          * Convert di_flags to xflags.
172          */
173         vap->va_xflags = xfs_ip2xflags(ip);
174
175         /*
176          * Exit for inode revalidate.  See if any of the rest of
177          * the fields to be filled in are needed.
178          */
179         if ((vap->va_mask &
180              (XFS_AT_EXTSIZE|XFS_AT_NEXTENTS|XFS_AT_ANEXTENTS|
181               XFS_AT_GENCOUNT|XFS_AT_VCODE)) == 0)
182                 goto all_done;
183
184         vap->va_extsize = ip->i_d.di_extsize << mp->m_sb.sb_blocklog;
185         vap->va_nextents =
186                 (ip->i_df.if_flags & XFS_IFEXTENTS) ?
187                         ip->i_df.if_bytes / sizeof(xfs_bmbt_rec_t) :
188                         ip->i_d.di_nextents;
189         if (ip->i_afp)
190                 vap->va_anextents =
191                         (ip->i_afp->if_flags & XFS_IFEXTENTS) ?
192                                 ip->i_afp->if_bytes / sizeof(xfs_bmbt_rec_t) :
193                                  ip->i_d.di_anextents;
194         else
195                 vap->va_anextents = 0;
196         vap->va_gen = ip->i_d.di_gen;
197
198  all_done:
199         if (!(flags & ATTR_LAZY))
200                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
201         return 0;
202 }
203
204
205 /*
206  * xfs_setattr
207  */
208 int
209 xfs_setattr(
210         xfs_inode_t             *ip,
211         bhv_vattr_t             *vap,
212         int                     flags,
213         cred_t                  *credp)
214 {
215         bhv_vnode_t             *vp = XFS_ITOV(ip);
216         xfs_mount_t             *mp = ip->i_mount;
217         xfs_trans_t             *tp;
218         int                     mask;
219         int                     code;
220         uint                    lock_flags;
221         uint                    commit_flags=0;
222         uid_t                   uid=0, iuid=0;
223         gid_t                   gid=0, igid=0;
224         int                     timeflags = 0;
225         xfs_prid_t              projid=0, iprojid=0;
226         int                     mandlock_before, mandlock_after;
227         struct xfs_dquot        *udqp, *gdqp, *olddquot1, *olddquot2;
228         int                     file_owner;
229         int                     need_iolock = 1;
230
231         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
232
233         if (mp->m_flags & XFS_MOUNT_RDONLY)
234                 return XFS_ERROR(EROFS);
235
236         /*
237          * Cannot set certain attributes.
238          */
239         mask = vap->va_mask;
240         if (mask & XFS_AT_NOSET) {
241                 return XFS_ERROR(EINVAL);
242         }
243
244         if (XFS_FORCED_SHUTDOWN(mp))
245                 return XFS_ERROR(EIO);
246
247         /*
248          * Timestamps do not need to be logged and hence do not
249          * need to be done within a transaction.
250          */
251         if (mask & XFS_AT_UPDTIMES) {
252                 ASSERT((mask & ~XFS_AT_UPDTIMES) == 0);
253                 timeflags = ((mask & XFS_AT_UPDATIME) ? XFS_ICHGTIME_ACC : 0) |
254                             ((mask & XFS_AT_UPDCTIME) ? XFS_ICHGTIME_CHG : 0) |
255                             ((mask & XFS_AT_UPDMTIME) ? XFS_ICHGTIME_MOD : 0);
256                 xfs_ichgtime(ip, timeflags);
257                 return 0;
258         }
259
260         olddquot1 = olddquot2 = NULL;
261         udqp = gdqp = NULL;
262
263         /*
264          * If disk quotas is on, we make sure that the dquots do exist on disk,
265          * before we start any other transactions. Trying to do this later
266          * is messy. We don't care to take a readlock to look at the ids
267          * in inode here, because we can't hold it across the trans_reserve.
268          * If the IDs do change before we take the ilock, we're covered
269          * because the i_*dquot fields will get updated anyway.
270          */
271         if (XFS_IS_QUOTA_ON(mp) &&
272             (mask & (XFS_AT_UID|XFS_AT_GID|XFS_AT_PROJID))) {
273                 uint    qflags = 0;
274
275                 if ((mask & XFS_AT_UID) && XFS_IS_UQUOTA_ON(mp)) {
276                         uid = vap->va_uid;
277                         qflags |= XFS_QMOPT_UQUOTA;
278                 } else {
279                         uid = ip->i_d.di_uid;
280                 }
281                 if ((mask & XFS_AT_GID) && XFS_IS_GQUOTA_ON(mp)) {
282                         gid = vap->va_gid;
283                         qflags |= XFS_QMOPT_GQUOTA;
284                 }  else {
285                         gid = ip->i_d.di_gid;
286                 }
287                 if ((mask & XFS_AT_PROJID) && XFS_IS_PQUOTA_ON(mp)) {
288                         projid = vap->va_projid;
289                         qflags |= XFS_QMOPT_PQUOTA;
290                 }  else {
291                         projid = ip->i_d.di_projid;
292                 }
293                 /*
294                  * We take a reference when we initialize udqp and gdqp,
295                  * so it is important that we never blindly double trip on
296                  * the same variable. See xfs_create() for an example.
297                  */
298                 ASSERT(udqp == NULL);
299                 ASSERT(gdqp == NULL);
300                 code = XFS_QM_DQVOPALLOC(mp, ip, uid, gid, projid, qflags,
301                                          &udqp, &gdqp);
302                 if (code)
303                         return code;
304         }
305
306         /*
307          * For the other attributes, we acquire the inode lock and
308          * first do an error checking pass.
309          */
310         tp = NULL;
311         lock_flags = XFS_ILOCK_EXCL;
312         if (flags & ATTR_NOLOCK)
313                 need_iolock = 0;
314         if (!(mask & XFS_AT_SIZE)) {
315                 if ((mask != (XFS_AT_CTIME|XFS_AT_ATIME|XFS_AT_MTIME)) ||
316                     (mp->m_flags & XFS_MOUNT_WSYNC)) {
317                         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
318                         commit_flags = 0;
319                         if ((code = xfs_trans_reserve(tp, 0,
320                                                      XFS_ICHANGE_LOG_RES(mp), 0,
321                                                      0, 0))) {
322                                 lock_flags = 0;
323                                 goto error_return;
324                         }
325                 }
326         } else {
327                 if (DM_EVENT_ENABLED(ip, DM_EVENT_TRUNCATE) &&
328                     !(flags & ATTR_DMI)) {
329                         int dmflags = AT_DELAY_FLAG(flags) | DM_SEM_FLAG_WR;
330                         code = XFS_SEND_DATA(mp, DM_EVENT_TRUNCATE, vp,
331                                 vap->va_size, 0, dmflags, NULL);
332                         if (code) {
333                                 lock_flags = 0;
334                                 goto error_return;
335                         }
336                 }
337                 if (need_iolock)
338                         lock_flags |= XFS_IOLOCK_EXCL;
339         }
340
341         xfs_ilock(ip, lock_flags);
342
343         /* boolean: are we the file owner? */
344         file_owner = (current_fsuid(credp) == ip->i_d.di_uid);
345
346         /*
347          * Change various properties of a file.
348          * Only the owner or users with CAP_FOWNER
349          * capability may do these things.
350          */
351         if (mask &
352             (XFS_AT_MODE|XFS_AT_XFLAGS|XFS_AT_EXTSIZE|XFS_AT_UID|
353              XFS_AT_GID|XFS_AT_PROJID)) {
354                 /*
355                  * CAP_FOWNER overrides the following restrictions:
356                  *
357                  * The user ID of the calling process must be equal
358                  * to the file owner ID, except in cases where the
359                  * CAP_FSETID capability is applicable.
360                  */
361                 if (!file_owner && !capable(CAP_FOWNER)) {
362                         code = XFS_ERROR(EPERM);
363                         goto error_return;
364                 }
365
366                 /*
367                  * CAP_FSETID overrides the following restrictions:
368                  *
369                  * The effective user ID of the calling process shall match
370                  * the file owner when setting the set-user-ID and
371                  * set-group-ID bits on that file.
372                  *
373                  * The effective group ID or one of the supplementary group
374                  * IDs of the calling process shall match the group owner of
375                  * the file when setting the set-group-ID bit on that file
376                  */
377                 if (mask & XFS_AT_MODE) {
378                         mode_t m = 0;
379
380                         if ((vap->va_mode & S_ISUID) && !file_owner)
381                                 m |= S_ISUID;
382                         if ((vap->va_mode & S_ISGID) &&
383                             !in_group_p((gid_t)ip->i_d.di_gid))
384                                 m |= S_ISGID;
385 #if 0
386                         /* Linux allows this, Irix doesn't. */
387                         if ((vap->va_mode & S_ISVTX) && !VN_ISDIR(vp))
388                                 m |= S_ISVTX;
389 #endif
390                         if (m && !capable(CAP_FSETID))
391                                 vap->va_mode &= ~m;
392                 }
393         }
394
395         /*
396          * Change file ownership.  Must be the owner or privileged.
397          * If the system was configured with the "restricted_chown"
398          * option, the owner is not permitted to give away the file,
399          * and can change the group id only to a group of which he
400          * or she is a member.
401          */
402         if (mask & (XFS_AT_UID|XFS_AT_GID|XFS_AT_PROJID)) {
403                 /*
404                  * These IDs could have changed since we last looked at them.
405                  * But, we're assured that if the ownership did change
406                  * while we didn't have the inode locked, inode's dquot(s)
407                  * would have changed also.
408                  */
409                 iuid = ip->i_d.di_uid;
410                 iprojid = ip->i_d.di_projid;
411                 igid = ip->i_d.di_gid;
412                 gid = (mask & XFS_AT_GID) ? vap->va_gid : igid;
413                 uid = (mask & XFS_AT_UID) ? vap->va_uid : iuid;
414                 projid = (mask & XFS_AT_PROJID) ? (xfs_prid_t)vap->va_projid :
415                          iprojid;
416
417                 /*
418                  * CAP_CHOWN overrides the following restrictions:
419                  *
420                  * If _POSIX_CHOWN_RESTRICTED is defined, this capability
421                  * shall override the restriction that a process cannot
422                  * change the user ID of a file it owns and the restriction
423                  * that the group ID supplied to the chown() function
424                  * shall be equal to either the group ID or one of the
425                  * supplementary group IDs of the calling process.
426                  */
427                 if (restricted_chown &&
428                     (iuid != uid || (igid != gid &&
429                                      !in_group_p((gid_t)gid))) &&
430                     !capable(CAP_CHOWN)) {
431                         code = XFS_ERROR(EPERM);
432                         goto error_return;
433                 }
434                 /*
435                  * Do a quota reservation only if uid/projid/gid is actually
436                  * going to change.
437                  */
438                 if ((XFS_IS_UQUOTA_ON(mp) && iuid != uid) ||
439                     (XFS_IS_PQUOTA_ON(mp) && iprojid != projid) ||
440                     (XFS_IS_GQUOTA_ON(mp) && igid != gid)) {
441                         ASSERT(tp);
442                         code = XFS_QM_DQVOPCHOWNRESV(mp, tp, ip, udqp, gdqp,
443                                                 capable(CAP_FOWNER) ?
444                                                 XFS_QMOPT_FORCE_RES : 0);
445                         if (code)       /* out of quota */
446                                 goto error_return;
447                 }
448         }
449
450         /*
451          * Truncate file.  Must have write permission and not be a directory.
452          */
453         if (mask & XFS_AT_SIZE) {
454                 /* Short circuit the truncate case for zero length files */
455                 if ((vap->va_size == 0) &&
456                    (ip->i_size == 0) && (ip->i_d.di_nextents == 0)) {
457                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
458                         lock_flags &= ~XFS_ILOCK_EXCL;
459                         if (mask & XFS_AT_CTIME)
460                                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
461                         code = 0;
462                         goto error_return;
463                 }
464
465                 if (VN_ISDIR(vp)) {
466                         code = XFS_ERROR(EISDIR);
467                         goto error_return;
468                 } else if (!VN_ISREG(vp)) {
469                         code = XFS_ERROR(EINVAL);
470                         goto error_return;
471                 }
472                 /*
473                  * Make sure that the dquots are attached to the inode.
474                  */
475                 if ((code = XFS_QM_DQATTACH(mp, ip, XFS_QMOPT_ILOCKED)))
476                         goto error_return;
477         }
478
479         /*
480          * Change file access or modified times.
481          */
482         if (mask & (XFS_AT_ATIME|XFS_AT_MTIME)) {
483                 if (!file_owner) {
484                         if ((flags & ATTR_UTIME) &&
485                             !capable(CAP_FOWNER)) {
486                                 code = XFS_ERROR(EPERM);
487                                 goto error_return;
488                         }
489                 }
490         }
491
492         /*
493          * Change extent size or realtime flag.
494          */
495         if (mask & (XFS_AT_EXTSIZE|XFS_AT_XFLAGS)) {
496                 /*
497                  * Can't change extent size if any extents are allocated.
498                  */
499                 if (ip->i_d.di_nextents && (mask & XFS_AT_EXTSIZE) &&
500                     ((ip->i_d.di_extsize << mp->m_sb.sb_blocklog) !=
501                      vap->va_extsize) ) {
502                         code = XFS_ERROR(EINVAL);       /* EFBIG? */
503                         goto error_return;
504                 }
505
506                 /*
507                  * Can't change realtime flag if any extents are allocated.
508                  */
509                 if ((ip->i_d.di_nextents || ip->i_delayed_blks) &&
510                     (mask & XFS_AT_XFLAGS) &&
511                     (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) !=
512                     (vap->va_xflags & XFS_XFLAG_REALTIME)) {
513                         code = XFS_ERROR(EINVAL);       /* EFBIG? */
514                         goto error_return;
515                 }
516                 /*
517                  * Extent size must be a multiple of the appropriate block
518                  * size, if set at all.
519                  */
520                 if ((mask & XFS_AT_EXTSIZE) && vap->va_extsize != 0) {
521                         xfs_extlen_t    size;
522
523                         if ((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ||
524                             ((mask & XFS_AT_XFLAGS) &&
525                             (vap->va_xflags & XFS_XFLAG_REALTIME))) {
526                                 size = mp->m_sb.sb_rextsize <<
527                                        mp->m_sb.sb_blocklog;
528                         } else {
529                                 size = mp->m_sb.sb_blocksize;
530                         }
531                         if (vap->va_extsize % size) {
532                                 code = XFS_ERROR(EINVAL);
533                                 goto error_return;
534                         }
535                 }
536                 /*
537                  * If realtime flag is set then must have realtime data.
538                  */
539                 if ((mask & XFS_AT_XFLAGS) &&
540                     (vap->va_xflags & XFS_XFLAG_REALTIME)) {
541                         if ((mp->m_sb.sb_rblocks == 0) ||
542                             (mp->m_sb.sb_rextsize == 0) ||
543                             (ip->i_d.di_extsize % mp->m_sb.sb_rextsize)) {
544                                 code = XFS_ERROR(EINVAL);
545                                 goto error_return;
546                         }
547                 }
548
549                 /*
550                  * Can't modify an immutable/append-only file unless
551                  * we have appropriate permission.
552                  */
553                 if ((mask & XFS_AT_XFLAGS) &&
554                     (ip->i_d.di_flags &
555                                 (XFS_DIFLAG_IMMUTABLE|XFS_DIFLAG_APPEND) ||
556                      (vap->va_xflags &
557                                 (XFS_XFLAG_IMMUTABLE | XFS_XFLAG_APPEND))) &&
558                     !capable(CAP_LINUX_IMMUTABLE)) {
559                         code = XFS_ERROR(EPERM);
560                         goto error_return;
561                 }
562         }
563
564         /*
565          * Now we can make the changes.  Before we join the inode
566          * to the transaction, if XFS_AT_SIZE is set then take care of
567          * the part of the truncation that must be done without the
568          * inode lock.  This needs to be done before joining the inode
569          * to the transaction, because the inode cannot be unlocked
570          * once it is a part of the transaction.
571          */
572         if (mask & XFS_AT_SIZE) {
573                 code = 0;
574                 if ((vap->va_size > ip->i_size) &&
575                     (flags & ATTR_NOSIZETOK) == 0) {
576                         code = xfs_igrow_start(ip, vap->va_size, credp);
577                 }
578                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
579
580                 /*
581                  * We are going to log the inode size change in this
582                  * transaction so any previous writes that are beyond the on
583                  * disk EOF and the new EOF that have not been written out need
584                  * to be written here. If we do not write the data out, we
585                  * expose ourselves to the null files problem.
586                  *
587                  * Only flush from the on disk size to the smaller of the in
588                  * memory file size or the new size as that's the range we
589                  * really care about here and prevents waiting for other data
590                  * not within the range we care about here.
591                  */
592                 if (!code &&
593                     (ip->i_size != ip->i_d.di_size) &&
594                     (vap->va_size > ip->i_d.di_size)) {
595                         code = xfs_flush_pages(ip,
596                                         ip->i_d.di_size, vap->va_size,
597                                         XFS_B_ASYNC, FI_NONE);
598                 }
599
600                 /* wait for all I/O to complete */
601                 vn_iowait(ip);
602
603                 if (!code)
604                         code = xfs_itruncate_data(ip, vap->va_size);
605                 if (code) {
606                         ASSERT(tp == NULL);
607                         lock_flags &= ~XFS_ILOCK_EXCL;
608                         ASSERT(lock_flags == XFS_IOLOCK_EXCL);
609                         goto error_return;
610                 }
611                 tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_SIZE);
612                 if ((code = xfs_trans_reserve(tp, 0,
613                                              XFS_ITRUNCATE_LOG_RES(mp), 0,
614                                              XFS_TRANS_PERM_LOG_RES,
615                                              XFS_ITRUNCATE_LOG_COUNT))) {
616                         xfs_trans_cancel(tp, 0);
617                         if (need_iolock)
618                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
619                         return code;
620                 }
621                 commit_flags = XFS_TRANS_RELEASE_LOG_RES;
622                 xfs_ilock(ip, XFS_ILOCK_EXCL);
623         }
624
625         if (tp) {
626                 xfs_trans_ijoin(tp, ip, lock_flags);
627                 xfs_trans_ihold(tp, ip);
628         }
629
630         /* determine whether mandatory locking mode changes */
631         mandlock_before = MANDLOCK(vp, ip->i_d.di_mode);
632
633         /*
634          * Truncate file.  Must have write permission and not be a directory.
