nfsd4: Fix filp leak
[linux-2.6.git] / fs / ocfs2 / file.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/capability.h>
27 #include <linux/fs.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
38 #include <linux/quotaops.h>
39 #include <linux/blkdev.h>
40
41 #include <cluster/masklog.h>
42
43 #include "ocfs2.h"
44
45 #include "alloc.h"
46 #include "aops.h"
47 #include "dir.h"
48 #include "dlmglue.h"
49 #include "extent_map.h"
50 #include "file.h"
51 #include "sysfile.h"
52 #include "inode.h"
53 #include "ioctl.h"
54 #include "journal.h"
55 #include "locks.h"
56 #include "mmap.h"
57 #include "suballoc.h"
58 #include "super.h"
59 #include "xattr.h"
60 #include "acl.h"
61 #include "quota.h"
62 #include "refcounttree.h"
63 #include "ocfs2_trace.h"
64
65 #include "buffer_head_io.h"
66
67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
68 {
69         struct ocfs2_file_private *fp;
70
71         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
72         if (!fp)
73                 return -ENOMEM;
74
75         fp->fp_file = file;
76         mutex_init(&fp->fp_mutex);
77         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78         file->private_data = fp;
79
80         return 0;
81 }
82
83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
84 {
85         struct ocfs2_file_private *fp = file->private_data;
86         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
87
88         if (fp) {
89                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90                 ocfs2_lock_res_free(&fp->fp_flock);
91                 kfree(fp);
92                 file->private_data = NULL;
93         }
94 }
95
96 static int ocfs2_file_open(struct inode *inode, struct file *file)
97 {
98         int status;
99         int mode = file->f_flags;
100         struct ocfs2_inode_info *oi = OCFS2_I(inode);
101
102         trace_ocfs2_file_open(inode, file, file->f_path.dentry,
103                               (unsigned long long)OCFS2_I(inode)->ip_blkno,
104                               file->f_path.dentry->d_name.len,
105                               file->f_path.dentry->d_name.name, mode);
106
107         if (file->f_mode & FMODE_WRITE)
108                 dquot_initialize(inode);
109
110         spin_lock(&oi->ip_lock);
111
112         /* Check that the inode hasn't been wiped from disk by another
113          * node. If it hasn't then we're safe as long as we hold the
114          * spin lock until our increment of open count. */
115         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
116                 spin_unlock(&oi->ip_lock);
117
118                 status = -ENOENT;
119                 goto leave;
120         }
121
122         if (mode & O_DIRECT)
123                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
124
125         oi->ip_open_count++;
126         spin_unlock(&oi->ip_lock);
127
128         status = ocfs2_init_file_private(inode, file);
129         if (status) {
130                 /*
131                  * We want to set open count back if we're failing the
132                  * open.
133                  */
134                 spin_lock(&oi->ip_lock);
135                 oi->ip_open_count--;
136                 spin_unlock(&oi->ip_lock);
137         }
138
139 leave:
140         return status;
141 }
142
143 static int ocfs2_file_release(struct inode *inode, struct file *file)
144 {
145         struct ocfs2_inode_info *oi = OCFS2_I(inode);
146
147         spin_lock(&oi->ip_lock);
148         if (!--oi->ip_open_count)
149                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
150
151         trace_ocfs2_file_release(inode, file, file->f_path.dentry,
152                                  oi->ip_blkno,
153                                  file->f_path.dentry->d_name.len,
154                                  file->f_path.dentry->d_name.name,
155                                  oi->ip_open_count);
156         spin_unlock(&oi->ip_lock);
157
158         ocfs2_free_file_private(inode, file);
159
160         return 0;
161 }
162
163 static int ocfs2_dir_open(struct inode *inode, struct file *file)
164 {
165         return ocfs2_init_file_private(inode, file);
166 }
167
168 static int ocfs2_dir_release(struct inode *inode, struct file *file)
169 {
170         ocfs2_free_file_private(inode, file);
171         return 0;
172 }
173
174 static int ocfs2_sync_file(struct file *file, int datasync)
175 {
176         int err = 0;
177         journal_t *journal;
178         struct inode *inode = file->f_mapping->host;
179         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
180
181         trace_ocfs2_sync_file(inode, file, file->f_path.dentry,
182                               OCFS2_I(inode)->ip_blkno,
183                               file->f_path.dentry->d_name.len,
184                               file->f_path.dentry->d_name.name,
185                               (unsigned long long)datasync);
186
187         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
188                 /*
189                  * We still have to flush drive's caches to get data to the
190                  * platter
191                  */
192                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
193                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
194                 goto bail;
195         }
196
197         journal = osb->journal->j_journal;
198         err = jbd2_journal_force_commit(journal);
199
200 bail:
201         if (err)
202                 mlog_errno(err);
203
204         return (err < 0) ? -EIO : 0;
205 }
206
207 int ocfs2_should_update_atime(struct inode *inode,
208                               struct vfsmount *vfsmnt)
209 {
210         struct timespec now;
211         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
212
213         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
214                 return 0;
215
216         if ((inode->i_flags & S_NOATIME) ||
217             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
218                 return 0;
219
220         /*
221          * We can be called with no vfsmnt structure - NFSD will
222          * sometimes do this.
223          *
224          * Note that our action here is different than touch_atime() -
225          * if we can't tell whether this is a noatime mount, then we
226          * don't know whether to trust the value of s_atime_quantum.
227          */
228         if (vfsmnt == NULL)
229                 return 0;
230
231         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
232             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
233                 return 0;
234
235         if (vfsmnt->mnt_flags & MNT_RELATIME) {
236                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
237                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
238                         return 1;
239
240                 return 0;
241         }
242
243         now = CURRENT_TIME;
244         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
245                 return 0;
246         else
247                 return 1;
248 }
249
250 int ocfs2_update_inode_atime(struct inode *inode,
251                              struct buffer_head *bh)
252 {
253         int ret;
254         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
255         handle_t *handle;
256         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
257
258         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
259         if (IS_ERR(handle)) {
260                 ret = PTR_ERR(handle);
261                 mlog_errno(ret);
262                 goto out;
263         }
264
265         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
266                                       OCFS2_JOURNAL_ACCESS_WRITE);
267         if (ret) {
268                 mlog_errno(ret);
269                 goto out_commit;
270         }
271
272         /*
273          * Don't use ocfs2_mark_inode_dirty() here as we don't always
274          * have i_mutex to guard against concurrent changes to other
275          * inode fields.
276          */
277         inode->i_atime = CURRENT_TIME;
278         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
279         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
280         ocfs2_journal_dirty(handle, bh);
281
282 out_commit:
283         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
284 out:
285         return ret;
286 }
287
288 static int ocfs2_set_inode_size(handle_t *handle,
289                                 struct inode *inode,
290                                 struct buffer_head *fe_bh,
291                                 u64 new_i_size)
292 {
293         int status;
294
295         i_size_write(inode, new_i_size);
296         inode->i_blocks = ocfs2_inode_sector_count(inode);
297         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
298
299         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
300         if (status < 0) {
301                 mlog_errno(status);
302                 goto bail;
303         }
304
305 bail:
306         return status;
307 }
308
309 int ocfs2_simple_size_update(struct inode *inode,
310                              struct buffer_head *di_bh,
311                              u64 new_i_size)
312 {
313         int ret;
314         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
315         handle_t *handle = NULL;
316
317         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
318         if (IS_ERR(handle)) {
319                 ret = PTR_ERR(handle);
320                 mlog_errno(ret);
321                 goto out;
322         }
323
324         ret = ocfs2_set_inode_size(handle, inode, di_bh,
325                                    new_i_size);
326         if (ret < 0)
327                 mlog_errno(ret);
328
329         ocfs2_commit_trans(osb, handle);
330 out:
331         return ret;
332 }
333
334 static int ocfs2_cow_file_pos(struct inode *inode,
335                               struct buffer_head *fe_bh,
336                               u64 offset)
337 {
338         int status;
339         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
340         unsigned int num_clusters = 0;
341         unsigned int ext_flags = 0;
342
343         /*
344          * If the new offset is aligned to the range of the cluster, there is
345          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
346          * CoW either.
