block: make gendisk hold a reference to its queue
[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                 if (!bp)
87                         return XFS_ERROR(ENOMEM);
88                 error = bp->b_error;
89                 if (error) {
90                         xfs_ioerror_alert("xfs_readlink",
91                                   ip->i_mount, bp, XFS_BUF_ADDR(bp));
92                         xfs_buf_relse(bp);
93                         goto out;
94                 }
95                 if (pathlen < byte_cnt)
96                         byte_cnt = pathlen;
97                 pathlen -= byte_cnt;
98
99                 memcpy(link, bp->b_addr, byte_cnt);
100                 xfs_buf_relse(bp);
101         }
102
103         link[ip->i_d.di_size] = '\0';
104         error = 0;
105
106  out:
107         return error;
108 }
109
110 int
111 xfs_readlink(
112         xfs_inode_t     *ip,
113         char            *link)
114 {
115         xfs_mount_t     *mp = ip->i_mount;
116         int             pathlen;
117         int             error = 0;
118
119         trace_xfs_readlink(ip);
120
121         if (XFS_FORCED_SHUTDOWN(mp))
122                 return XFS_ERROR(EIO);
123
124         xfs_ilock(ip, XFS_ILOCK_SHARED);
125
126         ASSERT(S_ISLNK(ip->i_d.di_mode));
127         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
128
129         pathlen = ip->i_d.di_size;
130         if (!pathlen)
131                 goto out;
132
133         if (ip->i_df.if_flags & XFS_IFINLINE) {
134                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
135                 link[pathlen] = '\0';
136         } else {
137                 error = xfs_readlink_bmap(ip, link);
138         }
139
140  out:
141         xfs_iunlock(ip, XFS_ILOCK_SHARED);
142         return error;
143 }
144
145 /*
146  * Flags for xfs_free_eofblocks
147  */
148 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
149
150 /*
151  * This is called by xfs_inactive to free any blocks beyond eof
152  * when the link count isn't zero and by xfs_dm_punch_hole() when
153  * punching a hole to EOF.
154  */
155 STATIC int
156 xfs_free_eofblocks(
157         xfs_mount_t     *mp,
158         xfs_inode_t     *ip,
159         int             flags)
160 {
161         xfs_trans_t     *tp;
162         int             error;
163         xfs_fileoff_t   end_fsb;
164         xfs_fileoff_t   last_fsb;
165         xfs_filblks_t   map_len;
166         int             nimaps;
167         xfs_bmbt_irec_t imap;
168
169         /*
170          * Figure out if there are any blocks beyond the end
171          * of the file.  If not, then there is nothing to do.
172          */
173         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
174         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
175         if (last_fsb <= end_fsb)
176                 return 0;
177         map_len = last_fsb - end_fsb;
178
179         nimaps = 1;
180         xfs_ilock(ip, XFS_ILOCK_SHARED);
181         error = xfs_bmapi(NULL, ip, end_fsb, map_len, 0,
182                           NULL, 0, &imap, &nimaps, NULL);
183         xfs_iunlock(ip, XFS_ILOCK_SHARED);
184
185         if (!error && (nimaps != 0) &&
186             (imap.br_startblock != HOLESTARTBLOCK ||
187              ip->i_delayed_blks)) {
188                 /*
189                  * Attach the dquots to the inode up front.
190                  */
191                 error = xfs_qm_dqattach(ip, 0);
192                 if (error)
193                         return error;
194
195                 /*
196                  * There are blocks after the end of file.
197                  * Free them up now by truncating the file to
198                  * its current size.
199                  */
200                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
201
202                 if (flags & XFS_FREE_EOF_TRYLOCK) {
203                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
204                                 xfs_trans_cancel(tp, 0);
205                                 return 0;
206                         }
207                 } else {
208                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
209                 }
210
211                 error = xfs_trans_reserve(tp, 0,
212                                           XFS_ITRUNCATE_LOG_RES(mp),
213                                           0, XFS_TRANS_PERM_LOG_RES,
214                                           XFS_ITRUNCATE_LOG_COUNT);
215                 if (error) {
216                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
217                         xfs_trans_cancel(tp, 0);
218                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
219                         return error;
220                 }
221
222                 xfs_ilock(ip, XFS_ILOCK_EXCL);
223                 xfs_trans_ijoin(tp, ip);
224
225                 error = xfs_itruncate_data(&tp, ip, ip->i_size);
226                 if (error) {
227                         /*
228                          * If we get an error at this point we simply don't
229                          * bother truncating the file.
230                          */
231                         xfs_trans_cancel(tp,
232                                          (XFS_TRANS_RELEASE_LOG_RES |
233                                           XFS_TRANS_ABORT));
234                 } else {
235                         error = xfs_trans_commit(tp,
236                                                 XFS_TRANS_RELEASE_LOG_RES);
237                 }
238                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
239         }
240         return error;
241 }
242
243 /*
244  * Free a symlink that has blocks associated with it.
245  */
246 STATIC int
247 xfs_inactive_symlink_rmt(
248         xfs_inode_t     *ip,
249         xfs_trans_t     **tpp)
250 {
251         xfs_buf_t       *bp;
252         int             committed;
253         int             done;
254         int             error;
255         xfs_fsblock_t   first_block;
256         xfs_bmap_free_t free_list;
257         int             i;
258         xfs_mount_t     *mp;
259         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
260         int             nmaps;
261         xfs_trans_t     *ntp;
262         int             size;
263         xfs_trans_t     *tp;
264
265         tp = *tpp;
266         mp = ip->i_mount;
267         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
268         /*
269          * We're freeing a symlink that has some
270          * blocks allocated to it.  Free the
271          * blocks here.  We know that we've got
272          * either 1 or 2 extents and that we can
273          * free them all in one bunmapi call.
274          */
275         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
276         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
277                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
278                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
279                 xfs_trans_cancel(tp, 0);
280                 *tpp = NULL;
281                 return error;
282         }
283         /*
284          * Lock the inode, fix the size, and join it to the transaction.
285          * Hold it so in the normal path, we still have it locked for
286          * the second transaction.  In the error paths we need it
287          * held so the cancel won't rele it, see below.
288          */
289         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
290         size = (int)ip->i_d.di_size;
291         ip->i_d.di_size = 0;
292         xfs_trans_ijoin(tp, ip);
293         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
294         /*
295          * Find the block(s) so we can inval and unmap them.
296          */
297         done = 0;
298         xfs_bmap_init(&free_list, &first_block);
299         nmaps = ARRAY_SIZE(mval);
300         if ((error = xfs_bmapi(tp, ip, 0, XFS_B_TO_FSB(mp, size),
301                         XFS_BMAPI_METADATA, &first_block, 0, mval, &nmaps,
302                         &free_list)))
303                 goto error0;
304         /*
305          * Invalidate the block(s).
306          */
307         for (i = 0; i < nmaps; i++) {
308                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
309                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
310                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
311                 xfs_trans_binval(tp, bp);
312         }
313         /*
314          * Unmap the dead block(s) to the free_list.
315          */
316         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
317                         &first_block, &free_list, &done)))
318                 goto error1;
319         ASSERT(done);
320         /*
321          * Commit the first transaction.  This logs the EFI and the inode.
322          */
323         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
324                 goto error1;
325         /*
326          * The transaction must have been committed, since there were
327          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
328          * The new tp has the extent freeing and EFDs.
329          */
330         ASSERT(committed);
331         /*
332          * The first xact was committed, so add the inode to the new one.
333          * Mark it dirty so it will be logged and moved forward in the log as
334          * part of every commit.
335          */
336         xfs_trans_ijoin(tp, ip);
337         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
338         /*
339          * Get a new, empty transaction to return to our caller.
340          */
341         ntp = xfs_trans_dup(tp);
342         /*
343          * Commit the transaction containing extent freeing and EFDs.
344          * If we get an error on the commit here or on the reserve below,
345          * we need to unlock the inode since the new transaction doesn't
346          * have the inode attached.
347          */
348         error = xfs_trans_commit(tp, 0);
349         tp = ntp;
350         if (error) {
351                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
352                 goto error0;
353         }
354         /*
355          * transaction commit worked ok so we can drop the extra ticket
356          * reference that we gained in xfs_trans_dup()
357          */
358         xfs_log_ticket_put(tp->t_ticket);
359
360         /*
361          * Remove the memory for extent descriptions (just bookkeeping).
362          */
363         if (ip->i_df.if_bytes)
364                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
365         ASSERT(ip->i_df.if_bytes == 0);
366         /*
367          * Put an itruncate log reservation in the new transaction
368          * for our caller.