635          */
636         if (mask & XFS_AT_SIZE) {
637                 if (vap->va_size > ip->i_size) {
638                         xfs_igrow_finish(tp, ip, vap->va_size,
639                             !(flags & ATTR_DMI));
640                 } else if ((vap->va_size <= ip->i_size) ||
641                            ((vap->va_size == 0) && ip->i_d.di_nextents)) {
642                         /*
643                          * signal a sync transaction unless
644                          * we're truncating an already unlinked
645                          * file on a wsync filesystem
646                          */
647                         code = xfs_itruncate_finish(&tp, ip,
648                                             (xfs_fsize_t)vap->va_size,
649                                             XFS_DATA_FORK,
650                                             ((ip->i_d.di_nlink != 0 ||
651                                               !(mp->m_flags & XFS_MOUNT_WSYNC))
652                                              ? 1 : 0));
653                         if (code)
654                                 goto abort_return;
655                         /*
656                          * Truncated "down", so we're removing references
657                          * to old data here - if we now delay flushing for
658                          * a long time, we expose ourselves unduly to the
659                          * notorious NULL files problem.  So, we mark this
660                          * vnode and flush it when the file is closed, and
661                          * do not wait the usual (long) time for writeout.
662                          */
663                         xfs_iflags_set(ip, XFS_ITRUNCATED);
664                 }
665                 /*
666                  * Have to do this even if the file's size doesn't change.
667                  */
668                 timeflags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
669         }
670
671         /*
672          * Change file access modes.
673          */
674         if (mask & XFS_AT_MODE) {
675                 ip->i_d.di_mode &= S_IFMT;
676                 ip->i_d.di_mode |= vap->va_mode & ~S_IFMT;
677
678                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
679                 timeflags |= XFS_ICHGTIME_CHG;
680         }
681
682         /*
683          * Change file ownership.  Must be the owner or privileged.
684          * If the system was configured with the "restricted_chown"
685          * option, the owner is not permitted to give away the file,
686          * and can change the group id only to a group of which he
687          * or she is a member.
688          */
689         if (mask & (XFS_AT_UID|XFS_AT_GID|XFS_AT_PROJID)) {
690                 /*
691                  * CAP_FSETID overrides the following restrictions:
692                  *
693                  * The set-user-ID and set-group-ID bits of a file will be
694                  * cleared upon successful return from chown()
695                  */
696                 if ((ip->i_d.di_mode & (S_ISUID|S_ISGID)) &&
697                     !capable(CAP_FSETID)) {
698                         ip->i_d.di_mode &= ~(S_ISUID|S_ISGID);
699                 }
700
701                 /*
702                  * Change the ownerships and register quota modifications
703                  * in the transaction.
704                  */
705                 if (iuid != uid) {
706                         if (XFS_IS_UQUOTA_ON(mp)) {
707                                 ASSERT(mask & XFS_AT_UID);
708                                 ASSERT(udqp);
709                                 olddquot1 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
710                                                         &ip->i_udquot, udqp);
711                         }
712                         ip->i_d.di_uid = uid;
713                 }
714                 if (igid != gid) {
715                         if (XFS_IS_GQUOTA_ON(mp)) {
716                                 ASSERT(!XFS_IS_PQUOTA_ON(mp));
717                                 ASSERT(mask & XFS_AT_GID);
718                                 ASSERT(gdqp);
719                                 olddquot2 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
720                                                         &ip->i_gdquot, gdqp);
721                         }
722                         ip->i_d.di_gid = gid;
723                 }
724                 if (iprojid != projid) {
725                         if (XFS_IS_PQUOTA_ON(mp)) {
726                                 ASSERT(!XFS_IS_GQUOTA_ON(mp));
727                                 ASSERT(mask & XFS_AT_PROJID);
728                                 ASSERT(gdqp);
729                                 olddquot2 = XFS_QM_DQVOPCHOWN(mp, tp, ip,
730                                                         &ip->i_gdquot, gdqp);
731                         }
732                         ip->i_d.di_projid = projid;
733                         /*
734                          * We may have to rev the inode as well as
735                          * the superblock version number since projids didn't
736                          * exist before DINODE_VERSION_2 and SB_VERSION_NLINK.
737                          */
738                         if (ip->i_d.di_version == XFS_DINODE_VERSION_1)
739                                 xfs_bump_ino_vers2(tp, ip);
740                 }
741
742                 xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
743                 timeflags |= XFS_ICHGTIME_CHG;
744         }
745
746
747         /*
748          * Change file access or modified times.
749          */
750         if (mask & (XFS_AT_ATIME|XFS_AT_MTIME)) {
751                 if (mask & XFS_AT_ATIME) {
752                         ip->i_d.di_atime.t_sec = vap->va_atime.tv_sec;
753                         ip->i_d.di_atime.t_nsec = vap->va_atime.tv_nsec;
754                         ip->i_update_core = 1;
755                         timeflags &= ~XFS_ICHGTIME_ACC;
756                 }
757                 if (mask & XFS_AT_MTIME) {
758                         ip->i_d.di_mtime.t_sec = vap->va_mtime.tv_sec;
759                         ip->i_d.di_mtime.t_nsec = vap->va_mtime.tv_nsec;
760                         timeflags &= ~XFS_ICHGTIME_MOD;
761                         timeflags |= XFS_ICHGTIME_CHG;
762                 }
763                 if (tp && (flags & ATTR_UTIME))
764                         xfs_trans_log_inode (tp, ip, XFS_ILOG_CORE);
765         }
766
767         /*
768          * Change XFS-added attributes.
769          */
770         if (mask & (XFS_AT_EXTSIZE|XFS_AT_XFLAGS)) {
771                 if (mask & XFS_AT_EXTSIZE) {
772                         /*
773                          * Converting bytes to fs blocks.
774                          */
775                         ip->i_d.di_extsize = vap->va_extsize >>
776                                 mp->m_sb.sb_blocklog;
777                 }
778                 if (mask & XFS_AT_XFLAGS) {
779                         uint    di_flags;
780
781                         /* can't set PREALLOC this way, just preserve it */
782                         di_flags = (ip->i_d.di_flags & XFS_DIFLAG_PREALLOC);
783                         if (vap->va_xflags & XFS_XFLAG_IMMUTABLE)
784                                 di_flags |= XFS_DIFLAG_IMMUTABLE;
785                         if (vap->va_xflags & XFS_XFLAG_APPEND)
786                                 di_flags |= XFS_DIFLAG_APPEND;
787                         if (vap->va_xflags & XFS_XFLAG_SYNC)
788                                 di_flags |= XFS_DIFLAG_SYNC;
789                         if (vap->va_xflags & XFS_XFLAG_NOATIME)
790                                 di_flags |= XFS_DIFLAG_NOATIME;
791                         if (vap->va_xflags & XFS_XFLAG_NODUMP)
792                                 di_flags |= XFS_DIFLAG_NODUMP;
793                         if (vap->va_xflags & XFS_XFLAG_PROJINHERIT)
794                                 di_flags |= XFS_DIFLAG_PROJINHERIT;
795                         if (vap->va_xflags & XFS_XFLAG_NODEFRAG)
796                                 di_flags |= XFS_DIFLAG_NODEFRAG;
797                         if (vap->va_xflags & XFS_XFLAG_FILESTREAM)
798                                 di_flags |= XFS_DIFLAG_FILESTREAM;
799                         if ((ip->i_d.di_mode & S_IFMT) == S_IFDIR) {
800                                 if (vap->va_xflags & XFS_XFLAG_RTINHERIT)
801                                         di_flags |= XFS_DIFLAG_RTINHERIT;
802                                 if (vap->va_xflags & XFS_XFLAG_NOSYMLINKS)
803                                         di_flags |= XFS_DIFLAG_NOSYMLINKS;
804                                 if (vap->va_xflags & XFS_XFLAG_EXTSZINHERIT)
805                                         di_flags |= XFS_DIFLAG_EXTSZINHERIT;
806                         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFREG) {
807                                 if (vap->va_xflags & XFS_XFLAG_REALTIME) {
808                                         di_flags |= XFS_DIFLAG_REALTIME;
809                                         ip->i_iocore.io_flags |= XFS_IOCORE_RT;
810                                 } else {
811                                         ip->i_iocore.io_flags &= ~XFS_IOCORE_RT;
812                                 }
813                                 if (vap->va_xflags & XFS_XFLAG_EXTSIZE)
814                                         di_flags |= XFS_DIFLAG_EXTSIZE;
815                         }
816                         ip->i_d.di_flags = di_flags;
817                 }
818                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
819                 timeflags |= XFS_ICHGTIME_CHG;
820         }
821
822         /*
823          * Change file inode change time only if XFS_AT_CTIME set
824          * AND we have been called by a DMI function.
825          */
826
827         if ( (flags & ATTR_DMI) && (mask & XFS_AT_CTIME) ) {
828                 ip->i_d.di_ctime.t_sec = vap->va_ctime.tv_sec;
829                 ip->i_d.di_ctime.t_nsec = vap->va_ctime.tv_nsec;
830                 ip->i_update_core = 1;
831                 timeflags &= ~XFS_ICHGTIME_CHG;
832         }
833
834         /*
835          * Send out timestamp changes that need to be set to the
836          * current time.  Not done when called by a DMI function.
837          */
838         if (timeflags && !(flags & ATTR_DMI))
839                 xfs_ichgtime(ip, timeflags);
840
841         XFS_STATS_INC(xs_ig_attrchg);
842
843         /*
844          * If this is a synchronous mount, make sure that the
845          * transaction goes to disk before returning to the user.
846          * This is slightly sub-optimal in that truncates require
847          * two sync transactions instead of one for wsync filesystems.
848          * One for the truncate and one for the timestamps since we
849          * don't want to change the timestamps unless we're sure the
850          * truncate worked.  Truncates are less than 1% of the laddis
851          * mix so this probably isn't worth the trouble to optimize.
852          */
853         code = 0;
854         if (tp) {
855                 if (mp->m_flags & XFS_MOUNT_WSYNC)
856                         xfs_trans_set_sync(tp);
857
858                 code = xfs_trans_commit(tp, commit_flags);
859         }
860
861         /*
862          * If the (regular) file's mandatory locking mode changed, then
863          * notify the vnode.  We do this under the inode lock to prevent
864          * racing calls to vop_vnode_change.
865          */
866         mandlock_after = MANDLOCK(vp, ip->i_d.di_mode);
867
868         xfs_iunlock(ip, lock_flags);
869
870         /*
871          * Release any dquot(s) the inode had kept before chown.
872          */
873         XFS_QM_DQRELE(mp, olddquot1);
874         XFS_QM_DQRELE(mp, olddquot2);
875         XFS_QM_DQRELE(mp, udqp);
876         XFS_QM_DQRELE(mp, gdqp);
877
878         if (code) {
879                 return code;
880         }
881
882         if (DM_EVENT_ENABLED(ip, DM_EVENT_ATTRIBUTE) &&
883             !(flags & ATTR_DMI)) {
884                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_ATTRIBUTE, vp, DM_RIGHT_NULL,
885                                         NULL, DM_RIGHT_NULL, NULL, NULL,
886                                         0, 0, AT_DELAY_FLAG(flags));
887         }
888         return 0;
889
890  abort_return:
891         commit_flags |= XFS_TRANS_ABORT;
892         /* FALLTHROUGH */
893  error_return:
894         XFS_QM_DQRELE(mp, udqp);
895         XFS_QM_DQRELE(mp, gdqp);
896         if (tp) {
897                 xfs_trans_cancel(tp, commit_flags);
898         }
899         if (lock_flags != 0) {
900                 xfs_iunlock(ip, lock_flags);
901         }
902         return code;
903 }
904
905
906 /*
907  * xfs_access
908  * Null conversion from vnode mode bits to inode mode bits, as in efs.
909  */
910 int
911 xfs_access(
912         xfs_inode_t     *ip,
913         int             mode,
914         cred_t          *credp)
915 {
916         int             error;
917
918         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
919
920         xfs_ilock(ip, XFS_ILOCK_SHARED);
921         error = xfs_iaccess(ip, mode, credp);
922         xfs_iunlock(ip, XFS_ILOCK_SHARED);
923         return error;
924 }
925
926
927 /*
928  * The maximum pathlen is 1024 bytes. Since the minimum file system
929  * blocksize is 512 bytes, we can get a max of 2 extents back from
930  * bmapi.
931  */
932 #define SYMLINK_MAPS 2
933
934 STATIC int
935 xfs_readlink_bmap(
936         xfs_inode_t     *ip,
937         char            *link)
938 {
939         xfs_mount_t     *mp = ip->i_mount;
940         int             pathlen = ip->i_d.di_size;
941         int             nmaps = SYMLINK_MAPS;
942         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
943         xfs_daddr_t     d;
944         int             byte_cnt;
945         int             n;
946         xfs_buf_t       *bp;
947         int             error = 0;
948
949         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
950                         mval, &nmaps, NULL, NULL);
951         if (error)
952                 goto out;
953
954         for (n = 0; n < nmaps; n++) {
955                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
956                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
957
958                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
959                 error = XFS_BUF_GETERROR(bp);
960                 if (error) {
961                         xfs_ioerror_alert("xfs_readlink",
962                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
963                         xfs_buf_relse(bp);
964                         goto out;
965                 }
966                 if (pathlen < byte_cnt)
967                         byte_cnt = pathlen;
968                 pathlen -= byte_cnt;
969
970                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
971                 xfs_buf_relse(bp);
972         }
973
974         link[ip->i_d.di_size] = '\0';
975         error = 0;
976
977  out:
978         return error;
979 }
980
981 int
982 xfs_readlink(
983         xfs_inode_t     *ip,
984         char            *link)
985 {
986         xfs_mount_t     *mp = ip->i_mount;
987         int             pathlen;
988         int             error = 0;
989
990         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
991
992         if (XFS_FORCED_SHUTDOWN(mp))
993                 return XFS_ERROR(EIO);
994
995         xfs_ilock(ip, XFS_ILOCK_SHARED);
996
997         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
998         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
999
1000         pathlen = ip->i_d.di_size;
1001         if (!pathlen)
1002                 goto out;
1003
1004         if (ip->i_df.if_flags & XFS_IFINLINE) {
1005                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
1006                 link[pathlen] = '\0';
1007         } else {
1008                 error = xfs_readlink_bmap(ip, link);
1009         }
1010
1011  out:
1012         xfs_iunlock(ip, XFS_ILOCK_SHARED);
1013         return error;
1014 }
1015
1016 /*
1017  * xfs_fsync
1018  *
1019  * This is called to sync the inode and its data out to disk.
1020  * We need to hold the I/O lock while flushing the data, and
1021  * the inode lock while flushing the inode.  The inode lock CANNOT
1022  * be held while flushing the data, so acquire after we're done
1023  * with that.
1024  */
1025 int
1026 xfs_fsync(
1027         xfs_inode_t     *ip,
1028         int             flag,
1029         xfs_off_t       start,
1030         xfs_off_t       stop)
1031 {
1032         xfs_trans_t     *tp;
1033         int             error;
1034         int             log_flushed = 0, changed = 1;
1035
1036         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
1037
1038         ASSERT(start >= 0 && stop >= -1);
1039
1040         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
1041                 return XFS_ERROR(EIO);
1042
1043         if (flag & FSYNC_DATA)
1044                 filemap_fdatawait(vn_to_inode(XFS_ITOV(ip))->i_mapping);
1045
1046         /*
1047          * We always need to make sure that the required inode state
1048          * is safe on disk.  The vnode might be clean but because
1049          * of committed transactions that haven't hit the disk yet.
1050          * Likewise, there could be unflushed non-transactional
1051          * changes to the inode core that have to go to disk.
1052          *
1053          * The following code depends on one assumption:  that
1054          * any transaction that changes an inode logs the core
1055          * because it has to change some field in the inode core
1056          * (typically nextents or nblocks).  That assumption
1057          * implies that any transactions against an inode will
1058          * catch any non-transactional updates.  If inode-altering
1059          * transactions exist that violate this assumption, the
1060          * code breaks.  Right now, it figures that if the involved
1061          * update_* field is clear and the inode is unpinned, the
1062          * inode is clean.  Either it's been flushed or it's been
1063          * committed and the commit has hit the disk unpinning the inode.
1064          * (Note that xfs_inode_item_format() called at commit clears
1065          * the update_* fields.)
1066          */
1067         xfs_ilock(ip, XFS_ILOCK_SHARED);
1068
1069         /* If we are flushing data then we care about update_size
1070          * being set, otherwise we care about update_core
1071          */
1072         if ((flag & FSYNC_DATA) ?
1073                         (ip->i_update_size == 0) :
1074                         (ip->i_update_core == 0)) {
1075                 /*
1076                  * Timestamps/size haven't changed since last inode
1077                  * flush or inode transaction commit.  That means
1078                  * either nothing got written or a transaction
1079                  * committed which caught the updates.  If the
1080                  * latter happened and the transaction hasn't
1081                  * hit the disk yet, the inode will be still
1082                  * be pinned.  If it is, force the log.
1083                  */
1084
1085                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
1086
1087                 if (xfs_ipincount(ip)) {
1088                         _xfs_log_force(ip->i_mount, (xfs_lsn_t)0,
1089                                       XFS_LOG_FORCE |
1090                                       ((flag & FSYNC_WAIT)
1091                                        ? XFS_LOG_SYNC : 0),
1092                                       &log_flushed);
1093                 } else {
1094                         /*
1095                          * If the inode is not pinned and nothing
1096                          * has changed we don't need to flush the
1097                          * cache.
1098                          */
1099                         changed = 0;
1100                 }
1101                 error = 0;
1102         } else  {
1103                 /*
1104                  * Kick off a transaction to log the inode
1105                  * core to get the updates.  Make it
1106                  * sync if FSYNC_WAIT is passed in (which
1107                  * is done by everybody but specfs).  The
1108                  * sync transaction will also force the log.
1109                  */
1110                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
1111                 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_FSYNC_TS);
1112                 if ((error = xfs_trans_reserve(tp, 0,
1113                                 XFS_FSYNC_TS_LOG_RES(ip->i_mount),
1114                                 0, 0, 0)))  {
1115                         xfs_trans_cancel(tp, 0);
1116                         return error;
1117                 }
1118                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1119
1120                 /*
1121                  * Note - it's possible that we might have pushed
1122                  * ourselves out of the way during trans_reserve
1123                  * which would flush the inode.  But there's no
1124                  * guarantee that the inode buffer has actually
1125                  * gone out yet (it's delwri).  Plus the buffer
1126                  * could be pinned anyway if it's part of an
1127                  * inode in another recent transaction.  So we
1128                  * play it safe and fire off the transaction anyway.
1129                  */
1130                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1131                 xfs_trans_ihold(tp, ip);
1132                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1133                 if (flag & FSYNC_WAIT)
1134                         xfs_trans_set_sync(tp);
1135                 error = _xfs_trans_commit(tp, 0, &log_flushed);
1136
1137                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1138         }
1139
1140         if ((ip->i_mount->m_flags & XFS_MOUNT_BARRIER) && changed) {
1141                 /*
1142                  * If the log write didn't issue an ordered tag we need
1143                  * to flush the disk cache for the data device now.