347          */
348         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
349                 return 0;
350
351         status = ocfs2_get_clusters(inode, cpos, &phys,
352                                     &num_clusters, &ext_flags);
353         if (status) {
354                 mlog_errno(status);
355                 goto out;
356         }
357
358         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
359                 goto out;
360
361         return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
362
363 out:
364         return status;
365 }
366
367 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
368                                      struct inode *inode,
369                                      struct buffer_head *fe_bh,
370                                      u64 new_i_size)
371 {
372         int status;
373         handle_t *handle;
374         struct ocfs2_dinode *di;
375         u64 cluster_bytes;
376
377         /*
378          * We need to CoW the cluster contains the offset if it is reflinked
379          * since we will call ocfs2_zero_range_for_truncate later which will
380          * write "0" from offset to the end of the cluster.
381          */
382         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
383         if (status) {
384                 mlog_errno(status);
385                 return status;
386         }
387
388         /* TODO: This needs to actually orphan the inode in this
389          * transaction. */
390
391         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
392         if (IS_ERR(handle)) {
393                 status = PTR_ERR(handle);
394                 mlog_errno(status);
395                 goto out;
396         }
397
398         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
399                                          OCFS2_JOURNAL_ACCESS_WRITE);
400         if (status < 0) {
401                 mlog_errno(status);
402                 goto out_commit;
403         }
404
405         /*
406          * Do this before setting i_size.
407          */
408         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
409         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
410                                                cluster_bytes);
411         if (status) {
412                 mlog_errno(status);
413                 goto out_commit;
414         }
415
416         i_size_write(inode, new_i_size);
417         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
418
419         di = (struct ocfs2_dinode *) fe_bh->b_data;
420         di->i_size = cpu_to_le64(new_i_size);
421         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
422         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
423
424         ocfs2_journal_dirty(handle, fe_bh);
425
426 out_commit:
427         ocfs2_commit_trans(osb, handle);
428 out:
429         return status;
430 }
431
432 static int ocfs2_truncate_file(struct inode *inode,
433                                struct buffer_head *di_bh,
434                                u64 new_i_size)
435 {
436         int status = 0;
437         struct ocfs2_dinode *fe = NULL;
438         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
439
440         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
441          * already validated it */
442         fe = (struct ocfs2_dinode *) di_bh->b_data;
443
444         trace_ocfs2_truncate_file((unsigned long long)OCFS2_I(inode)->ip_blkno,
445                                   (unsigned long long)le64_to_cpu(fe->i_size),
446                                   (unsigned long long)new_i_size);
447
448         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
449                         "Inode %llu, inode i_size = %lld != di "
450                         "i_size = %llu, i_flags = 0x%x\n",
451                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
452                         i_size_read(inode),
453                         (unsigned long long)le64_to_cpu(fe->i_size),
454                         le32_to_cpu(fe->i_flags));
455
456         if (new_i_size > le64_to_cpu(fe->i_size)) {
457                 trace_ocfs2_truncate_file_error(
458                         (unsigned long long)le64_to_cpu(fe->i_size),
459                         (unsigned long long)new_i_size);
460                 status = -EINVAL;
461                 mlog_errno(status);
462                 goto bail;
463         }
464
465         /* lets handle the simple truncate cases before doing any more
466          * cluster locking. */
467         if (new_i_size == le64_to_cpu(fe->i_size))
468                 goto bail;
469
470         down_write(&OCFS2_I(inode)->ip_alloc_sem);
471
472         ocfs2_resv_discard(&osb->osb_la_resmap,
473                            &OCFS2_I(inode)->ip_la_data_resv);
474
475         /*
476          * The inode lock forced other nodes to sync and drop their
477          * pages, which (correctly) happens even if we have a truncate
478          * without allocation change - ocfs2 cluster sizes can be much
479          * greater than page size, so we have to truncate them
480          * anyway.
481          */
482         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
483         truncate_inode_pages(inode->i_mapping, new_i_size);
484
485         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
486                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
487                                                i_size_read(inode), 1);
488                 if (status)
489                         mlog_errno(status);
490
491                 goto bail_unlock_sem;
492         }
493
494         /* alright, we're going to need to do a full blown alloc size
495          * change. Orphan the inode so that recovery can complete the
496          * truncate if necessary. This does the task of marking
497          * i_size. */
498         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
499         if (status < 0) {
500                 mlog_errno(status);
501                 goto bail_unlock_sem;
502         }
503
504         status = ocfs2_commit_truncate(osb, inode, di_bh);
505         if (status < 0) {
506                 mlog_errno(status);
507                 goto bail_unlock_sem;
508         }
509
510         /* TODO: orphan dir cleanup here. */
511 bail_unlock_sem:
512         up_write(&OCFS2_I(inode)->ip_alloc_sem);
513
514 bail:
515         if (!status && OCFS2_I(inode)->ip_clusters == 0)
516                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
517
518         return status;
519 }
520
521 /*
522  * extend file allocation only here.
523  * we'll update all the disk stuff, and oip->alloc_size
524  *
525  * expect stuff to be locked, a transaction started and enough data /
526  * metadata reservations in the contexts.
527  *
528  * Will return -EAGAIN, and a reason if a restart is needed.
529  * If passed in, *reason will always be set, even in error.
530  */
531 int ocfs2_add_inode_data(struct ocfs2_super *osb,
532                          struct inode *inode,
533                          u32 *logical_offset,
534                          u32 clusters_to_add,
535                          int mark_unwritten,
536                          struct buffer_head *fe_bh,
537                          handle_t *handle,
538                          struct ocfs2_alloc_context *data_ac,
539                          struct ocfs2_alloc_context *meta_ac,
540                          enum ocfs2_alloc_restarted *reason_ret)
541 {
542         int ret;
543         struct ocfs2_extent_tree et;
544
545         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
546         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
547                                           clusters_to_add, mark_unwritten,
548                                           data_ac, meta_ac, reason_ret);
549
550         return ret;
551 }
552
553 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
554                                      u32 clusters_to_add, int mark_unwritten)
555 {
556         int status = 0;
557         int restart_func = 0;
558         int credits;
559         u32 prev_clusters;
560         struct buffer_head *bh = NULL;
561         struct ocfs2_dinode *fe = NULL;
562         handle_t *handle = NULL;
563         struct ocfs2_alloc_context *data_ac = NULL;
564         struct ocfs2_alloc_context *meta_ac = NULL;
565         enum ocfs2_alloc_restarted why;
566         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
567         struct ocfs2_extent_tree et;
568         int did_quota = 0;
569
570         /*
571          * This function only exists for file systems which don't
572          * support holes.