369          */
370         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
371                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
372                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
373                 goto error0;
374         }
375         /*
376          * Return with the inode locked but not joined to the transaction.
377          */
378         *tpp = tp;
379         return 0;
380
381  error1:
382         xfs_bmap_cancel(&free_list);
383  error0:
384         /*
385          * Have to come here with the inode locked and either
386          * (held and in the transaction) or (not in the transaction).
387          * If the inode isn't held then cancel would iput it, but
388          * that's wrong since this is inactive and the vnode ref
389          * count is 0 already.
390          * Cancel won't do anything to the inode if held, but it still
391          * needs to be locked until the cancel is done, if it was
392          * joined to the transaction.
393          */
394         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
395         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
396         *tpp = NULL;
397         return error;
398
399 }
400
401 STATIC int
402 xfs_inactive_symlink_local(
403         xfs_inode_t     *ip,
404         xfs_trans_t     **tpp)
405 {
406         int             error;
407
408         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
409         /*
410          * We're freeing a symlink which fit into
411          * the inode.  Just free the memory used
412          * to hold the old symlink.
413          */
414         error = xfs_trans_reserve(*tpp, 0,
415                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
416                                   0, XFS_TRANS_PERM_LOG_RES,
417                                   XFS_ITRUNCATE_LOG_COUNT);
418
419         if (error) {
420                 xfs_trans_cancel(*tpp, 0);
421                 *tpp = NULL;
422                 return error;
423         }
424         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
425
426         /*
427          * Zero length symlinks _can_ exist.
428          */
429         if (ip->i_df.if_bytes > 0) {
430                 xfs_idata_realloc(ip,
431                                   -(ip->i_df.if_bytes),
432                                   XFS_DATA_FORK);
433                 ASSERT(ip->i_df.if_bytes == 0);
434         }
435         return 0;
436 }
437
438 STATIC int
439 xfs_inactive_attrs(
440         xfs_inode_t     *ip,
441         xfs_trans_t     **tpp)
442 {
443         xfs_trans_t     *tp;
444         int             error;
445         xfs_mount_t     *mp;
446
447         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
448         tp = *tpp;
449         mp = ip->i_mount;
450         ASSERT(ip->i_d.di_forkoff != 0);
451         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
452         xfs_iunlock(ip, XFS_ILOCK_EXCL);
453         if (error)
454                 goto error_unlock;
455
456         error = xfs_attr_inactive(ip);
457         if (error)
458                 goto error_unlock;
459
460         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
461         error = xfs_trans_reserve(tp, 0,
462                                   XFS_IFREE_LOG_RES(mp),
463                                   0, XFS_TRANS_PERM_LOG_RES,
464                                   XFS_INACTIVE_LOG_COUNT);
465         if (error)
466                 goto error_cancel;
467
468         xfs_ilock(ip, XFS_ILOCK_EXCL);
469         xfs_trans_ijoin(tp, ip);
470         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
471
472         ASSERT(ip->i_d.di_anextents == 0);
473
474         *tpp = tp;
475         return 0;
476
477 error_cancel:
478         ASSERT(XFS_FORCED_SHUTDOWN(mp));
479         xfs_trans_cancel(tp, 0);
480 error_unlock:
481         *tpp = NULL;
482         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
483         return error;
484 }
485
486 int
487 xfs_release(
488         xfs_inode_t     *ip)
489 {
490         xfs_mount_t     *mp = ip->i_mount;
491         int             error;
492
493         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
494                 return 0;
495
496         /* If this is a read-only mount, don't do this (would generate I/O) */
497         if (mp->m_flags & XFS_MOUNT_RDONLY)
498                 return 0;
499
500         if (!XFS_FORCED_SHUTDOWN(mp)) {
501                 int truncated;
502
503                 /*
504                  * If we are using filestreams, and we have an unlinked
505                  * file that we are processing the last close on, then nothing
506                  * will be able to reopen and write to this file. Purge this
507                  * inode from the filestreams cache so that it doesn't delay
508                  * teardown of the inode.
509                  */
510                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
511                         xfs_filestream_deassociate(ip);
512
513                 /*
514                  * If we previously truncated this file and removed old data
515                  * in the process, we want to initiate "early" writeout on
516                  * the last close.  This is an attempt to combat the notorious
517                  * NULL files problem which is particularly noticeable from a
518                  * truncate down, buffered (re-)write (delalloc), followed by
519                  * a crash.  What we are effectively doing here is
520                  * significantly reducing the time window where we'd otherwise
521                  * be exposed to that problem.
522                  */
523                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
524                 if (truncated) {
525                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
526                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
527                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
528                 }
529         }
530
531         if (ip->i_d.di_nlink == 0)
532                 return 0;
533
534         if ((S_ISREG(ip->i_d.di_mode) &&
535              ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
536                ip->i_delayed_blks > 0)) &&
537              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
538             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
539
540                 /*
541                  * If we can't get the iolock just skip truncating the blocks
542                  * past EOF because we could deadlock with the mmap_sem
543                  * otherwise.  We'll get another chance to drop them once the
544                  * last reference to the inode is dropped, so we'll never leak
545                  * blocks permanently.
546                  *
547                  * Further, check if the inode is being opened, written and
548                  * closed frequently and we have delayed allocation blocks
549                  * outstanding (e.g. streaming writes from the NFS server),
550                  * truncating the blocks past EOF will cause fragmentation to
551                  * occur.
552                  *
553                  * In this case don't do the truncation, either, but we have to
554                  * be careful how we detect this case. Blocks beyond EOF show
555                  * up as i_delayed_blks even when the inode is clean, so we
556                  * need to truncate them away first before checking for a dirty
557                  * release. Hence on the first dirty close we will still remove
558                  * the speculative allocation, but after that we will leave it
559                  * in place.
560                  */
561                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
562                         return 0;
563
564                 error = xfs_free_eofblocks(mp, ip,
565                                            XFS_FREE_EOF_TRYLOCK);
566                 if (error)
567                         return error;
568
569                 /* delalloc blocks after truncation means it really is dirty */
570                 if (ip->i_delayed_blks)
571                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
572         }
573         return 0;
574 }
575
576 /*
577  * xfs_inactive
578  *
579  * This is called when the vnode reference count for the vnode
580  * goes to zero.  If the file has been unlinked, then it must
581  * now be truncated.  Also, we clear all of the read-ahead state
582  * kept for the inode here since the file is now closed.
583  */
584 int
585 xfs_inactive(
586         xfs_inode_t     *ip)
587 {
588         xfs_bmap_free_t free_list;
589         xfs_fsblock_t   first_block;
590         int             committed;
591         xfs_trans_t     *tp;
592         xfs_mount_t     *mp;
593         int             error;
594         int             truncate;
595
596         /*
597          * If the inode is already free, then there can be nothing
598          * to clean up here.
599          */
600         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
601                 ASSERT(ip->i_df.if_real_bytes == 0);
602                 ASSERT(ip->i_df.if_broot_bytes == 0);
603                 return VN_INACTIVE_CACHE;
604         }
605
606         /*
607          * Only do a truncate if it's a regular file with
608          * some actual space in it.  It's OK to look at the
609          * inode's fields without the lock because we're the
610          * only one with a reference to the inode.
611          */
612         truncate = ((ip->i_d.di_nlink == 0) &&
613             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
614              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
615             S_ISREG(ip->i_d.di_mode));
616
617         mp = ip->i_mount;
618
619         error = 0;
620
621         /* If this is a read-only mount, don't do this (would generate I/O) */
622         if (mp->m_flags & XFS_MOUNT_RDONLY)
623                 goto out;
624
625         if (ip->i_d.di_nlink != 0) {
626                 if ((S_ISREG(ip->i_d.di_mode) &&
627                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
628                        ip->i_delayed_blks > 0)) &&
629                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
630                      (!(ip->i_d.di_flags &
631                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
632                       (ip->i_delayed_blks != 0)))) {
633                         error = xfs_free_eofblocks(mp, ip, 0);
634                         if (error)
635                                 return VN_INACTIVE_CACHE;
636                 }
637                 goto out;
638         }
639
640         ASSERT(ip->i_d.di_nlink == 0);
641
642         error = xfs_qm_dqattach(ip, 0);
643         if (error)
644                 return VN_INACTIVE_CACHE;
645
646         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
647         if (truncate) {
648                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
649
650                 xfs_ioend_wait(ip);
651
652                 error = xfs_trans_reserve(tp, 0,
653                                           XFS_ITRUNCATE_LOG_RES(mp),
654                                           0, XFS_TRANS_PERM_LOG_RES,
655                                           XFS_ITRUNCATE_LOG_COUNT);
656                 if (error) {
657                         /* Don't call itruncate_cleanup */
658                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
659                         xfs_trans_cancel(tp, 0);
660                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
661                         return VN_INACTIVE_CACHE;
662                 }
663
664                 xfs_ilock(ip, XFS_ILOCK_EXCL);
665                 xfs_trans_ijoin(tp, ip);
666
667                 error = xfs_itruncate_data(&tp, ip, 0);
668                 if (error) {
669                         xfs_trans_cancel(tp,
670                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
671                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
672                         return VN_INACTIVE_CACHE;
673                 }
674         } else if (S_ISLNK(ip->i_d.di_mode)) {
675
676                 /*
677                  * If we get an error while cleaning up a
678                  * symlink we bail out.