1144                  */
1145                 if (!log_flushed)
1146                         xfs_blkdev_issue_flush(ip->i_mount->m_ddev_targp);
1147
1148                 /*
1149                  * If this inode is on the RT dev we need to flush that
1150                  * cache as well.
1151                  */
1152                 if (ip->i_d.di_flags & XFS_DIFLAG_REALTIME)
1153                         xfs_blkdev_issue_flush(ip->i_mount->m_rtdev_targp);
1154         }
1155
1156         return error;
1157 }
1158
1159 /*
1160  * This is called by xfs_inactive to free any blocks beyond eof
1161  * when the link count isn't zero and by xfs_dm_punch_hole() when
1162  * punching a hole to EOF.
1163  */
1164 int
1165 xfs_free_eofblocks(
1166         xfs_mount_t     *mp,
1167         xfs_inode_t     *ip,
1168         int             flags)
1169 {
1170         xfs_trans_t     *tp;
1171         int             error;
1172         xfs_fileoff_t   end_fsb;
1173         xfs_fileoff_t   last_fsb;
1174         xfs_filblks_t   map_len;
1175         int             nimaps;
1176         xfs_bmbt_irec_t imap;
1177         int             use_iolock = (flags & XFS_FREE_EOF_LOCK);
1178
1179         /*
1180          * Figure out if there are any blocks beyond the end
1181          * of the file.  If not, then there is nothing to do.
1182          */
1183         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
1184         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
1185         map_len = last_fsb - end_fsb;
1186         if (map_len <= 0)
1187                 return 0;
1188
1189         nimaps = 1;
1190         xfs_ilock(ip, XFS_ILOCK_SHARED);
1191         error = XFS_BMAPI(mp, NULL, &ip->i_iocore, end_fsb, map_len, 0,
1192                           NULL, 0, &imap, &nimaps, NULL, NULL);
1193         xfs_iunlock(ip, XFS_ILOCK_SHARED);
1194
1195         if (!error && (nimaps != 0) &&
1196             (imap.br_startblock != HOLESTARTBLOCK ||
1197              ip->i_delayed_blks)) {
1198                 /*
1199                  * Attach the dquots to the inode up front.
1200                  */
1201                 if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
1202                         return error;
1203
1204                 /*
1205                  * There are blocks after the end of file.
1206                  * Free them up now by truncating the file to
1207                  * its current size.
1208                  */
1209                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1210
1211                 /*
1212                  * Do the xfs_itruncate_start() call before
1213                  * reserving any log space because
1214                  * itruncate_start will call into the buffer
1215                  * cache and we can't
1216                  * do that within a transaction.
1217                  */
1218                 if (use_iolock)
1219                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
1220                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
1221                                     ip->i_size);
1222                 if (error) {
1223                         xfs_trans_cancel(tp, 0);
1224                         if (use_iolock)
1225                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1226                         return error;
1227                 }
1228
1229                 error = xfs_trans_reserve(tp, 0,
1230                                           XFS_ITRUNCATE_LOG_RES(mp),
1231                                           0, XFS_TRANS_PERM_LOG_RES,
1232                                           XFS_ITRUNCATE_LOG_COUNT);
1233                 if (error) {
1234                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1235                         xfs_trans_cancel(tp, 0);
1236                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1237                         return error;
1238                 }
1239
1240                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1241                 xfs_trans_ijoin(tp, ip,
1242                                 XFS_IOLOCK_EXCL |
1243                                 XFS_ILOCK_EXCL);
1244                 xfs_trans_ihold(tp, ip);
1245
1246                 error = xfs_itruncate_finish(&tp, ip,
1247                                              ip->i_size,
1248                                              XFS_DATA_FORK,
1249                                              0);
1250                 /*
1251                  * If we get an error at this point we
1252                  * simply don't bother truncating the file.
1253                  */
1254                 if (error) {
1255                         xfs_trans_cancel(tp,
1256                                          (XFS_TRANS_RELEASE_LOG_RES |
1257                                           XFS_TRANS_ABORT));
1258                 } else {
1259                         error = xfs_trans_commit(tp,
1260                                                 XFS_TRANS_RELEASE_LOG_RES);
1261                 }
1262                 xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
1263                                             : XFS_ILOCK_EXCL));
1264         }
1265         return error;
1266 }
1267
1268 /*
1269  * Free a symlink that has blocks associated with it.
1270  */
1271 STATIC int
1272 xfs_inactive_symlink_rmt(
1273         xfs_inode_t     *ip,
1274         xfs_trans_t     **tpp)
1275 {
1276         xfs_buf_t       *bp;
1277         int             committed;
1278         int             done;
1279         int             error;
1280         xfs_fsblock_t   first_block;
1281         xfs_bmap_free_t free_list;
1282         int             i;
1283         xfs_mount_t     *mp;
1284         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
1285         int             nmaps;
1286         xfs_trans_t     *ntp;
1287         int             size;
1288         xfs_trans_t     *tp;
1289
1290         tp = *tpp;
1291         mp = ip->i_mount;
1292         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
1293         /*
1294          * We're freeing a symlink that has some
1295          * blocks allocated to it.  Free the
1296          * blocks here.  We know that we've got
1297          * either 1 or 2 extents and that we can
1298          * free them all in one bunmapi call.
1299          */
1300         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
1301         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
1302                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
1303                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1304                 xfs_trans_cancel(tp, 0);
1305                 *tpp = NULL;
1306                 return error;
1307         }
1308         /*
1309          * Lock the inode, fix the size, and join it to the transaction.
1310          * Hold it so in the normal path, we still have it locked for
1311          * the second transaction.  In the error paths we need it
1312          * held so the cancel won't rele it, see below.
1313          */
1314         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1315         size = (int)ip->i_d.di_size;
1316         ip->i_d.di_size = 0;
1317         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1318         xfs_trans_ihold(tp, ip);
1319         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1320         /*
1321          * Find the block(s) so we can inval and unmap them.
1322          */
1323         done = 0;
1324         XFS_BMAP_INIT(&free_list, &first_block);
1325         nmaps = ARRAY_SIZE(mval);
1326         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
1327                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
1328                         &free_list, NULL)))
1329                 goto error0;
1330         /*
1331          * Invalidate the block(s).
1332          */
1333         for (i = 0; i < nmaps; i++) {
1334                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
1335                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
1336                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
1337                 xfs_trans_binval(tp, bp);
1338         }
1339         /*
1340          * Unmap the dead block(s) to the free_list.
1341          */
1342         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
1343                         &first_block, &free_list, NULL, &done)))
1344                 goto error1;
1345         ASSERT(done);
1346         /*
1347          * Commit the first transaction.  This logs the EFI and the inode.
1348          */
1349         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
1350                 goto error1;
1351         /*
1352          * The transaction must have been committed, since there were
1353          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
1354          * The new tp has the extent freeing and EFDs.
1355          */
1356         ASSERT(committed);
1357         /*
1358          * The first xact was committed, so add the inode to the new one.
1359          * Mark it dirty so it will be logged and moved forward in the log as
1360          * part of every commit.
1361          */
1362         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1363         xfs_trans_ihold(tp, ip);
1364         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1365         /*
1366          * Get a new, empty transaction to return to our caller.
1367          */
1368         ntp = xfs_trans_dup(tp);
1369         /*
1370          * Commit the transaction containing extent freeing and EFDs.
1371          * If we get an error on the commit here or on the reserve below,
1372          * we need to unlock the inode since the new transaction doesn't
1373          * have the inode attached.
1374          */
1375         error = xfs_trans_commit(tp, 0);
1376         tp = ntp;
1377         if (error) {
1378                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1379                 goto error0;
1380         }
1381         /*
1382          * Remove the memory for extent descriptions (just bookkeeping).
1383          */
1384         if (ip->i_df.if_bytes)
1385                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
1386         ASSERT(ip->i_df.if_bytes == 0);
1387         /*
1388          * Put an itruncate log reservation in the new transaction
1389          * for our caller.
1390          */
1391         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
1392                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
1393                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1394                 goto error0;
1395         }
1396         /*
1397          * Return with the inode locked but not joined to the transaction.
1398          */
1399         *tpp = tp;
1400         return 0;
1401
1402  error1:
1403         xfs_bmap_cancel(&free_list);
1404  error0:
1405         /*
1406          * Have to come here with the inode locked and either
1407          * (held and in the transaction) or (not in the transaction).
1408          * If the inode isn't held then cancel would iput it, but
1409          * that's wrong since this is inactive and the vnode ref
1410          * count is 0 already.
1411          * Cancel won't do anything to the inode if held, but it still
1412          * needs to be locked until the cancel is done, if it was
1413          * joined to the transaction.
1414          */
1415         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1416         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1417         *tpp = NULL;
1418         return error;
1419
1420 }
1421
1422 STATIC int
1423 xfs_inactive_symlink_local(
1424         xfs_inode_t     *ip,
1425         xfs_trans_t     **tpp)
1426 {
1427         int             error;
1428
1429         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
1430         /*
1431          * We're freeing a symlink which fit into
1432          * the inode.  Just free the memory used
1433          * to hold the old symlink.
1434          */
1435         error = xfs_trans_reserve(*tpp, 0,
1436                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
1437                                   0, XFS_TRANS_PERM_LOG_RES,
1438                                   XFS_ITRUNCATE_LOG_COUNT);
1439
1440         if (error) {
1441                 xfs_trans_cancel(*tpp, 0);
1442                 *tpp = NULL;
1443                 return error;
1444         }
1445         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1446
1447         /*
1448          * Zero length symlinks _can_ exist.
1449          */
1450         if (ip->i_df.if_bytes > 0) {
1451                 xfs_idata_realloc(ip,
1452                                   -(ip->i_df.if_bytes),
1453                                   XFS_DATA_FORK);
1454                 ASSERT(ip->i_df.if_bytes == 0);
1455         }
1456         return 0;
1457 }
1458
1459 STATIC int
1460 xfs_inactive_attrs(
1461         xfs_inode_t     *ip,
1462         xfs_trans_t     **tpp)
1463 {
1464         xfs_trans_t     *tp;
1465         int             error;
1466         xfs_mount_t     *mp;
1467
1468         ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
1469         tp = *tpp;
1470         mp = ip->i_mount;
1471         ASSERT(ip->i_d.di_forkoff != 0);
1472         xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1473         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1474
1475         error = xfs_attr_inactive(ip);
1476         if (error) {
1477                 *tpp = NULL;
1478                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1479                 return error; /* goto out */
1480         }
1481
1482         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1483         error = xfs_trans_reserve(tp, 0,
1484                                   XFS_IFREE_LOG_RES(mp),
1485                                   0, XFS_TRANS_PERM_LOG_RES,
1486                                   XFS_INACTIVE_LOG_COUNT);
1487         if (error) {
1488                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1489                 xfs_trans_cancel(tp, 0);
1490                 *tpp = NULL;
1491                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1492                 return error;
1493         }
1494
1495         xfs_ilock(ip, XFS_ILOCK_EXCL);
1496         xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1497         xfs_trans_ihold(tp, ip);
1498         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1499
1500         ASSERT(ip->i_d.di_anextents == 0);
1501
1502         *tpp = tp;
1503         return 0;
1504 }
1505
1506 int
1507 xfs_release(
1508         xfs_inode_t     *ip)
1509 {
1510         bhv_vnode_t     *vp = XFS_ITOV(ip);
1511         xfs_mount_t     *mp = ip->i_mount;
1512         int             error;
1513
1514         if (!VN_ISREG(vp) || (ip->i_d.di_mode == 0))
1515                 return 0;
1516
1517         /* If this is a read-only mount, don't do this (would generate I/O) */
1518         if (mp->m_flags & XFS_MOUNT_RDONLY)
1519                 return 0;
1520
1521         if (!XFS_FORCED_SHUTDOWN(mp)) {
1522                 int truncated;
1523
1524                 /*
1525                  * If we are using filestreams, and we have an unlinked
1526                  * file that we are processing the last close on, then nothing
1527                  * will be able to reopen and write to this file. Purge this
1528                  * inode from the filestreams cache so that it doesn't delay
1529                  * teardown of the inode.
1530                  */
1531                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
1532                         xfs_filestream_deassociate(ip);
1533
1534                 /*
1535                  * If we previously truncated this file and removed old data
1536                  * in the process, we want to initiate "early" writeout on
1537                  * the last close.  This is an attempt to combat the notorious
1538                  * NULL files problem which is particularly noticable from a
1539                  * truncate down, buffered (re-)write (delalloc), followed by
1540                  * a crash.  What we are effectively doing here is
1541                  * significantly reducing the time window where we'd otherwise
1542                  * be exposed to that problem.
1543                  */
1544                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1545                 if (truncated && VN_DIRTY(vp) && ip->i_delayed_blks > 0)
1546                         xfs_flush_pages(ip, 0, -1, XFS_B_ASYNC, FI_NONE);
1547         }
1548
1549 #ifdef HAVE_REFCACHE
1550         /* If we are in the NFS reference cache then don't do this now */
1551         if (ip->i_refcache)
1552                 return 0;
1553 #endif
1554
1555         if (ip->i_d.di_nlink != 0) {
1556                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1557                      ((ip->i_size > 0) || (VN_CACHED(vp) > 0 ||
1558                        ip->i_delayed_blks > 0)) &&
1559                      (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
1560                     (!(ip->i_d.di_flags &
1561                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
1562                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1563                         if (error)
1564                                 return error;
1565                         /* Update linux inode block count after free above */
1566                         vn_to_inode(vp)->i_blocks = XFS_FSB_TO_BB(mp,
1567                                 ip->i_d.di_nblocks + ip->i_delayed_blks);
1568                 }
1569         }
1570
1571         return 0;
1572 }
1573
1574 /*
1575  * xfs_inactive
1576  *
1577  * This is called when the vnode reference count for the vnode
1578  * goes to zero.  If the file has been unlinked, then it must
1579  * now be truncated.  Also, we clear all of the read-ahead state
1580  * kept for the inode here since the file is now closed.
1581  */
1582 int
1583 xfs_inactive(
1584         xfs_inode_t     *ip)
1585 {
1586         bhv_vnode_t     *vp = XFS_ITOV(ip);
1587         xfs_bmap_free_t free_list;
1588         xfs_fsblock_t   first_block;
1589         int             committed;
1590         xfs_trans_t     *tp;
1591         xfs_mount_t     *mp;
1592         int             error;
1593         int             truncate;
1594
1595         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
1596
1597         /*
1598          * If the inode is already free, then there can be nothing
1599          * to clean up here.
1600          */
1601         if (ip->i_d.di_mode == 0 || VN_BAD(vp)) {
1602                 ASSERT(ip->i_df.if_real_bytes == 0);
1603                 ASSERT(ip->i_df.if_broot_bytes == 0);
1604                 return VN_INACTIVE_CACHE;
1605         }
1606
1607         /*
1608          * Only do a truncate if it's a regular file with
1609          * some actual space in it.  It's OK to look at the
1610          * inode's fields without the lock because we're the
1611          * only one with a reference to the inode.
1612          */
1613         truncate = ((ip->i_d.di_nlink == 0) &&
1614             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
1615              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
1616             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
1617
1618         mp = ip->i_mount;
1619
1620         if (ip->i_d.di_nlink == 0 && DM_EVENT_ENABLED(ip, DM_EVENT_DESTROY)) {
1621                 (void) XFS_SEND_DESTROY(mp, vp, DM_RIGHT_NULL);
1622         }
1623
1624         error = 0;
1625
1626         /* If this is a read-only mount, don't do this (would generate I/O) */
1627         if (mp->m_flags & XFS_MOUNT_RDONLY)
1628                 goto out;
1629
1630         if (ip->i_d.di_nlink != 0) {
1631                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
1632                      ((ip->i_size > 0) || (VN_CACHED(vp) > 0 ||
1633                        ip->i_delayed_blks > 0)) &&
1634                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
1635                      (!(ip->i_d.di_flags &
1636                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
1637                       (ip->i_delayed_blks != 0)))) {
1638                         error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
1639                         if (error)
1640                                 return VN_INACTIVE_CACHE;
1641                         /* Update linux inode block count after free above */
1642                         vn_to_inode(vp)->i_blocks = XFS_FSB_TO_BB(mp,
1643                                 ip->i_d.di_nblocks + ip->i_delayed_blks);
1644                 }
1645                 goto out;
1646         }
1647
1648         ASSERT(ip->i_d.di_nlink == 0);
1649
1650         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
1651                 return VN_INACTIVE_CACHE;
1652
1653         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1654         if (truncate) {
1655                 /*
1656                  * Do the xfs_itruncate_start() call before
1657                  * reserving any log space because itruncate_start
1658                  * will call into the buffer cache and we can't
1659                  * do that within a transaction.
1660                  */
1661                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1662
1663                 error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE, 0);
1664                 if (error) {
1665                         xfs_trans_cancel(tp, 0);
1666                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1667                         return VN_INACTIVE_CACHE;
1668                 }
1669
1670                 error = xfs_trans_reserve(tp, 0,
1671                                           XFS_ITRUNCATE_LOG_RES(mp),
1672                                           0, XFS_TRANS_PERM_LOG_RES,
1673                                           XFS_ITRUNCATE_LOG_COUNT);
1674                 if (error) {
1675                         /* Don't call itruncate_cleanup */
1676                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1677                         xfs_trans_cancel(tp, 0);
1678                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1679                         return VN_INACTIVE_CACHE;
1680                 }
1681
1682                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1683                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1684                 xfs_trans_ihold(tp, ip);
1685
1686                 /*
1687                  * normally, we have to run xfs_itruncate_finish sync.
1688                  * But if filesystem is wsync and we're in the inactive
1689                  * path, then we know that nlink == 0, and that the
1690                  * xaction that made nlink == 0 is permanently committed
1691                  * since xfs_remove runs as a synchronous transaction.
1692                  */
1693                 error = xfs_itruncate_finish(&tp, ip, 0, XFS_DATA_FORK,
1694                                 (!(mp->m_flags & XFS_MOUNT_WSYNC) ? 1 : 0));
1695
1696                 if (error) {
1697                         xfs_trans_cancel(tp,
1698                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1699                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1700                         return VN_INACTIVE_CACHE;
1701                 }
1702         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
1703
1704                 /*
1705                  * If we get an error while cleaning up a
1706                  * symlink we bail out.
