xfs: return the buffer locked from xfs_buf_get_uncached
[linux-2.6.git] / fs / xfs / xfs_vnodeops.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18
19 #include "xfs.h"
20 #include "xfs_fs.h"
21 #include "xfs_types.h"
22 #include "xfs_bit.h"
23 #include "xfs_log.h"
24 #include "xfs_inum.h"
25 #include "xfs_trans.h"
26 #include "xfs_sb.h"
27 #include "xfs_ag.h"
28 #include "xfs_dir2.h"
29 #include "xfs_mount.h"
30 #include "xfs_da_btree.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_dinode.h"
34 #include "xfs_inode.h"
35 #include "xfs_inode_item.h"
36 #include "xfs_itable.h"
37 #include "xfs_ialloc.h"
38 #include "xfs_alloc.h"
39 #include "xfs_bmap.h"
40 #include "xfs_acl.h"
41 #include "xfs_attr.h"
42 #include "xfs_rw.h"
43 #include "xfs_error.h"
44 #include "xfs_quota.h"
45 #include "xfs_utils.h"
46 #include "xfs_rtalloc.h"
47 #include "xfs_trans_space.h"
48 #include "xfs_log_priv.h"
49 #include "xfs_filestream.h"
50 #include "xfs_vnodeops.h"
51 #include "xfs_trace.h"
52
53 /*
54  * The maximum pathlen is 1024 bytes. Since the minimum file system
55  * blocksize is 512 bytes, we can get a max of 2 extents back from
56  * bmapi.
57  */
58 #define SYMLINK_MAPS 2
59
60 STATIC int
61 xfs_readlink_bmap(
62         xfs_inode_t     *ip,
63         char            *link)
64 {
65         xfs_mount_t     *mp = ip->i_mount;
66         int             pathlen = ip->i_d.di_size;
67         int             nmaps = SYMLINK_MAPS;
68         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
69         xfs_daddr_t     d;
70         int             byte_cnt;
71         int             n;
72         xfs_buf_t       *bp;
73         int             error = 0;
74
75         error = xfs_bmapi(NULL, ip, 0, XFS_B_TO_FSB(mp, pathlen), 0, NULL, 0,
76                         mval, &nmaps, NULL);
77         if (error)
78                 goto out;
79
80         for (n = 0; n < nmaps; n++) {
81                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
82                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
83
84                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt),
85                                   XBF_LOCK | XBF_MAPPED | XBF_DONT_BLOCK);
86                 error = XFS_BUF_GETERROR(bp);
87                 if (error) {
88                         xfs_ioerror_alert("xfs_readlink",
89                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
90                         xfs_buf_relse(bp);
91                         goto out;
92                 }
93                 if (pathlen < byte_cnt)
94                         byte_cnt = pathlen;
95                 pathlen -= byte_cnt;
96
97                 memcpy(link, XFS_BUF_PTR(bp), byte_cnt);
98                 xfs_buf_relse(bp);
99         }
100
101         link[ip->i_d.di_size] = '\0';
102         error = 0;
103
104  out:
105         return error;
106 }
107
108 int
109 xfs_readlink(
110         xfs_inode_t     *ip,
111         char            *link)
112 {
113         xfs_mount_t     *mp = ip->i_mount;
114         int             pathlen;
115         int             error = 0;
116
117         trace_xfs_readlink(ip);
118
119         if (XFS_FORCED_SHUTDOWN(mp))
120                 return XFS_ERROR(EIO);
121
122         xfs_ilock(ip, XFS_ILOCK_SHARED);
123
124         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFLNK);
125         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
126
127         pathlen = ip->i_d.di_size;
128         if (!pathlen)
129                 goto out;
130
131         if (ip->i_df.if_flags & XFS_IFINLINE) {
132                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
133                 link[pathlen] = '\0';
134         } else {
135                 error = xfs_readlink_bmap(ip, link);
136         }
137
138  out:
139         xfs_iunlock(ip, XFS_ILOCK_SHARED);
140         return error;
141 }
142
143 /*
144  * Flags for xfs_free_eofblocks
145  */
146 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
147
148 /*
149  * This is called by xfs_inactive to free any blocks beyond eof
150  * when the link count isn't zero and by xfs_dm_punch_hole() when
151  * punching a hole to EOF.
152  */
153 STATIC int
154 xfs_free_eofblocks(
155         xfs_mount_t     *mp,
156         xfs_inode_t     *ip,
157         int             flags)
158 {
159         xfs_trans_t     *tp;
160         int             error;
161         xfs_fileoff_t   end_fsb;
162         xfs_fileoff_t   last_fsb;
163         xfs_filblks_t   map_len;
164         int             nimaps;
165         xfs_bmbt_irec_t imap;
166
167         /*
168          * Figure out if there are any blocks beyond the end
169          * of the file.  If not, then there is nothing to do.
170          */
171         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
172         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
173         if (last_fsb <= end_fsb)
174                 return 0;
175         map_len = last_fsb - end_fsb;
176
177         nimaps = 1;
178         xfs_ilock(ip, XFS_ILOCK_SHARED);
179         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
180                           NULL, 0, &imap, &nimaps, NULL);
181         xfs_iunlock(ip, XFS_ILOCK_SHARED);
182
183         if (!error && (nimaps != 0) &&
184             (imap.br_startblock != HOLESTARTBLOCK ||
185              ip->i_delayed_blks)) {
186                 /*
187                  * Attach the dquots to the inode up front.
188                  */
189                 error = xfs_qm_dqattach(ip, 0);
190                 if (error)
191                         return error;
192
193                 /*
194                  * There are blocks after the end of file.
195                  * Free them up now by truncating the file to
196                  * its current size.
197                  */
198                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
199
200                 if (flags & XFS_FREE_EOF_TRYLOCK) {
201                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
202                                 xfs_trans_cancel(tp, 0);
203                                 return 0;
204                         }
205                 } else {
206                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
207                 }
208
209                 error = xfs_trans_reserve(tp, 0,
210                                           XFS_ITRUNCATE_LOG_RES(mp),
211                                           0, XFS_TRANS_PERM_LOG_RES,
212                                           XFS_ITRUNCATE_LOG_COUNT);
213                 if (error) {
214                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
215                         xfs_trans_cancel(tp, 0);
216                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
217                         return error;
218                 }
219
220                 xfs_ilock(ip, XFS_ILOCK_EXCL);
221                 xfs_trans_ijoin(tp, ip);
222
223                 error = xfs_itruncate_data(&tp, ip, ip->i_size);
224                 if (error) {
225                         /*
226                          * If we get an error at this point we simply don't
227                          * bother truncating the file.
228                          */
229                         xfs_trans_cancel(tp,
230                                          (XFS_TRANS_RELEASE_LOG_RES |
231                                           XFS_TRANS_ABORT));
232                 } else {
233                         error = xfs_trans_commit(tp,
234                                                 XFS_TRANS_RELEASE_LOG_RES);
235                 }
236                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
237         }
238         return error;
239 }
240
241 /*
242  * Free a symlink that has blocks associated with it.
243  */
244 STATIC int
245 xfs_inactive_symlink_rmt(
246         xfs_inode_t     *ip,
247         xfs_trans_t     **tpp)
248 {
249         xfs_buf_t       *bp;
250         int             committed;
251         int             done;
252         int             error;
253         xfs_fsblock_t   first_block;
254         xfs_bmap_free_t free_list;
255         int             i;
256         xfs_mount_t     *mp;
257         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
258         int             nmaps;
259         xfs_trans_t     *ntp;
260         int             size;
261         xfs_trans_t     *tp;
262
263         tp = *tpp;
264         mp = ip->i_mount;
265         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
266         /*
267          * We're freeing a symlink that has some
268          * blocks allocated to it.  Free the
269          * blocks here.  We know that we've got
270          * either 1 or 2 extents and that we can
271          * free them all in one bunmapi call.
272          */
273         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
274         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
275                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
276                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
277                 xfs_trans_cancel(tp, 0);
278                 *tpp = NULL;
279                 return error;
280         }
281         /*
282          * Lock the inode, fix the size, and join it to the transaction.
283          * Hold it so in the normal path, we still have it locked for
284          * the second transaction.  In the error paths we need it
285          * held so the cancel won't rele it, see below.
286          */
287         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
288         size = (int)ip->i_d.di_size;
289         ip->i_d.di_size = 0;
290         xfs_trans_ijoin(tp, ip);
291         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
292         /*
293          * Find the block(s) so we can inval and unmap them.
294          */
295         done = 0;
296         xfs_bmap_init(&free_list, &first_block);
297         nmaps = ARRAY_SIZE(mval);
298         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
299                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
300                         &free_list)))
301                 goto error0;
302         /*
303          * Invalidate the block(s).