573          */
574         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
575
576         status = ocfs2_read_inode_block(inode, &bh);
577         if (status < 0) {
578                 mlog_errno(status);
579                 goto leave;
580         }
581         fe = (struct ocfs2_dinode *) bh->b_data;
582
583 restart_all:
584         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
585
586         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
587         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
588                                        &data_ac, &meta_ac);
589         if (status) {
590                 mlog_errno(status);
591                 goto leave;
592         }
593
594         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
595                                             clusters_to_add);
596         handle = ocfs2_start_trans(osb, credits);
597         if (IS_ERR(handle)) {
598                 status = PTR_ERR(handle);
599                 handle = NULL;
600                 mlog_errno(status);
601                 goto leave;
602         }
603
604 restarted_transaction:
605         trace_ocfs2_extend_allocation(
606                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
607                 (unsigned long long)i_size_read(inode),
608                 le32_to_cpu(fe->i_clusters), clusters_to_add,
609                 why, restart_func);
610
611         status = dquot_alloc_space_nodirty(inode,
612                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
613         if (status)
614                 goto leave;
615         did_quota = 1;
616
617         /* reserve a write to the file entry early on - that we if we
618          * run out of credits in the allocation path, we can still
619          * update i_size. */
620         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
621                                          OCFS2_JOURNAL_ACCESS_WRITE);
622         if (status < 0) {
623                 mlog_errno(status);
624                 goto leave;
625         }
626
627         prev_clusters = OCFS2_I(inode)->ip_clusters;
628
629         status = ocfs2_add_inode_data(osb,
630                                       inode,
631                                       &logical_start,
632                                       clusters_to_add,
633                                       mark_unwritten,
634                                       bh,
635                                       handle,
636                                       data_ac,
637                                       meta_ac,
638                                       &why);
639         if ((status < 0) && (status != -EAGAIN)) {
640                 if (status != -ENOSPC)
641                         mlog_errno(status);
642                 goto leave;
643         }
644
645         ocfs2_journal_dirty(handle, bh);
646
647         spin_lock(&OCFS2_I(inode)->ip_lock);
648         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
649         spin_unlock(&OCFS2_I(inode)->ip_lock);
650         /* Release unused quota reservation */
651         dquot_free_space(inode,
652                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
653         did_quota = 0;
654
655         if (why != RESTART_NONE && clusters_to_add) {
656                 if (why == RESTART_META) {
657                         restart_func = 1;
658                         status = 0;
659                 } else {
660                         BUG_ON(why != RESTART_TRANS);
661
662                         /* TODO: This can be more intelligent. */
663                         credits = ocfs2_calc_extend_credits(osb->sb,
664                                                             &fe->id2.i_list,
665                                                             clusters_to_add);
666                         status = ocfs2_extend_trans(handle, credits);
667                         if (status < 0) {
668                                 /* handle still has to be committed at
669                                  * this point. */
670                                 status = -ENOMEM;
671                                 mlog_errno(status);
672                                 goto leave;
673                         }
674                         goto restarted_transaction;
675                 }
676         }
677
678         trace_ocfs2_extend_allocation_end(OCFS2_I(inode)->ip_blkno,
679              le32_to_cpu(fe->i_clusters),
680              (unsigned long long)le64_to_cpu(fe->i_size),
681              OCFS2_I(inode)->ip_clusters,
682              (unsigned long long)i_size_read(inode));
683
684 leave:
685         if (status < 0 && did_quota)
686                 dquot_free_space(inode,
687                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
688         if (handle) {
689                 ocfs2_commit_trans(osb, handle);
690                 handle = NULL;
691         }
692         if (data_ac) {
693                 ocfs2_free_alloc_context(data_ac);
694                 data_ac = NULL;
695         }
696         if (meta_ac) {
697                 ocfs2_free_alloc_context(meta_ac);
698                 meta_ac = NULL;
699         }
700         if ((!status) && restart_func) {
701                 restart_func = 0;
702                 goto restart_all;
703         }
704         brelse(bh);
705         bh = NULL;
706
707         return status;
708 }
709
710 /*
711  * While a write will already be ordering the data, a truncate will not.
712  * Thus, we need to explicitly order the zeroed pages.
713  */
714 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
715 {
716         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
717         handle_t *handle = NULL;
718         int ret = 0;
719
720         if (!ocfs2_should_order_data(inode))
721                 goto out;
722
723         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
724         if (IS_ERR(handle)) {
725                 ret = -ENOMEM;
726                 mlog_errno(ret);
727                 goto out;
728         }
729
730         ret = ocfs2_jbd2_file_inode(handle, inode);
731         if (ret < 0)
732                 mlog_errno(ret);
733
734 out:
735         if (ret) {
736                 if (!IS_ERR(handle))
737                         ocfs2_commit_trans(osb, handle);
738                 handle = ERR_PTR(ret);
739         }
740         return handle;
741 }
742
743 /* Some parts of this taken from generic_cont_expand, which turned out
744  * to be too fragile to do exactly what we need without us having to
745  * worry about recursive locking in ->write_begin() and ->write_end(). */
746 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
747                                  u64 abs_to)
748 {
749         struct address_space *mapping = inode->i_mapping;
750         struct page *page;
751         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
752         handle_t *handle = NULL;
753         int ret = 0;
754         unsigned zero_from, zero_to, block_start, block_end;
755
756         BUG_ON(abs_from >= abs_to);
757         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
758         BUG_ON(abs_from & (inode->i_blkbits - 1));
759
760         page = find_or_create_page(mapping, index, GFP_NOFS);
761         if (!page) {
762                 ret = -ENOMEM;
763                 mlog_errno(ret);
764                 goto out;
765         }
766
767         /* Get the offsets within the page that we want to zero */
768         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
769         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
770         if (!zero_to)
771                 zero_to = PAGE_CACHE_SIZE;
772
773         trace_ocfs2_write_zero_page(
774                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
775                         (unsigned long long)abs_from,
776                         (unsigned long long)abs_to,
777                         index, zero_from, zero_to);
778
779         /* We know that zero_from is block aligned */
780         for (block_start = zero_from; block_start < zero_to;
781              block_start = block_end) {
782                 block_end = block_start + (1 << inode->i_blkbits);
783
784                 /*
785                  * block_start is block-aligned.  Bump it by one to force
786                  * __block_write_begin and block_commit_write to zero the
787                  * whole block.
788                  */
789                 ret = __block_write_begin(page, block_start + 1, 0,
790                                           ocfs2_get_block);
791                 if (ret < 0) {
792                         mlog_errno(ret);
793                         goto out_unlock;
794                 }
795
796                 if (!handle) {
797                         handle = ocfs2_zero_start_ordered_transaction(inode);
798                         if (IS_ERR(handle)) {
799                                 ret = PTR_ERR(handle);
800                                 handle = NULL;
801                                 break;
802                         }
803                 }
804
805                 /* must not update i_size! */
806                 ret = block_commit_write(page, block_start + 1,
807                                          block_start + 1);
808                 if (ret < 0)
809                         mlog_errno(ret);
810                 else
811                         ret = 0;
812         }
813
814         if (handle)
815                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
816
817 out_unlock:
818         unlock_page(page);
819         page_cache_release(page);
820 out:
821         return ret;
822 }
823
824 /*
825  * Find the next range to zero.  We do this in terms of bytes because
826  * that's what ocfs2_zero_extend() wants, and it is dealing with the
827  * pagecache.  We may return multiple extents.
828  *
829  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
830  * needs to be zeroed.  range_start and range_end return the next zeroing
831  * range.  A subsequent call should pass the previous range_end as its
832  * zero_start.  If range_end is 0, there's nothing to do.
833  *
834  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
835  */
836 static int ocfs2_zero_extend_get_range(struct inode *inode,
837                                        struct buffer_head *di_bh,
838                                        u64 zero_start, u64 zero_end,
839                                        u64 *range_start, u64 *range_end)
840 {
841         int rc = 0, needs_cow = 0;
842         u32 p_cpos, zero_clusters = 0;
843         u32 zero_cpos =
844                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
845         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
846         unsigned int num_clusters = 0;
847         unsigned int ext_flags = 0;
848
849         while (zero_cpos < last_cpos) {
850                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
851                                         &num_clusters, &ext_flags);
852                 if (rc) {
853                         mlog_errno(rc);
854                         goto out;
855                 }
856
857                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
858                         zero_clusters = num_clusters;
859                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
860                                 needs_cow = 1;
861                         break;
862                 }
863
864                 zero_cpos += num_clusters;
865         }
866         if (!zero_clusters) {
867                 *range_end = 0;
868                 goto out;
869         }
870
871         while ((zero_cpos + zero_clusters) < last_cpos) {
872                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
873                                         &p_cpos, &num_clusters,
874                                         &ext_flags);
875                 if (rc) {
876                         mlog_errno(rc);
877                         goto out;
878                 }
879
880                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
881                         break;
882                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
883                         needs_cow = 1;
884                 zero_clusters += num_clusters;
885         }
886         if ((zero_cpos + zero_clusters) > last_cpos)
887                 zero_clusters = last_cpos - zero_cpos;
888
889         if (needs_cow) {
890                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
891                                         zero_clusters, UINT_MAX);
892                 if (rc) {
893                         mlog_errno(rc);
894                         goto out;
895                 }
896         }
897
898         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
899         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
900                                              zero_cpos + zero_clusters);
901
902 out:
903         return rc;
904 }
905
906 /*
907  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
908  * has made sure that the entire range needs zeroing.