1707                  */
1708                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
1709                         xfs_inactive_symlink_rmt(ip, &tp) :
1710                         xfs_inactive_symlink_local(ip, &tp);
1711
1712                 if (error) {
1713                         ASSERT(tp == NULL);
1714                         return VN_INACTIVE_CACHE;
1715                 }
1716
1717                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1718                 xfs_trans_ihold(tp, ip);
1719         } else {
1720                 error = xfs_trans_reserve(tp, 0,
1721                                           XFS_IFREE_LOG_RES(mp),
1722                                           0, XFS_TRANS_PERM_LOG_RES,
1723                                           XFS_INACTIVE_LOG_COUNT);
1724                 if (error) {
1725                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
1726                         xfs_trans_cancel(tp, 0);
1727                         return VN_INACTIVE_CACHE;
1728                 }
1729
1730                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
1731                 xfs_trans_ijoin(tp, ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1732                 xfs_trans_ihold(tp, ip);
1733         }
1734
1735         /*
1736          * If there are attributes associated with the file
1737          * then blow them away now.  The code calls a routine
1738          * that recursively deconstructs the attribute fork.
1739          * We need to just commit the current transaction
1740          * because we can't use it for xfs_attr_inactive().
1741          */
1742         if (ip->i_d.di_anextents > 0) {
1743                 error = xfs_inactive_attrs(ip, &tp);
1744                 /*
1745                  * If we got an error, the transaction is already
1746                  * cancelled, and the inode is unlocked. Just get out.
1747                  */
1748                  if (error)
1749                          return VN_INACTIVE_CACHE;
1750         } else if (ip->i_afp) {
1751                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1752         }
1753
1754         /*
1755          * Free the inode.
1756          */
1757         XFS_BMAP_INIT(&free_list, &first_block);
1758         error = xfs_ifree(tp, ip, &free_list);
1759         if (error) {
1760                 /*
1761                  * If we fail to free the inode, shut down.  The cancel
1762                  * might do that, we need to make sure.  Otherwise the
1763                  * inode might be lost for a long time or forever.
1764                  */
1765                 if (!XFS_FORCED_SHUTDOWN(mp)) {
1766                         cmn_err(CE_NOTE,
1767                 "xfs_inactive:  xfs_ifree() returned an error = %d on %s",
1768                                 error, mp->m_fsname);
1769                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1770                 }
1771                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1772         } else {
1773                 /*
1774                  * Credit the quota account(s). The inode is gone.
1775                  */
1776                 XFS_TRANS_MOD_DQUOT_BYINO(mp, tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1777
1778                 /*
1779                  * Just ignore errors at this point.  There is
1780                  * nothing we can do except to try to keep going.
1781                  */
1782                 (void) xfs_bmap_finish(&tp,  &free_list, &committed);
1783                 (void) xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1784         }
1785         /*
1786          * Release the dquots held by inode, if any.
1787          */
1788         XFS_QM_DQDETACH(mp, ip);
1789
1790         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1791
1792  out:
1793         return VN_INACTIVE_CACHE;
1794 }
1795
1796
1797 int
1798 xfs_lookup(
1799         xfs_inode_t             *dp,
1800         bhv_vname_t             *dentry,
1801         bhv_vnode_t             **vpp)
1802 {
1803         xfs_inode_t             *ip;
1804         xfs_ino_t               e_inum;
1805         int                     error;
1806         uint                    lock_mode;
1807
1808         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
1809
1810         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
1811                 return XFS_ERROR(EIO);
1812
1813         lock_mode = xfs_ilock_map_shared(dp);
1814         error = xfs_dir_lookup_int(dp, lock_mode, dentry, &e_inum, &ip);
1815         if (!error) {
1816                 *vpp = XFS_ITOV(ip);
1817                 ITRACE(ip);
1818         }
1819         xfs_iunlock_map_shared(dp, lock_mode);
1820         return error;
1821 }
1822
1823 int
1824 xfs_create(
1825         xfs_inode_t             *dp,
1826         bhv_vname_t             *dentry,
1827         mode_t                  mode,
1828         xfs_dev_t               rdev,
1829         bhv_vnode_t             **vpp,
1830         cred_t                  *credp)
1831 {
1832         char                    *name = VNAME(dentry);
1833         xfs_mount_t             *mp = dp->i_mount;
1834         bhv_vnode_t             *dir_vp = XFS_ITOV(dp);
1835         xfs_inode_t             *ip;
1836         bhv_vnode_t             *vp = NULL;
1837         xfs_trans_t             *tp;
1838         int                     error;
1839         xfs_bmap_free_t         free_list;
1840         xfs_fsblock_t           first_block;
1841         boolean_t               unlock_dp_on_error = B_FALSE;
1842         int                     dm_event_sent = 0;
1843         uint                    cancel_flags;
1844         int                     committed;
1845         xfs_prid_t              prid;
1846         struct xfs_dquot        *udqp, *gdqp;
1847         uint                    resblks;
1848         int                     namelen;
1849
1850         ASSERT(!*vpp);
1851         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
1852
1853         namelen = VNAMELEN(dentry);
1854
1855         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
1856                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
1857                                 dir_vp, DM_RIGHT_NULL, NULL,
1858                                 DM_RIGHT_NULL, name, NULL,
1859                                 mode, 0, 0);
1860
1861                 if (error)
1862                         return error;
1863                 dm_event_sent = 1;
1864         }
1865
1866         if (XFS_FORCED_SHUTDOWN(mp))
1867                 return XFS_ERROR(EIO);
1868
1869         /* Return through std_return after this point. */
1870
1871         udqp = gdqp = NULL;
1872         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1873                 prid = dp->i_d.di_projid;
1874         else
1875                 prid = (xfs_prid_t)dfltprid;
1876
1877         /*
1878          * Make sure that we have allocated dquot(s) on disk.
1879          */
1880         error = XFS_QM_DQVOPALLOC(mp, dp,
1881                         current_fsuid(credp), current_fsgid(credp), prid,
1882                         XFS_QMOPT_QUOTALL|XFS_QMOPT_INHERIT, &udqp, &gdqp);
1883         if (error)
1884                 goto std_return;
1885
1886         ip = NULL;
1887
1888         tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1889         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1890         resblks = XFS_CREATE_SPACE_RES(mp, namelen);
1891         /*
1892          * Initially assume that the file does not exist and
1893          * reserve the resources for that case.  If that is not
1894          * the case we'll drop the one we have and get a more
1895          * appropriate transaction later.
1896          */
1897         error = xfs_trans_reserve(tp, resblks, XFS_CREATE_LOG_RES(mp), 0,
1898                         XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1899         if (error == ENOSPC) {
1900                 resblks = 0;
1901                 error = xfs_trans_reserve(tp, 0, XFS_CREATE_LOG_RES(mp), 0,
1902                                 XFS_TRANS_PERM_LOG_RES, XFS_CREATE_LOG_COUNT);
1903         }
1904         if (error) {
1905                 cancel_flags = 0;
1906                 goto error_return;
1907         }
1908
1909         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1910         unlock_dp_on_error = B_TRUE;
1911
1912         XFS_BMAP_INIT(&free_list, &first_block);
1913
1914         ASSERT(ip == NULL);
1915
1916         /*
1917          * Reserve disk quota and the inode.
1918          */
1919         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
1920         if (error)
1921                 goto error_return;
1922
1923         if (resblks == 0 && (error = xfs_dir_canenter(tp, dp, name, namelen)))
1924                 goto error_return;
1925         error = xfs_dir_ialloc(&tp, dp, mode, 1,
1926                         rdev, credp, prid, resblks > 0,
1927                         &ip, &committed);
1928         if (error) {
1929                 if (error == ENOSPC)
1930                         goto error_return;
1931                 goto abort_return;
1932         }
1933         ITRACE(ip);
1934
1935         /*
1936          * At this point, we've gotten a newly allocated inode.
1937          * It is locked (and joined to the transaction).
1938          */
1939
1940         ASSERT(ismrlocked (&ip->i_lock, MR_UPDATE));
1941
1942         /*
1943          * Now we join the directory inode to the transaction.  We do not do it
1944          * earlier because xfs_dir_ialloc might commit the previous transaction
1945          * (and release all the locks).  An error from here on will result in
1946          * the transaction cancel unlocking dp so don't do it explicitly in the
1947          * error path.
1948          */
1949         VN_HOLD(dir_vp);
1950         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1951         unlock_dp_on_error = B_FALSE;
1952
1953         error = xfs_dir_createname(tp, dp, name, namelen, ip->i_ino,
1954                                         &first_block, &free_list, resblks ?
1955                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1956         if (error) {
1957                 ASSERT(error != ENOSPC);
1958                 goto abort_return;
1959         }
1960         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1961         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1962
1963         /*
1964          * If this is a synchronous mount, make sure that the
1965          * create transaction goes to disk before returning to
1966          * the user.
1967          */
1968         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1969                 xfs_trans_set_sync(tp);
1970         }
1971
1972         dp->i_gen++;
1973
1974         /*
1975          * Attach the dquot(s) to the inodes and modify them incore.
1976          * These ids of the inode couldn't have changed since the new
1977          * inode has been locked ever since it was created.
1978          */
1979         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
1980
1981         /*
1982          * xfs_trans_commit normally decrements the vnode ref count
1983          * when it unlocks the inode. Since we want to return the
1984          * vnode to the caller, we bump the vnode ref count now.
1985          */
1986         IHOLD(ip);
1987         vp = XFS_ITOV(ip);
1988
1989         error = xfs_bmap_finish(&tp, &free_list, &committed);
1990         if (error) {
1991                 xfs_bmap_cancel(&free_list);
1992                 goto abort_rele;
1993         }
1994
1995         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1996         if (error) {
1997                 IRELE(ip);
1998                 tp = NULL;
1999                 goto error_return;
2000         }
2001
2002         XFS_QM_DQRELE(mp, udqp);
2003         XFS_QM_DQRELE(mp, gdqp);
2004
2005         *vpp = vp;
2006
2007         /* Fallthrough to std_return with error = 0  */
2008
2009 std_return:
2010         if ((*vpp || (error != 0 && dm_event_sent != 0)) &&
2011             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
2012                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
2013                         dir_vp, DM_RIGHT_NULL,
2014                         *vpp ? vp:NULL,
2015                         DM_RIGHT_NULL, name, NULL,
2016                         mode, error, 0);
2017         }
2018         return error;
2019
2020  abort_return:
2021         cancel_flags |= XFS_TRANS_ABORT;
2022         /* FALLTHROUGH */
2023
2024  error_return:
2025         if (tp != NULL)
2026                 xfs_trans_cancel(tp, cancel_flags);
2027
2028         XFS_QM_DQRELE(mp, udqp);
2029         XFS_QM_DQRELE(mp, gdqp);
2030
2031         if (unlock_dp_on_error)
2032                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2033
2034         goto std_return;
2035
2036  abort_rele:
2037         /*
2038          * Wait until after the current transaction is aborted to
2039          * release the inode.  This prevents recursive transactions
2040          * and deadlocks from xfs_inactive.
2041          */
2042         cancel_flags |= XFS_TRANS_ABORT;
2043         xfs_trans_cancel(tp, cancel_flags);
2044         IRELE(ip);
2045
2046         XFS_QM_DQRELE(mp, udqp);
2047         XFS_QM_DQRELE(mp, gdqp);
2048
2049         goto std_return;
2050 }
2051
2052 #ifdef DEBUG
2053 /*
2054  * Some counters to see if (and how often) we are hitting some deadlock
2055  * prevention code paths.
2056  */
2057
2058 int xfs_rm_locks;
2059 int xfs_rm_lock_delays;
2060 int xfs_rm_attempts;
2061 #endif
2062
2063 /*
2064  * The following routine will lock the inodes associated with the
2065  * directory and the named entry in the directory. The locks are
2066  * acquired in increasing inode number.
2067  *
2068  * If the entry is "..", then only the directory is locked. The
2069  * vnode ref count will still include that from the .. entry in
2070  * this case.
2071  *
2072  * There is a deadlock we need to worry about. If the locked directory is
2073  * in the AIL, it might be blocking up the log. The next inode we lock
2074  * could be already locked by another thread waiting for log space (e.g
2075  * a permanent log reservation with a long running transaction (see
2076  * xfs_itruncate_finish)). To solve this, we must check if the directory
2077  * is in the ail and use lock_nowait. If we can't lock, we need to
2078  * drop the inode lock on the directory and try again. xfs_iunlock will
2079  * potentially push the tail if we were holding up the log.
2080  */
2081 STATIC int
2082 xfs_lock_dir_and_entry(
2083         xfs_inode_t     *dp,
2084         xfs_inode_t     *ip)    /* inode of entry 'name' */
2085 {
2086         int             attempts;
2087         xfs_ino_t       e_inum;
2088         xfs_inode_t     *ips[2];
2089         xfs_log_item_t  *lp;
2090
2091 #ifdef DEBUG
2092         xfs_rm_locks++;
2093 #endif
2094         attempts = 0;
2095
2096 again:
2097         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2098
2099         e_inum = ip->i_ino;
2100
2101         ITRACE(ip);
2102
2103         /*
2104          * We want to lock in increasing inum. Since we've already
2105          * acquired the lock on the directory, we may need to release
2106          * if if the inum of the entry turns out to be less.
2107          */
2108         if (e_inum > dp->i_ino) {
2109                 /*
2110                  * We are already in the right order, so just
2111                  * lock on the inode of the entry.
2112                  * We need to use nowait if dp is in the AIL.
2113                  */
2114
2115                 lp = (xfs_log_item_t *)dp->i_itemp;
2116                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
2117                         if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2118                                 attempts++;
2119 #ifdef DEBUG
2120                                 xfs_rm_attempts++;
2121 #endif
2122
2123                                 /*
2124                                  * Unlock dp and try again.
2125                                  * xfs_iunlock will try to push the tail
2126                                  * if the inode is in the AIL.
2127                                  */
2128
2129                                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2130
2131                                 if ((attempts % 5) == 0) {
2132                                         delay(1); /* Don't just spin the CPU */
2133 #ifdef DEBUG
2134                                         xfs_rm_lock_delays++;
2135 #endif
2136                                 }
2137                                 goto again;
2138                         }
2139                 } else {
2140                         xfs_ilock(ip, XFS_ILOCK_EXCL);
2141                 }
2142         } else if (e_inum < dp->i_ino) {
2143                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2144
2145                 ips[0] = ip;
2146                 ips[1] = dp;
2147                 xfs_lock_inodes(ips, 2, 0, XFS_ILOCK_EXCL);
2148         }
2149         /* else  e_inum == dp->i_ino */
2150         /*     This can happen if we're asked to lock /x/..
2151          *     the entry is "..", which is also the parent directory.
2152          */
2153
2154         return 0;
2155 }
2156
2157 #ifdef DEBUG
2158 int xfs_locked_n;
2159 int xfs_small_retries;
2160 int xfs_middle_retries;
2161 int xfs_lots_retries;
2162 int xfs_lock_delays;
2163 #endif
2164
2165 /*
2166  * Bump the subclass so xfs_lock_inodes() acquires each lock with
2167  * a different value
2168  */
2169 static inline int
2170 xfs_lock_inumorder(int lock_mode, int subclass)
2171 {
2172         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
2173                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
2174         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
2175                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
2176
2177         return lock_mode;
2178 }
2179
2180 /*
2181  * The following routine will lock n inodes in exclusive mode.
2182  * We assume the caller calls us with the inodes in i_ino order.
2183  *
2184  * We need to detect deadlock where an inode that we lock
2185  * is in the AIL and we start waiting for another inode that is locked
2186  * by a thread in a long running transaction (such as truncate). This can
2187  * result in deadlock since the long running trans might need to wait
2188  * for the inode we just locked in order to push the tail and free space
2189  * in the log.
2190  */
2191 void
2192 xfs_lock_inodes(
2193         xfs_inode_t     **ips,
2194         int             inodes,
2195         int             first_locked,
2196         uint            lock_mode)
2197 {
2198         int             attempts = 0, i, j, try_lock;
2199         xfs_log_item_t  *lp;
2200
2201         ASSERT(ips && (inodes >= 2)); /* we need at least two */
2202
2203         if (first_locked) {
2204                 try_lock = 1;
2205                 i = 1;
2206         } else {
2207                 try_lock = 0;
2208                 i = 0;
2209         }
2210
2211 again:
2212         for (; i < inodes; i++) {
2213                 ASSERT(ips[i]);
2214
2215                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
2216                         continue;
2217
2218                 /*
2219                  * If try_lock is not set yet, make sure all locked inodes
2220                  * are not in the AIL.
2221                  * If any are, set try_lock to be used later.
2222                  */
2223
2224                 if (!try_lock) {
2225                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
2226                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
2227                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
2228                                         try_lock++;
2229                                 }
2230                         }
2231                 }
2232
2233                 /*
2234                  * If any of the previous locks we have locked is in the AIL,
2235                  * we must TRY to get the second and subsequent locks. If
2236                  * we can't get any, we must release all we have
2237                  * and try again.
2238                  */
2239
2240                 if (try_lock) {
2241                         /* try_lock must be 0 if i is 0. */
2242                         /*
2243                          * try_lock means we have an inode locked
2244                          * that is in the AIL.
2245                          */
2246                         ASSERT(i != 0);
2247                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
2248                                 attempts++;
2249
2250                                 /*
2251                                  * Unlock all previous guys and try again.
2252                                  * xfs_iunlock will try to push the tail
2253                                  * if the inode is in the AIL.
2254                                  */
2255
2256                                 for(j = i - 1; j >= 0; j--) {
2257
2258                                         /*
2259                                          * Check to see if we've already
2260                                          * unlocked this one.
2261                                          * Not the first one going back,
2262                                          * and the inode ptr is the same.