304          */
305         for (i = 0; i < nmaps; i++) {
306                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
307                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
308                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
309                 xfs_trans_binval(tp, bp);
310         }
311         /*
312          * Unmap the dead block(s) to the free_list.
313          */
314         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
315                         &first_block, &free_list, &done)))
316                 goto error1;
317         ASSERT(done);
318         /*
319          * Commit the first transaction.  This logs the EFI and the inode.
320          */
321         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
322                 goto error1;
323         /*
324          * The transaction must have been committed, since there were
325          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
326          * The new tp has the extent freeing and EFDs.
327          */
328         ASSERT(committed);
329         /*
330          * The first xact was committed, so add the inode to the new one.
331          * Mark it dirty so it will be logged and moved forward in the log as
332          * part of every commit.
333          */
334         xfs_trans_ijoin(tp, ip);
335         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
336         /*
337          * Get a new, empty transaction to return to our caller.
338          */
339         ntp = xfs_trans_dup(tp);
340         /*
341          * Commit the transaction containing extent freeing and EFDs.
342          * If we get an error on the commit here or on the reserve below,
343          * we need to unlock the inode since the new transaction doesn't
344          * have the inode attached.
345          */
346         error = xfs_trans_commit(tp, 0);
347         tp = ntp;
348         if (error) {
349                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
350                 goto error0;
351         }
352         /*
353          * transaction commit worked ok so we can drop the extra ticket
354          * reference that we gained in xfs_trans_dup()
355          */
356         xfs_log_ticket_put(tp->t_ticket);
357
358         /*
359          * Remove the memory for extent descriptions (just bookkeeping).
360          */
361         if (ip->i_df.if_bytes)
362                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
363         ASSERT(ip->i_df.if_bytes == 0);
364         /*
365          * Put an itruncate log reservation in the new transaction
366          * for our caller.
367          */
368         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
369                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
370                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
371                 goto error0;
372         }
373         /*
374          * Return with the inode locked but not joined to the transaction.
375          */
376         *tpp = tp;
377         return 0;
378
379  error1:
380         xfs_bmap_cancel(&free_list);
381  error0:
382         /*
383          * Have to come here with the inode locked and either
384          * (held and in the transaction) or (not in the transaction).
385          * If the inode isn't held then cancel would iput it, but
386          * that's wrong since this is inactive and the vnode ref
387          * count is 0 already.
388          * Cancel won't do anything to the inode if held, but it still
389          * needs to be locked until the cancel is done, if it was
390          * joined to the transaction.
391          */
392         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
393         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
394         *tpp = NULL;
395         return error;
396
397 }
398
399 STATIC int
400 xfs_inactive_symlink_local(
401         xfs_inode_t     *ip,
402         xfs_trans_t     **tpp)
403 {
404         int             error;
405
406         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
407         /*
408          * We're freeing a symlink which fit into
409          * the inode.  Just free the memory used
410          * to hold the old symlink.
411          */
412         error = xfs_trans_reserve(*tpp, 0,
413                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
414                                   0, XFS_TRANS_PERM_LOG_RES,
415                                   XFS_ITRUNCATE_LOG_COUNT);
416
417         if (error) {
418                 xfs_trans_cancel(*tpp, 0);
419                 *tpp = NULL;
420                 return error;
421         }
422         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
423
424         /*
425          * Zero length symlinks _can_ exist.
426          */
427         if (ip->i_df.if_bytes > 0) {
428                 xfs_idata_realloc(ip,
429                                   -(ip->i_df.if_bytes),
430                                   XFS_DATA_FORK);
431                 ASSERT(ip->i_df.if_bytes == 0);
432         }
433         return 0;
434 }
435
436 STATIC int
437 xfs_inactive_attrs(
438         xfs_inode_t     *ip,
439         xfs_trans_t     **tpp)
440 {
441         xfs_trans_t     *tp;
442         int             error;
443         xfs_mount_t     *mp;
444
445         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
446         tp = *tpp;
447         mp = ip->i_mount;
448         ASSERT(ip->i_d.di_forkoff != 0);
449         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
450         xfs_iunlock(ip, XFS_ILOCK_EXCL);
451         if (error)
452                 goto error_unlock;
453
454         error = xfs_attr_inactive(ip);
455         if (error)
456                 goto error_unlock;
457
458         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
459         error = xfs_trans_reserve(tp, 0,
460                                   XFS_IFREE_LOG_RES(mp),
461                                   0, XFS_TRANS_PERM_LOG_RES,
462                                   XFS_INACTIVE_LOG_COUNT);
463         if (error)
464                 goto error_cancel;
465
466         xfs_ilock(ip, XFS_ILOCK_EXCL);
467         xfs_trans_ijoin(tp, ip);
468         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
469
470         ASSERT(ip->i_d.di_anextents == 0);
471
472         *tpp = tp;
473         return 0;
474
475 error_cancel:
476         ASSERT(XFS_FORCED_SHUTDOWN(mp));
477         xfs_trans_cancel(tp, 0);
478 error_unlock:
479         *tpp = NULL;
480         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
481         return error;
482 }
483
484 int
485 xfs_release(
486         xfs_inode_t     *ip)
487 {
488         xfs_mount_t     *mp = ip->i_mount;
489         int             error;
490
491         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
492                 return 0;
493
494         /* If this is a read-only mount, don't do this (would generate I/O) */
495         if (mp->m_flags & XFS_MOUNT_RDONLY)
496                 return 0;
497
498         if (!XFS_FORCED_SHUTDOWN(mp)) {
499                 int truncated;
500
501                 /*
502                  * If we are using filestreams, and we have an unlinked
503                  * file that we are processing the last close on, then nothing
504                  * will be able to reopen and write to this file. Purge this
505                  * inode from the filestreams cache so that it doesn't delay
506                  * teardown of the inode.
507                  */
508                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
509                         xfs_filestream_deassociate(ip);
510
511                 /*
512                  * If we previously truncated this file and removed old data
513                  * in the process, we want to initiate "early" writeout on
514                  * the last close.  This is an attempt to combat the notorious
515                  * NULL files problem which is particularly noticeable from a
516                  * truncate down, buffered (re-)write (delalloc), followed by
517                  * a crash.  What we are effectively doing here is
518                  * significantly reducing the time window where we'd otherwise
519                  * be exposed to that problem.
520                  */
521                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
522                 if (truncated) {
523                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
524                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
525                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
526                 }
527         }
528
529         if (ip->i_d.di_nlink == 0)
530                 return 0;
531
532         if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
533              ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
534                ip->i_delayed_blks > 0)) &&
535              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
536             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
537
538                 /*
539                  * If we can't get the iolock just skip truncating the blocks
540                  * past EOF because we could deadlock with the mmap_sem
541                  * otherwise.  We'll get another chance to drop them once the
542                  * last reference to the inode is dropped, so we'll never leak
543                  * blocks permanently.
544                  *
545                  * Further, check if the inode is being opened, written and
546                  * closed frequently and we have delayed allocation blocks
547                  * outstanding (e.g. streaming writes from the NFS server),
548                  * truncating the blocks past EOF will cause fragmentation to
549                  * occur.
550                  *
551                  * In this case don't do the truncation, either, but we have to
552                  * be careful how we detect this case. Blocks beyond EOF show
553                  * up as i_delayed_blks even when the inode is clean, so we
554                  * need to truncate them away first before checking for a dirty
555                  * release. Hence on the first dirty close we will still remove
556                  * the speculative allocation, but after that we will leave it
557                  * in place.
558                  */
559                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
560                         return 0;
561
562                 error = xfs_free_eofblocks(mp, ip,
563                                            XFS_FREE_EOF_TRYLOCK);
564                 if (error)
565                         return error;
566
567                 /* delalloc blocks after truncation means it really is dirty */
568                 if (ip->i_delayed_blks)
569                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
570         }
571         return 0;
572 }
573
574 /*
575  * xfs_inactive
576  *
577  * This is called when the vnode reference count for the vnode
578  * goes to zero.  If the file has been unlinked, then it must
579  * now be truncated.  Also, we clear all of the read-ahead state
580  * kept for the inode here since the file is now closed.
581  */
582 int
583 xfs_inactive(
584         xfs_inode_t     *ip)
585 {
586         xfs_bmap_free_t free_list;
587         xfs_fsblock_t   first_block;
588         int             committed;
589         xfs_trans_t     *tp;
590         xfs_mount_t     *mp;
591         int             error;
592         int             truncate;
593
594         /*
595          * If the inode is already free, then there can be nothing
596          * to clean up here.
597          */
598         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
599                 ASSERT(ip->i_df.if_real_bytes == 0);
600                 ASSERT(ip->i_df.if_broot_bytes == 0);
601                 return VN_INACTIVE_CACHE;
602         }
603
604         /*
605          * Only do a truncate if it's a regular file with
606          * some actual space in it.  It's OK to look at the
607          * inode's fields without the lock because we're the
608          * only one with a reference to the inode.