679                  */
680                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
681                         xfs_inactive_symlink_rmt(ip, &tp) :
682                         xfs_inactive_symlink_local(ip, &tp);
683
684                 if (error) {
685                         ASSERT(tp == NULL);
686                         return VN_INACTIVE_CACHE;
687                 }
688
689                 xfs_trans_ijoin(tp, ip);
690         } else {
691                 error = xfs_trans_reserve(tp, 0,
692                                           XFS_IFREE_LOG_RES(mp),
693                                           0, XFS_TRANS_PERM_LOG_RES,
694                                           XFS_INACTIVE_LOG_COUNT);
695                 if (error) {
696                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
697                         xfs_trans_cancel(tp, 0);
698                         return VN_INACTIVE_CACHE;
699                 }
700
701                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
702                 xfs_trans_ijoin(tp, ip);
703         }
704
705         /*
706          * If there are attributes associated with the file
707          * then blow them away now.  The code calls a routine
708          * that recursively deconstructs the attribute fork.
709          * We need to just commit the current transaction
710          * because we can't use it for xfs_attr_inactive().
711          */
712         if (ip->i_d.di_anextents > 0) {
713                 error = xfs_inactive_attrs(ip, &tp);
714                 /*
715                  * If we got an error, the transaction is already
716                  * cancelled, and the inode is unlocked. Just get out.
717                  */
718                  if (error)
719                          return VN_INACTIVE_CACHE;
720         } else if (ip->i_afp) {
721                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
722         }
723
724         /*
725          * Free the inode.
726          */
727         xfs_bmap_init(&free_list, &first_block);
728         error = xfs_ifree(tp, ip, &free_list);
729         if (error) {
730                 /*
731                  * If we fail to free the inode, shut down.  The cancel
732                  * might do that, we need to make sure.  Otherwise the
733                  * inode might be lost for a long time or forever.
734                  */
735                 if (!XFS_FORCED_SHUTDOWN(mp)) {
736                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
737                                 __func__, error);
738                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
739                 }
740                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
741         } else {
742                 /*
743                  * Credit the quota account(s). The inode is gone.
744                  */
745                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
746
747                 /*
748                  * Just ignore errors at this point.  There is nothing we can
749                  * do except to try to keep going. Make sure it's not a silent
750                  * error.
751                  */
752                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
753                 if (error)
754                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
755                                 __func__, error);
756                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
757                 if (error)
758                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
759                                 __func__, error);
760         }
761
762         /*
763          * Release the dquots held by inode, if any.
764          */
765         xfs_qm_dqdetach(ip);
766         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
767
768  out:
769         return VN_INACTIVE_CACHE;
770 }
771
772 /*
773  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
774  * is allowed, otherwise it has to be an exact match. If a CI match is found,
775  * ci_name->name will point to a the actual name (caller must free) or
776  * will be set to NULL if an exact match is found.
777  */
778 int
779 xfs_lookup(
780         xfs_inode_t             *dp,
781         struct xfs_name         *name,
782         xfs_inode_t             **ipp,
783         struct xfs_name         *ci_name)
784 {
785         xfs_ino_t               inum;
786         int                     error;
787         uint                    lock_mode;
788
789         trace_xfs_lookup(dp, name);
790
791         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
792                 return XFS_ERROR(EIO);
793
794         lock_mode = xfs_ilock_map_shared(dp);
795         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
796         xfs_iunlock_map_shared(dp, lock_mode);
797
798         if (error)
799                 goto out;
800
801         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
802         if (error)
803                 goto out_free_name;
804
805         return 0;
806
807 out_free_name:
808         if (ci_name)
809                 kmem_free(ci_name->name);
810 out:
811         *ipp = NULL;
812         return error;
813 }
814
815 int
816 xfs_create(
817         xfs_inode_t             *dp,
818         struct xfs_name         *name,
819         mode_t                  mode,
820         xfs_dev_t               rdev,
821         xfs_inode_t             **ipp)
822 {
823         int                     is_dir = S_ISDIR(mode);
824         struct xfs_mount        *mp = dp->i_mount;
825         struct xfs_inode        *ip = NULL;
826         struct xfs_trans        *tp = NULL;
827         int                     error;
828         xfs_bmap_free_t         free_list;
829         xfs_fsblock_t           first_block;
830         boolean_t               unlock_dp_on_error = B_FALSE;
831         uint                    cancel_flags;
832         int                     committed;
833         prid_t                  prid;
834         struct xfs_dquot        *udqp = NULL;
835         struct xfs_dquot        *gdqp = NULL;
836         uint                    resblks;
837         uint                    log_res;
838         uint                    log_count;
839
840         trace_xfs_create(dp, name);
841
842         if (XFS_FORCED_SHUTDOWN(mp))
843                 return XFS_ERROR(EIO);
844
845         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
846                 prid = xfs_get_projid(dp);
847         else
848                 prid = XFS_PROJID_DEFAULT;
849
850         /*
851          * Make sure that we have allocated dquot(s) on disk.
852          */
853         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
854                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
855         if (error)
856                 return error;
857
858         if (is_dir) {
859                 rdev = 0;
860                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
861                 log_res = XFS_MKDIR_LOG_RES(mp);
862                 log_count = XFS_MKDIR_LOG_COUNT;
863                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
864         } else {
865                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
866                 log_res = XFS_CREATE_LOG_RES(mp);
867                 log_count = XFS_CREATE_LOG_COUNT;
868                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
869         }
870
871         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
872
873         /*
874          * Initially assume that the file does not exist and
875          * reserve the resources for that case.  If that is not
876          * the case we'll drop the one we have and get a more
877          * appropriate transaction later.
878          */
879         error = xfs_trans_reserve(tp, resblks, log_res, 0,
880                         XFS_TRANS_PERM_LOG_RES, log_count);
881         if (error == ENOSPC) {
882                 /* flush outstanding delalloc blocks and retry */
883                 xfs_flush_inodes(dp);
884                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
885                                 XFS_TRANS_PERM_LOG_RES, log_count);
886         }
887         if (error == ENOSPC) {
888                 /* No space at all so try a "no-allocation" reservation */
889                 resblks = 0;
890                 error = xfs_trans_reserve(tp, 0, log_res, 0,
891                                 XFS_TRANS_PERM_LOG_RES, log_count);
892         }
893         if (error) {
894                 cancel_flags = 0;
895                 goto out_trans_cancel;
896         }
897
898         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
899         unlock_dp_on_error = B_TRUE;
900
901         /*
902          * Check for directory link count overflow.
903          */
904         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
905                 error = XFS_ERROR(EMLINK);
906                 goto out_trans_cancel;
907         }
908
909         xfs_bmap_init(&free_list, &first_block);
910
911         /*
912          * Reserve disk quota and the inode.
913          */
914         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
915         if (error)
916                 goto out_trans_cancel;
917
918         error = xfs_dir_canenter(tp, dp, name, resblks);
919         if (error)
920                 goto out_trans_cancel;
921
922         /*
923          * A newly created regular or special file just has one directory
924          * entry pointing to them, but a directory also the "." entry
925          * pointing to itself.
926          */
927         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
928                                prid, resblks > 0, &ip, &committed);
929         if (error) {
930                 if (error == ENOSPC)
931                         goto out_trans_cancel;
932                 goto out_trans_abort;
933         }
934
935         /*
936          * Now we join the directory inode to the transaction.  We do not do it
937          * earlier because xfs_dir_ialloc might commit the previous transaction
938          * (and release all the locks).  An error from here on will result in
939          * the transaction cancel unlocking dp so don't do it explicitly in the
940          * error path.
941          */
942         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
943         unlock_dp_on_error = B_FALSE;
944
945         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
946                                         &first_block, &free_list, resblks ?