2263                                          */
2264                                         if ((j != (i - 1)) && ips[j] ==
2265                                                                 ips[j+1])
2266                                                 continue;
2267
2268                                         xfs_iunlock(ips[j], lock_mode);
2269                                 }
2270
2271                                 if ((attempts % 5) == 0) {
2272                                         delay(1); /* Don't just spin the CPU */
2273 #ifdef DEBUG
2274                                         xfs_lock_delays++;
2275 #endif
2276                                 }
2277                                 i = 0;
2278                                 try_lock = 0;
2279                                 goto again;
2280                         }
2281                 } else {
2282                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
2283                 }
2284         }
2285
2286 #ifdef DEBUG
2287         if (attempts) {
2288                 if (attempts < 5) xfs_small_retries++;
2289                 else if (attempts < 100) xfs_middle_retries++;
2290                 else xfs_lots_retries++;
2291         } else {
2292                 xfs_locked_n++;
2293         }
2294 #endif
2295 }
2296
2297 #ifdef  DEBUG
2298 #define REMOVE_DEBUG_TRACE(x)   {remove_which_error_return = (x);}
2299 int remove_which_error_return = 0;
2300 #else /* ! DEBUG */
2301 #define REMOVE_DEBUG_TRACE(x)
2302 #endif  /* ! DEBUG */
2303
2304 int
2305 xfs_remove(
2306         xfs_inode_t             *dp,
2307         bhv_vname_t             *dentry)
2308 {
2309         bhv_vnode_t             *dir_vp = XFS_ITOV(dp);
2310         char                    *name = VNAME(dentry);
2311         xfs_mount_t             *mp = dp->i_mount;
2312         xfs_inode_t             *ip;
2313         xfs_trans_t             *tp = NULL;
2314         int                     error = 0;
2315         xfs_bmap_free_t         free_list;
2316         xfs_fsblock_t           first_block;
2317         int                     cancel_flags;
2318         int                     committed;
2319         int                     dm_di_mode = 0;
2320         int                     link_zero;
2321         uint                    resblks;
2322         int                     namelen;
2323
2324         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
2325
2326         if (XFS_FORCED_SHUTDOWN(mp))
2327                 return XFS_ERROR(EIO);
2328
2329         namelen = VNAMELEN(dentry);
2330
2331         if (!xfs_get_dir_entry(dentry, &ip)) {
2332                 dm_di_mode = ip->i_d.di_mode;
2333                 IRELE(ip);
2334         }
2335
2336         if (DM_EVENT_ENABLED(dp, DM_EVENT_REMOVE)) {
2337                 error = XFS_SEND_NAMESP(mp, DM_EVENT_REMOVE, dir_vp,
2338                                         DM_RIGHT_NULL, NULL, DM_RIGHT_NULL,
2339                                         name, NULL, dm_di_mode, 0, 0);
2340                 if (error)
2341                         return error;
2342         }
2343
2344         /* From this point on, return through std_return */
2345         ip = NULL;
2346
2347         /*
2348          * We need to get a reference to ip before we get our log
2349          * reservation. The reason for this is that we cannot call
2350          * xfs_iget for an inode for which we do not have a reference
2351          * once we've acquired a log reservation. This is because the
2352          * inode we are trying to get might be in xfs_inactive going
2353          * for a log reservation. Since we'll have to wait for the
2354          * inactive code to complete before returning from xfs_iget,
2355          * we need to make sure that we don't have log space reserved
2356          * when we call xfs_iget.  Instead we get an unlocked reference
2357          * to the inode before getting our log reservation.
2358          */
2359         error = xfs_get_dir_entry(dentry, &ip);
2360         if (error) {
2361                 REMOVE_DEBUG_TRACE(__LINE__);
2362                 goto std_return;
2363         }
2364
2365         dm_di_mode = ip->i_d.di_mode;
2366
2367         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
2368
2369         ITRACE(ip);
2370
2371         error = XFS_QM_DQATTACH(mp, dp, 0);
2372         if (!error && dp != ip)
2373                 error = XFS_QM_DQATTACH(mp, ip, 0);
2374         if (error) {
2375                 REMOVE_DEBUG_TRACE(__LINE__);
2376                 IRELE(ip);
2377                 goto std_return;
2378         }
2379
2380         tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
2381         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2382         /*
2383          * We try to get the real space reservation first,
2384          * allowing for directory btree deletion(s) implying
2385          * possible bmap insert(s).  If we can't get the space
2386          * reservation then we use 0 instead, and avoid the bmap
2387          * btree insert(s) in the directory code by, if the bmap
2388          * insert tries to happen, instead trimming the LAST
2389          * block from the directory.
2390          */
2391         resblks = XFS_REMOVE_SPACE_RES(mp);
2392         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
2393                         XFS_TRANS_PERM_LOG_RES, XFS_REMOVE_LOG_COUNT);
2394         if (error == ENOSPC) {
2395                 resblks = 0;
2396                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
2397                                 XFS_TRANS_PERM_LOG_RES, XFS_REMOVE_LOG_COUNT);
2398         }
2399         if (error) {
2400                 ASSERT(error != ENOSPC);
2401                 REMOVE_DEBUG_TRACE(__LINE__);
2402                 xfs_trans_cancel(tp, 0);
2403                 IRELE(ip);
2404                 return error;
2405         }
2406
2407         error = xfs_lock_dir_and_entry(dp, ip);
2408         if (error) {
2409                 REMOVE_DEBUG_TRACE(__LINE__);
2410                 xfs_trans_cancel(tp, cancel_flags);
2411                 IRELE(ip);
2412                 goto std_return;
2413         }
2414
2415         /*
2416          * At this point, we've gotten both the directory and the entry
2417          * inodes locked.
2418          */
2419         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2420         if (dp != ip) {
2421                 /*
2422                  * Increment vnode ref count only in this case since
2423                  * there's an extra vnode reference in the case where
2424                  * dp == ip.
2425                  */
2426                 IHOLD(dp);
2427                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2428         }
2429
2430         /*
2431          * Entry must exist since we did a lookup in xfs_lock_dir_and_entry.
2432          */
2433         XFS_BMAP_INIT(&free_list, &first_block);
2434         error = xfs_dir_removename(tp, dp, name, namelen, ip->i_ino,
2435                                         &first_block, &free_list, 0);
2436         if (error) {
2437                 ASSERT(error != ENOENT);
2438                 REMOVE_DEBUG_TRACE(__LINE__);
2439                 goto error1;
2440         }
2441         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2442
2443         dp->i_gen++;
2444         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2445
2446         error = xfs_droplink(tp, ip);
2447         if (error) {
2448                 REMOVE_DEBUG_TRACE(__LINE__);
2449                 goto error1;
2450         }
2451
2452         /* Determine if this is the last link while
2453          * we are in the transaction.
2454          */
2455         link_zero = (ip)->i_d.di_nlink==0;
2456
2457         /*
2458          * Take an extra ref on the inode so that it doesn't
2459          * go to xfs_inactive() from within the commit.
2460          */
2461         IHOLD(ip);
2462
2463         /*
2464          * If this is a synchronous mount, make sure that the
2465          * remove transaction goes to disk before returning to
2466          * the user.
2467          */
2468         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2469                 xfs_trans_set_sync(tp);
2470         }
2471
2472         error = xfs_bmap_finish(&tp, &free_list, &committed);
2473         if (error) {
2474                 REMOVE_DEBUG_TRACE(__LINE__);
2475                 goto error_rele;
2476         }
2477
2478         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2479         if (error) {
2480                 IRELE(ip);
2481                 goto std_return;
2482         }
2483
2484         /*
2485          * Before we drop our extra reference to the inode, purge it
2486          * from the refcache if it is there.  By waiting until afterwards
2487          * to do the IRELE, we ensure that we won't go inactive in the
2488          * xfs_refcache_purge_ip routine (although that would be OK).
2489          */
2490         xfs_refcache_purge_ip(ip);
2491
2492         /*
2493          * If we are using filestreams, kill the stream association.
2494          * If the file is still open it may get a new one but that
2495          * will get killed on last close in xfs_close() so we don't
2496          * have to worry about that.
2497          */
2498         if (link_zero && xfs_inode_is_filestream(ip))
2499                 xfs_filestream_deassociate(ip);
2500
2501         vn_trace_exit(ip, __FUNCTION__, (inst_t *)__return_address);
2502
2503         IRELE(ip);
2504
2505 /*      Fall through to std_return with error = 0 */
2506  std_return:
2507         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTREMOVE)) {
2508                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTREMOVE,
2509                                 dir_vp, DM_RIGHT_NULL,
2510                                 NULL, DM_RIGHT_NULL,
2511                                 name, NULL, dm_di_mode, error, 0);
2512         }
2513         return error;
2514
2515  error1:
2516         xfs_bmap_cancel(&free_list);
2517         cancel_flags |= XFS_TRANS_ABORT;
2518         xfs_trans_cancel(tp, cancel_flags);
2519         goto std_return;
2520
2521  error_rele:
2522         /*
2523          * In this case make sure to not release the inode until after
2524          * the current transaction is aborted.  Releasing it beforehand
2525          * can cause us to go to xfs_inactive and start a recursive
2526          * transaction which can easily deadlock with the current one.
2527          */
2528         xfs_bmap_cancel(&free_list);
2529         cancel_flags |= XFS_TRANS_ABORT;
2530         xfs_trans_cancel(tp, cancel_flags);
2531
2532         /*
2533          * Before we drop our extra reference to the inode, purge it
2534          * from the refcache if it is there.  By waiting until afterwards
2535          * to do the IRELE, we ensure that we won't go inactive in the
2536          * xfs_refcache_purge_ip routine (although that would be OK).
2537          */
2538         xfs_refcache_purge_ip(ip);
2539
2540         IRELE(ip);
2541
2542         goto std_return;
2543 }
2544
2545 int
2546 xfs_link(
2547         xfs_inode_t             *tdp,
2548         bhv_vnode_t             *src_vp,
2549         bhv_vname_t             *dentry)
2550 {
2551         bhv_vnode_t             *target_dir_vp = XFS_ITOV(tdp);
2552         xfs_mount_t             *mp = tdp->i_mount;
2553         xfs_inode_t             *sip = xfs_vtoi(src_vp);
2554         xfs_trans_t             *tp;
2555         xfs_inode_t             *ips[2];
2556         int                     error;
2557         xfs_bmap_free_t         free_list;
2558         xfs_fsblock_t           first_block;
2559         int                     cancel_flags;
2560         int                     committed;
2561         int                     resblks;
2562         char                    *target_name = VNAME(dentry);
2563         int                     target_namelen;
2564
2565         vn_trace_entry(tdp, __FUNCTION__, (inst_t *)__return_address);
2566         vn_trace_entry(xfs_vtoi(src_vp), __FUNCTION__, (inst_t *)__return_address);
2567
2568         target_namelen = VNAMELEN(dentry);
2569         ASSERT(!VN_ISDIR(src_vp));
2570
2571         if (XFS_FORCED_SHUTDOWN(mp))
2572                 return XFS_ERROR(EIO);
2573
2574         if (DM_EVENT_ENABLED(tdp, DM_EVENT_LINK)) {
2575                 error = XFS_SEND_NAMESP(mp, DM_EVENT_LINK,
2576                                         target_dir_vp, DM_RIGHT_NULL,
2577                                         src_vp, DM_RIGHT_NULL,
2578                                         target_name, NULL, 0, 0, 0);
2579                 if (error)
2580                         return error;
2581         }
2582
2583         /* Return through std_return after this point. */
2584
2585         error = XFS_QM_DQATTACH(mp, sip, 0);
2586         if (!error && sip != tdp)
2587                 error = XFS_QM_DQATTACH(mp, tdp, 0);
2588         if (error)
2589                 goto std_return;
2590
2591         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
2592         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2593         resblks = XFS_LINK_SPACE_RES(mp, target_namelen);
2594         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
2595                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2596         if (error == ENOSPC) {
2597                 resblks = 0;
2598                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
2599                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
2600         }
2601         if (error) {
2602                 cancel_flags = 0;
2603                 goto error_return;
2604         }
2605
2606         if (sip->i_ino < tdp->i_ino) {
2607                 ips[0] = sip;
2608                 ips[1] = tdp;
2609         } else {
2610                 ips[0] = tdp;
2611                 ips[1] = sip;
2612         }
2613
2614         xfs_lock_inodes(ips, 2, 0, XFS_ILOCK_EXCL);
2615
2616         /*
2617          * Increment vnode ref counts since xfs_trans_commit &
2618          * xfs_trans_cancel will both unlock the inodes and
2619          * decrement the associated ref counts.
2620          */
2621         VN_HOLD(src_vp);
2622         VN_HOLD(target_dir_vp);
2623         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
2624         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
2625
2626         /*
2627          * If the source has too many links, we can't make any more to it.
2628          */
2629         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
2630                 error = XFS_ERROR(EMLINK);
2631                 goto error_return;
2632         }
2633
2634         /*
2635          * If we are using project inheritance, we only allow hard link
2636          * creation in our tree when the project IDs are the same; else
2637          * the tree quota mechanism could be circumvented.
2638          */
2639         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2640                      (tdp->i_d.di_projid != sip->i_d.di_projid))) {
2641                 error = XFS_ERROR(EXDEV);
2642                 goto error_return;
2643         }
2644
2645         if (resblks == 0 &&
2646             (error = xfs_dir_canenter(tp, tdp, target_name, target_namelen)))
2647                 goto error_return;
2648
2649         XFS_BMAP_INIT(&free_list, &first_block);
2650
2651         error = xfs_dir_createname(tp, tdp, target_name, target_namelen,
2652                                    sip->i_ino, &first_block, &free_list,
2653                                    resblks);
2654         if (error)
2655                 goto abort_return;
2656         xfs_ichgtime(tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2657         tdp->i_gen++;
2658         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
2659
2660         error = xfs_bumplink(tp, sip);
2661         if (error)
2662                 goto abort_return;
2663
2664         /*
2665          * If this is a synchronous mount, make sure that the
2666          * link transaction goes to disk before returning to
2667          * the user.
2668          */
2669         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2670                 xfs_trans_set_sync(tp);
2671         }
2672
2673         error = xfs_bmap_finish (&tp, &free_list, &committed);
2674         if (error) {
2675                 xfs_bmap_cancel(&free_list);
2676                 goto abort_return;
2677         }
2678
2679         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2680         if (error)
2681                 goto std_return;
2682
2683         /* Fall through to std_return with error = 0. */
2684 std_return:
2685         if (DM_EVENT_ENABLED(sip, DM_EVENT_POSTLINK)) {
2686                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTLINK,
2687                                 target_dir_vp, DM_RIGHT_NULL,
2688                                 src_vp, DM_RIGHT_NULL,
2689                                 target_name, NULL, 0, error, 0);
2690         }
2691         return error;
2692
2693  abort_return:
2694         cancel_flags |= XFS_TRANS_ABORT;
2695         /* FALLTHROUGH */
2696
2697  error_return:
2698         xfs_trans_cancel(tp, cancel_flags);
2699         goto std_return;
2700 }
2701
2702
2703 int
2704 xfs_mkdir(
2705         xfs_inode_t             *dp,
2706         bhv_vname_t             *dentry,
2707         mode_t                  mode,
2708         bhv_vnode_t             **vpp,
2709         cred_t                  *credp)
2710 {
2711         bhv_vnode_t             *dir_vp = XFS_ITOV(dp);
2712         char                    *dir_name = VNAME(dentry);
2713         int                     dir_namelen = VNAMELEN(dentry);
2714         xfs_mount_t             *mp = dp->i_mount;
2715         xfs_inode_t             *cdp;   /* inode of created dir */
2716         bhv_vnode_t             *cvp;   /* vnode of created dir */
2717         xfs_trans_t             *tp;
2718         int                     cancel_flags;
2719         int                     error;
2720         int                     committed;
2721         xfs_bmap_free_t         free_list;
2722         xfs_fsblock_t           first_block;
2723         boolean_t               unlock_dp_on_error = B_FALSE;
2724         boolean_t               created = B_FALSE;
2725         int                     dm_event_sent = 0;
2726         xfs_prid_t              prid;
2727         struct xfs_dquot        *udqp, *gdqp;
2728         uint                    resblks;
2729
2730         if (XFS_FORCED_SHUTDOWN(mp))
2731                 return XFS_ERROR(EIO);
2732
2733         tp = NULL;
2734
2735         if (DM_EVENT_ENABLED(dp, DM_EVENT_CREATE)) {
2736                 error = XFS_SEND_NAMESP(mp, DM_EVENT_CREATE,
2737                                         dir_vp, DM_RIGHT_NULL, NULL,
2738                                         DM_RIGHT_NULL, dir_name, NULL,
2739                                         mode, 0, 0);
2740                 if (error)
2741                         return error;
2742                 dm_event_sent = 1;
2743         }
2744
2745         /* Return through std_return after this point. */
2746
2747         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
2748
2749         mp = dp->i_mount;
2750         udqp = gdqp = NULL;
2751         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
2752                 prid = dp->i_d.di_projid;
2753         else
2754                 prid = (xfs_prid_t)dfltprid;
2755
2756         /*
2757          * Make sure that we have allocated dquot(s) on disk.
2758          */
2759         error = XFS_QM_DQVOPALLOC(mp, dp,
2760                         current_fsuid(credp), current_fsgid(credp), prid,
2761                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
2762         if (error)
2763                 goto std_return;
2764
2765         tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
2766         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2767         resblks = XFS_MKDIR_SPACE_RES(mp, dir_namelen);
2768         error = xfs_trans_reserve(tp, resblks, XFS_MKDIR_LOG_RES(mp), 0,
2769                                   XFS_TRANS_PERM_LOG_RES, XFS_MKDIR_LOG_COUNT);
2770         if (error == ENOSPC) {
2771                 resblks = 0;
2772                 error = xfs_trans_reserve(tp, 0, XFS_MKDIR_LOG_RES(mp), 0,
2773                                           XFS_TRANS_PERM_LOG_RES,
2774                                           XFS_MKDIR_LOG_COUNT);
2775         }
2776         if (error) {
2777                 cancel_flags = 0;
2778                 goto error_return;
2779         }
2780
2781         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
2782         unlock_dp_on_error = B_TRUE;
2783
2784         /*
2785          * Check for directory link count overflow.
2786          */
2787         if (dp->i_d.di_nlink >= XFS_MAXLINK) {
2788                 error = XFS_ERROR(EMLINK);
2789                 goto error_return;
2790         }
2791
2792         /*
2793          * Reserve disk quota and the inode.
2794          */
2795         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
2796         if (error)
2797                 goto error_return;
2798
2799         if (resblks == 0 &&
2800             (error = xfs_dir_canenter(tp, dp, dir_name, dir_namelen)))
2801                 goto error_return;
2802         /*
2803          * create the directory inode.
2804          */
2805         error = xfs_dir_ialloc(&tp, dp, mode, 2,
2806                         0, credp, prid, resblks > 0,
2807                 &cdp, NULL);
2808         if (error) {
2809                 if (error == ENOSPC)
2810                         goto error_return;
2811                 goto abort_return;
2812         }
2813         ITRACE(cdp);
2814
2815         /*
2816          * Now we add the directory inode to the transaction.
2817          * We waited until now since xfs_dir_ialloc might start
2818          * a new transaction.  Had we joined the transaction
2819          * earlier, the locks might have gotten released. An error
2820          * from here on will result in the transaction cancel
2821          * unlocking dp so don't do it explicitly in the error path.
2822          */
2823         VN_HOLD(dir_vp);
2824         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2825         unlock_dp_on_error = B_FALSE;
2826
2827         XFS_BMAP_INIT(&free_list, &first_block);
2828
2829         error = xfs_dir_createname(tp, dp, dir_name, dir_namelen, cdp->i_ino,
2830                                    &first_block, &free_list, resblks ?
2831                                    resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
2832         if (error) {
2833                 ASSERT(error != ENOSPC);
2834                 goto error1;
2835         }
2836         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2837
2838         /*
2839          * Bump the in memory version number of the parent directory
2840          * so that other processes accessing it will recognize that
2841          * the directory has changed.