609          */
610         truncate = ((ip->i_d.di_nlink == 0) &&
611             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
612              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
613             ((ip->i_d.di_mode & S_IFMT) == S_IFREG));
614
615         mp = ip->i_mount;
616
617         error = 0;
618
619         /* If this is a read-only mount, don't do this (would generate I/O) */
620         if (mp->m_flags & XFS_MOUNT_RDONLY)
621                 goto out;
622
623         if (ip->i_d.di_nlink != 0) {
624                 if ((((ip->i_d.di_mode & S_IFMT) == S_IFREG) &&
625                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
626                        ip->i_delayed_blks > 0)) &&
627                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
628                      (!(ip->i_d.di_flags &
629                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
630                       (ip->i_delayed_blks != 0)))) {
631                         error = xfs_free_eofblocks(mp, ip, 0);
632                         if (error)
633                                 return VN_INACTIVE_CACHE;
634                 }
635                 goto out;
636         }
637
638         ASSERT(ip->i_d.di_nlink == 0);
639
640         error = xfs_qm_dqattach(ip, 0);
641         if (error)
642                 return VN_INACTIVE_CACHE;
643
644         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
645         if (truncate) {
646                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
647
648                 xfs_ioend_wait(ip);
649
650                 error = xfs_trans_reserve(tp, 0,
651                                           XFS_ITRUNCATE_LOG_RES(mp),
652                                           0, XFS_TRANS_PERM_LOG_RES,
653                                           XFS_ITRUNCATE_LOG_COUNT);
654                 if (error) {
655                         /* Don't call itruncate_cleanup */
656                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
657                         xfs_trans_cancel(tp, 0);
658                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
659                         return VN_INACTIVE_CACHE;
660                 }
661
662                 xfs_ilock(ip, XFS_ILOCK_EXCL);
663                 xfs_trans_ijoin(tp, ip);
664
665                 error = xfs_itruncate_data(&tp, ip, 0);
666                 if (error) {
667                         xfs_trans_cancel(tp,
668                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
669                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
670                         return VN_INACTIVE_CACHE;
671                 }
672         } else if ((ip->i_d.di_mode & S_IFMT) == S_IFLNK) {
673
674                 /*
675                  * If we get an error while cleaning up a
676                  * symlink we bail out.
677                  */
678                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
679                         xfs_inactive_symlink_rmt(ip, &tp) :
680                         xfs_inactive_symlink_local(ip, &tp);
681
682                 if (error) {
683                         ASSERT(tp == NULL);
684                         return VN_INACTIVE_CACHE;
685                 }
686
687                 xfs_trans_ijoin(tp, ip);
688         } else {
689                 error = xfs_trans_reserve(tp, 0,
690                                           XFS_IFREE_LOG_RES(mp),
691                                           0, XFS_TRANS_PERM_LOG_RES,
692                                           XFS_INACTIVE_LOG_COUNT);
693                 if (error) {
694                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
695                         xfs_trans_cancel(tp, 0);
696                         return VN_INACTIVE_CACHE;
697                 }
698
699                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
700                 xfs_trans_ijoin(tp, ip);
701         }
702
703         /*
704          * If there are attributes associated with the file
705          * then blow them away now.  The code calls a routine
706          * that recursively deconstructs the attribute fork.
707          * We need to just commit the current transaction
708          * because we can't use it for xfs_attr_inactive().
709          */
710         if (ip->i_d.di_anextents > 0) {
711                 error = xfs_inactive_attrs(ip, &tp);
712                 /*
713                  * If we got an error, the transaction is already
714                  * cancelled, and the inode is unlocked. Just get out.
715                  */
716                  if (error)
717                          return VN_INACTIVE_CACHE;
718         } else if (ip->i_afp) {
719                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
720         }
721
722         /*
723          * Free the inode.
724          */
725         xfs_bmap_init(&free_list, &first_block);
726         error = xfs_ifree(tp, ip, &free_list);
727         if (error) {
728                 /*
729                  * If we fail to free the inode, shut down.  The cancel
730                  * might do that, we need to make sure.  Otherwise the
731                  * inode might be lost for a long time or forever.
732                  */
733                 if (!XFS_FORCED_SHUTDOWN(mp)) {
734                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
735                                 __func__, error);
736                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
737                 }
738                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
739         } else {
740                 /*
741                  * Credit the quota account(s). The inode is gone.
742                  */
743                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
744
745                 /*
746                  * Just ignore errors at this point.  There is nothing we can
747                  * do except to try to keep going. Make sure it's not a silent
748                  * error.
749                  */
750                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
751                 if (error)
752                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
753                                 __func__, error);
754                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
755                 if (error)
756                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
757                                 __func__, error);
758         }
759
760         /*
761          * Release the dquots held by inode, if any.
762          */
763         xfs_qm_dqdetach(ip);
764         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
765
766  out:
767         return VN_INACTIVE_CACHE;
768 }
769
770 /*
771  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
772  * is allowed, otherwise it has to be an exact match. If a CI match is found,
773  * ci_name->name will point to a the actual name (caller must free) or
774  * will be set to NULL if an exact match is found.
775  */
776 int
777 xfs_lookup(
778         xfs_inode_t             *dp,
779         struct xfs_name         *name,
780         xfs_inode_t             **ipp,
781         struct xfs_name         *ci_name)
782 {
783         xfs_ino_t               inum;
784         int                     error;
785         uint                    lock_mode;
786
787         trace_xfs_lookup(dp, name);
788
789         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
790                 return XFS_ERROR(EIO);
791
792         lock_mode = xfs_ilock_map_shared(dp);
793         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
794         xfs_iunlock_map_shared(dp, lock_mode);
795
796         if (error)
797                 goto out;
798
799         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
800         if (error)
801                 goto out_free_name;
802
803         return 0;
804
805 out_free_name:
806         if (ci_name)
807                 kmem_free(ci_name->name);
808 out:
809         *ipp = NULL;
810         return error;
811 }
812
813 int
814 xfs_create(
815         xfs_inode_t             *dp,
816         struct xfs_name         *name,
817         mode_t                  mode,
818         xfs_dev_t               rdev,
819         xfs_inode_t             **ipp)
820 {
821         int                     is_dir = S_ISDIR(mode);
822         struct xfs_mount        *mp = dp->i_mount;
823         struct xfs_inode        *ip = NULL;
824         struct xfs_trans        *tp = NULL;
825         int                     error;
826         xfs_bmap_free_t         free_list;
827         xfs_fsblock_t           first_block;
828         boolean_t               unlock_dp_on_error = B_FALSE;
829         uint                    cancel_flags;
830         int                     committed;
831         prid_t                  prid;
832         struct xfs_dquot        *udqp = NULL;
833         struct xfs_dquot        *gdqp = NULL;
834         uint                    resblks;
835         uint                    log_res;
836         uint                    log_count;
837
838         trace_xfs_create(dp, name);
839
840         if (XFS_FORCED_SHUTDOWN(mp))
841                 return XFS_ERROR(EIO);
842
843         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
844                 prid = xfs_get_projid(dp);
845         else
846                 prid = XFS_PROJID_DEFAULT;
847
848         /*
849          * Make sure that we have allocated dquot(s) on disk.
850          */
851         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
852                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
853         if (error)
854                 return error;
855
856         if (is_dir) {
857                 rdev = 0;
858                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
859                 log_res = XFS_MKDIR_LOG_RES(mp);
860                 log_count = XFS_MKDIR_LOG_COUNT;
861                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
862         } else {
863                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
864                 log_res = XFS_CREATE_LOG_RES(mp);
865                 log_count = XFS_CREATE_LOG_COUNT;
866                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
867         }
868
869         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
870
871         /*
872          * Initially assume that the file does not exist and
873          * reserve the resources for that case.  If that is not
874          * the case we'll drop the one we have and get a more
875          * appropriate transaction later.
876          */
877         error = xfs_trans_reserve(tp, resblks, log_res, 0,
878                         XFS_TRANS_PERM_LOG_RES, log_count);
879         if (error == ENOSPC) {
880                 /* flush outstanding delalloc blocks and retry */
881                 xfs_flush_inodes(dp);
882                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
883                                 XFS_TRANS_PERM_LOG_RES, log_count);
884         }
885         if (error == ENOSPC) {
886                 /* No space at all so try a "no-allocation" reservation */
887                 resblks = 0;
888                 error = xfs_trans_reserve(tp, 0, log_res, 0,
889                                 XFS_TRANS_PERM_LOG_RES, log_count);
890         }
891         if (error) {
892                 cancel_flags = 0;
893                 goto out_trans_cancel;
894         }
895
896         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
897         unlock_dp_on_error = B_TRUE;
898
899         /*
900          * Check for directory link count overflow.