909  */
910 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
911                                    u64 range_end)
912 {
913         int rc = 0;
914         u64 next_pos;
915         u64 zero_pos = range_start;
916
917         trace_ocfs2_zero_extend_range(
918                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
919                         (unsigned long long)range_start,
920                         (unsigned long long)range_end);
921         BUG_ON(range_start >= range_end);
922
923         while (zero_pos < range_end) {
924                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
925                 if (next_pos > range_end)
926                         next_pos = range_end;
927                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
928                 if (rc < 0) {
929                         mlog_errno(rc);
930                         break;
931                 }
932                 zero_pos = next_pos;
933
934                 /*
935                  * Very large extends have the potential to lock up
936                  * the cpu for extended periods of time.
937                  */
938                 cond_resched();
939         }
940
941         return rc;
942 }
943
944 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
945                       loff_t zero_to_size)
946 {
947         int ret = 0;
948         u64 zero_start, range_start = 0, range_end = 0;
949         struct super_block *sb = inode->i_sb;
950
951         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
952         trace_ocfs2_zero_extend((unsigned long long)OCFS2_I(inode)->ip_blkno,
953                                 (unsigned long long)zero_start,
954                                 (unsigned long long)i_size_read(inode));
955         while (zero_start < zero_to_size) {
956                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
957                                                   zero_to_size,
958                                                   &range_start,
959                                                   &range_end);
960                 if (ret) {
961                         mlog_errno(ret);
962                         break;
963                 }
964                 if (!range_end)
965                         break;
966                 /* Trim the ends */
967                 if (range_start < zero_start)
968                         range_start = zero_start;
969                 if (range_end > zero_to_size)
970                         range_end = zero_to_size;
971
972                 ret = ocfs2_zero_extend_range(inode, range_start,
973                                               range_end);
974                 if (ret) {
975                         mlog_errno(ret);
976                         break;
977                 }
978                 zero_start = range_end;
979         }
980
981         return ret;
982 }
983
984 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
985                           u64 new_i_size, u64 zero_to)
986 {
987         int ret;
988         u32 clusters_to_add;
989         struct ocfs2_inode_info *oi = OCFS2_I(inode);
990
991         /*
992          * Only quota files call this without a bh, and they can't be
993          * refcounted.
994          */
995         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
996         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
997
998         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
999         if (clusters_to_add < oi->ip_clusters)
1000                 clusters_to_add = 0;
1001         else
1002                 clusters_to_add -= oi->ip_clusters;
1003
1004         if (clusters_to_add) {
1005                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1006                                                 clusters_to_add, 0);
1007                 if (ret) {
1008                         mlog_errno(ret);
1009                         goto out;
1010                 }
1011         }
1012
1013         /*
1014          * Call this even if we don't add any clusters to the tree. We
1015          * still need to zero the area between the old i_size and the
1016          * new i_size.
1017          */
1018         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1019         if (ret < 0)
1020                 mlog_errno(ret);
1021
1022 out:
1023         return ret;
1024 }
1025
1026 static int ocfs2_extend_file(struct inode *inode,
1027                              struct buffer_head *di_bh,
1028                              u64 new_i_size)
1029 {
1030         int ret = 0;
1031         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1032
1033         BUG_ON(!di_bh);
1034
1035         /* setattr sometimes calls us like this. */
1036         if (new_i_size == 0)
1037                 goto out;
1038
1039         if (i_size_read(inode) == new_i_size)
1040                 goto out;
1041         BUG_ON(new_i_size < i_size_read(inode));
1042
1043         /*
1044          * The alloc sem blocks people in read/write from reading our
1045          * allocation until we're done changing it. We depend on
1046          * i_mutex to block other extend/truncate calls while we're
1047          * here.  We even have to hold it for sparse files because there
1048          * might be some tail zeroing.
1049          */
1050         down_write(&oi->ip_alloc_sem);
1051
1052         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1053                 /*
1054                  * We can optimize small extends by keeping the inodes
1055                  * inline data.
1056                  */
1057                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1058                         up_write(&oi->ip_alloc_sem);
1059                         goto out_update_size;
1060                 }
1061
1062                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1063                 if (ret) {
1064                         up_write(&oi->ip_alloc_sem);
1065                         mlog_errno(ret);
1066                         goto out;
1067                 }
1068         }
1069
1070         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1071                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1072         else
1073                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1074                                             new_i_size);
1075
1076         up_write(&oi->ip_alloc_sem);
1077
1078         if (ret < 0) {
1079                 mlog_errno(ret);
1080                 goto out;
1081         }
1082
1083 out_update_size:
1084         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1085         if (ret < 0)
1086                 mlog_errno(ret);
1087
1088 out:
1089         return ret;
1090 }
1091
1092 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1093 {
1094         int status = 0, size_change;
1095         struct inode *inode = dentry->d_inode;
1096         struct super_block *sb = inode->i_sb;
1097         struct ocfs2_super *osb = OCFS2_SB(sb);
1098         struct buffer_head *bh = NULL;
1099         handle_t *handle = NULL;
1100         struct dquot *transfer_to[MAXQUOTAS] = { };
1101         int qtype;
1102
1103         trace_ocfs2_setattr(inode, dentry,
1104                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1105                             dentry->d_name.len, dentry->d_name.name,
1106                             attr->ia_valid, attr->ia_mode,
1107                             attr->ia_uid, attr->ia_gid);
1108
1109         /* ensuring we don't even attempt to truncate a symlink */
1110         if (S_ISLNK(inode->i_mode))
1111                 attr->ia_valid &= ~ATTR_SIZE;
1112
1113 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1114                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1115         if (!(attr->ia_valid & OCFS2_VALID_ATTRS))
1116                 return 0;
1117
1118         status = inode_change_ok(inode, attr);
1119         if (status)
1120                 return status;
1121
1122         if (is_quota_modification(inode, attr))
1123                 dquot_initialize(inode);
1124         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1125         if (size_change) {
1126                 status = ocfs2_rw_lock(inode, 1);
1127                 if (status < 0) {
1128                         mlog_errno(status);
1129                         goto bail;
1130                 }
1131         }
1132
1133         status = ocfs2_inode_lock(inode, &bh, 1);
1134         if (status < 0) {
1135                 if (status != -ENOENT)
1136                         mlog_errno(status);
1137                 goto bail_unlock_rw;
1138         }
1139
1140         if (size_change && attr->ia_size != i_size_read(inode)) {
1141                 status = inode_newsize_ok(inode, attr->ia_size);
1142                 if (status)
1143                         goto bail_unlock;
1144
1145                 if (i_size_read(inode) > attr->ia_size) {
1146                         if (ocfs2_should_order_data(inode)) {
1147                                 status = ocfs2_begin_ordered_truncate(inode,
1148                                                                       attr->ia_size);
1149                                 if (status)
1150                                         goto bail_unlock;
1151                         }
1152                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1153                 } else
1154                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1155                 if (status < 0) {
1156                         if (status != -ENOSPC)
1157                                 mlog_errno(status);
1158                         status = -ENOSPC;
1159                         goto bail_unlock;
1160                 }
1161         }
1162
1163         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1164             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1165                 /*
1166                  * Gather pointers to quota structures so that allocation /
1167                  * freeing of quota structures happens here and not inside
1168                  * dquot_transfer() where we have problems with lock ordering
1169                  */
1170                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1171                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1172                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1173                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1174                                                       USRQUOTA);
1175                         if (!transfer_to[USRQUOTA]) {
1176                                 status = -ESRCH;
1177                                 goto bail_unlock;
1178                         }
1179                 }
1180                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1181                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1182                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1183                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1184                                                       GRPQUOTA);
1185                         if (!transfer_to[GRPQUOTA]) {
1186                                 status = -ESRCH;
1187                                 goto bail_unlock;
1188                         }
1189                 }
1190                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1191                                            2 * ocfs2_quota_trans_credits(sb));
1192                 if (IS_ERR(handle)) {
1193                         status = PTR_ERR(handle);
1194                         mlog_errno(status);
1195                         goto bail_unlock;
1196                 }
1197                 status = __dquot_transfer(inode, transfer_to);
1198                 if (status < 0)
1199                         goto bail_commit;
1200         } else {
1201                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1202                 if (IS_ERR(handle)) {
1203                         status = PTR_ERR(handle);
1204                         mlog_errno(status);
1205                         goto bail_unlock;
1206                 }
1207         }
1208
1209         /*
1210          * This will intentionally not wind up calling truncate_setsize(),
1211          * since all the work for a size change has been done above.