947                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
948         if (error) {
949                 ASSERT(error != ENOSPC);
950                 goto out_trans_abort;
951         }
952         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
953         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
954
955         if (is_dir) {
956                 error = xfs_dir_init(tp, ip, dp);
957                 if (error)
958                         goto out_bmap_cancel;
959
960                 error = xfs_bumplink(tp, dp);
961                 if (error)
962                         goto out_bmap_cancel;
963         }
964
965         /*
966          * If this is a synchronous mount, make sure that the
967          * create transaction goes to disk before returning to
968          * the user.
969          */
970         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
971                 xfs_trans_set_sync(tp);
972
973         /*
974          * Attach the dquot(s) to the inodes and modify them incore.
975          * These ids of the inode couldn't have changed since the new
976          * inode has been locked ever since it was created.
977          */
978         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
979
980         error = xfs_bmap_finish(&tp, &free_list, &committed);
981         if (error)
982                 goto out_bmap_cancel;
983
984         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
985         if (error)
986                 goto out_release_inode;
987
988         xfs_qm_dqrele(udqp);
989         xfs_qm_dqrele(gdqp);
990
991         *ipp = ip;
992         return 0;
993
994  out_bmap_cancel:
995         xfs_bmap_cancel(&free_list);
996  out_trans_abort:
997         cancel_flags |= XFS_TRANS_ABORT;
998  out_trans_cancel:
999         xfs_trans_cancel(tp, cancel_flags);
1000  out_release_inode:
1001         /*
1002          * Wait until after the current transaction is aborted to
1003          * release the inode.  This prevents recursive transactions
1004          * and deadlocks from xfs_inactive.
1005          */
1006         if (ip)
1007                 IRELE(ip);
1008
1009         xfs_qm_dqrele(udqp);
1010         xfs_qm_dqrele(gdqp);
1011
1012         if (unlock_dp_on_error)
1013                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1014         return error;
1015 }
1016
1017 #ifdef DEBUG
1018 int xfs_locked_n;
1019 int xfs_small_retries;
1020 int xfs_middle_retries;
1021 int xfs_lots_retries;
1022 int xfs_lock_delays;
1023 #endif
1024
1025 /*
1026  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1027  * a different value
1028  */
1029 static inline int
1030 xfs_lock_inumorder(int lock_mode, int subclass)
1031 {
1032         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1033                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1034         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1035                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1036
1037         return lock_mode;
1038 }
1039
1040 /*
1041  * The following routine will lock n inodes in exclusive mode.
1042  * We assume the caller calls us with the inodes in i_ino order.
1043  *
1044  * We need to detect deadlock where an inode that we lock
1045  * is in the AIL and we start waiting for another inode that is locked
1046  * by a thread in a long running transaction (such as truncate). This can
1047  * result in deadlock since the long running trans might need to wait
1048  * for the inode we just locked in order to push the tail and free space
1049  * in the log.
1050  */
1051 void
1052 xfs_lock_inodes(
1053         xfs_inode_t     **ips,
1054         int             inodes,
1055         uint            lock_mode)
1056 {
1057         int             attempts = 0, i, j, try_lock;
1058         xfs_log_item_t  *lp;
1059
1060         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1061
1062         try_lock = 0;
1063         i = 0;
1064
1065 again:
1066         for (; i < inodes; i++) {
1067                 ASSERT(ips[i]);
1068
1069                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1070                         continue;
1071
1072                 /*
1073                  * If try_lock is not set yet, make sure all locked inodes
1074                  * are not in the AIL.
1075                  * If any are, set try_lock to be used later.
1076                  */
1077
1078                 if (!try_lock) {
1079                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1080                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1081                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1082                                         try_lock++;
1083                                 }
1084                         }
1085                 }
1086
1087                 /*
1088                  * If any of the previous locks we have locked is in the AIL,
1089                  * we must TRY to get the second and subsequent locks. If
1090                  * we can't get any, we must release all we have
1091                  * and try again.
1092                  */
1093
1094                 if (try_lock) {
1095                         /* try_lock must be 0 if i is 0. */
1096                         /*
1097                          * try_lock means we have an inode locked
1098                          * that is in the AIL.
1099                          */
1100                         ASSERT(i != 0);
1101                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1102                                 attempts++;
1103
1104                                 /*
1105                                  * Unlock all previous guys and try again.
1106                                  * xfs_iunlock will try to push the tail
1107                                  * if the inode is in the AIL.
1108                                  */
1109
1110                                 for(j = i - 1; j >= 0; j--) {
1111
1112                                         /*
1113                                          * Check to see if we've already
1114                                          * unlocked this one.
1115                                          * Not the first one going back,
1116                                          * and the inode ptr is the same.
1117                                          */
1118                                         if ((j != (i - 1)) && ips[j] ==
1119                                                                 ips[j+1])
1120                                                 continue;
1121
1122                                         xfs_iunlock(ips[j], lock_mode);
1123                                 }
1124
1125                                 if ((attempts % 5) == 0) {
1126                                         delay(1); /* Don't just spin the CPU */
1127 #ifdef DEBUG
1128                                         xfs_lock_delays++;
1129 #endif
1130                                 }
1131                                 i = 0;
1132                                 try_lock = 0;
1133                                 goto again;
1134                         }
1135                 } else {
1136                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1137                 }
1138         }
1139
1140 #ifdef DEBUG
1141         if (attempts) {
1142                 if (attempts < 5) xfs_small_retries++;
1143                 else if (attempts < 100) xfs_middle_retries++;
1144                 else xfs_lots_retries++;
1145         } else {
1146                 xfs_locked_n++;
1147         }
1148 #endif
1149 }
1150
1151 /*
1152  * xfs_lock_two_inodes() can only be used to lock one type of lock
1153  * at a time - the iolock or the ilock, but not both at once. If
1154  * we lock both at once, lockdep will report false positives saying
1155  * we have violated locking orders.
1156  */
1157 void
1158 xfs_lock_two_inodes(
1159         xfs_inode_t             *ip0,
1160         xfs_inode_t             *ip1,
1161         uint                    lock_mode)
1162 {
1163         xfs_inode_t             *temp;
1164         int                     attempts = 0;
1165         xfs_log_item_t          *lp;
1166
1167         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1168                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1169         ASSERT(ip0->i_ino != ip1->i_ino);
1170
1171         if (ip0->i_ino > ip1->i_ino) {
1172                 temp = ip0;
1173                 ip0 = ip1;
1174                 ip1 = temp;
1175         }
1176
1177  again:
1178         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1179
1180         /*
1181          * If the first lock we have locked is in the AIL, we must TRY to get
1182          * the second lock. If we can't get it, we must release the first one
1183          * and try again.
1184          */
1185         lp = (xfs_log_item_t *)ip0->i_itemp;
1186         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1187                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1188                         xfs_iunlock(ip0, lock_mode);
1189                         if ((++attempts % 5) == 0)
1190                                 delay(1); /* Don't just spin the CPU */
1191                         goto again;
1192                 }
1193         } else {
1194                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1195         }
1196 }
1197
1198 int
1199 xfs_remove(
1200         xfs_inode_t             *dp,
1201         struct xfs_name         *name,
1202         xfs_inode_t             *ip)
1203 {
1204         xfs_mount_t             *mp = dp->i_mount;
1205         xfs_trans_t             *tp = NULL;
1206         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1207         int                     error = 0;
1208         xfs_bmap_free_t         free_list;
1209         xfs_fsblock_t           first_block;
1210         int                     cancel_flags;
1211         int                     committed;
1212         int                     link_zero;
1213         uint                    resblks;
1214         uint                    log_count;
1215
1216         trace_xfs_remove(dp, name);
1217
1218         if (XFS_FORCED_SHUTDOWN(mp))
1219                 return XFS_ERROR(EIO);
1220
1221         error = xfs_qm_dqattach(dp, 0);
1222         if (error)
1223                 goto std_return;
1224
1225         error = xfs_qm_dqattach(ip, 0);
1226         if (error)
1227                 goto std_return;
1228
1229         if (is_dir) {
1230                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1231                 log_count = XFS_DEFAULT_LOG_COUNT;
1232         } else {
1233                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1234                 log_count = XFS_REMOVE_LOG_COUNT;
1235         }
1236         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1237
1238         /*
1239          * We try to get the real space reservation first,
1240          * allowing for directory btree deletion(s) implying
1241          * possible bmap insert(s).  If we can't get the space
1242          * reservation then we use 0 instead, and avoid the bmap
1243          * btree insert(s) in the directory code by, if the bmap
1244          * insert tries to happen, instead trimming the LAST
1245          * block from the directory.