901          */
902         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
903                 error = XFS_ERROR(EMLINK);
904                 goto out_trans_cancel;
905         }
906
907         xfs_bmap_init(&free_list, &first_block);
908
909         /*
910          * Reserve disk quota and the inode.
911          */
912         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
913         if (error)
914                 goto out_trans_cancel;
915
916         error = xfs_dir_canenter(tp, dp, name, resblks);
917         if (error)
918                 goto out_trans_cancel;
919
920         /*
921          * A newly created regular or special file just has one directory
922          * entry pointing to them, but a directory also the "." entry
923          * pointing to itself.
924          */
925         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
926                                prid, resblks > 0, &ip, &committed);
927         if (error) {
928                 if (error == ENOSPC)
929                         goto out_trans_cancel;
930                 goto out_trans_abort;
931         }
932
933         /*
934          * Now we join the directory inode to the transaction.  We do not do it
935          * earlier because xfs_dir_ialloc might commit the previous transaction
936          * (and release all the locks).  An error from here on will result in
937          * the transaction cancel unlocking dp so don't do it explicitly in the
938          * error path.
939          */
940         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
941         unlock_dp_on_error = B_FALSE;
942
943         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
944                                         &first_block, &free_list, resblks ?
945                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
946         if (error) {
947                 ASSERT(error != ENOSPC);
948                 goto out_trans_abort;
949         }
950         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
951         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
952
953         if (is_dir) {
954                 error = xfs_dir_init(tp, ip, dp);
955                 if (error)
956                         goto out_bmap_cancel;
957
958                 error = xfs_bumplink(tp, dp);
959                 if (error)
960                         goto out_bmap_cancel;
961         }
962
963         /*
964          * If this is a synchronous mount, make sure that the
965          * create transaction goes to disk before returning to
966          * the user.
967          */
968         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
969                 xfs_trans_set_sync(tp);
970
971         /*
972          * Attach the dquot(s) to the inodes and modify them incore.
973          * These ids of the inode couldn't have changed since the new
974          * inode has been locked ever since it was created.
975          */
976         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
977
978         error = xfs_bmap_finish(&tp, &free_list, &committed);
979         if (error)
980                 goto out_bmap_cancel;
981
982         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
983         if (error)
984                 goto out_release_inode;
985
986         xfs_qm_dqrele(udqp);
987         xfs_qm_dqrele(gdqp);
988
989         *ipp = ip;
990         return 0;
991
992  out_bmap_cancel:
993         xfs_bmap_cancel(&free_list);
994  out_trans_abort:
995         cancel_flags |= XFS_TRANS_ABORT;
996  out_trans_cancel:
997         xfs_trans_cancel(tp, cancel_flags);
998  out_release_inode:
999         /*
1000          * Wait until after the current transaction is aborted to
1001          * release the inode.  This prevents recursive transactions
1002          * and deadlocks from xfs_inactive.
1003          */
1004         if (ip)
1005                 IRELE(ip);
1006
1007         xfs_qm_dqrele(udqp);
1008         xfs_qm_dqrele(gdqp);
1009
1010         if (unlock_dp_on_error)
1011                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1012         return error;
1013 }
1014
1015 #ifdef DEBUG
1016 int xfs_locked_n;
1017 int xfs_small_retries;
1018 int xfs_middle_retries;
1019 int xfs_lots_retries;
1020 int xfs_lock_delays;
1021 #endif
1022
1023 /*
1024  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1025  * a different value
1026  */
1027 static inline int
1028 xfs_lock_inumorder(int lock_mode, int subclass)
1029 {
1030         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1031                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1032         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1033                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1034
1035         return lock_mode;
1036 }
1037
1038 /*
1039  * The following routine will lock n inodes in exclusive mode.
1040  * We assume the caller calls us with the inodes in i_ino order.
1041  *
1042  * We need to detect deadlock where an inode that we lock
1043  * is in the AIL and we start waiting for another inode that is locked
1044  * by a thread in a long running transaction (such as truncate). This can
1045  * result in deadlock since the long running trans might need to wait
1046  * for the inode we just locked in order to push the tail and free space
1047  * in the log.
1048  */
1049 void
1050 xfs_lock_inodes(
1051         xfs_inode_t     **ips,
1052         int             inodes,
1053         uint            lock_mode)
1054 {
1055         int             attempts = 0, i, j, try_lock;
1056         xfs_log_item_t  *lp;
1057
1058         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1059
1060         try_lock = 0;
1061         i = 0;
1062
1063 again:
1064         for (; i < inodes; i++) {
1065                 ASSERT(ips[i]);
1066
1067                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1068                         continue;
1069
1070                 /*
1071                  * If try_lock is not set yet, make sure all locked inodes
1072                  * are not in the AIL.
1073                  * If any are, set try_lock to be used later.
1074                  */
1075
1076                 if (!try_lock) {
1077                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1078                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1079                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1080                                         try_lock++;
1081                                 }
1082                         }
1083                 }
1084
1085                 /*
1086                  * If any of the previous locks we have locked is in the AIL,
1087                  * we must TRY to get the second and subsequent locks. If
1088                  * we can't get any, we must release all we have
1089                  * and try again.
1090                  */
1091
1092                 if (try_lock) {
1093                         /* try_lock must be 0 if i is 0. */
1094                         /*
1095                          * try_lock means we have an inode locked
1096                          * that is in the AIL.
1097                          */
1098                         ASSERT(i != 0);
1099                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1100                                 attempts++;
1101
1102                                 /*
1103                                  * Unlock all previous guys and try again.
1104                                  * xfs_iunlock will try to push the tail
1105                                  * if the inode is in the AIL.
1106                                  */
1107
1108                                 for(j = i - 1; j >= 0; j--) {
1109
1110                                         /*
1111                                          * Check to see if we've already
1112                                          * unlocked this one.
1113                                          * Not the first one going back,
1114                                          * and the inode ptr is the same.
1115                                          */
1116                                         if ((j != (i - 1)) && ips[j] ==
1117                                                                 ips[j+1])
1118                                                 continue;
1119
1120                                         xfs_iunlock(ips[j], lock_mode);
1121                                 }
1122
1123                                 if ((attempts % 5) == 0) {
1124                                         delay(1); /* Don't just spin the CPU */
1125 #ifdef DEBUG
1126                                         xfs_lock_delays++;
1127 #endif
1128                                 }
1129                                 i = 0;
1130                                 try_lock = 0;
1131                                 goto again;
1132                         }
1133                 } else {
1134                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1135                 }
1136         }
1137
1138 #ifdef DEBUG
1139         if (attempts) {
1140                 if (attempts < 5) xfs_small_retries++;
1141                 else if (attempts < 100) xfs_middle_retries++;
1142                 else xfs_lots_retries++;
1143         } else {
1144                 xfs_locked_n++;
1145         }
1146 #endif
1147 }
1148
1149 /*
1150  * xfs_lock_two_inodes() can only be used to lock one type of lock
1151  * at a time - the iolock or the ilock, but not both at once. If
1152  * we lock both at once, lockdep will report false positives saying
1153  * we have violated locking orders.
1154  */
1155 void
1156 xfs_lock_two_inodes(
1157         xfs_inode_t             *ip0,
1158         xfs_inode_t             *ip1,
1159         uint                    lock_mode)
1160 {
1161         xfs_inode_t             *temp;
1162         int                     attempts = 0;
1163         xfs_log_item_t          *lp;
1164
1165         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1166                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1167         ASSERT(ip0->i_ino != ip1->i_ino);
1168
1169         if (ip0->i_ino > ip1->i_ino) {
1170                 temp = ip0;
1171                 ip0 = ip1;
1172                 ip1 = temp;
1173         }
1174
1175  again:
1176         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1177
1178         /*
1179          * If the first lock we have locked is in the AIL, we must TRY to get
1180          * the second lock. If we can't get it, we must release the first one
1181          * and try again.