2842          */
2843         dp->i_gen++;
2844
2845         error = xfs_dir_init(tp, cdp, dp);
2846         if (error)
2847                 goto error2;
2848
2849         cdp->i_gen = 1;
2850         error = xfs_bumplink(tp, dp);
2851         if (error)
2852                 goto error2;
2853
2854         cvp = XFS_ITOV(cdp);
2855
2856         created = B_TRUE;
2857
2858         *vpp = cvp;
2859         IHOLD(cdp);
2860
2861         /*
2862          * Attach the dquots to the new inode and modify the icount incore.
2863          */
2864         XFS_QM_DQVOPCREATE(mp, tp, cdp, udqp, gdqp);
2865
2866         /*
2867          * If this is a synchronous mount, make sure that the
2868          * mkdir transaction goes to disk before returning to
2869          * the user.
2870          */
2871         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2872                 xfs_trans_set_sync(tp);
2873         }
2874
2875         error = xfs_bmap_finish(&tp, &free_list, &committed);
2876         if (error) {
2877                 IRELE(cdp);
2878                 goto error2;
2879         }
2880
2881         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2882         XFS_QM_DQRELE(mp, udqp);
2883         XFS_QM_DQRELE(mp, gdqp);
2884         if (error) {
2885                 IRELE(cdp);
2886         }
2887
2888         /* Fall through to std_return with error = 0 or errno from
2889          * xfs_trans_commit. */
2890
2891 std_return:
2892         if ((created || (error != 0 && dm_event_sent != 0)) &&
2893             DM_EVENT_ENABLED(dp, DM_EVENT_POSTCREATE)) {
2894                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTCREATE,
2895                                         dir_vp, DM_RIGHT_NULL,
2896                                         created ? XFS_ITOV(cdp):NULL,
2897                                         DM_RIGHT_NULL,
2898                                         dir_name, NULL,
2899                                         mode, error, 0);
2900         }
2901         return error;
2902
2903  error2:
2904  error1:
2905         xfs_bmap_cancel(&free_list);
2906  abort_return:
2907         cancel_flags |= XFS_TRANS_ABORT;
2908  error_return:
2909         xfs_trans_cancel(tp, cancel_flags);
2910         XFS_QM_DQRELE(mp, udqp);
2911         XFS_QM_DQRELE(mp, gdqp);
2912
2913         if (unlock_dp_on_error)
2914                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
2915
2916         goto std_return;
2917 }
2918
2919 int
2920 xfs_rmdir(
2921         xfs_inode_t             *dp,
2922         bhv_vname_t             *dentry)
2923 {
2924         bhv_vnode_t             *dir_vp = XFS_ITOV(dp);
2925         char                    *name = VNAME(dentry);
2926         int                     namelen = VNAMELEN(dentry);
2927         xfs_mount_t             *mp = dp->i_mount;
2928         xfs_inode_t             *cdp;   /* child directory */
2929         xfs_trans_t             *tp;
2930         int                     error;
2931         xfs_bmap_free_t         free_list;
2932         xfs_fsblock_t           first_block;
2933         int                     cancel_flags;
2934         int                     committed;
2935         int                     dm_di_mode = S_IFDIR;
2936         int                     last_cdp_link;
2937         uint                    resblks;
2938
2939         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
2940
2941         if (XFS_FORCED_SHUTDOWN(mp))
2942                 return XFS_ERROR(EIO);
2943
2944         if (!xfs_get_dir_entry(dentry, &cdp)) {
2945                 dm_di_mode = cdp->i_d.di_mode;
2946                 IRELE(cdp);
2947         }
2948
2949         if (DM_EVENT_ENABLED(dp, DM_EVENT_REMOVE)) {
2950                 error = XFS_SEND_NAMESP(mp, DM_EVENT_REMOVE,
2951                                         dir_vp, DM_RIGHT_NULL,
2952                                         NULL, DM_RIGHT_NULL,
2953                                         name, NULL, dm_di_mode, 0, 0);
2954                 if (error)
2955                         return XFS_ERROR(error);
2956         }
2957
2958         /* Return through std_return after this point. */
2959
2960         cdp = NULL;
2961
2962         /*
2963          * We need to get a reference to cdp before we get our log
2964          * reservation.  The reason for this is that we cannot call
2965          * xfs_iget for an inode for which we do not have a reference
2966          * once we've acquired a log reservation.  This is because the
2967          * inode we are trying to get might be in xfs_inactive going
2968          * for a log reservation.  Since we'll have to wait for the
2969          * inactive code to complete before returning from xfs_iget,
2970          * we need to make sure that we don't have log space reserved
2971          * when we call xfs_iget.  Instead we get an unlocked reference
2972          * to the inode before getting our log reservation.
2973          */
2974         error = xfs_get_dir_entry(dentry, &cdp);
2975         if (error) {
2976                 REMOVE_DEBUG_TRACE(__LINE__);
2977                 goto std_return;
2978         }
2979         mp = dp->i_mount;
2980         dm_di_mode = cdp->i_d.di_mode;
2981
2982         /*
2983          * Get the dquots for the inodes.
2984          */
2985         error = XFS_QM_DQATTACH(mp, dp, 0);
2986         if (!error && dp != cdp)
2987                 error = XFS_QM_DQATTACH(mp, cdp, 0);
2988         if (error) {
2989                 IRELE(cdp);
2990                 REMOVE_DEBUG_TRACE(__LINE__);
2991                 goto std_return;
2992         }
2993
2994         tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
2995         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2996         /*
2997          * We try to get the real space reservation first,
2998          * allowing for directory btree deletion(s) implying
2999          * possible bmap insert(s).  If we can't get the space
3000          * reservation then we use 0 instead, and avoid the bmap
3001          * btree insert(s) in the directory code by, if the bmap
3002          * insert tries to happen, instead trimming the LAST
3003          * block from the directory.
3004          */
3005         resblks = XFS_REMOVE_SPACE_RES(mp);
3006         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
3007                         XFS_TRANS_PERM_LOG_RES, XFS_DEFAULT_LOG_COUNT);
3008         if (error == ENOSPC) {
3009                 resblks = 0;
3010                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
3011                                 XFS_TRANS_PERM_LOG_RES, XFS_DEFAULT_LOG_COUNT);
3012         }
3013         if (error) {
3014                 ASSERT(error != ENOSPC);
3015                 cancel_flags = 0;
3016                 IRELE(cdp);
3017                 goto error_return;
3018         }
3019         XFS_BMAP_INIT(&free_list, &first_block);
3020
3021         /*
3022          * Now lock the child directory inode and the parent directory
3023          * inode in the proper order.  This will take care of validating
3024          * that the directory entry for the child directory inode has
3025          * not changed while we were obtaining a log reservation.
3026          */
3027         error = xfs_lock_dir_and_entry(dp, cdp);
3028         if (error) {
3029                 xfs_trans_cancel(tp, cancel_flags);
3030                 IRELE(cdp);
3031                 goto std_return;
3032         }
3033
3034         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
3035         if (dp != cdp) {
3036                 /*
3037                  * Only increment the parent directory vnode count if
3038                  * we didn't bump it in looking up cdp.  The only time
3039                  * we don't bump it is when we're looking up ".".
3040                  */
3041                 VN_HOLD(dir_vp);
3042         }
3043
3044         ITRACE(cdp);
3045         xfs_trans_ijoin(tp, cdp, XFS_ILOCK_EXCL);
3046
3047         ASSERT(cdp->i_d.di_nlink >= 2);
3048         if (cdp->i_d.di_nlink != 2) {
3049                 error = XFS_ERROR(ENOTEMPTY);
3050                 goto error_return;
3051         }
3052         if (!xfs_dir_isempty(cdp)) {
3053                 error = XFS_ERROR(ENOTEMPTY);
3054                 goto error_return;
3055         }
3056
3057         error = xfs_dir_removename(tp, dp, name, namelen, cdp->i_ino,
3058                                         &first_block, &free_list, resblks);
3059         if (error)
3060                 goto error1;
3061
3062         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3063
3064         /*
3065          * Bump the in memory generation count on the parent
3066          * directory so that other can know that it has changed.
3067          */
3068         dp->i_gen++;
3069
3070         /*
3071          * Drop the link from cdp's "..".
3072          */
3073         error = xfs_droplink(tp, dp);
3074         if (error) {
3075                 goto error1;
3076         }
3077
3078         /*
3079          * Drop the link from dp to cdp.
3080          */
3081         error = xfs_droplink(tp, cdp);
3082         if (error) {
3083                 goto error1;
3084         }
3085
3086         /*
3087          * Drop the "." link from cdp to self.
3088          */
3089         error = xfs_droplink(tp, cdp);
3090         if (error) {
3091                 goto error1;
3092         }
3093
3094         /* Determine these before committing transaction */
3095         last_cdp_link = (cdp)->i_d.di_nlink==0;
3096
3097         /*
3098          * Take an extra ref on the child vnode so that it
3099          * does not go to xfs_inactive() from within the commit.
3100          */
3101         IHOLD(cdp);
3102
3103         /*
3104          * If this is a synchronous mount, make sure that the
3105          * rmdir transaction goes to disk before returning to
3106          * the user.
3107          */
3108         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
3109                 xfs_trans_set_sync(tp);
3110         }
3111
3112         error = xfs_bmap_finish (&tp, &free_list, &committed);
3113         if (error) {
3114                 xfs_bmap_cancel(&free_list);
3115                 xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES |
3116                                  XFS_TRANS_ABORT));
3117                 IRELE(cdp);
3118                 goto std_return;
3119         }
3120
3121         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3122         if (error) {
3123                 IRELE(cdp);
3124                 goto std_return;
3125         }
3126
3127
3128         IRELE(cdp);
3129
3130         /* Fall through to std_return with error = 0 or the errno
3131          * from xfs_trans_commit. */
3132  std_return:
3133         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTREMOVE)) {
3134                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTREMOVE,
3135                                         dir_vp, DM_RIGHT_NULL,
3136                                         NULL, DM_RIGHT_NULL,
3137                                         name, NULL, dm_di_mode,
3138                                         error, 0);
3139         }
3140         return error;
3141
3142  error1:
3143         xfs_bmap_cancel(&free_list);
3144         cancel_flags |= XFS_TRANS_ABORT;
3145         /* FALLTHROUGH */
3146
3147  error_return:
3148         xfs_trans_cancel(tp, cancel_flags);
3149         goto std_return;
3150 }
3151
3152 int
3153 xfs_symlink(
3154         xfs_inode_t             *dp,
3155         bhv_vname_t             *dentry,
3156         char                    *target_path,
3157         mode_t                  mode,
3158         bhv_vnode_t             **vpp,
3159         cred_t                  *credp)
3160 {
3161         bhv_vnode_t             *dir_vp = XFS_ITOV(dp);
3162         xfs_mount_t             *mp = dp->i_mount;
3163         xfs_trans_t             *tp;
3164         xfs_inode_t             *ip;
3165         int                     error;
3166         int                     pathlen;
3167         xfs_bmap_free_t         free_list;
3168         xfs_fsblock_t           first_block;
3169         boolean_t               unlock_dp_on_error = B_FALSE;
3170         uint                    cancel_flags;
3171         int                     committed;
3172         xfs_fileoff_t           first_fsb;
3173         xfs_filblks_t           fs_blocks;
3174         int                     nmaps;
3175         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
3176         xfs_daddr_t             d;
3177         char                    *cur_chunk;
3178         int                     byte_cnt;
3179         int                     n;
3180         xfs_buf_t               *bp;
3181         xfs_prid_t              prid;
3182         struct xfs_dquot        *udqp, *gdqp;
3183         uint                    resblks;
3184         char                    *link_name = VNAME(dentry);
3185         int                     link_namelen;
3186
3187         *vpp = NULL;
3188         error = 0;
3189         ip = NULL;
3190         tp = NULL;
3191
3192         vn_trace_entry(dp, __FUNCTION__, (inst_t *)__return_address);
3193
3194
3195         if (XFS_FORCED_SHUTDOWN(mp))
3196                 return XFS_ERROR(EIO);
3197
3198         link_namelen = VNAMELEN(dentry);
3199
3200         /*
3201          * Check component lengths of the target path name.
3202          */
3203         pathlen = strlen(target_path);
3204         if (pathlen >= MAXPATHLEN)      /* total string too long */
3205                 return XFS_ERROR(ENAMETOOLONG);
3206         if (pathlen >= MAXNAMELEN) {    /* is any component too long? */
3207                 int len, total;
3208                 char *path;
3209
3210                 for (total = 0, path = target_path; total < pathlen;) {
3211                         /*
3212                          * Skip any slashes.
3213                          */
3214                         while(*path == '/') {
3215                                 total++;
3216                                 path++;
3217                         }
3218
3219                         /*
3220                          * Count up to the next slash or end of path.
3221                          * Error out if the component is bigger than MAXNAMELEN.
3222                          */
3223                         for(len = 0; *path != '/' && total < pathlen;total++, path++) {
3224                                 if (++len >= MAXNAMELEN) {
3225                                         error = ENAMETOOLONG;
3226                                         return error;
3227                                 }
3228                         }
3229                 }
3230         }
3231
3232         if (DM_EVENT_ENABLED(dp, DM_EVENT_SYMLINK)) {
3233                 error = XFS_SEND_NAMESP(mp, DM_EVENT_SYMLINK, dir_vp,
3234                                         DM_RIGHT_NULL, NULL, DM_RIGHT_NULL,
3235                                         link_name, target_path, 0, 0, 0);
3236                 if (error)
3237                         return error;
3238         }
3239
3240         /* Return through std_return after this point. */
3241
3242         udqp = gdqp = NULL;
3243         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
3244                 prid = dp->i_d.di_projid;
3245         else
3246                 prid = (xfs_prid_t)dfltprid;
3247
3248         /*
3249          * Make sure that we have allocated dquot(s) on disk.
3250          */
3251         error = XFS_QM_DQVOPALLOC(mp, dp,
3252                         current_fsuid(credp), current_fsgid(credp), prid,
3253                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
3254         if (error)
3255                 goto std_return;
3256
3257         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
3258         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
3259         /*
3260          * The symlink will fit into the inode data fork?
3261          * There can't be any attributes so we get the whole variable part.
3262          */
3263         if (pathlen <= XFS_LITINO(mp))
3264                 fs_blocks = 0;
3265         else
3266                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
3267         resblks = XFS_SYMLINK_SPACE_RES(mp, link_namelen, fs_blocks);
3268         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
3269                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
3270         if (error == ENOSPC && fs_blocks == 0) {
3271                 resblks = 0;
3272                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
3273                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
3274         }
3275         if (error) {
3276                 cancel_flags = 0;
3277                 goto error_return;
3278         }
3279
3280         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
3281         unlock_dp_on_error = B_TRUE;
3282
3283         /*
3284          * Check whether the directory allows new symlinks or not.
3285          */
3286         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
3287                 error = XFS_ERROR(EPERM);
3288                 goto error_return;
3289         }
3290
3291         /*
3292          * Reserve disk quota : blocks and inode.
3293          */
3294         error = XFS_TRANS_RESERVE_QUOTA(mp, tp, udqp, gdqp, resblks, 1, 0);
3295         if (error)
3296                 goto error_return;
3297
3298         /*
3299          * Check for ability to enter directory entry, if no space reserved.
3300          */
3301         if (resblks == 0 &&
3302             (error = xfs_dir_canenter(tp, dp, link_name, link_namelen)))
3303                 goto error_return;
3304         /*
3305          * Initialize the bmap freelist prior to calling either
3306          * bmapi or the directory create code.
3307          */
3308         XFS_BMAP_INIT(&free_list, &first_block);
3309
3310         /*
3311          * Allocate an inode for the symlink.
3312          */
3313         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT),
3314                                1, 0, credp, prid, resblks > 0, &ip, NULL);
3315         if (error) {
3316                 if (error == ENOSPC)
3317                         goto error_return;
3318                 goto error1;
3319         }
3320         ITRACE(ip);
3321
3322         /*
3323          * An error after we've joined dp to the transaction will result in the
3324          * transaction cancel unlocking dp so don't do it explicitly in the
3325          * error path.
3326          */
3327         VN_HOLD(dir_vp);
3328         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
3329         unlock_dp_on_error = B_FALSE;
3330
3331         /*
3332          * Also attach the dquot(s) to it, if applicable.
3333          */
3334         XFS_QM_DQVOPCREATE(mp, tp, ip, udqp, gdqp);
3335
3336         if (resblks)
3337                 resblks -= XFS_IALLOC_SPACE_RES(mp);
3338         /*
3339          * If the symlink will fit into the inode, write it inline.
3340          */
3341         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
3342                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
3343                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
3344                 ip->i_d.di_size = pathlen;
3345
3346                 /*
3347                  * The inode was initially created in extent format.
3348                  */
3349                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
3350                 ip->i_df.if_flags |= XFS_IFINLINE;
3351
3352                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
3353                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
3354
3355         } else {
3356                 first_fsb = 0;
3357                 nmaps = SYMLINK_MAPS;
3358
3359                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
3360                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
3361                                   &first_block, resblks, mval, &nmaps,
3362                                   &free_list, NULL);
3363                 if (error) {
3364                         goto error1;
3365                 }
3366
3367                 if (resblks)
3368                         resblks -= fs_blocks;
3369                 ip->i_d.di_size = pathlen;
3370                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
3371
3372                 cur_chunk = target_path;
3373                 for (n = 0; n < nmaps; n++) {
3374                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
3375                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
3376                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
3377                                                BTOBB(byte_cnt), 0);
3378                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
3379                         if (pathlen < byte_cnt) {
3380                                 byte_cnt = pathlen;
3381                         }
3382                         pathlen -= byte_cnt;
3383
3384                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
3385                         cur_chunk += byte_cnt;
3386
3387                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
3388                 }
3389         }
3390
3391         /*
3392          * Create the directory entry for the symlink.
3393          */
3394         error = xfs_dir_createname(tp, dp, link_name, link_namelen, ip->i_ino,
3395                                    &first_block, &free_list, resblks);
3396         if (error)
3397                 goto error1;
3398         xfs_ichgtime(dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3399         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
3400
3401         /*
3402          * Bump the in memory version number of the parent directory
3403          * so that other processes accessing it will recognize that
3404          * the directory has changed.
3405          */
3406         dp->i_gen++;
3407
3408         /*
3409          * If this is a synchronous mount, make sure that the
3410          * symlink transaction goes to disk before returning to
3411          * the user.
3412          */
3413         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
3414                 xfs_trans_set_sync(tp);
3415         }
3416
3417         /*
3418          * xfs_trans_commit normally decrements the vnode ref count
3419          * when it unlocks the inode. Since we want to return the
3420          * vnode to the caller, we bump the vnode ref count now.