1212          * Otherwise, we could get into problems with truncate as
1213          * ip_alloc_sem is used there to protect against i_size
1214          * changes.
1215          *
1216          * XXX: this means the conditional below can probably be removed.
1217          */
1218         if ((attr->ia_valid & ATTR_SIZE) &&
1219             attr->ia_size != i_size_read(inode)) {
1220                 status = vmtruncate(inode, attr->ia_size);
1221                 if (status) {
1222                         mlog_errno(status);
1223                         goto bail_commit;
1224                 }
1225         }
1226
1227         setattr_copy(inode, attr);
1228         mark_inode_dirty(inode);
1229
1230         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1231         if (status < 0)
1232                 mlog_errno(status);
1233
1234 bail_commit:
1235         ocfs2_commit_trans(osb, handle);
1236 bail_unlock:
1237         ocfs2_inode_unlock(inode, 1);
1238 bail_unlock_rw:
1239         if (size_change)
1240                 ocfs2_rw_unlock(inode, 1);
1241 bail:
1242         brelse(bh);
1243
1244         /* Release quota pointers in case we acquired them */
1245         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1246                 dqput(transfer_to[qtype]);
1247
1248         if (!status && attr->ia_valid & ATTR_MODE) {
1249                 status = ocfs2_acl_chmod(inode);
1250                 if (status < 0)
1251                         mlog_errno(status);
1252         }
1253
1254         return status;
1255 }
1256
1257 int ocfs2_getattr(struct vfsmount *mnt,
1258                   struct dentry *dentry,
1259                   struct kstat *stat)
1260 {
1261         struct inode *inode = dentry->d_inode;
1262         struct super_block *sb = dentry->d_inode->i_sb;
1263         struct ocfs2_super *osb = sb->s_fs_info;
1264         int err;
1265
1266         err = ocfs2_inode_revalidate(dentry);
1267         if (err) {
1268                 if (err != -ENOENT)
1269                         mlog_errno(err);
1270                 goto bail;
1271         }
1272
1273         generic_fillattr(inode, stat);
1274
1275         /* We set the blksize from the cluster size for performance */
1276         stat->blksize = osb->s_clustersize;
1277
1278 bail:
1279         return err;
1280 }
1281
1282 int ocfs2_permission(struct inode *inode, int mask, unsigned int flags)
1283 {
1284         int ret;
1285
1286         if (flags & IPERM_FLAG_RCU)
1287                 return -ECHILD;
1288
1289         ret = ocfs2_inode_lock(inode, NULL, 0);
1290         if (ret) {
1291                 if (ret != -ENOENT)
1292                         mlog_errno(ret);
1293                 goto out;
1294         }
1295
1296         ret = generic_permission(inode, mask, flags, ocfs2_check_acl);
1297
1298         ocfs2_inode_unlock(inode, 0);
1299 out:
1300         return ret;
1301 }
1302
1303 static int __ocfs2_write_remove_suid(struct inode *inode,
1304                                      struct buffer_head *bh)
1305 {
1306         int ret;
1307         handle_t *handle;
1308         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1309         struct ocfs2_dinode *di;
1310
1311         trace_ocfs2_write_remove_suid(
1312                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1313                         inode->i_mode);
1314
1315         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1316         if (IS_ERR(handle)) {
1317                 ret = PTR_ERR(handle);
1318                 mlog_errno(ret);
1319                 goto out;
1320         }
1321
1322         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1323                                       OCFS2_JOURNAL_ACCESS_WRITE);
1324         if (ret < 0) {
1325                 mlog_errno(ret);
1326                 goto out_trans;
1327         }
1328
1329         inode->i_mode &= ~S_ISUID;
1330         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1331                 inode->i_mode &= ~S_ISGID;
1332
1333         di = (struct ocfs2_dinode *) bh->b_data;
1334         di->i_mode = cpu_to_le16(inode->i_mode);
1335
1336         ocfs2_journal_dirty(handle, bh);
1337
1338 out_trans:
1339         ocfs2_commit_trans(osb, handle);
1340 out:
1341         return ret;
1342 }
1343
1344 /*
1345  * Will look for holes and unwritten extents in the range starting at
1346  * pos for count bytes (inclusive).
1347  */
1348 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1349                                        size_t count)
1350 {
1351         int ret = 0;
1352         unsigned int extent_flags;
1353         u32 cpos, clusters, extent_len, phys_cpos;
1354         struct super_block *sb = inode->i_sb;
1355
1356         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1357         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1358
1359         while (clusters) {
1360                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1361                                          &extent_flags);
1362                 if (ret < 0) {
1363                         mlog_errno(ret);
1364                         goto out;
1365                 }
1366
1367                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1368                         ret = 1;
1369                         break;
1370                 }
1371
1372                 if (extent_len > clusters)
1373                         extent_len = clusters;
1374
1375                 clusters -= extent_len;
1376                 cpos += extent_len;
1377         }
1378 out:
1379         return ret;
1380 }
1381
1382 static int ocfs2_write_remove_suid(struct inode *inode)
1383 {
1384         int ret;
1385         struct buffer_head *bh = NULL;
1386
1387         ret = ocfs2_read_inode_block(inode, &bh);
1388         if (ret < 0) {
1389                 mlog_errno(ret);
1390                 goto out;
1391         }
1392
1393         ret =  __ocfs2_write_remove_suid(inode, bh);
1394 out:
1395         brelse(bh);
1396         return ret;
1397 }
1398
1399 /*
1400  * Allocate enough extents to cover the region starting at byte offset
1401  * start for len bytes. Existing extents are skipped, any extents
1402  * added are marked as "unwritten".
1403  */
1404 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1405                                             u64 start, u64 len)
1406 {
1407         int ret;
1408         u32 cpos, phys_cpos, clusters, alloc_size;
1409         u64 end = start + len;
1410         struct buffer_head *di_bh = NULL;
1411
1412         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1413                 ret = ocfs2_read_inode_block(inode, &di_bh);
1414                 if (ret) {
1415                         mlog_errno(ret);
1416                         goto out;
1417                 }
1418
1419                 /*
1420                  * Nothing to do if the requested reservation range
1421                  * fits within the inode.