1246          */
1247         resblks = XFS_REMOVE_SPACE_RES(mp);
1248         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1249                                   XFS_TRANS_PERM_LOG_RES, log_count);
1250         if (error == ENOSPC) {
1251                 resblks = 0;
1252                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1253                                           XFS_TRANS_PERM_LOG_RES, log_count);
1254         }
1255         if (error) {
1256                 ASSERT(error != ENOSPC);
1257                 cancel_flags = 0;
1258                 goto out_trans_cancel;
1259         }
1260
1261         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1262
1263         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1264         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1265
1266         /*
1267          * If we're removing a directory perform some additional validation.
1268          */
1269         if (is_dir) {
1270                 ASSERT(ip->i_d.di_nlink >= 2);
1271                 if (ip->i_d.di_nlink != 2) {
1272                         error = XFS_ERROR(ENOTEMPTY);
1273                         goto out_trans_cancel;
1274                 }
1275                 if (!xfs_dir_isempty(ip)) {
1276                         error = XFS_ERROR(ENOTEMPTY);
1277                         goto out_trans_cancel;
1278                 }
1279         }
1280
1281         xfs_bmap_init(&free_list, &first_block);
1282         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1283                                         &first_block, &free_list, resblks);
1284         if (error) {
1285                 ASSERT(error != ENOENT);
1286                 goto out_bmap_cancel;
1287         }
1288         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1289
1290         if (is_dir) {
1291                 /*
1292                  * Drop the link from ip's "..".
1293                  */
1294                 error = xfs_droplink(tp, dp);
1295                 if (error)
1296                         goto out_bmap_cancel;
1297
1298                 /*
1299                  * Drop the "." link from ip to self.
1300                  */
1301                 error = xfs_droplink(tp, ip);
1302                 if (error)
1303                         goto out_bmap_cancel;
1304         } else {
1305                 /*
1306                  * When removing a non-directory we need to log the parent
1307                  * inode here.  For a directory this is done implicitly
1308                  * by the xfs_droplink call for the ".." entry.
1309                  */
1310                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1311         }
1312
1313         /*
1314          * Drop the link from dp to ip.
1315          */
1316         error = xfs_droplink(tp, ip);
1317         if (error)
1318                 goto out_bmap_cancel;
1319
1320         /*
1321          * Determine if this is the last link while
1322          * we are in the transaction.
1323          */
1324         link_zero = (ip->i_d.di_nlink == 0);
1325
1326         /*
1327          * If this is a synchronous mount, make sure that the
1328          * remove transaction goes to disk before returning to
1329          * the user.
1330          */
1331         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1332                 xfs_trans_set_sync(tp);
1333
1334         error = xfs_bmap_finish(&tp, &free_list, &committed);
1335         if (error)
1336                 goto out_bmap_cancel;
1337
1338         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1339         if (error)
1340                 goto std_return;
1341
1342         /*
1343          * If we are using filestreams, kill the stream association.
1344          * If the file is still open it may get a new one but that
1345          * will get killed on last close in xfs_close() so we don't
1346          * have to worry about that.
1347          */
1348         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1349                 xfs_filestream_deassociate(ip);
1350
1351         return 0;
1352
1353  out_bmap_cancel:
1354         xfs_bmap_cancel(&free_list);
1355         cancel_flags |= XFS_TRANS_ABORT;
1356  out_trans_cancel:
1357         xfs_trans_cancel(tp, cancel_flags);
1358  std_return:
1359         return error;
1360 }
1361
1362 int
1363 xfs_link(
1364         xfs_inode_t             *tdp,
1365         xfs_inode_t             *sip,
1366         struct xfs_name         *target_name)
1367 {
1368         xfs_mount_t             *mp = tdp->i_mount;
1369         xfs_trans_t             *tp;
1370         int                     error;
1371         xfs_bmap_free_t         free_list;
1372         xfs_fsblock_t           first_block;
1373         int                     cancel_flags;
1374         int                     committed;
1375         int                     resblks;
1376
1377         trace_xfs_link(tdp, target_name);
1378
1379         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1380
1381         if (XFS_FORCED_SHUTDOWN(mp))
1382                 return XFS_ERROR(EIO);
1383
1384         error = xfs_qm_dqattach(sip, 0);
1385         if (error)
1386                 goto std_return;
1387
1388         error = xfs_qm_dqattach(tdp, 0);
1389         if (error)
1390                 goto std_return;
1391
1392         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1393         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1394         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1395         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1396                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1397         if (error == ENOSPC) {
1398                 resblks = 0;
1399                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1400                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1401         }
1402         if (error) {
1403                 cancel_flags = 0;
1404                 goto error_return;
1405         }
1406
1407         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1408
1409         xfs_trans_ijoin_ref(tp, sip, XFS_ILOCK_EXCL);
1410         xfs_trans_ijoin_ref(tp, tdp, XFS_ILOCK_EXCL);
1411
1412         /*
1413          * If the source has too many links, we can't make any more to it.
1414          */
1415         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1416                 error = XFS_ERROR(EMLINK);
1417                 goto error_return;
1418         }
1419
1420         /*
1421          * If we are using project inheritance, we only allow hard link
1422          * creation in our tree when the project IDs are the same; else
1423          * the tree quota mechanism could be circumvented.
1424          */
1425         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1426                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1427                 error = XFS_ERROR(EXDEV);
1428                 goto error_return;
1429         }
1430
1431         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1432         if (error)
1433                 goto error_return;
1434
1435         xfs_bmap_init(&free_list, &first_block);
1436
1437         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1438                                         &first_block, &free_list, resblks);
1439         if (error)
1440                 goto abort_return;
1441         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1442         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1443
1444         error = xfs_bumplink(tp, sip);
1445         if (error)
1446                 goto abort_return;
1447
1448         /*
1449          * If this is a synchronous mount, make sure that the
1450          * link transaction goes to disk before returning to
1451          * the user.
1452          */
1453         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1454                 xfs_trans_set_sync(tp);
1455         }
1456
1457         error = xfs_bmap_finish (&tp, &free_list, &committed);
1458         if (error) {
1459                 xfs_bmap_cancel(&free_list);
1460                 goto abort_return;
1461         }
1462
1463         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1464
1465  abort_return:
1466         cancel_flags |= XFS_TRANS_ABORT;
1467  error_return:
1468         xfs_trans_cancel(tp, cancel_flags);
1469  std_return:
1470         return error;
1471 }
1472
1473 int
1474 xfs_symlink(
1475         xfs_inode_t             *dp,
1476         struct xfs_name         *link_name,
1477         const char              *target_path,
1478         mode_t                  mode,
1479         xfs_inode_t             **ipp)
1480 {
1481         xfs_mount_t             *mp = dp->i_mount;
1482         xfs_trans_t             *tp;
1483         xfs_inode_t             *ip;
1484         int                     error;
1485         int                     pathlen;
1486         xfs_bmap_free_t         free_list;
1487         xfs_fsblock_t           first_block;
1488         boolean_t               unlock_dp_on_error = B_FALSE;
1489         uint                    cancel_flags;
1490         int                     committed;
1491         xfs_fileoff_t           first_fsb;
1492         xfs_filblks_t           fs_blocks;
1493         int                     nmaps;
1494         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1495         xfs_daddr_t             d;
1496         const char              *cur_chunk;
1497         int                     byte_cnt;
1498         int                     n;
1499         xfs_buf_t               *bp;
1500         prid_t                  prid;
1501         struct xfs_dquot        *udqp, *gdqp;
1502         uint                    resblks;
1503
1504         *ipp = NULL;
1505         error = 0;
1506         ip = NULL;
1507         tp = NULL;
1508
1509         trace_xfs_symlink(dp, link_name);
1510
1511         if (XFS_FORCED_SHUTDOWN(mp))
1512                 return XFS_ERROR(EIO);
1513
1514         /*
1515          * Check component lengths of the target path name.
1516          */
1517         pathlen = strlen(target_path);
1518         if (pathlen >= MAXPATHLEN)      /* total string too long */
1519                 return XFS_ERROR(ENAMETOOLONG);
1520
1521         udqp = gdqp = NULL;
1522         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1523                 prid = xfs_get_projid(dp);
1524         else
1525                 prid = XFS_PROJID_DEFAULT;
1526
1527         /*
1528          * Make sure that we have allocated dquot(s) on disk.
1529          */
1530         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1531                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1532         if (error)
1533                 goto std_return;
1534
1535         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1536         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1537         /*
1538          * The symlink will fit into the inode data fork?
1539          * There can't be any attributes so we get the whole variable part.