1182          */
1183         lp = (xfs_log_item_t *)ip0->i_itemp;
1184         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1185                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1186                         xfs_iunlock(ip0, lock_mode);
1187                         if ((++attempts % 5) == 0)
1188                                 delay(1); /* Don't just spin the CPU */
1189                         goto again;
1190                 }
1191         } else {
1192                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1193         }
1194 }
1195
1196 int
1197 xfs_remove(
1198         xfs_inode_t             *dp,
1199         struct xfs_name         *name,
1200         xfs_inode_t             *ip)
1201 {
1202         xfs_mount_t             *mp = dp->i_mount;
1203         xfs_trans_t             *tp = NULL;
1204         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1205         int                     error = 0;
1206         xfs_bmap_free_t         free_list;
1207         xfs_fsblock_t           first_block;
1208         int                     cancel_flags;
1209         int                     committed;
1210         int                     link_zero;
1211         uint                    resblks;
1212         uint                    log_count;
1213
1214         trace_xfs_remove(dp, name);
1215
1216         if (XFS_FORCED_SHUTDOWN(mp))
1217                 return XFS_ERROR(EIO);
1218
1219         error = xfs_qm_dqattach(dp, 0);
1220         if (error)
1221                 goto std_return;
1222
1223         error = xfs_qm_dqattach(ip, 0);
1224         if (error)
1225                 goto std_return;
1226
1227         if (is_dir) {
1228                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1229                 log_count = XFS_DEFAULT_LOG_COUNT;
1230         } else {
1231                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1232                 log_count = XFS_REMOVE_LOG_COUNT;
1233         }
1234         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1235
1236         /*
1237          * We try to get the real space reservation first,
1238          * allowing for directory btree deletion(s) implying
1239          * possible bmap insert(s).  If we can't get the space
1240          * reservation then we use 0 instead, and avoid the bmap
1241          * btree insert(s) in the directory code by, if the bmap
1242          * insert tries to happen, instead trimming the LAST
1243          * block from the directory.
1244          */
1245         resblks = XFS_REMOVE_SPACE_RES(mp);
1246         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1247                                   XFS_TRANS_PERM_LOG_RES, log_count);
1248         if (error == ENOSPC) {
1249                 resblks = 0;
1250                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1251                                           XFS_TRANS_PERM_LOG_RES, log_count);
1252         }
1253         if (error) {
1254                 ASSERT(error != ENOSPC);
1255                 cancel_flags = 0;
1256                 goto out_trans_cancel;
1257         }
1258
1259         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1260
1261         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1262         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1263
1264         /*
1265          * If we're removing a directory perform some additional validation.
1266          */
1267         if (is_dir) {
1268                 ASSERT(ip->i_d.di_nlink >= 2);
1269                 if (ip->i_d.di_nlink != 2) {
1270                         error = XFS_ERROR(ENOTEMPTY);
1271                         goto out_trans_cancel;
1272                 }
1273                 if (!xfs_dir_isempty(ip)) {
1274                         error = XFS_ERROR(ENOTEMPTY);
1275                         goto out_trans_cancel;
1276                 }
1277         }
1278
1279         xfs_bmap_init(&free_list, &first_block);
1280         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1281                                         &first_block, &free_list, resblks);
1282         if (error) {
1283                 ASSERT(error != ENOENT);
1284                 goto out_bmap_cancel;
1285         }
1286         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1287
1288         if (is_dir) {
1289                 /*
1290                  * Drop the link from ip's "..".
1291                  */
1292                 error = xfs_droplink(tp, dp);
1293                 if (error)
1294                         goto out_bmap_cancel;
1295
1296                 /*
1297                  * Drop the "." link from ip to self.
1298                  */
1299                 error = xfs_droplink(tp, ip);
1300                 if (error)
1301                         goto out_bmap_cancel;
1302         } else {
1303                 /*
1304                  * When removing a non-directory we need to log the parent
1305                  * inode here.  For a directory this is done implicitly
1306                  * by the xfs_droplink call for the ".." entry.
1307                  */
1308                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1309         }
1310
1311         /*
1312          * Drop the link from dp to ip.
1313          */
1314         error = xfs_droplink(tp, ip);
1315         if (error)
1316                 goto out_bmap_cancel;
1317
1318         /*
1319          * Determine if this is the last link while
1320          * we are in the transaction.
1321          */
1322         link_zero = (ip->i_d.di_nlink == 0);
1323
1324         /*
1325          * If this is a synchronous mount, make sure that the
1326          * remove transaction goes to disk before returning to
1327          * the user.
1328          */
1329         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1330                 xfs_trans_set_sync(tp);
1331
1332         error = xfs_bmap_finish(&tp, &free_list, &committed);
1333         if (error)
1334                 goto out_bmap_cancel;
1335
1336         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1337         if (error)
1338                 goto std_return;
1339
1340         /*
1341          * If we are using filestreams, kill the stream association.
1342          * If the file is still open it may get a new one but that
1343          * will get killed on last close in xfs_close() so we don't
1344          * have to worry about that.
1345          */
1346         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1347                 xfs_filestream_deassociate(ip);
1348
1349         return 0;
1350
1351  out_bmap_cancel:
1352         xfs_bmap_cancel(&free_list);
1353         cancel_flags |= XFS_TRANS_ABORT;
1354  out_trans_cancel:
1355         xfs_trans_cancel(tp, cancel_flags);
1356  std_return:
1357         return error;
1358 }
1359
1360 int
1361 xfs_link(
1362         xfs_inode_t             *tdp,
1363         xfs_inode_t             *sip,
1364         struct xfs_name         *target_name)
1365 {
1366         xfs_mount_t             *mp = tdp->i_mount;
1367         xfs_trans_t             *tp;
1368         int                     error;
1369         xfs_bmap_free_t         free_list;
1370         xfs_fsblock_t           first_block;
1371         int                     cancel_flags;
1372         int                     committed;
1373         int                     resblks;
1374
1375         trace_xfs_link(tdp, target_name);
1376
1377         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1378
1379         if (XFS_FORCED_SHUTDOWN(mp))
1380                 return XFS_ERROR(EIO);
1381
1382         error = xfs_qm_dqattach(sip, 0);
1383         if (error)
1384                 goto std_return;
1385
1386         error = xfs_qm_dqattach(tdp, 0);
1387         if (error)
1388                 goto std_return;
1389
1390         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1391         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1392         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1393         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1394                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1395         if (error == ENOSPC) {
1396                 resblks = 0;
1397                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1398                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1399         }
1400         if (error) {
1401                 cancel_flags = 0;
1402                 goto error_return;
1403         }
1404
1405         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1406
1407         xfs_trans_ijoin_ref(tp, sip, XFS_ILOCK_EXCL);
1408         xfs_trans_ijoin_ref(tp, tdp, XFS_ILOCK_EXCL);
1409
1410         /*
1411          * If the source has too many links, we can't make any more to it.
1412          */
1413         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1414                 error = XFS_ERROR(EMLINK);
1415                 goto error_return;
1416         }
1417
1418         /*
1419          * If we are using project inheritance, we only allow hard link
1420          * creation in our tree when the project IDs are the same; else
1421          * the tree quota mechanism could be circumvented.
1422          */
1423         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1424                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1425                 error = XFS_ERROR(EXDEV);
1426                 goto error_return;
1427         }
1428
1429         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1430         if (error)
1431                 goto error_return;
1432
1433         xfs_bmap_init(&free_list, &first_block);
1434
1435         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1436                                         &first_block, &free_list, resblks);
1437         if (error)
1438                 goto abort_return;
1439         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1440         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1441
1442         error = xfs_bumplink(tp, sip);
1443         if (error)
1444                 goto abort_return;
1445
1446         /*
1447          * If this is a synchronous mount, make sure that the
1448          * link transaction goes to disk before returning to
1449          * the user.
1450          */
1451         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1452                 xfs_trans_set_sync(tp);
1453         }
1454
1455         error = xfs_bmap_finish (&tp, &free_list, &committed);
1456         if (error) {
1457                 xfs_bmap_cancel(&free_list);
1458                 goto abort_return;
1459         }
1460
1461         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1462
1463  abort_return:
1464         cancel_flags |= XFS_TRANS_ABORT;
1465  error_return:
1466         xfs_trans_cancel(tp, cancel_flags);
1467  std_return:
1468         return error;
1469 }
1470
1471 int
1472 xfs_symlink(
1473         xfs_inode_t             *dp,
1474         struct xfs_name         *link_name,
1475         const char              *target_path,
1476         mode_t                  mode,
1477         xfs_inode_t             **ipp)
1478 {
1479         xfs_mount_t             *mp = dp->i_mount;
1480         xfs_trans_t             *tp;
1481         xfs_inode_t             *ip;
1482         int                     error;
1483         int                     pathlen;
1484         xfs_bmap_free_t         free_list;
1485         xfs_fsblock_t           first_block;
1486         boolean_t               unlock_dp_on_error = B_FALSE;
1487         uint                    cancel_flags;
1488         int                     committed;
1489         xfs_fileoff_t           first_fsb;
1490         xfs_filblks_t           fs_blocks;
1491         int                     nmaps;
1492         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1493         xfs_daddr_t             d;
1494         const char              *cur_chunk;
1495         int                     byte_cnt;
1496         int                     n;
1497         xfs_buf_t               *bp;
1498         prid_t                  prid;
1499         struct xfs_dquot        *udqp, *gdqp;
1500         uint                    resblks;
1501
1502         *ipp = NULL;
1503         error = 0;
1504         ip = NULL;
1505         tp = NULL;
1506
1507         trace_xfs_symlink(dp, link_name);
1508
1509         if (XFS_FORCED_SHUTDOWN(mp))
1510                 return XFS_ERROR(EIO);
1511
1512         /*
1513          * Check component lengths of the target path name.