1422                  */
1423                 if (ocfs2_size_fits_inline_data(di_bh, end))
1424                         goto out;
1425
1426                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1427                 if (ret) {
1428                         mlog_errno(ret);
1429                         goto out;
1430                 }
1431         }
1432
1433         /*
1434          * We consider both start and len to be inclusive.
1435          */
1436         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1437         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1438         clusters -= cpos;
1439
1440         while (clusters) {
1441                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1442                                          &alloc_size, NULL);
1443                 if (ret) {
1444                         mlog_errno(ret);
1445                         goto out;
1446                 }
1447
1448                 /*
1449                  * Hole or existing extent len can be arbitrary, so
1450                  * cap it to our own allocation request.
1451                  */
1452                 if (alloc_size > clusters)
1453                         alloc_size = clusters;
1454
1455                 if (phys_cpos) {
1456                         /*
1457                          * We already have an allocation at this
1458                          * region so we can safely skip it.
1459                          */
1460                         goto next;
1461                 }
1462
1463                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1464                 if (ret) {
1465                         if (ret != -ENOSPC)
1466                                 mlog_errno(ret);
1467                         goto out;
1468                 }
1469
1470 next:
1471                 cpos += alloc_size;
1472                 clusters -= alloc_size;
1473         }
1474
1475         ret = 0;
1476 out:
1477
1478         brelse(di_bh);
1479         return ret;
1480 }
1481
1482 /*
1483  * Truncate a byte range, avoiding pages within partial clusters. This
1484  * preserves those pages for the zeroing code to write to.
1485  */
1486 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1487                                          u64 byte_len)
1488 {
1489         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1490         loff_t start, end;
1491         struct address_space *mapping = inode->i_mapping;
1492
1493         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1494         end = byte_start + byte_len;
1495         end = end & ~(osb->s_clustersize - 1);
1496
1497         if (start < end) {
1498                 unmap_mapping_range(mapping, start, end - start, 0);
1499                 truncate_inode_pages_range(mapping, start, end - 1);
1500         }
1501 }
1502
1503 static int ocfs2_zero_partial_clusters(struct inode *inode,
1504                                        u64 start, u64 len)
1505 {
1506         int ret = 0;
1507         u64 tmpend, end = start + len;
1508         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1509         unsigned int csize = osb->s_clustersize;
1510         handle_t *handle;
1511
1512         /*
1513          * The "start" and "end" values are NOT necessarily part of
1514          * the range whose allocation is being deleted. Rather, this
1515          * is what the user passed in with the request. We must zero
1516          * partial clusters here. There's no need to worry about
1517          * physical allocation - the zeroing code knows to skip holes.
1518          */
1519         trace_ocfs2_zero_partial_clusters(
1520                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1521                 (unsigned long long)start, (unsigned long long)end);
1522
1523         /*
1524          * If both edges are on a cluster boundary then there's no
1525          * zeroing required as the region is part of the allocation to
1526          * be truncated.
1527          */
1528         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1529                 goto out;
1530
1531         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1532         if (IS_ERR(handle)) {
1533                 ret = PTR_ERR(handle);
1534                 mlog_errno(ret);
1535                 goto out;
1536         }
1537
1538         /*
1539          * We want to get the byte offset of the end of the 1st cluster.
1540          */
1541         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1542         if (tmpend > end)
1543                 tmpend = end;
1544
1545         trace_ocfs2_zero_partial_clusters_range1((unsigned long long)start,
1546                                                  (unsigned long long)tmpend);
1547
1548         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1549         if (ret)
1550                 mlog_errno(ret);
1551
1552         if (tmpend < end) {
1553                 /*
1554                  * This may make start and end equal, but the zeroing
1555                  * code will skip any work in that case so there's no
1556                  * need to catch it up here.
1557                  */
1558                 start = end & ~(osb->s_clustersize - 1);
1559
1560                 trace_ocfs2_zero_partial_clusters_range2(
1561                         (unsigned long long)start, (unsigned long long)end);
1562
1563                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1564                 if (ret)
1565                         mlog_errno(ret);
1566         }
1567
1568         ocfs2_commit_trans(osb, handle);
1569 out:
1570         return ret;
1571 }
1572
1573 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1574 {
1575         int i;
1576         struct ocfs2_extent_rec *rec = NULL;
1577
1578         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1579
1580                 rec = &el->l_recs[i];
1581
1582                 if (le32_to_cpu(rec->e_cpos) < pos)
1583                         break;
1584         }
1585
1586         return i;
1587 }
1588
1589 /*
1590  * Helper to calculate the punching pos and length in one run, we handle the
1591  * following three cases in order:
1592  *
1593  * - remove the entire record
1594  * - remove a partial record
1595  * - no record needs to be removed (hole-punching completed)
1596 */
1597 static void ocfs2_calc_trunc_pos(struct inode *inode,
1598                                  struct ocfs2_extent_list *el,
1599                                  struct ocfs2_extent_rec *rec,
1600                                  u32 trunc_start, u32 *trunc_cpos,
1601                                  u32 *trunc_len, u32 *trunc_end,
1602                                  u64 *blkno, int *done)
1603 {
1604         int ret = 0;
1605         u32 coff, range;
1606
1607         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1608
1609         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1610                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1611                 /*
1612                  * Skip holes if any.
1613                  */
1614                 if (range < *trunc_end)
1615                         *trunc_end = range;
1616                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1617                 *blkno = le64_to_cpu(rec->e_blkno);
1618                 *trunc_end = le32_to_cpu(rec->e_cpos);
1619         } else if (range > trunc_start) {
1620                 *trunc_cpos = trunc_start;
1621                 *trunc_len = *trunc_end - trunc_start;
1622                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1623                 *blkno = le64_to_cpu(rec->e_blkno) +
1624                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1625                 *trunc_end = trunc_start;
1626         } else {
1627                 /*
1628                  * It may have two following possibilities:
1629                  *
1630                  * - last record has been removed
1631                  * - trunc_start was within a hole
1632                  *
1633                  * both two cases mean the completion of hole punching.
1634                  */
1635                 ret = 1;
1636         }
1637
1638         *done = ret;
1639 }
1640
1641 static int ocfs2_remove_inode_range(struct inode *inode,
1642                                     struct buffer_head *di_bh, u64 byte_start,
1643                                     u64 byte_len)
1644 {
1645         int ret = 0, flags = 0, done = 0, i;
1646         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1647         u32 cluster_in_el;
1648         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1649         struct ocfs2_cached_dealloc_ctxt dealloc;
1650         struct address_space *mapping = inode->i_mapping;
1651         struct ocfs2_extent_tree et;
1652         struct ocfs2_path *path = NULL;
1653         struct ocfs2_extent_list *el = NULL;
1654         struct ocfs2_extent_rec *rec = NULL;
1655         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1656         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1657
1658         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1659         ocfs2_init_dealloc_ctxt(&dealloc);
1660
1661         trace_ocfs2_remove_inode_range(
1662                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1663                         (unsigned long long)byte_start,
1664                         (unsigned long long)byte_len);
1665
1666         if (byte_len == 0)
1667                 return 0;
1668
1669         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1670                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1671                                             byte_start + byte_len, 0);
1672                 if (ret) {
1673                         mlog_errno(ret);
1674                         goto out;
1675                 }
1676                 /*
1677                  * There's no need to get fancy with the page cache
1678                  * truncate of an inline-data inode. We're talking
1679                  * about less than a page here, which will be cached
1680                  * in the dinode buffer anyway.
1681                  */
1682                 unmap_mapping_range(mapping, 0, 0, 0);
1683                 truncate_inode_pages(mapping, 0);
1684                 goto out;
1685         }
1686
1687         /*
1688          * For reflinks, we may need to CoW 2 clusters which might be
1689          * partially zero'd later, if hole's start and end offset were
1690          * within one cluster(means is not exactly aligned to clustersize).