1540          */
1541         if (pathlen <= XFS_LITINO(mp))
1542                 fs_blocks = 0;
1543         else
1544                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
1545         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
1546         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
1547                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1548         if (error == ENOSPC && fs_blocks == 0) {
1549                 resblks = 0;
1550                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
1551                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1552         }
1553         if (error) {
1554                 cancel_flags = 0;
1555                 goto error_return;
1556         }
1557
1558         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1559         unlock_dp_on_error = B_TRUE;
1560
1561         /*
1562          * Check whether the directory allows new symlinks or not.
1563          */
1564         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
1565                 error = XFS_ERROR(EPERM);
1566                 goto error_return;
1567         }
1568
1569         /*
1570          * Reserve disk quota : blocks and inode.
1571          */
1572         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1573         if (error)
1574                 goto error_return;
1575
1576         /*
1577          * Check for ability to enter directory entry, if no space reserved.
1578          */
1579         error = xfs_dir_canenter(tp, dp, link_name, resblks);
1580         if (error)
1581                 goto error_return;
1582         /*
1583          * Initialize the bmap freelist prior to calling either
1584          * bmapi or the directory create code.
1585          */
1586         xfs_bmap_init(&free_list, &first_block);
1587
1588         /*
1589          * Allocate an inode for the symlink.
1590          */
1591         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
1592                                prid, resblks > 0, &ip, NULL);
1593         if (error) {
1594                 if (error == ENOSPC)
1595                         goto error_return;
1596                 goto error1;
1597         }
1598
1599         /*
1600          * An error after we've joined dp to the transaction will result in the
1601          * transaction cancel unlocking dp so don't do it explicitly in the
1602          * error path.
1603          */
1604         xfs_trans_ijoin_ref(tp, dp, XFS_ILOCK_EXCL);
1605         unlock_dp_on_error = B_FALSE;
1606
1607         /*
1608          * Also attach the dquot(s) to it, if applicable.
1609          */
1610         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1611
1612         if (resblks)
1613                 resblks -= XFS_IALLOC_SPACE_RES(mp);
1614         /*
1615          * If the symlink will fit into the inode, write it inline.
1616          */
1617         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
1618                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
1619                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
1620                 ip->i_d.di_size = pathlen;
1621
1622                 /*
1623                  * The inode was initially created in extent format.
1624                  */
1625                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
1626                 ip->i_df.if_flags |= XFS_IFINLINE;
1627
1628                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
1629                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
1630
1631         } else {
1632                 first_fsb = 0;
1633                 nmaps = SYMLINK_MAPS;
1634
1635                 error = xfs_bmapi(tp, ip, first_fsb, fs_blocks,
1636                                   XFS_BMAPI_WRITE | XFS_BMAPI_METADATA,
1637                                   &first_block, resblks, mval, &nmaps,
1638                                   &free_list);
1639                 if (error)
1640                         goto error2;
1641
1642                 if (resblks)
1643                         resblks -= fs_blocks;
1644                 ip->i_d.di_size = pathlen;
1645                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1646
1647                 cur_chunk = target_path;
1648                 for (n = 0; n < nmaps; n++) {
1649                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
1650                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
1651                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
1652                                                BTOBB(byte_cnt), 0);
1653                         ASSERT(!xfs_buf_geterror(bp));
1654                         if (pathlen < byte_cnt) {
1655                                 byte_cnt = pathlen;
1656                         }
1657                         pathlen -= byte_cnt;
1658
1659                         memcpy(bp->b_addr, cur_chunk, byte_cnt);
1660                         cur_chunk += byte_cnt;
1661
1662                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
1663                 }
1664         }
1665
1666         /*
1667          * Create the directory entry for the symlink.
1668          */
1669         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
1670                                         &first_block, &free_list, resblks);
1671         if (error)
1672                 goto error2;
1673         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1674         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1675
1676         /*
1677          * If this is a synchronous mount, make sure that the
1678          * symlink transaction goes to disk before returning to
1679          * the user.
1680          */
1681         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1682                 xfs_trans_set_sync(tp);
1683         }
1684
1685         error = xfs_bmap_finish(&tp, &free_list, &committed);
1686         if (error) {
1687                 goto error2;
1688         }
1689         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1690         xfs_qm_dqrele(udqp);
1691         xfs_qm_dqrele(gdqp);
1692
1693         *ipp = ip;
1694         return 0;
1695
1696  error2:
1697         IRELE(ip);
1698  error1:
1699         xfs_bmap_cancel(&free_list);
1700         cancel_flags |= XFS_TRANS_ABORT;
1701  error_return:
1702         xfs_trans_cancel(tp, cancel_flags);
1703         xfs_qm_dqrele(udqp);
1704         xfs_qm_dqrele(gdqp);
1705
1706         if (unlock_dp_on_error)
1707                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1708  std_return:
1709         return error;
1710 }
1711
1712 int
1713 xfs_set_dmattrs(
1714         xfs_inode_t     *ip,
1715         u_int           evmask,
1716         u_int16_t       state)
1717 {
1718         xfs_mount_t     *mp = ip->i_mount;
1719         xfs_trans_t     *tp;
1720         int             error;
1721
1722         if (!capable(CAP_SYS_ADMIN))
1723                 return XFS_ERROR(EPERM);
1724
1725         if (XFS_FORCED_SHUTDOWN(mp))
1726                 return XFS_ERROR(EIO);
1727
1728         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1729         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1730         if (error) {
1731                 xfs_trans_cancel(tp, 0);
1732                 return error;
1733         }
1734         xfs_ilock(ip, XFS_ILOCK_EXCL);
1735         xfs_trans_ijoin_ref(tp, ip, XFS_ILOCK_EXCL);
1736
1737         ip->i_d.di_dmevmask = evmask;
1738         ip->i_d.di_dmstate  = state;
1739
1740         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1741         error = xfs_trans_commit(tp, 0);
1742
1743         return error;
1744 }
1745
1746 /*
1747  * xfs_alloc_file_space()
1748  *      This routine allocates disk space for the given file.
1749  *
1750  *      If alloc_type == 0, this request is for an ALLOCSP type
1751  *      request which will change the file size.  In this case, no
1752  *      DMAPI event will be generated by the call.  A TRUNCATE event
1753  *      will be generated later by xfs_setattr.
1754  *
1755  *      If alloc_type != 0, this request is for a RESVSP type
1756  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1757  *      lower block boundary byte address is less than the file's
1758  *      length.
1759  *
1760  * RETURNS:
1761  *       0 on success
1762  *      errno on error
1763  *
1764  */
1765 STATIC int
1766 xfs_alloc_file_space(
1767         xfs_inode_t             *ip,
1768         xfs_off_t               offset,
1769         xfs_off_t               len,
1770         int                     alloc_type,
1771         int                     attr_flags)
1772 {
1773         xfs_mount_t             *mp = ip->i_mount;
1774         xfs_off_t               count;
1775         xfs_filblks_t           allocated_fsb;
1776         xfs_filblks_t           allocatesize_fsb;
1777         xfs_extlen_t            extsz, temp;
1778         xfs_fileoff_t           startoffset_fsb;
1779         xfs_fsblock_t           firstfsb;
1780         int                     nimaps;
1781         int                     bmapi_flag;
1782         int                     quota_flag;
1783         int                     rt;
1784         xfs_trans_t             *tp;
1785         xfs_bmbt_irec_t         imaps[1], *imapp;
1786         xfs_bmap_free_t         free_list;
1787         uint                    qblocks, resblks, resrtextents;
1788         int                     committed;
1789         int                     error;
1790
1791         trace_xfs_alloc_file_space(ip);
1792
1793         if (XFS_FORCED_SHUTDOWN(mp))
1794                 return XFS_ERROR(EIO);
1795
1796         error = xfs_qm_dqattach(ip, 0);
1797         if (error)
1798                 return error;
1799
1800         if (len <= 0)
1801                 return XFS_ERROR(EINVAL);
1802
1803         rt = XFS_IS_REALTIME_INODE(ip);
1804         extsz = xfs_get_extsz_hint(ip);
1805
1806         count = len;
1807         imapp = &imaps[0];
1808         nimaps = 1;
1809         bmapi_flag = XFS_BMAPI_WRITE | alloc_type;
1810         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1811         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1812
1813         /*
1814          * Allocate file space until done or until there is an error
1815          */
1816         while (allocatesize_fsb && !error) {
1817                 xfs_fileoff_t   s, e;
1818
1819                 /*
1820                  * Determine space reservations for data/realtime.