1514          */
1515         pathlen = strlen(target_path);
1516         if (pathlen >= MAXPATHLEN)      /* total string too long */
1517                 return XFS_ERROR(ENAMETOOLONG);
1518
1519         udqp = gdqp = NULL;
1520         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1521                 prid = xfs_get_projid(dp);
1522         else
1523                 prid = XFS_PROJID_DEFAULT;
1524
1525         /*
1526          * Make sure that we have allocated dquot(s) on disk.
1527          */
1528         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1529                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1530         if (error)
1531                 goto std_return;
1532
1533         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1534         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1535         /*
1536          * The symlink will fit into the inode data fork?
1537          * There can't be any attributes so we get the whole variable part.
1538          */
1539         if (pathlen <= XFS_LITINO(mp))
1540                 fs_blocks = 0;
1541         else
1542                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
1543         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
1544         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
1545                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1546         if (error == ENOSPC && fs_blocks == 0) {
1547                 resblks = 0;
1548                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
1549                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1550         }
1551         if (error) {
1552                 cancel_flags = 0;
1553                 goto error_return;
1554         }
1555
1556         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1557         unlock_dp_on_error = B_TRUE;
1558
1559         /*
1560          * Check whether the directory allows new symlinks or not.
1561          */
1562         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
1563                 error = XFS_ERROR(EPERM);
1564                 goto error_return;
1565         }
1566
1567         /*
1568          * Reserve disk quota : blocks and inode.
1569          */
1570         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1571         if (error)
1572                 goto error_return;
1573
1574         /*
1575          * Check for ability to enter directory entry, if no space reserved.
1576          */
1577         error = xfs_dir_canenter(tp, dp, link_name, resblks);
1578         if (error)
1579                 goto error_return;
1580         /*
1581          * Initialize the bmap freelist prior to calling either
1582          * bmapi or the directory create code.
1583          */
1584         xfs_bmap_init(&free_list, &first_block);
1585
1586         /*
1587          * Allocate an inode for the symlink.
1588          */
1589         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
1590                                prid, resblks > 0, &ip, NULL);
1591         if (error) {
1592                 if (error == ENOSPC)
1593                         goto error_return;
1594                 goto error1;
1595         }
1596
1597         /*
1598          * An error after we've joined dp to the transaction will result in the
1599          * transaction cancel unlocking dp so don't do it explicitly in the
1600          * error path.
1601          */
1602         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1603         unlock_dp_on_error = B_FALSE;
1604
1605         /*
1606          * Also attach the dquot(s) to it, if applicable.
1607          */
1608         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1609
1610         if (resblks)
1611                 resblks -= XFS_IALLOC_SPACE_RES(mp);
1612         /*
1613          * If the symlink will fit into the inode, write it inline.
1614          */
1615         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
1616                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
1617                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
1618                 ip->i_d.di_size = pathlen;
1619
1620                 /*
1621                  * The inode was initially created in extent format.
1622                  */
1623                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
1624                 ip->i_df.if_flags |= XFS_IFINLINE;
1625
1626                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
1627                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
1628
1629         } else {
1630                 first_fsb = 0;
1631                 nmaps = SYMLINK_MAPS;
1632
1633                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
1634                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
1635                                   &first_block, resblks, mval, &nmaps,
1636                                   &free_list);
1637                 if (error)
1638                         goto error2;
1639
1640                 if (resblks)
1641                         resblks -= fs_blocks;
1642                 ip->i_d.di_size = pathlen;
1643                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1644
1645                 cur_chunk = target_path;
1646                 for (n = 0; n < nmaps; n++) {
1647                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
1648                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
1649                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
1650                                                BTOBB(byte_cnt), 0);
1651                         ASSERT(bp && !XFS_BUF_GETERROR(bp));
1652                         if (pathlen < byte_cnt) {
1653                                 byte_cnt = pathlen;
1654                         }
1655                         pathlen -= byte_cnt;
1656
1657                         memcpy(XFS_BUF_PTR(bp), cur_chunk, byte_cnt);
1658                         cur_chunk += byte_cnt;
1659
1660                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
1661                 }
1662         }
1663
1664         /*
1665          * Create the directory entry for the symlink.
1666          */
1667         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
1668                                         &first_block, &free_list, resblks);
1669         if (error)
1670                 goto error2;
1671         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1672         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1673
1674         /*
1675          * If this is a synchronous mount, make sure that the
1676          * symlink transaction goes to disk before returning to
1677          * the user.
1678          */
1679         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1680                 xfs_trans_set_sync(tp);
1681         }
1682
1683         error = xfs_bmap_finish(&tp, &free_list, &committed);
1684         if (error) {
1685                 goto error2;
1686         }
1687         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1688         xfs_qm_dqrele(udqp);
1689         xfs_qm_dqrele(gdqp);
1690
1691         *ipp = ip;
1692         return 0;
1693
1694  error2:
1695         IRELE(ip);
1696  error1:
1697         xfs_bmap_cancel(&free_list);
1698         cancel_flags |= XFS_TRANS_ABORT;
1699  error_return:
1700         xfs_trans_cancel(tp, cancel_flags);
1701         xfs_qm_dqrele(udqp);
1702         xfs_qm_dqrele(gdqp);
1703
1704         if (unlock_dp_on_error)
1705                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1706  std_return:
1707         return error;
1708 }
1709
1710 int
1711 xfs_set_dmattrs(
1712         xfs_inode_t     *ip,
1713         u_int           evmask,
1714         u_int16_t       state)
1715 {
1716         xfs_mount_t     *mp = ip->i_mount;
1717         xfs_trans_t     *tp;
1718         int             error;
1719
1720         if (!capable(CAP_SYS_ADMIN))
1721                 return XFS_ERROR(EPERM);
1722
1723         if (XFS_FORCED_SHUTDOWN(mp))
1724                 return XFS_ERROR(EIO);
1725
1726         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1727         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1728         if (error) {
1729                 xfs_trans_cancel(tp, 0);
1730                 return error;
1731         }
1732         xfs_ilock(ip, XFS_ILOCK_EXCL);
1733         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1734
1735         ip->i_d.di_dmevmask = evmask;
1736         ip->i_d.di_dmstate  = state;
1737
1738         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1739         error = xfs_trans_commit(tp, 0);
1740
1741         return error;
1742 }
1743
1744 /*
1745  * xfs_alloc_file_space()
1746  *      This routine allocates disk space for the given file.
1747  *
1748  *      If alloc_type == 0, this request is for an ALLOCSP type
1749  *      request which will change the file size.  In this case, no
1750  *      DMAPI event will be generated by the call.  A TRUNCATE event
1751  *      will be generated later by xfs_setattr.
1752  *
1753  *      If alloc_type != 0, this request is for a RESVSP type
1754  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1755  *      lower block boundary byte address is less than the file's
1756  *      length.
1757  *
1758  * RETURNS:
1759  *       0 on success
1760  *      errno on error
1761  *
1762  */
1763 STATIC int
1764 xfs_alloc_file_space(
1765         xfs_inode_t             *ip,
1766         xfs_off_t               offset,
1767         xfs_off_t               len,
1768         int                     alloc_type,
1769         int                     attr_flags)
1770 {
1771         xfs_mount_t             *mp = ip->i_mount;
1772         xfs_off_t               count;
1773         xfs_filblks_t           allocated_fsb;
1774         xfs_filblks_t           allocatesize_fsb;
1775         xfs_extlen_t            extsz, temp;
1776         xfs_fileoff_t           startoffset_fsb;
1777         xfs_fsblock_t           firstfsb;
1778         int                     nimaps;
1779         int                     bmapi_flag;
1780         int                     quota_flag;
1781         int                     rt;
1782         xfs_trans_t             *tp;
1783         xfs_bmbt_irec_t         imaps[1], *imapp;
1784         xfs_bmap_free_t         free_list;
1785         uint                    qblocks, resblks, resrtextents;
1786         int                     committed;
1787         int                     error;
1788
1789         trace_xfs_alloc_file_space(ip);
1790
1791         if (XFS_FORCED_SHUTDOWN(mp))
1792                 return XFS_ERROR(EIO);
1793
1794         error = xfs_qm_dqattach(ip, 0);
1795         if (error)
1796                 return error;
1797
1798         if (len <= 0)
1799                 return XFS_ERROR(EINVAL);
1800
1801         rt = XFS_IS_REALTIME_INODE(ip);
1802         extsz = xfs_get_extsz_hint(ip);
1803
1804         count = len;
1805         imapp = &imaps[0];
1806         nimaps = 1;
1807         bmapi_flag = XFS_BMAPI_WRITE | alloc_type;
1808         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1809         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1810
1811         /*
1812          * Allocate file space until done or until there is an error
1813          */
1814         while (allocatesize_fsb && !error) {
1815                 xfs_fileoff_t   s, e;
1816
1817                 /*
1818                  * Determine space reservations for data/realtime.