1691          */
1692
1693         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1694
1695                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1696                 if (ret) {
1697                         mlog_errno(ret);
1698                         goto out;
1699                 }
1700
1701                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1702                 if (ret) {
1703                         mlog_errno(ret);
1704                         goto out;
1705                 }
1706         }
1707
1708         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1709         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1710         cluster_in_el = trunc_end;
1711
1712         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1713         if (ret) {
1714                 mlog_errno(ret);
1715                 goto out;
1716         }
1717
1718         path = ocfs2_new_path_from_et(&et);
1719         if (!path) {
1720                 ret = -ENOMEM;
1721                 mlog_errno(ret);
1722                 goto out;
1723         }
1724
1725         while (trunc_end > trunc_start) {
1726
1727                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1728                                       cluster_in_el);
1729                 if (ret) {
1730                         mlog_errno(ret);
1731                         goto out;
1732                 }
1733
1734                 el = path_leaf_el(path);
1735
1736                 i = ocfs2_find_rec(el, trunc_end);
1737                 /*
1738                  * Need to go to previous extent block.
1739                  */
1740                 if (i < 0) {
1741                         if (path->p_tree_depth == 0)
1742                                 break;
1743
1744                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1745                                                             path,
1746                                                             &cluster_in_el);
1747                         if (ret) {
1748                                 mlog_errno(ret);
1749                                 goto out;
1750                         }
1751
1752                         /*
1753                          * We've reached the leftmost extent block,
1754                          * it's safe to leave.
1755                          */
1756                         if (cluster_in_el == 0)
1757                                 break;
1758
1759                         /*
1760                          * The 'pos' searched for previous extent block is
1761                          * always one cluster less than actual trunc_end.
1762                          */
1763                         trunc_end = cluster_in_el + 1;
1764
1765                         ocfs2_reinit_path(path, 1);
1766
1767                         continue;
1768
1769                 } else
1770                         rec = &el->l_recs[i];
1771
1772                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1773                                      &trunc_len, &trunc_end, &blkno, &done);
1774                 if (done)
1775                         break;
1776
1777                 flags = rec->e_flags;
1778                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1779
1780                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1781                                                phys_cpos, trunc_len, flags,
1782                                                &dealloc, refcount_loc);
1783                 if (ret < 0) {
1784                         mlog_errno(ret);
1785                         goto out;
1786                 }
1787
1788                 cluster_in_el = trunc_end;
1789
1790                 ocfs2_reinit_path(path, 1);
1791         }
1792
1793         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1794
1795 out:
1796         ocfs2_schedule_truncate_log_flush(osb, 1);
1797         ocfs2_run_deallocs(osb, &dealloc);
1798
1799         return ret;
1800 }
1801
1802 /*
1803  * Parts of this function taken from xfs_change_file_space()
1804  */
1805 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1806                                      loff_t f_pos, unsigned int cmd,
1807                                      struct ocfs2_space_resv *sr,
1808                                      int change_size)
1809 {
1810         int ret;
1811         s64 llen;
1812         loff_t size;
1813         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1814         struct buffer_head *di_bh = NULL;
1815         handle_t *handle;
1816         unsigned long long max_off = inode->i_sb->s_maxbytes;
1817
1818         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1819                 return -EROFS;
1820
1821         mutex_lock(&inode->i_mutex);
1822
1823         /*
1824          * This prevents concurrent writes on other nodes
1825          */
1826         ret = ocfs2_rw_lock(inode, 1);
1827         if (ret) {
1828                 mlog_errno(ret);
1829                 goto out;
1830         }
1831
1832         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1833         if (ret) {
1834                 mlog_errno(ret);
1835                 goto out_rw_unlock;
1836         }
1837
1838         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1839                 ret = -EPERM;
1840                 goto out_inode_unlock;
1841         }
1842
1843         switch (sr->l_whence) {
1844         case 0: /*SEEK_SET*/
1845                 break;
1846         case 1: /*SEEK_CUR*/
1847                 sr->l_start += f_pos;
1848                 break;
1849         case 2: /*SEEK_END*/
1850                 sr->l_start += i_size_read(inode);
1851                 break;
1852         default:
1853                 ret = -EINVAL;
1854                 goto out_inode_unlock;
1855         }
1856         sr->l_whence = 0;
1857
1858         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1859
1860         if (sr->l_start < 0
1861             || sr->l_start > max_off
1862             || (sr->l_start + llen) < 0
1863             || (sr->l_start + llen) > max_off) {
1864                 ret = -EINVAL;
1865                 goto out_inode_unlock;
1866         }
1867         size = sr->l_start + sr->l_len;
1868
1869         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1870                 if (sr->l_len <= 0) {
1871                         ret = -EINVAL;
1872                         goto out_inode_unlock;
1873                 }
1874         }
1875
1876         if (file && should_remove_suid(file->f_path.dentry)) {
1877                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1878                 if (ret) {
1879                         mlog_errno(ret);
1880                         goto out_inode_unlock;
1881                 }
1882         }
1883
1884         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1885         switch (cmd) {
1886         case OCFS2_IOC_RESVSP:
1887         case OCFS2_IOC_RESVSP64:
1888                 /*
1889                  * This takes unsigned offsets, but the signed ones we
1890                  * pass have been checked against overflow above.
1891                  */
1892                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1893                                                        sr->l_len);
1894                 break;
1895         case OCFS2_IOC_UNRESVSP:
1896         case OCFS2_IOC_UNRESVSP64:
1897                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1898                                                sr->l_len);
1899                 break;
1900         default:
1901                 ret = -EINVAL;
1902         }
1903         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1904         if (ret) {
1905                 mlog_errno(ret);
1906                 goto out_inode_unlock;
1907         }
1908
1909         /*
1910          * We update c/mtime for these changes
1911          */
1912         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1913         if (IS_ERR(handle)) {
1914                 ret = PTR_ERR(handle);
1915                 mlog_errno(ret);
1916                 goto out_inode_unlock;
1917         }
1918
1919         if (change_size && i_size_read(inode) < size)
1920                 i_size_write(inode, size);
1921
1922         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1923         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1924         if (ret < 0)
1925                 mlog_errno(ret);
1926
1927         ocfs2_commit_trans(osb, handle);
1928
1929 out_inode_unlock:
1930         brelse(di_bh);
1931         ocfs2_inode_unlock(inode, 1);
1932 out_rw_unlock:
1933         ocfs2_rw_unlock(inode, 1);
1934
1935 out:
1936         mutex_unlock(&inode->i_mutex);
1937         return ret;
1938 }
1939
1940 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1941                             struct ocfs2_space_resv *sr)
1942 {
1943         struct inode *inode = file->f_path.dentry->d_inode;
1944         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1945
1946         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1947             !ocfs2_writes_unwritten_extents(osb))
1948                 return -ENOTTY;
1949         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1950                  !ocfs2_sparse_alloc(osb))
1951                 return -ENOTTY;
1952
1953         if (!S_ISREG(inode->i_mode))
1954                 return -EINVAL;
1955
1956         if (!(file->f_mode & FMODE_WRITE))
1957                 return -EBADF;
1958
1959         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1960 }
1961
1962 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
1963                             loff_t len)
1964 {
1965         struct inode *inode = file->f_path.dentry->d_inode;
1966         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1967         struct ocfs2_space_resv sr;
1968         int change_size = 1;
1969         int cmd = OCFS2_IOC_RESVSP64;
1970
1971         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1972                 return -EOPNOTSUPP;
1973         if (!ocfs2_writes_unwritten_extents(osb))
1974                 return -EOPNOTSUPP;
1975
1976         if (mode & FALLOC_FL_KEEP_SIZE)
1977                 change_size = 0;
1978
1979         if (mode & FALLOC_FL_PUNCH_HOLE)
1980                 cmd = OCFS2_IOC_UNRESVSP64;
1981
1982         sr.l_whence = 0;
1983         sr.l_start = (s64)offset;
1984         sr.l_len = (s64)len;
1985
1986         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
1987                                          change_size);
1988 }
1989
1990 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1991                                    size_t count)
1992 {
1993         int ret = 0;
1994         unsigned int extent_flags;
1995         u32 cpos, clusters, extent_len, phys_cpos;
1996         struct super_block *sb = inode->i_sb;
1997
1998         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1999             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2000             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2001                 return 0;
2002
2003         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2004         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2005
2006         while (clusters) {
2007                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2008                                          &extent_flags);
2009                 if (ret < 0) {
2010                         mlog_errno(ret);
2011                         goto out;
2012                 }
2013
2014                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2015                         ret = 1;
2016                         break;
2017                 }
2018
2019                 if (extent_len > clusters)
2020                         extent_len = clusters;
2021
2022                 clusters -= extent_len;
2023                 cpos += extent_len;
2024         }
2025 out:
2026         return ret;
2027 }
2028
2029 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2030                                             struct file *file,
2031                                             loff_t pos, size_t count,
2032                                             int *meta_level)
2033 {
2034         int ret;
2035         struct buffer_head *di_bh = NULL;
2036         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2037         u32 clusters =
2038                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2039
2040         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2041         if (ret) {
2042                 mlog_errno(ret);
2043                 goto out;
2044         }
2045
2046         *meta_level = 1;
2047
2048         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2049         if (ret)
2050                 mlog_errno(ret);
2051 out:
2052         brelse(di_bh);
2053         return ret;
2054 }
2055
2056 static int ocfs2_prepare_inode_for_write(struct file *file,
2057                                          loff_t *ppos,
2058                                          size_t count,
2059                                          int appending,
2060                                          int *direct_io,
2061                                          int *has_refcount)
2062 {
2063         int ret = 0, meta_level = 0;
2064         struct dentry *dentry = file->f_path.dentry;
2065         struct inode *inode = dentry->d_inode;
2066         loff_t saved_pos = 0, end;
2067
2068         /*
2069          * We start with a read level meta lock and only jump to an ex
2070          * if we need to make modifications here.