1821                  */
1822                 if (unlikely(extsz)) {
1823                         s = startoffset_fsb;
1824                         do_div(s, extsz);
1825                         s *= extsz;
1826                         e = startoffset_fsb + allocatesize_fsb;
1827                         if ((temp = do_mod(startoffset_fsb, extsz)))
1828                                 e += temp;
1829                         if ((temp = do_mod(e, extsz)))
1830                                 e += extsz - temp;
1831                 } else {
1832                         s = 0;
1833                         e = allocatesize_fsb;
1834                 }
1835
1836                 /*
1837                  * The transaction reservation is limited to a 32-bit block
1838                  * count, hence we need to limit the number of blocks we are
1839                  * trying to reserve to avoid an overflow. We can't allocate
1840                  * more than @nimaps extents, and an extent is limited on disk
1841                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1842                  */
1843                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1844                 if (unlikely(rt)) {
1845                         resrtextents = qblocks = resblks;
1846                         resrtextents /= mp->m_sb.sb_rextsize;
1847                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1848                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1849                 } else {
1850                         resrtextents = 0;
1851                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1852                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1853                 }
1854
1855                 /*
1856                  * Allocate and setup the transaction.
1857                  */
1858                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1859                 error = xfs_trans_reserve(tp, resblks,
1860                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1861                                           XFS_TRANS_PERM_LOG_RES,
1862                                           XFS_WRITE_LOG_COUNT);
1863                 /*
1864                  * Check for running out of space
1865                  */
1866                 if (error) {
1867                         /*
1868                          * Free the transaction structure.
1869                          */
1870                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1871                         xfs_trans_cancel(tp, 0);
1872                         break;
1873                 }
1874                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1875                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1876                                                       0, quota_flag);
1877                 if (error)
1878                         goto error1;
1879
1880                 xfs_trans_ijoin(tp, ip);
1881
1882                 /*
1883                  * Issue the xfs_bmapi() call to allocate the blocks
1884                  */
1885                 xfs_bmap_init(&free_list, &firstfsb);
1886                 error = xfs_bmapi(tp, ip, startoffset_fsb,
1887                                   allocatesize_fsb, bmapi_flag,
1888                                   &firstfsb, 0, imapp, &nimaps,
1889                                   &free_list);
1890                 if (error) {
1891                         goto error0;
1892                 }
1893
1894                 /*
1895                  * Complete the transaction
1896                  */
1897                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1898                 if (error) {
1899                         goto error0;
1900                 }
1901
1902                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1903                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1904                 if (error) {
1905                         break;
1906                 }
1907
1908                 allocated_fsb = imapp->br_blockcount;
1909
1910                 if (nimaps == 0) {
1911                         error = XFS_ERROR(ENOSPC);
1912                         break;
1913                 }
1914
1915                 startoffset_fsb += allocated_fsb;
1916                 allocatesize_fsb -= allocated_fsb;
1917         }
1918
1919         return error;
1920
1921 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1922         xfs_bmap_cancel(&free_list);
1923         xfs_trans_unreserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
1924
1925 error1: /* Just cancel transaction */
1926         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1927         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1928         return error;
1929 }
1930
1931 /*
1932  * Zero file bytes between startoff and endoff inclusive.
1933  * The iolock is held exclusive and no blocks are buffered.
1934  *
1935  * This function is used by xfs_free_file_space() to zero
1936  * partial blocks when the range to free is not block aligned.
1937  * When unreserving space with boundaries that are not block
1938  * aligned we round up the start and round down the end
1939  * boundaries and then use this function to zero the parts of
1940  * the blocks that got dropped during the rounding.
1941  */
1942 STATIC int
1943 xfs_zero_remaining_bytes(
1944         xfs_inode_t             *ip,
1945         xfs_off_t               startoff,
1946         xfs_off_t               endoff)
1947 {
1948         xfs_bmbt_irec_t         imap;
1949         xfs_fileoff_t           offset_fsb;
1950         xfs_off_t               lastoffset;
1951         xfs_off_t               offset;
1952         xfs_buf_t               *bp;
1953         xfs_mount_t             *mp = ip->i_mount;
1954         int                     nimap;
1955         int                     error = 0;
1956
1957         /*
1958          * Avoid doing I/O beyond eof - it's not necessary
1959          * since nothing can read beyond eof.  The space will
1960          * be zeroed when the file is extended anyway.
1961          */
1962         if (startoff >= ip->i_size)
1963                 return 0;
1964
1965         if (endoff > ip->i_size)
1966                 endoff = ip->i_size;
1967
1968         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1969                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1970                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
1971         if (!bp)
1972                 return XFS_ERROR(ENOMEM);
1973
1974         xfs_buf_unlock(bp);
1975
1976         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1977                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1978                 nimap = 1;
1979                 error = xfs_bmapi(NULL, ip, offset_fsb, 1, 0,
1980                         NULL, 0, &imap, &nimap, NULL);
1981                 if (error || nimap < 1)
1982                         break;
1983                 ASSERT(imap.br_blockcount >= 1);
1984                 ASSERT(imap.br_startoff == offset_fsb);
1985                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1986                 if (lastoffset > endoff)
1987                         lastoffset = endoff;
1988                 if (imap.br_startblock == HOLESTARTBLOCK)
1989                         continue;
1990                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1991                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1992                         continue;
1993                 XFS_BUF_UNDONE(bp);
1994                 XFS_BUF_UNWRITE(bp);
1995                 XFS_BUF_READ(bp);
1996                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1997                 xfsbdstrat(mp, bp);
1998                 error = xfs_buf_iowait(bp);
1999                 if (error) {
2000                         xfs_ioerror_alert("xfs_zero_remaining_bytes(read)",
2001                                           mp, bp, XFS_BUF_ADDR(bp));
2002                         break;
2003                 }
2004                 memset(bp->b_addr +
2005                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2006                       0, lastoffset - offset + 1);
2007                 XFS_BUF_UNDONE(bp);
2008                 XFS_BUF_UNREAD(bp);
2009                 XFS_BUF_WRITE(bp);
2010                 xfsbdstrat(mp, bp);
2011                 error = xfs_buf_iowait(bp);
2012                 if (error) {
2013                         xfs_ioerror_alert("xfs_zero_remaining_bytes(write)",
2014                                           mp, bp, XFS_BUF_ADDR(bp));
2015                         break;
2016                 }
2017         }
2018         xfs_buf_free(bp);
2019         return error;
2020 }
2021
2022 /*
2023  * xfs_free_file_space()
2024  *      This routine frees disk space for the given file.
2025  *
2026  *      This routine is only called by xfs_change_file_space
2027  *      for an UNRESVSP type call.
2028  *
2029  * RETURNS:
2030  *       0 on success
2031  *      errno on error
2032  *
2033  */
2034 STATIC int
2035 xfs_free_file_space(
2036         xfs_inode_t             *ip,
2037         xfs_off_t               offset,
2038         xfs_off_t               len,
2039         int                     attr_flags)
2040 {
2041         int                     committed;
2042         int                     done;
2043         xfs_fileoff_t           endoffset_fsb;
2044         int                     error;
2045         xfs_fsblock_t           firstfsb;
2046         xfs_bmap_free_t         free_list;
2047         xfs_bmbt_irec_t         imap;
2048         xfs_off_t               ioffset;
2049         xfs_extlen_t            mod=0;
2050         xfs_mount_t             *mp;
2051         int                     nimap;
2052         uint                    resblks;
2053         uint                    rounding;
2054         int                     rt;
2055         xfs_fileoff_t           startoffset_fsb;
2056         xfs_trans_t             *tp;
2057         int                     need_iolock = 1;
2058
2059         mp = ip->i_mount;
2060
2061         trace_xfs_free_file_space(ip);
2062
2063         error = xfs_qm_dqattach(ip, 0);
2064         if (error)
2065                 return error;
2066
2067         error = 0;
2068         if (len <= 0)   /* if nothing being freed */
2069                 return error;
2070         rt = XFS_IS_REALTIME_INODE(ip);
2071         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2072         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2073
2074         if (attr_flags & XFS_ATTR_NOLOCK)
2075                 need_iolock = 0;
2076         if (need_iolock) {
2077                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2078                 /* wait for the completion of any pending DIOs */
2079                 xfs_ioend_wait(ip);
2080         }
2081
2082         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2083         ioffset = offset & ~(rounding - 1);
2084
2085         if (VN_CACHED(VFS_I(ip)) != 0) {
2086                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2087                 if (error)
2088                         goto out_unlock_iolock;
2089         }
2090
2091         /*
2092          * Need to zero the stuff we're not freeing, on disk.
2093          * If it's a realtime file & can't use unwritten extents then we
2094          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2095          * will take care of it for us.