1819                  */
1820                 if (unlikely(extsz)) {
1821                         s = startoffset_fsb;
1822                         do_div(s, extsz);
1823                         s *= extsz;
1824                         e = startoffset_fsb + allocatesize_fsb;
1825                         if ((temp = do_mod(startoffset_fsb, extsz)))
1826                                 e += temp;
1827                         if ((temp = do_mod(e, extsz)))
1828                                 e += extsz - temp;
1829                 } else {
1830                         s = 0;
1831                         e = allocatesize_fsb;
1832                 }
1833
1834                 /*
1835                  * The transaction reservation is limited to a 32-bit block
1836                  * count, hence we need to limit the number of blocks we are
1837                  * trying to reserve to avoid an overflow. We can't allocate
1838                  * more than @nimaps extents, and an extent is limited on disk
1839                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1840                  */
1841                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1842                 if (unlikely(rt)) {
1843                         resrtextents = qblocks = resblks;
1844                         resrtextents /= mp->m_sb.sb_rextsize;
1845                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1846                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1847                 } else {
1848                         resrtextents = 0;
1849                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1850                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1851                 }
1852
1853                 /*
1854                  * Allocate and setup the transaction.
1855                  */
1856                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1857                 error = xfs_trans_reserve(tp, resblks,
1858                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1859                                           XFS_TRANS_PERM_LOG_RES,
1860                                           XFS_WRITE_LOG_COUNT);
1861                 /*
1862                  * Check for running out of space
1863                  */
1864                 if (error) {
1865                         /*
1866                          * Free the transaction structure.
1867                          */
1868                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1869                         xfs_trans_cancel(tp, 0);
1870                         break;
1871                 }
1872                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1873                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1874                                                       0, quota_flag);
1875                 if (error)
1876                         goto error1;
1877
1878                 xfs_trans_ijoin(tp, ip);
1879
1880                 /*
1881                  * Issue the xfs_bmapi() call to allocate the blocks
1882                  */
1883                 xfs_bmap_init(&free_list, &firstfsb);
1884                 error = xfs_bmapi(tp, ip, startoffset_fsb,
1885                                   allocatesize_fsb, bmapi_flag,
1886                                   &firstfsb, 0, imapp, &nimaps,
1887                                   &free_list);
1888                 if (error) {
1889                         goto error0;
1890                 }
1891
1892                 /*
1893                  * Complete the transaction
1894                  */
1895                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1896                 if (error) {
1897                         goto error0;
1898                 }
1899
1900                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1901                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1902                 if (error) {
1903                         break;
1904                 }
1905
1906                 allocated_fsb = imapp->br_blockcount;
1907
1908                 if (nimaps == 0) {
1909                         error = XFS_ERROR(ENOSPC);
1910                         break;
1911                 }
1912
1913                 startoffset_fsb += allocated_fsb;
1914                 allocatesize_fsb -= allocated_fsb;
1915         }
1916
1917         return error;
1918
1919 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1920         xfs_bmap_cancel(&free_list);
1921         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
1922
1923 error1: /* Just cancel transaction */
1924         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1925         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1926         return error;
1927 }
1928
1929 /*
1930  * Zero file bytes between startoff and endoff inclusive.
1931  * The iolock is held exclusive and no blocks are buffered.
1932  *
1933  * This function is used by xfs_free_file_space() to zero
1934  * partial blocks when the range to free is not block aligned.
1935  * When unreserving space with boundaries that are not block
1936  * aligned we round up the start and round down the end
1937  * boundaries and then use this function to zero the parts of
1938  * the blocks that got dropped during the rounding.
1939  */
1940 STATIC int
1941 xfs_zero_remaining_bytes(
1942         xfs_inode_t             *ip,
1943         xfs_off_t               startoff,
1944         xfs_off_t               endoff)
1945 {
1946         xfs_bmbt_irec_t         imap;
1947         xfs_fileoff_t           offset_fsb;
1948         xfs_off_t               lastoffset;
1949         xfs_off_t               offset;
1950         xfs_buf_t               *bp;
1951         xfs_mount_t             *mp = ip->i_mount;
1952         int                     nimap;
1953         int                     error = 0;
1954
1955         /*
1956          * Avoid doing I/O beyond eof - it's not necessary
1957          * since nothing can read beyond eof.  The space will
1958          * be zeroed when the file is extended anyway.
1959          */
1960         if (startoff >= ip->i_size)
1961                 return 0;
1962
1963         if (endoff > ip->i_size)
1964                 endoff = ip->i_size;
1965
1966         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1967                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1968                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
1969         if (!bp)
1970                 return XFS_ERROR(ENOMEM);
1971
1972         xfs_buf_unlock(bp);
1973
1974         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1975                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1976                 nimap = 1;
1977                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
1978                         NULL, 0, &imap, &nimap, NULL);
1979                 if (error || nimap < 1)
1980                         break;
1981                 ASSERT(imap.br_blockcount >= 1);
1982                 ASSERT(imap.br_startoff == offset_fsb);
1983                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1984                 if (lastoffset > endoff)
1985                         lastoffset = endoff;
1986                 if (imap.br_startblock == HOLESTARTBLOCK)
1987                         continue;
1988                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1989                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1990                         continue;
1991                 XFS_BUF_UNDONE(bp);
1992                 XFS_BUF_UNWRITE(bp);
1993                 XFS_BUF_READ(bp);
1994                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1995                 xfsbdstrat(mp, bp);
1996                 error = xfs_buf_iowait(bp);
1997                 if (error) {
1998                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
1999                                           mp, bp, XFS_BUF_ADDR(bp));
2000                         break;
2001                 }
2002                 memset(XFS_BUF_PTR(bp) +
2003                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2004                       0, lastoffset - offset + 1);
2005                 XFS_BUF_UNDONE(bp);
2006                 XFS_BUF_UNREAD(bp);
2007                 XFS_BUF_WRITE(bp);
2008                 xfsbdstrat(mp, bp);
2009                 error = xfs_buf_iowait(bp);
2010                 if (error) {
2011                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
2012                                           mp, bp, XFS_BUF_ADDR(bp));
2013                         break;
2014                 }
2015         }
2016         xfs_buf_free(bp);
2017         return error;
2018 }
2019
2020 /*
2021  * xfs_free_file_space()
2022  *      This routine frees disk space for the given file.
2023  *
2024  *      This routine is only called by xfs_change_file_space
2025  *      for an UNRESVSP type call.
2026  *
2027  * RETURNS:
2028  *       0 on success
2029  *      errno on error
2030  *
2031  */
2032 STATIC int
2033 xfs_free_file_space(
2034         xfs_inode_t             *ip,
2035         xfs_off_t               offset,
2036         xfs_off_t               len,
2037         int                     attr_flags)
2038 {
2039         int                     committed;
2040         int                     done;
2041         xfs_fileoff_t           endoffset_fsb;
2042         int                     error;
2043         xfs_fsblock_t           firstfsb;
2044         xfs_bmap_free_t         free_list;
2045         xfs_bmbt_irec_t         imap;
2046         xfs_off_t               ioffset;
2047         xfs_extlen_t            mod=0;
2048         xfs_mount_t             *mp;
2049         int                     nimap;
2050         uint                    resblks;
2051         uint                    rounding;
2052         int                     rt;
2053         xfs_fileoff_t           startoffset_fsb;
2054         xfs_trans_t             *tp;
2055         int                     need_iolock = 1;
2056
2057         mp = ip->i_mount;
2058
2059         trace_xfs_free_file_space(ip);
2060
2061         error = xfs_qm_dqattach(ip, 0);
2062         if (error)
2063                 return error;
2064
2065         error = 0;
2066         if (len <= 0)   /* if nothing being freed */
2067                 return error;
2068         rt = XFS_IS_REALTIME_INODE(ip);
2069         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2070         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2071
2072         if (attr_flags & XFS_ATTR_NOLOCK)
2073                 need_iolock = 0;
2074         if (need_iolock) {
2075                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2076                 /* wait for the completion of any pending DIOs */
2077                 xfs_ioend_wait(ip);
2078         }
2079
2080         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2081         ioffset = offset & ~(rounding - 1);
2082
2083         if (VN_CACHED(VFS_I(ip)) != 0) {
2084                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2085                 if (error)
2086                         goto out_unlock_iolock;
2087         }
2088
2089         /*
2090          * Need to zero the stuff we're not freeing, on disk.