3421          */
3422         IHOLD(ip);
3423
3424         error = xfs_bmap_finish(&tp, &free_list, &committed);
3425         if (error) {
3426                 goto error2;
3427         }
3428         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
3429         XFS_QM_DQRELE(mp, udqp);
3430         XFS_QM_DQRELE(mp, gdqp);
3431
3432         /* Fall through to std_return with error = 0 or errno from
3433          * xfs_trans_commit     */
3434 std_return:
3435         if (DM_EVENT_ENABLED(dp, DM_EVENT_POSTSYMLINK)) {
3436                 (void) XFS_SEND_NAMESP(mp, DM_EVENT_POSTSYMLINK,
3437                                         dir_vp, DM_RIGHT_NULL,
3438                                         error ? NULL : XFS_ITOV(ip),
3439                                         DM_RIGHT_NULL, link_name, target_path,
3440                                         0, error, 0);
3441         }
3442
3443         if (!error) {
3444                 bhv_vnode_t *vp;
3445
3446                 ASSERT(ip);
3447                 vp = XFS_ITOV(ip);
3448                 *vpp = vp;
3449         }
3450         return error;
3451
3452  error2:
3453         IRELE(ip);
3454  error1:
3455         xfs_bmap_cancel(&free_list);
3456         cancel_flags |= XFS_TRANS_ABORT;
3457  error_return:
3458         xfs_trans_cancel(tp, cancel_flags);
3459         XFS_QM_DQRELE(mp, udqp);
3460         XFS_QM_DQRELE(mp, gdqp);
3461
3462         if (unlock_dp_on_error)
3463                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
3464
3465         goto std_return;
3466 }
3467
3468
3469 /*
3470  * xfs_fid2
3471  *
3472  * A fid routine that takes a pointer to a previously allocated
3473  * fid structure (like xfs_fast_fid) but uses a 64 bit inode number.
3474  */
3475 int
3476 xfs_fid2(
3477         xfs_inode_t     *ip,
3478         fid_t           *fidp)
3479 {
3480         xfs_fid2_t      *xfid = (xfs_fid2_t *)fidp;
3481
3482         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
3483         ASSERT(sizeof(fid_t) >= sizeof(xfs_fid2_t));
3484
3485         xfid->fid_len = sizeof(xfs_fid2_t) - sizeof(xfid->fid_len);
3486         xfid->fid_pad = 0;
3487         /*
3488          * use memcpy because the inode is a long long and there's no
3489          * assurance that xfid->fid_ino is properly aligned.
3490          */
3491         memcpy(&xfid->fid_ino, &ip->i_ino, sizeof(xfid->fid_ino));
3492         xfid->fid_gen = ip->i_d.di_gen;
3493
3494         return 0;
3495 }
3496
3497
3498 int
3499 xfs_rwlock(
3500         xfs_inode_t     *ip,
3501         bhv_vrwlock_t   locktype)
3502 {
3503         if (S_ISDIR(ip->i_d.di_mode))
3504                 return 1;
3505         if (locktype == VRWLOCK_WRITE) {
3506                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
3507         } else if (locktype == VRWLOCK_TRY_READ) {
3508                 return xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED);
3509         } else if (locktype == VRWLOCK_TRY_WRITE) {
3510                 return xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL);
3511         } else {
3512                 ASSERT((locktype == VRWLOCK_READ) ||
3513                        (locktype == VRWLOCK_WRITE_DIRECT));
3514                 xfs_ilock(ip, XFS_IOLOCK_SHARED);
3515         }
3516
3517         return 1;
3518 }
3519
3520
3521 void
3522 xfs_rwunlock(
3523         xfs_inode_t     *ip,
3524         bhv_vrwlock_t   locktype)
3525 {
3526         if (S_ISDIR(ip->i_d.di_mode))
3527                 return;
3528         if (locktype == VRWLOCK_WRITE) {
3529                 /*
3530                  * In the write case, we may have added a new entry to
3531                  * the reference cache.  This might store a pointer to
3532                  * an inode to be released in this inode.  If it is there,
3533                  * clear the pointer and release the inode after unlocking
3534                  * this one.
3535                  */
3536                 xfs_refcache_iunlock(ip, XFS_IOLOCK_EXCL);
3537         } else {
3538                 ASSERT((locktype == VRWLOCK_READ) ||
3539                        (locktype == VRWLOCK_WRITE_DIRECT));
3540                 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
3541         }
3542         return;
3543 }
3544
3545
3546 int
3547 xfs_inode_flush(
3548         xfs_inode_t     *ip,
3549         int             flags)
3550 {
3551         xfs_mount_t     *mp = ip->i_mount;
3552         xfs_inode_log_item_t *iip = ip->i_itemp;
3553         int             error = 0;
3554
3555         if (XFS_FORCED_SHUTDOWN(mp))
3556                 return XFS_ERROR(EIO);
3557
3558         /*
3559          * Bypass inodes which have already been cleaned by
3560          * the inode flush clustering code inside xfs_iflush
3561          */
3562         if ((ip->i_update_core == 0) &&
3563             ((iip == NULL) || !(iip->ili_format.ilf_fields & XFS_ILOG_ALL)))
3564                 return 0;
3565
3566         if (flags & FLUSH_LOG) {
3567                 if (iip && iip->ili_last_lsn) {
3568                         xlog_t          *log = mp->m_log;
3569                         xfs_lsn_t       sync_lsn;
3570                         int             s, log_flags = XFS_LOG_FORCE;
3571
3572                         s = GRANT_LOCK(log);
3573                         sync_lsn = log->l_last_sync_lsn;
3574                         GRANT_UNLOCK(log, s);
3575
3576                         if ((XFS_LSN_CMP(iip->ili_last_lsn, sync_lsn) > 0)) {
3577                                 if (flags & FLUSH_SYNC)
3578                                         log_flags |= XFS_LOG_SYNC;
3579                                 error = xfs_log_force(mp, iip->ili_last_lsn, log_flags);
3580                                 if (error)
3581                                         return error;
3582                         }
3583
3584                         if (ip->i_update_core == 0)
3585                                 return 0;
3586                 }
3587         }
3588
3589         /*
3590          * We make this non-blocking if the inode is contended,
3591          * return EAGAIN to indicate to the caller that they
3592          * did not succeed. This prevents the flush path from
3593          * blocking on inodes inside another operation right
3594          * now, they get caught later by xfs_sync.
3595          */
3596         if (flags & FLUSH_INODE) {
3597                 int     flush_flags;
3598
3599                 if (flags & FLUSH_SYNC) {
3600                         xfs_ilock(ip, XFS_ILOCK_SHARED);
3601                         xfs_iflock(ip);
3602                 } else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3603                         if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
3604                                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3605                                 return EAGAIN;
3606                         }
3607                 } else {
3608                         return EAGAIN;
3609                 }
3610
3611                 if (flags & FLUSH_SYNC)
3612                         flush_flags = XFS_IFLUSH_SYNC;
3613                 else
3614                         flush_flags = XFS_IFLUSH_ASYNC;
3615
3616                 error = xfs_iflush(ip, flush_flags);
3617                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3618         }
3619
3620         return error;
3621 }
3622
3623
3624 int
3625 xfs_set_dmattrs(
3626         xfs_inode_t     *ip,
3627         u_int           evmask,
3628         u_int16_t       state)
3629 {
3630         xfs_mount_t     *mp = ip->i_mount;
3631         xfs_trans_t     *tp;
3632         int             error;
3633
3634         if (!capable(CAP_SYS_ADMIN))
3635                 return XFS_ERROR(EPERM);
3636
3637         if (XFS_FORCED_SHUTDOWN(mp))
3638                 return XFS_ERROR(EIO);
3639
3640         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
3641         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
3642         if (error) {
3643                 xfs_trans_cancel(tp, 0);
3644                 return error;
3645         }
3646         xfs_ilock(ip, XFS_ILOCK_EXCL);
3647         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3648
3649         ip->i_iocore.io_dmevmask = ip->i_d.di_dmevmask = evmask;
3650         ip->i_iocore.io_dmstate  = ip->i_d.di_dmstate  = state;
3651
3652         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
3653         IHOLD(ip);
3654         error = xfs_trans_commit(tp, 0);
3655
3656         return error;
3657 }
3658
3659 int
3660 xfs_reclaim(
3661         xfs_inode_t     *ip)
3662 {
3663         bhv_vnode_t     *vp = XFS_ITOV(ip);
3664
3665         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
3666
3667         ASSERT(!VN_MAPPED(vp));
3668
3669         /* bad inode, get out here ASAP */
3670         if (VN_BAD(vp)) {
3671                 xfs_ireclaim(ip);
3672                 return 0;
3673         }
3674
3675         vn_iowait(ip);
3676
3677         ASSERT(XFS_FORCED_SHUTDOWN(ip->i_mount) || ip->i_delayed_blks == 0);
3678
3679         /*
3680          * Make sure the atime in the XFS inode is correct before freeing the
3681          * Linux inode.
3682          */
3683         xfs_synchronize_atime(ip);
3684
3685         /*
3686          * If we have nothing to flush with this inode then complete the
3687          * teardown now, otherwise break the link between the xfs inode and the
3688          * linux inode and clean up the xfs inode later. This avoids flushing
3689          * the inode to disk during the delete operation itself.
3690          *
3691          * When breaking the link, we need to set the XFS_IRECLAIMABLE flag
3692          * first to ensure that xfs_iunpin() will never see an xfs inode
3693          * that has a linux inode being reclaimed. Synchronisation is provided
3694          * by the i_flags_lock.
3695          */
3696         if (!ip->i_update_core && (ip->i_itemp == NULL)) {
3697                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3698                 xfs_iflock(ip);
3699                 return xfs_finish_reclaim(ip, 1, XFS_IFLUSH_DELWRI_ELSE_SYNC);
3700         } else {
3701                 xfs_mount_t     *mp = ip->i_mount;
3702
3703                 /* Protect sync and unpin from us */
3704                 XFS_MOUNT_ILOCK(mp);
3705                 spin_lock(&ip->i_flags_lock);
3706                 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
3707                 vn_to_inode(vp)->i_private = NULL;
3708                 ip->i_vnode = NULL;
3709                 spin_unlock(&ip->i_flags_lock);
3710                 list_add_tail(&ip->i_reclaim, &mp->m_del_inodes);
3711                 XFS_MOUNT_IUNLOCK(mp);
3712         }
3713         return 0;
3714 }
3715
3716 int
3717 xfs_finish_reclaim(
3718         xfs_inode_t     *ip,
3719         int             locked,
3720         int             sync_mode)
3721 {
3722         xfs_perag_t     *pag = xfs_get_perag(ip->i_mount, ip->i_ino);
3723         bhv_vnode_t     *vp = XFS_ITOV_NULL(ip);
3724         int             error;
3725
3726         if (vp && VN_BAD(vp))
3727                 goto reclaim;
3728
3729         /* The hash lock here protects a thread in xfs_iget_core from
3730          * racing with us on linking the inode back with a vnode.
3731          * Once we have the XFS_IRECLAIM flag set it will not touch
3732          * us.
3733          */
3734         write_lock(&pag->pag_ici_lock);
3735         spin_lock(&ip->i_flags_lock);
3736         if (__xfs_iflags_test(ip, XFS_IRECLAIM) ||
3737             (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) && vp == NULL)) {
3738                 spin_unlock(&ip->i_flags_lock);
3739                 write_unlock(&pag->pag_ici_lock);
3740                 if (locked) {
3741                         xfs_ifunlock(ip);
3742                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3743                 }
3744                 return 1;
3745         }
3746         __xfs_iflags_set(ip, XFS_IRECLAIM);
3747         spin_unlock(&ip->i_flags_lock);
3748         write_unlock(&pag->pag_ici_lock);
3749         xfs_put_perag(ip->i_mount, pag);
3750
3751         /*
3752          * If the inode is still dirty, then flush it out.  If the inode
3753          * is not in the AIL, then it will be OK to flush it delwri as
3754          * long as xfs_iflush() does not keep any references to the inode.
3755          * We leave that decision up to xfs_iflush() since it has the
3756          * knowledge of whether it's OK to simply do a delwri flush of
3757          * the inode or whether we need to wait until the inode is
3758          * pulled from the AIL.
3759          * We get the flush lock regardless, though, just to make sure
3760          * we don't free it while it is being flushed.
3761          */
3762         if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
3763                 if (!locked) {
3764                         xfs_ilock(ip, XFS_ILOCK_EXCL);
3765                         xfs_iflock(ip);
3766                 }
3767
3768                 if (ip->i_update_core ||
3769                     ((ip->i_itemp != NULL) &&
3770                      (ip->i_itemp->ili_format.ilf_fields != 0))) {
3771                         error = xfs_iflush(ip, sync_mode);
3772                         /*
3773                          * If we hit an error, typically because of filesystem
3774                          * shutdown, we don't need to let vn_reclaim to know
3775                          * because we're gonna reclaim the inode anyway.
3776                          */
3777                         if (error) {
3778                                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3779                                 goto reclaim;
3780                         }
3781                         xfs_iflock(ip); /* synchronize with xfs_iflush_done */
3782                 }
3783
3784                 ASSERT(ip->i_update_core == 0);
3785                 ASSERT(ip->i_itemp == NULL ||
3786                        ip->i_itemp->ili_format.ilf_fields == 0);
3787                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3788         } else if (locked) {
3789                 /*
3790                  * We are not interested in doing an iflush if we're
3791                  * in the process of shutting down the filesystem forcibly.
3792                  * So, just reclaim the inode.
3793                  */
3794                 xfs_ifunlock(ip);
3795                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
3796         }
3797
3798  reclaim:
3799         xfs_ireclaim(ip);
3800         return 0;
3801 }
3802
3803 int
3804 xfs_finish_reclaim_all(xfs_mount_t *mp, int noblock)
3805 {
3806         int             purged;
3807         xfs_inode_t     *ip, *n;
3808         int             done = 0;
3809
3810         while (!done) {
3811                 purged = 0;
3812                 XFS_MOUNT_ILOCK(mp);
3813                 list_for_each_entry_safe(ip, n, &mp->m_del_inodes, i_reclaim) {
3814                         if (noblock) {
3815                                 if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0)
3816                                         continue;
3817                                 if (xfs_ipincount(ip) ||
3818                                     !xfs_iflock_nowait(ip)) {
3819                                         xfs_iunlock(ip, XFS_ILOCK_EXCL);
3820                                         continue;
3821                                 }
3822                         }
3823                         XFS_MOUNT_IUNLOCK(mp);
3824                         if (xfs_finish_reclaim(ip, noblock,
3825                                         XFS_IFLUSH_DELWRI_ELSE_ASYNC))
3826                                 delay(1);
3827                         purged = 1;
3828                         break;
3829                 }
3830
3831                 done = !purged;
3832         }
3833
3834         XFS_MOUNT_IUNLOCK(mp);
3835         return 0;
3836 }
3837
3838 /*
3839  * xfs_alloc_file_space()
3840  *      This routine allocates disk space for the given file.
3841  *
3842  *      If alloc_type == 0, this request is for an ALLOCSP type
3843  *      request which will change the file size.  In this case, no
3844  *      DMAPI event will be generated by the call.  A TRUNCATE event
3845  *      will be generated later by xfs_setattr.
3846  *
3847  *      If alloc_type != 0, this request is for a RESVSP type
3848  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
3849  *      lower block boundary byte address is less than the file's
3850  *      length.
3851  *
3852  * RETURNS:
3853  *       0 on success
3854  *      errno on error
3855  *
3856  */
3857 STATIC int
3858 xfs_alloc_file_space(
3859         xfs_inode_t             *ip,
3860         xfs_off_t               offset,
3861         xfs_off_t               len,
3862         int                     alloc_type,
3863         int                     attr_flags)
3864 {
3865         xfs_mount_t             *mp = ip->i_mount;
3866         xfs_off_t               count;
3867         xfs_filblks_t           allocated_fsb;
3868         xfs_filblks_t           allocatesize_fsb;
3869         xfs_extlen_t            extsz, temp;
3870         xfs_fileoff_t           startoffset_fsb;
3871         xfs_fsblock_t           firstfsb;
3872         int                     nimaps;
3873         int                     bmapi_flag;
3874         int                     quota_flag;
3875         int                     rt;
3876         xfs_trans_t             *tp;
3877         xfs_bmbt_irec_t         imaps[1], *imapp;
3878         xfs_bmap_free_t         free_list;
3879         uint                    qblocks, resblks, resrtextents;
3880         int                     committed;
3881         int                     error;
3882
3883         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
3884
3885         if (XFS_FORCED_SHUTDOWN(mp))
3886                 return XFS_ERROR(EIO);
3887
3888         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
3889                 return error;
3890
3891         if (len <= 0)
3892                 return XFS_ERROR(EINVAL);
3893
3894         rt = XFS_IS_REALTIME_INODE(ip);
3895         extsz = xfs_get_extsz_hint(ip);
3896
3897         count = len;
3898         imapp = &imaps[0];
3899         nimaps = 1;
3900         bmapi_flag = XFS_BMAPI_WRITE | (alloc_type ? XFS_BMAPI_PREALLOC : 0);
3901         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
3902         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
3903
3904         /*      Generate a DMAPI event if needed.       */
3905         if (alloc_type != 0 && offset < ip->i_size &&
3906                         (attr_flags&ATTR_DMI) == 0  &&
3907                         DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
3908                 xfs_off_t           end_dmi_offset;
3909
3910                 end_dmi_offset = offset+len;
3911                 if (end_dmi_offset > ip->i_size)
3912                         end_dmi_offset = ip->i_size;
3913                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, XFS_ITOV(ip),
3914                         offset, end_dmi_offset - offset,
3915                         0, NULL);
3916                 if (error)
3917                         return error;
3918         }
3919
3920         /*
3921          * Allocate file space until done or until there is an error
3922          */
3923 retry:
3924         while (allocatesize_fsb && !error) {
3925                 xfs_fileoff_t   s, e;
3926
3927                 /*
3928                  * Determine space reservations for data/realtime.
3929                  */
3930                 if (unlikely(extsz)) {
3931                         s = startoffset_fsb;
3932                         do_div(s, extsz);
3933                         s *= extsz;
3934                         e = startoffset_fsb + allocatesize_fsb;
3935                         if ((temp = do_mod(startoffset_fsb, extsz)))
3936                                 e += temp;
3937                         if ((temp = do_mod(e, extsz)))
3938                                 e += extsz - temp;
3939                 } else {
3940                         s = 0;
3941                         e = allocatesize_fsb;
3942                 }
3943
3944                 if (unlikely(rt)) {
3945                         resrtextents = qblocks = (uint)(e - s);
3946                         resrtextents /= mp->m_sb.sb_rextsize;
3947                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
3948                         quota_flag = XFS_QMOPT_RES_RTBLKS;
3949                 } else {
3950                         resrtextents = 0;
3951                         resblks = qblocks = \
3952                                 XFS_DIOSTRAT_SPACE_RES(mp, (uint)(e - s));
3953                         quota_flag = XFS_QMOPT_RES_REGBLKS;
3954                 }
3955
3956                 /*
3957                  * Allocate and setup the transaction.