2096          */
2097         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2098                 nimap = 1;
2099                 error = xfs_bmapi(NULL, ip, startoffset_fsb,
2100                         1, 0, NULL, 0, &imap, &nimap, NULL);
2101                 if (error)
2102                         goto out_unlock_iolock;
2103                 ASSERT(nimap == 0 || nimap == 1);
2104                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2105                         xfs_daddr_t     block;
2106
2107                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2108                         block = imap.br_startblock;
2109                         mod = do_div(block, mp->m_sb.sb_rextsize);
2110                         if (mod)
2111                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2112                 }
2113                 nimap = 1;
2114                 error = xfs_bmapi(NULL, ip, endoffset_fsb - 1,
2115                         1, 0, NULL, 0, &imap, &nimap, NULL);
2116                 if (error)
2117                         goto out_unlock_iolock;
2118                 ASSERT(nimap == 0 || nimap == 1);
2119                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2120                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2121                         mod++;
2122                         if (mod && (mod != mp->m_sb.sb_rextsize))
2123                                 endoffset_fsb -= mod;
2124                 }
2125         }
2126         if ((done = (endoffset_fsb <= startoffset_fsb)))
2127                 /*
2128                  * One contiguous piece to clear
2129                  */
2130                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2131         else {
2132                 /*
2133                  * Some full blocks, possibly two pieces to clear
2134                  */
2135                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2136                         error = xfs_zero_remaining_bytes(ip, offset,
2137                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2138                 if (!error &&
2139                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2140                         error = xfs_zero_remaining_bytes(ip,
2141                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2142                                 offset + len - 1);
2143         }
2144
2145         /*
2146          * free file space until done or until there is an error
2147          */
2148         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2149         while (!error && !done) {
2150
2151                 /*
2152                  * allocate and setup the transaction. Allow this
2153                  * transaction to dip into the reserve blocks to ensure
2154                  * the freeing of the space succeeds at ENOSPC.
2155                  */
2156                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2157                 tp->t_flags |= XFS_TRANS_RESERVE;
2158                 error = xfs_trans_reserve(tp,
2159                                           resblks,
2160                                           XFS_WRITE_LOG_RES(mp),
2161                                           0,
2162                                           XFS_TRANS_PERM_LOG_RES,
2163                                           XFS_WRITE_LOG_COUNT);
2164
2165                 /*
2166                  * check for running out of space
2167                  */
2168                 if (error) {
2169                         /*
2170                          * Free the transaction structure.
2171                          */
2172                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2173                         xfs_trans_cancel(tp, 0);
2174                         break;
2175                 }
2176                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2177                 error = xfs_trans_reserve_quota(tp, mp,
2178                                 ip->i_udquot, ip->i_gdquot,
2179                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2180                 if (error)
2181                         goto error1;
2182
2183                 xfs_trans_ijoin(tp, ip);
2184
2185                 /*
2186                  * issue the bunmapi() call to free the blocks
2187                  */
2188                 xfs_bmap_init(&free_list, &firstfsb);
2189                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2190                                   endoffset_fsb - startoffset_fsb,
2191                                   0, 2, &firstfsb, &free_list, &done);
2192                 if (error) {
2193                         goto error0;
2194                 }
2195
2196                 /*
2197                  * complete the transaction
2198                  */
2199                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2200                 if (error) {
2201                         goto error0;
2202                 }
2203
2204                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2205                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2206         }
2207
2208  out_unlock_iolock:
2209         if (need_iolock)
2210                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2211         return error;
2212
2213  error0:
2214         xfs_bmap_cancel(&free_list);
2215  error1:
2216         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2217         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2218                     XFS_ILOCK_EXCL);
2219         return error;
2220 }
2221
2222 /*
2223  * xfs_change_file_space()
2224  *      This routine allocates or frees disk space for the given file.
2225  *      The user specified parameters are checked for alignment and size
2226  *      limitations.
2227  *
2228  * RETURNS:
2229  *       0 on success
2230  *      errno on error
2231  *
2232  */
2233 int
2234 xfs_change_file_space(
2235         xfs_inode_t     *ip,
2236         int             cmd,
2237         xfs_flock64_t   *bf,
2238         xfs_off_t       offset,
2239         int             attr_flags)
2240 {
2241         xfs_mount_t     *mp = ip->i_mount;
2242         int             clrprealloc;
2243         int             error;
2244         xfs_fsize_t     fsize;
2245         int             setprealloc;
2246         xfs_off_t       startoffset;
2247         xfs_off_t       llen;
2248         xfs_trans_t     *tp;
2249         struct iattr    iattr;
2250         int             prealloc_type;
2251
2252         if (!S_ISREG(ip->i_d.di_mode))
2253                 return XFS_ERROR(EINVAL);
2254
2255         switch (bf->l_whence) {
2256         case 0: /*SEEK_SET*/
2257                 break;
2258         case 1: /*SEEK_CUR*/
2259                 bf->l_start += offset;
2260                 break;
2261         case 2: /*SEEK_END*/
2262                 bf->l_start += ip->i_size;
2263                 break;
2264         default:
2265                 return XFS_ERROR(EINVAL);
2266         }
2267
2268         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2269
2270         if (   (bf->l_start < 0)
2271             || (bf->l_start > XFS_MAXIOFFSET(mp))
2272             || (bf->l_start + llen < 0)
2273             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2274                 return XFS_ERROR(EINVAL);
2275
2276         bf->l_whence = 0;
2277
2278         startoffset = bf->l_start;
2279         fsize = ip->i_size;
2280
2281         /*
2282          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2283          * file space.
2284          * These calls do NOT zero the data space allocated to the file,
2285          * nor do they change the file size.
2286          *
2287          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2288          * space.
2289          * These calls cause the new file data to be zeroed and the file
2290          * size to be changed.
2291          */
2292         setprealloc = clrprealloc = 0;
2293         prealloc_type = XFS_BMAPI_PREALLOC;
2294
2295         switch (cmd) {
2296         case XFS_IOC_ZERO_RANGE:
2297                 prealloc_type |= XFS_BMAPI_CONVERT;
2298                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2299                 /* FALLTHRU */
2300         case XFS_IOC_RESVSP:
2301         case XFS_IOC_RESVSP64:
2302                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2303                                                 prealloc_type, attr_flags);
2304                 if (error)
2305                         return error;
2306                 setprealloc = 1;
2307                 break;
2308
2309         case XFS_IOC_UNRESVSP:
2310         case XFS_IOC_UNRESVSP64:
2311                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2312                                                                 attr_flags)))
2313                         return error;
2314                 break;
2315
2316         case XFS_IOC_ALLOCSP:
2317         case XFS_IOC_ALLOCSP64:
2318         case XFS_IOC_FREESP:
2319         case XFS_IOC_FREESP64:
2320                 if (startoffset > fsize) {
2321                         error = xfs_alloc_file_space(ip, fsize,
2322                                         startoffset - fsize, 0, attr_flags);
2323                         if (error)
2324                                 break;
2325                 }
2326
2327                 iattr.ia_valid = ATTR_SIZE;
2328                 iattr.ia_size = startoffset;
2329
2330                 error = xfs_setattr_size(ip, &iattr, attr_flags);
2331
2332                 if (error)
2333                         return error;
2334
2335                 clrprealloc = 1;
2336                 break;
2337
2338         default:
2339                 ASSERT(0);
2340                 return XFS_ERROR(EINVAL);
2341         }
2342
2343         /*
2344          * update the inode timestamp, mode, and prealloc flag bits
2345          */
2346         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2347
2348         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2349                                       0, 0, 0))) {
2350                 /* ASSERT(0); */
2351                 xfs_trans_cancel(tp, 0);
2352                 return error;
2353         }
2354
2355         xfs_ilock(ip, XFS_ILOCK_EXCL);
2356
2357         xfs_trans_ijoin(tp, ip);
2358
2359         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2360                 ip->i_d.di_mode &= ~S_ISUID;
2361
2362                 /*
2363                  * Note that we don't have to worry about mandatory
2364                  * file locking being disabled here because we only
2365                  * clear the S_ISGID bit if the Group execute bit is
2366                  * on, but if it was on then mandatory locking wouldn't
2367                  * have been enabled.
2368                  */
2369                 if (ip->i_d.di_mode & S_IXGRP)
2370                         ip->i_d.di_mode &= ~S_ISGID;
2371
2372                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2373         }
2374         if (setprealloc)
2375                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2376         else if (clrprealloc)
2377                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2378
2379         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2380         if (attr_flags & XFS_ATTR_SYNC)
2381                 xfs_trans_set_sync(tp);
2382
2383         error = xfs_trans_commit(tp, 0);
2384
2385         xfs_iunlock(ip, XFS_ILOCK_EXCL);
2386
2387         return error;
2388 }