2091          * If it's a realtime file & can't use unwritten extents then we
2092          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2093          * will take care of it for us.
2094          */
2095         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2096                 nimap = 1;
2097                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
2098                         1, 0, NULL, 0, &imap, &nimap, NULL);
2099                 if (error)
2100                         goto out_unlock_iolock;
2101                 ASSERT(nimap == 0 || nimap == 1);
2102                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2103                         xfs_daddr_t     block;
2104
2105                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2106                         block = imap.br_startblock;
2107                         mod = do_div(block, mp->m_sb.sb_rextsize);
2108                         if (mod)
2109                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2110                 }
2111                 nimap = 1;
2112                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
2113                         1, 0, NULL, 0, &imap, &nimap, NULL);
2114                 if (error)
2115                         goto out_unlock_iolock;
2116                 ASSERT(nimap == 0 || nimap == 1);
2117                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2118                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2119                         mod++;
2120                         if (mod && (mod != mp->m_sb.sb_rextsize))
2121                                 endoffset_fsb -= mod;
2122                 }
2123         }
2124         if ((done = (endoffset_fsb <= startoffset_fsb)))
2125                 /*
2126                  * One contiguous piece to clear
2127                  */
2128                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2129         else {
2130                 /*
2131                  * Some full blocks, possibly two pieces to clear
2132                  */
2133                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2134                         error = xfs_zero_remaining_bytes(ip, offset,
2135                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2136                 if (!error &&
2137                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2138                         error = xfs_zero_remaining_bytes(ip,
2139                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2140                                 offset + len - 1);
2141         }
2142
2143         /*
2144          * free file space until done or until there is an error
2145          */
2146         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2147         while (!error && !done) {
2148
2149                 /*
2150                  * allocate and setup the transaction. Allow this
2151                  * transaction to dip into the reserve blocks to ensure
2152                  * the freeing of the space succeeds at ENOSPC.
2153                  */
2154                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2155                 tp->t_flags |= XFS_TRANS_RESERVE;
2156                 error = xfs_trans_reserve(tp,
2157                                           resblks,
2158                                           XFS_WRITE_LOG_RES(mp),
2159                                           0,
2160                                           XFS_TRANS_PERM_LOG_RES,
2161                                           XFS_WRITE_LOG_COUNT);
2162
2163                 /*
2164                  * check for running out of space
2165                  */
2166                 if (error) {
2167                         /*
2168                          * Free the transaction structure.
2169                          */
2170                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2171                         xfs_trans_cancel(tp, 0);
2172                         break;
2173                 }
2174                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2175                 error = xfs_trans_reserve_quota(tp, mp,
2176                                 ip->i_udquot, ip->i_gdquot,
2177                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2178                 if (error)
2179                         goto error1;
2180
2181                 xfs_trans_ijoin(tp, ip);
2182
2183                 /*
2184                  * issue the bunmapi() call to free the blocks
2185                  */
2186                 xfs_bmap_init(&free_list, &firstfsb);
2187                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2188                                   endoffset_fsb - startoffset_fsb,
2189                                   0, 2, &firstfsb, &free_list, &done);
2190                 if (error) {
2191                         goto error0;
2192                 }
2193
2194                 /*
2195                  * complete the transaction
2196                  */
2197                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2198                 if (error) {
2199                         goto error0;
2200                 }
2201
2202                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2203                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2204         }
2205
2206  out_unlock_iolock:
2207         if (need_iolock)
2208                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2209         return error;
2210
2211  error0:
2212         xfs_bmap_cancel(&free_list);
2213  error1:
2214         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2215         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2216                     XFS_ILOCK_EXCL);
2217         return error;
2218 }
2219
2220 /*
2221  * xfs_change_file_space()
2222  *      This routine allocates or frees disk space for the given file.
2223  *      The user specified parameters are checked for alignment and size
2224  *      limitations.
2225  *
2226  * RETURNS:
2227  *       0 on success
2228  *      errno on error
2229  *
2230  */
2231 int
2232 xfs_change_file_space(
2233         xfs_inode_t     *ip,
2234         int             cmd,
2235         xfs_flock64_t   *bf,
2236         xfs_off_t       offset,
2237         int             attr_flags)
2238 {
2239         xfs_mount_t     *mp = ip->i_mount;
2240         int             clrprealloc;
2241         int             error;
2242         xfs_fsize_t     fsize;
2243         int             setprealloc;
2244         xfs_off_t       startoffset;
2245         xfs_off_t       llen;
2246         xfs_trans_t     *tp;
2247         struct iattr    iattr;
2248         int             prealloc_type;
2249
2250         if (!S_ISREG(ip->i_d.di_mode))
2251                 return XFS_ERROR(EINVAL);
2252
2253         switch (bf->l_whence) {
2254         case 0: /*SEEK_SET*/
2255                 break;
2256         case 1: /*SEEK_CUR*/
2257                 bf->l_start += offset;
2258                 break;
2259         case 2: /*SEEK_END*/
2260                 bf->l_start += ip->i_size;
2261                 break;
2262         default:
2263                 return XFS_ERROR(EINVAL);
2264         }
2265
2266         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2267
2268         if (   (bf->l_start < 0)
2269             || (bf->l_start > XFS_MAXIOFFSET(mp))
2270             || (bf->l_start + llen < 0)
2271             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2272                 return XFS_ERROR(EINVAL);
2273
2274         bf->l_whence = 0;
2275
2276         startoffset = bf->l_start;
2277         fsize = ip->i_size;
2278
2279         /*
2280          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2281          * file space.
2282          * These calls do NOT zero the data space allocated to the file,
2283          * nor do they change the file size.
2284          *
2285          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2286          * space.
2287          * These calls cause the new file data to be zeroed and the file
2288          * size to be changed.
2289          */
2290         setprealloc = clrprealloc = 0;
2291         prealloc_type = XFS_BMAPI_PREALLOC;
2292
2293         switch (cmd) {
2294         case XFS_IOC_ZERO_RANGE:
2295                 prealloc_type |= XFS_BMAPI_CONVERT;
2296                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2297                 /* FALLTHRU */
2298         case XFS_IOC_RESVSP:
2299         case XFS_IOC_RESVSP64:
2300                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2301                                                 prealloc_type, attr_flags);
2302                 if (error)
2303                         return error;
2304                 setprealloc = 1;
2305                 break;
2306
2307         case XFS_IOC_UNRESVSP:
2308         case XFS_IOC_UNRESVSP64:
2309                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2310                                                                 attr_flags)))
2311                         return error;
2312                 break;
2313
2314         case XFS_IOC_ALLOCSP:
2315         case XFS_IOC_ALLOCSP64:
2316         case XFS_IOC_FREESP:
2317         case XFS_IOC_FREESP64:
2318                 if (startoffset > fsize) {
2319                         error = xfs_alloc_file_space(ip, fsize,
2320                                         startoffset - fsize, 0, attr_flags);
2321                         if (error)
2322                                 break;
2323                 }
2324
2325                 iattr.ia_valid = ATTR_SIZE;
2326                 iattr.ia_size = startoffset;
2327
2328                 error = xfs_setattr_size(ip, &iattr, attr_flags);
2329
2330                 if (error)
2331                         return error;
2332
2333                 clrprealloc = 1;
2334                 break;
2335
2336         default:
2337                 ASSERT(0);
2338                 return XFS_ERROR(EINVAL);
2339         }
2340
2341         /*
2342          * update the inode timestamp, mode, and prealloc flag bits
2343          */
2344         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2345
2346         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2347                                       0, 0, 0))) {
2348                 /* ASSERT(0); */
2349                 xfs_trans_cancel(tp, 0);
2350                 return error;
2351         }
2352
2353         xfs_ilock(ip, XFS_ILOCK_EXCL);
2354
2355         xfs_trans_ijoin(tp, ip);
2356
2357         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2358                 ip->i_d.di_mode &= ~S_ISUID;
2359
2360                 /*
2361                  * Note that we don't have to worry about mandatory
2362                  * file locking being disabled here because we only
2363                  * clear the S_ISGID bit if the Group execute bit is
2364                  * on, but if it was on then mandatory locking wouldn't
2365                  * have been enabled.
2366                  */
2367                 if (ip->i_d.di_mode & S_IXGRP)
2368                         ip->i_d.di_mode &= ~S_ISGID;
2369
2370                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2371         }
2372         if (setprealloc)
2373                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2374         else if (clrprealloc)
2375                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2376
2377         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2378         if (attr_flags & XFS_ATTR_SYNC)
2379                 xfs_trans_set_sync(tp);
2380
2381         error = xfs_trans_commit(tp, 0);
2382
2383         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2384
2385         return error;
2386 }