3958                  */
3959                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
3960                 error = xfs_trans_reserve(tp, resblks,
3961                                           XFS_WRITE_LOG_RES(mp), resrtextents,
3962                                           XFS_TRANS_PERM_LOG_RES,
3963                                           XFS_WRITE_LOG_COUNT);
3964                 /*
3965                  * Check for running out of space
3966                  */
3967                 if (error) {
3968                         /*
3969                          * Free the transaction structure.
3970                          */
3971                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
3972                         xfs_trans_cancel(tp, 0);
3973                         break;
3974                 }
3975                 xfs_ilock(ip, XFS_ILOCK_EXCL);
3976                 error = XFS_TRANS_RESERVE_QUOTA_NBLKS(mp, tp, ip,
3977                                                       qblocks, 0, quota_flag);
3978                 if (error)
3979                         goto error1;
3980
3981                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
3982                 xfs_trans_ihold(tp, ip);
3983
3984                 /*
3985                  * Issue the xfs_bmapi() call to allocate the blocks
3986                  */
3987                 XFS_BMAP_INIT(&free_list, &firstfsb);
3988                 error = XFS_BMAPI(mp, tp, &ip->i_iocore, startoffset_fsb,
3989                                   allocatesize_fsb, bmapi_flag,
3990                                   &firstfsb, 0, imapp, &nimaps,
3991                                   &free_list, NULL);
3992                 if (error) {
3993                         goto error0;
3994                 }
3995
3996                 /*
3997                  * Complete the transaction
3998                  */
3999                 error = xfs_bmap_finish(&tp, &free_list, &committed);
4000                 if (error) {
4001                         goto error0;
4002                 }
4003
4004                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
4005                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
4006                 if (error) {
4007                         break;
4008                 }
4009
4010                 allocated_fsb = imapp->br_blockcount;
4011
4012                 if (nimaps == 0) {
4013                         error = XFS_ERROR(ENOSPC);
4014                         break;
4015                 }
4016
4017                 startoffset_fsb += allocated_fsb;
4018                 allocatesize_fsb -= allocated_fsb;
4019         }
4020 dmapi_enospc_check:
4021         if (error == ENOSPC && (attr_flags & ATTR_DMI) == 0 &&
4022             DM_EVENT_ENABLED(ip, DM_EVENT_NOSPACE)) {
4023                 error = XFS_SEND_NAMESP(mp, DM_EVENT_NOSPACE,
4024                                 XFS_ITOV(ip), DM_RIGHT_NULL,
4025                                 XFS_ITOV(ip), DM_RIGHT_NULL,
4026                                 NULL, NULL, 0, 0, 0); /* Delay flag intentionally unused */
4027                 if (error == 0)
4028                         goto retry;     /* Maybe DMAPI app. has made space */
4029                 /* else fall through with error from XFS_SEND_DATA */
4030         }
4031
4032         return error;
4033
4034 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
4035         xfs_bmap_cancel(&free_list);
4036         XFS_TRANS_UNRESERVE_QUOTA_NBLKS(mp, tp, ip, qblocks, 0, quota_flag);
4037
4038 error1: /* Just cancel transaction */
4039         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
4040         xfs_iunlock(ip, XFS_ILOCK_EXCL);
4041         goto dmapi_enospc_check;
4042 }
4043
4044 /*
4045  * Zero file bytes between startoff and endoff inclusive.
4046  * The iolock is held exclusive and no blocks are buffered.
4047  */
4048 STATIC int
4049 xfs_zero_remaining_bytes(
4050         xfs_inode_t             *ip,
4051         xfs_off_t               startoff,
4052         xfs_off_t               endoff)
4053 {
4054         xfs_bmbt_irec_t         imap;
4055         xfs_fileoff_t           offset_fsb;
4056         xfs_off_t               lastoffset;
4057         xfs_off_t               offset;
4058         xfs_buf_t               *bp;
4059         xfs_mount_t             *mp = ip->i_mount;
4060         int                     nimap;
4061         int                     error = 0;
4062
4063         bp = xfs_buf_get_noaddr(mp->m_sb.sb_blocksize,
4064                                 ip->i_d.di_flags & XFS_DIFLAG_REALTIME ?
4065                                 mp->m_rtdev_targp : mp->m_ddev_targp);
4066
4067         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
4068                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
4069                 nimap = 1;
4070                 error = XFS_BMAPI(mp, NULL, &ip->i_iocore, offset_fsb, 1, 0,
4071                         NULL, 0, &imap, &nimap, NULL, NULL);
4072                 if (error || nimap < 1)
4073                         break;
4074                 ASSERT(imap.br_blockcount >= 1);
4075                 ASSERT(imap.br_startoff == offset_fsb);
4076                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
4077                 if (lastoffset > endoff)
4078                         lastoffset = endoff;
4079                 if (imap.br_startblock == HOLESTARTBLOCK)
4080                         continue;
4081                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
4082                 if (imap.br_state == XFS_EXT_UNWRITTEN)
4083                         continue;
4084                 XFS_BUF_UNDONE(bp);
4085                 XFS_BUF_UNWRITE(bp);
4086                 XFS_BUF_READ(bp);
4087                 XFS_BUF_SET_ADDR(bp, XFS_FSB_TO_DB(ip, imap.br_startblock));
4088                 xfsbdstrat(mp, bp);
4089                 if ((error = xfs_iowait(bp))) {
4090                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
4091                                           mp, bp, XFS_BUF_ADDR(bp));
4092                         break;
4093                 }
4094                 memset(XFS_BUF_PTR(bp) +
4095                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
4096                       0, lastoffset - offset + 1);
4097                 XFS_BUF_UNDONE(bp);
4098                 XFS_BUF_UNREAD(bp);
4099                 XFS_BUF_WRITE(bp);
4100                 xfsbdstrat(mp, bp);
4101                 if ((error = xfs_iowait(bp))) {
4102                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
4103                                           mp, bp, XFS_BUF_ADDR(bp));
4104                         break;
4105                 }
4106         }
4107         xfs_buf_free(bp);
4108         return error;
4109 }
4110
4111 /*
4112  * xfs_free_file_space()
4113  *      This routine frees disk space for the given file.
4114  *
4115  *      This routine is only called by xfs_change_file_space
4116  *      for an UNRESVSP type call.
4117  *
4118  * RETURNS:
4119  *       0 on success
4120  *      errno on error
4121  *
4122  */
4123 STATIC int
4124 xfs_free_file_space(
4125         xfs_inode_t             *ip,
4126         xfs_off_t               offset,
4127         xfs_off_t               len,
4128         int                     attr_flags)
4129 {
4130         bhv_vnode_t             *vp;
4131         int                     committed;
4132         int                     done;
4133         xfs_off_t               end_dmi_offset;
4134         xfs_fileoff_t           endoffset_fsb;
4135         int                     error;
4136         xfs_fsblock_t           firstfsb;
4137         xfs_bmap_free_t         free_list;
4138         xfs_bmbt_irec_t         imap;
4139         xfs_off_t               ioffset;
4140         xfs_extlen_t            mod=0;
4141         xfs_mount_t             *mp;
4142         int                     nimap;
4143         uint                    resblks;
4144         uint                    rounding;
4145         int                     rt;
4146         xfs_fileoff_t           startoffset_fsb;
4147         xfs_trans_t             *tp;
4148         int                     need_iolock = 1;
4149
4150         vp = XFS_ITOV(ip);
4151         mp = ip->i_mount;
4152
4153         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
4154
4155         if ((error = XFS_QM_DQATTACH(mp, ip, 0)))
4156                 return error;
4157
4158         error = 0;
4159         if (len <= 0)   /* if nothing being freed */
4160                 return error;
4161         rt = (ip->i_d.di_flags & XFS_DIFLAG_REALTIME);
4162         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
4163         end_dmi_offset = offset + len;
4164         endoffset_fsb = XFS_B_TO_FSBT(mp, end_dmi_offset);
4165
4166         if (offset < ip->i_size && (attr_flags & ATTR_DMI) == 0 &&
4167             DM_EVENT_ENABLED(ip, DM_EVENT_WRITE)) {
4168                 if (end_dmi_offset > ip->i_size)
4169                         end_dmi_offset = ip->i_size;
4170                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, vp,
4171                                 offset, end_dmi_offset - offset,
4172                                 AT_DELAY_FLAG(attr_flags), NULL);
4173                 if (error)
4174                         return error;
4175         }
4176
4177         if (attr_flags & ATTR_NOLOCK)
4178                 need_iolock = 0;
4179         if (need_iolock) {
4180                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
4181                 vn_iowait(ip);  /* wait for the completion of any pending DIOs */
4182         }
4183
4184         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, NBPP);
4185         ioffset = offset & ~(rounding - 1);
4186
4187         if (VN_CACHED(vp) != 0) {
4188                 xfs_inval_cached_trace(&ip->i_iocore, ioffset, -1,
4189                                 ctooff(offtoct(ioffset)), -1);
4190                 error = xfs_flushinval_pages(ip,
4191                                 ctooff(offtoct(ioffset)),
4192                                 -1, FI_REMAPF_LOCKED);
4193                 if (error)
4194                         goto out_unlock_iolock;
4195         }
4196
4197         /*
4198          * Need to zero the stuff we're not freeing, on disk.
4199          * If its a realtime file & can't use unwritten extents then we
4200          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
4201          * will take care of it for us.
4202          */
4203         if (rt && !XFS_SB_VERSION_HASEXTFLGBIT(&mp->m_sb)) {
4204                 nimap = 1;
4205                 error = XFS_BMAPI(mp, NULL, &ip->i_iocore, startoffset_fsb,
4206                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
4207                 if (error)
4208                         goto out_unlock_iolock;
4209                 ASSERT(nimap == 0 || nimap == 1);
4210                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
4211                         xfs_daddr_t     block;
4212
4213                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
4214                         block = imap.br_startblock;
4215                         mod = do_div(block, mp->m_sb.sb_rextsize);
4216                         if (mod)
4217                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
4218                 }
4219                 nimap = 1;
4220                 error = XFS_BMAPI(mp, NULL, &ip->i_iocore, endoffset_fsb - 1,
4221                         1, 0, NULL, 0, &imap, &nimap, NULL, NULL);
4222                 if (error)
4223                         goto out_unlock_iolock;
4224                 ASSERT(nimap == 0 || nimap == 1);
4225                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
4226                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
4227                         mod++;
4228                         if (mod && (mod != mp->m_sb.sb_rextsize))
4229                                 endoffset_fsb -= mod;
4230                 }
4231         }
4232         if ((done = (endoffset_fsb <= startoffset_fsb)))
4233                 /*
4234                  * One contiguous piece to clear
4235                  */
4236                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
4237         else {
4238                 /*
4239                  * Some full blocks, possibly two pieces to clear
4240                  */
4241                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
4242                         error = xfs_zero_remaining_bytes(ip, offset,
4243                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
4244                 if (!error &&
4245                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
4246                         error = xfs_zero_remaining_bytes(ip,
4247                                 XFS_FSB_TO_B(mp, endoffset_fsb),
4248                                 offset + len - 1);
4249         }
4250
4251         /*
4252          * free file space until done or until there is an error
4253          */
4254         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
4255         while (!error && !done) {
4256
4257                 /*
4258                  * allocate and setup the transaction. Allow this
4259                  * transaction to dip into the reserve blocks to ensure
4260                  * the freeing of the space succeeds at ENOSPC.
4261                  */
4262                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
4263                 tp->t_flags |= XFS_TRANS_RESERVE;
4264                 error = xfs_trans_reserve(tp,
4265                                           resblks,
4266                                           XFS_WRITE_LOG_RES(mp),
4267                                           0,
4268                                           XFS_TRANS_PERM_LOG_RES,
4269                                           XFS_WRITE_LOG_COUNT);
4270
4271                 /*
4272                  * check for running out of space
4273                  */
4274                 if (error) {
4275                         /*
4276                          * Free the transaction structure.
4277                          */
4278                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
4279                         xfs_trans_cancel(tp, 0);
4280                         break;
4281                 }
4282                 xfs_ilock(ip, XFS_ILOCK_EXCL);
4283                 error = XFS_TRANS_RESERVE_QUOTA(mp, tp,
4284                                 ip->i_udquot, ip->i_gdquot, resblks, 0,
4285                                 XFS_QMOPT_RES_REGBLKS);
4286                 if (error)
4287                         goto error1;
4288
4289                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
4290                 xfs_trans_ihold(tp, ip);
4291
4292                 /*
4293                  * issue the bunmapi() call to free the blocks
4294                  */
4295                 XFS_BMAP_INIT(&free_list, &firstfsb);
4296                 error = XFS_BUNMAPI(mp, tp, &ip->i_iocore, startoffset_fsb,
4297                                   endoffset_fsb - startoffset_fsb,
4298                                   0, 2, &firstfsb, &free_list, NULL, &done);
4299                 if (error) {
4300                         goto error0;
4301                 }
4302
4303                 /*
4304                  * complete the transaction
4305                  */
4306                 error = xfs_bmap_finish(&tp, &free_list, &committed);
4307                 if (error) {
4308                         goto error0;
4309                 }
4310
4311                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
4312                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
4313         }
4314
4315  out_unlock_iolock:
4316         if (need_iolock)
4317                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
4318         return error;
4319
4320  error0:
4321         xfs_bmap_cancel(&free_list);
4322  error1:
4323         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
4324         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
4325                     XFS_ILOCK_EXCL);
4326         return error;
4327 }
4328
4329 /*
4330  * xfs_change_file_space()
4331  *      This routine allocates or frees disk space for the given file.
4332  *      The user specified parameters are checked for alignment and size
4333  *      limitations.
4334  *
4335  * RETURNS:
4336  *       0 on success
4337  *      errno on error
4338  *
4339  */
4340 int
4341 xfs_change_file_space(
4342         xfs_inode_t     *ip,
4343         int             cmd,
4344         xfs_flock64_t   *bf,
4345         xfs_off_t       offset,
4346         cred_t          *credp,
4347         int             attr_flags)
4348 {
4349         xfs_mount_t     *mp = ip->i_mount;
4350         int             clrprealloc;
4351         int             error;
4352         xfs_fsize_t     fsize;
4353         int             setprealloc;
4354         xfs_off_t       startoffset;
4355         xfs_off_t       llen;
4356         xfs_trans_t     *tp;
4357         bhv_vattr_t     va;
4358
4359         vn_trace_entry(ip, __FUNCTION__, (inst_t *)__return_address);
4360
4361         /*
4362          * must be a regular file and have write permission
4363          */
4364         if (!S_ISREG(ip->i_d.di_mode))
4365                 return XFS_ERROR(EINVAL);
4366
4367         xfs_ilock(ip, XFS_ILOCK_SHARED);
4368
4369         if ((error = xfs_iaccess(ip, S_IWUSR, credp))) {
4370                 xfs_iunlock(ip, XFS_ILOCK_SHARED);
4371                 return error;
4372         }
4373
4374         xfs_iunlock(ip, XFS_ILOCK_SHARED);
4375
4376         switch (bf->l_whence) {
4377         case 0: /*SEEK_SET*/
4378                 break;
4379         case 1: /*SEEK_CUR*/
4380                 bf->l_start += offset;
4381                 break;
4382         case 2: /*SEEK_END*/
4383                 bf->l_start += ip->i_size;
4384                 break;
4385         default:
4386                 return XFS_ERROR(EINVAL);
4387         }
4388
4389         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
4390
4391         if (   (bf->l_start < 0)
4392             || (bf->l_start > XFS_MAXIOFFSET(mp))
4393             || (bf->l_start + llen < 0)
4394             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
4395                 return XFS_ERROR(EINVAL);
4396
4397         bf->l_whence = 0;
4398
4399         startoffset = bf->l_start;
4400         fsize = ip->i_size;
4401
4402         /*
4403          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
4404          * file space.
4405          * These calls do NOT zero the data space allocated to the file,
4406          * nor do they change the file size.
4407          *
4408          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
4409          * space.
4410          * These calls cause the new file data to be zeroed and the file
4411          * size to be changed.
4412          */
4413         setprealloc = clrprealloc = 0;
4414
4415         switch (cmd) {
4416         case XFS_IOC_RESVSP:
4417         case XFS_IOC_RESVSP64:
4418                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
4419                                                                 1, attr_flags);
4420                 if (error)
4421                         return error;
4422                 setprealloc = 1;
4423                 break;
4424
4425         case XFS_IOC_UNRESVSP:
4426         case XFS_IOC_UNRESVSP64:
4427                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
4428                                                                 attr_flags)))
4429                         return error;
4430                 break;
4431
4432         case XFS_IOC_ALLOCSP:
4433         case XFS_IOC_ALLOCSP64:
4434         case XFS_IOC_FREESP:
4435         case XFS_IOC_FREESP64:
4436                 if (startoffset > fsize) {
4437                         error = xfs_alloc_file_space(ip, fsize,
4438                                         startoffset - fsize, 0, attr_flags);
4439                         if (error)
4440                                 break;
4441                 }
4442
4443                 va.va_mask = XFS_AT_SIZE;
4444                 va.va_size = startoffset;
4445
4446                 error = xfs_setattr(ip, &va, attr_flags, credp);
4447
4448                 if (error)
4449                         return error;
4450
4451                 clrprealloc = 1;
4452                 break;
4453
4454         default:
4455                 ASSERT(0);
4456                 return XFS_ERROR(EINVAL);
4457         }
4458
4459         /*
4460          * update the inode timestamp, mode, and prealloc flag bits
4461          */
4462         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
4463
4464         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
4465                                       0, 0, 0))) {
4466                 /* ASSERT(0); */
4467                 xfs_trans_cancel(tp, 0);
4468                 return error;
4469         }
4470
4471         xfs_ilock(ip, XFS_ILOCK_EXCL);
4472
4473         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
4474         xfs_trans_ihold(tp, ip);
4475
4476         if ((attr_flags & ATTR_DMI) == 0) {
4477                 ip->i_d.di_mode &= ~S_ISUID;
4478
4479                 /*
4480                  * Note that we don't have to worry about mandatory
4481                  * file locking being disabled here because we only
4482                  * clear the S_ISGID bit if the Group execute bit is
4483                  * on, but if it was on then mandatory locking wouldn't
4484                  * have been enabled.
4485                  */
4486                 if (ip->i_d.di_mode & S_IXGRP)
4487                         ip->i_d.di_mode &= ~S_ISGID;
4488
4489                 xfs_ichgtime(ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
4490         }
4491         if (setprealloc)
4492                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
4493         else if (clrprealloc)
4494                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
4495
4496         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
4497         xfs_trans_set_sync(tp);
4498
4499         error = xfs_trans_commit(tp, 0);
4500
4501         xfs_iunlock(ip, XFS_ILOCK_EXCL);
4502
4503         return error;
4504 }