2071          */
2072         for(;;) {
2073                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2074                 if (ret < 0) {
2075                         meta_level = -1;
2076                         mlog_errno(ret);
2077                         goto out;
2078                 }
2079
2080                 /* Clear suid / sgid if necessary. We do this here
2081                  * instead of later in the write path because
2082                  * remove_suid() calls ->setattr without any hint that
2083                  * we may have already done our cluster locking. Since
2084                  * ocfs2_setattr() *must* take cluster locks to
2085                  * proceeed, this will lead us to recursively lock the
2086                  * inode. There's also the dinode i_size state which
2087                  * can be lost via setattr during extending writes (we
2088                  * set inode->i_size at the end of a write. */
2089                 if (should_remove_suid(dentry)) {
2090                         if (meta_level == 0) {
2091                                 ocfs2_inode_unlock(inode, meta_level);
2092                                 meta_level = 1;
2093                                 continue;
2094                         }
2095
2096                         ret = ocfs2_write_remove_suid(inode);
2097                         if (ret < 0) {
2098                                 mlog_errno(ret);
2099                                 goto out_unlock;
2100                         }
2101                 }
2102
2103                 /* work on a copy of ppos until we're sure that we won't have
2104                  * to recalculate it due to relocking. */
2105                 if (appending)
2106                         saved_pos = i_size_read(inode);
2107                 else
2108                         saved_pos = *ppos;
2109
2110                 end = saved_pos + count;
2111
2112                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2113                 if (ret == 1) {
2114                         ocfs2_inode_unlock(inode, meta_level);
2115                         meta_level = -1;
2116
2117                         ret = ocfs2_prepare_inode_for_refcount(inode,
2118                                                                file,
2119                                                                saved_pos,
2120                                                                count,
2121                                                                &meta_level);
2122                         if (has_refcount)
2123                                 *has_refcount = 1;
2124                         if (direct_io)
2125                                 *direct_io = 0;
2126                 }
2127
2128                 if (ret < 0) {
2129                         mlog_errno(ret);
2130                         goto out_unlock;
2131                 }
2132
2133                 /*
2134                  * Skip the O_DIRECT checks if we don't need
2135                  * them.
2136                  */
2137                 if (!direct_io || !(*direct_io))
2138                         break;
2139
2140                 /*
2141                  * There's no sane way to do direct writes to an inode
2142                  * with inline data.
2143                  */
2144                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2145                         *direct_io = 0;
2146                         break;
2147                 }
2148
2149                 /*
2150                  * Allowing concurrent direct writes means
2151                  * i_size changes wouldn't be synchronized, so
2152                  * one node could wind up truncating another
2153                  * nodes writes.
2154                  */
2155                 if (end > i_size_read(inode)) {
2156                         *direct_io = 0;
2157                         break;
2158                 }
2159
2160                 /*
2161                  * We don't fill holes during direct io, so
2162                  * check for them here. If any are found, the
2163                  * caller will have to retake some cluster
2164                  * locks and initiate the io as buffered.
2165                  */
2166                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2167                 if (ret == 1) {
2168                         *direct_io = 0;
2169                         ret = 0;
2170                 } else if (ret < 0)
2171                         mlog_errno(ret);
2172                 break;
2173         }
2174
2175         if (appending)
2176                 *ppos = saved_pos;
2177
2178 out_unlock:
2179         trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2180                                             saved_pos, appending, count,
2181                                             direct_io, has_refcount);
2182
2183         if (meta_level >= 0)
2184                 ocfs2_inode_unlock(inode, meta_level);
2185
2186 out:
2187         return ret;
2188 }
2189
2190 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2191                                     const struct iovec *iov,
2192                                     unsigned long nr_segs,
2193                                     loff_t pos)
2194 {
2195         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2196         int can_do_direct, has_refcount = 0;
2197         ssize_t written = 0;
2198         size_t ocount;          /* original count */
2199         size_t count;           /* after file limit checks */
2200         loff_t old_size, *ppos = &iocb->ki_pos;
2201         u32 old_clusters;
2202         struct file *file = iocb->ki_filp;
2203         struct inode *inode = file->f_path.dentry->d_inode;
2204         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2205         int full_coherency = !(osb->s_mount_opt &
2206                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2207
2208         trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2209                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2210                 file->f_path.dentry->d_name.len,
2211                 file->f_path.dentry->d_name.name,
2212                 (unsigned int)nr_segs);
2213
2214         if (iocb->ki_left == 0)
2215                 return 0;
2216
2217         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2218
2219         appending = file->f_flags & O_APPEND ? 1 : 0;
2220         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2221
2222         mutex_lock(&inode->i_mutex);
2223
2224         ocfs2_iocb_clear_sem_locked(iocb);
2225
2226 relock:
2227         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2228         if (direct_io) {
2229                 down_read(&inode->i_alloc_sem);
2230                 have_alloc_sem = 1;
2231                 /* communicate with ocfs2_dio_end_io */
2232                 ocfs2_iocb_set_sem_locked(iocb);
2233         }
2234
2235         /*
2236          * Concurrent O_DIRECT writes are allowed with
2237          * mount_option "coherency=buffered".
2238          */
2239         rw_level = (!direct_io || full_coherency);
2240
2241         ret = ocfs2_rw_lock(inode, rw_level);
2242         if (ret < 0) {
2243                 mlog_errno(ret);
2244                 goto out_sems;
2245         }
2246
2247         /*
2248          * O_DIRECT writes with "coherency=full" need to take EX cluster
2249          * inode_lock to guarantee coherency.
2250          */
2251         if (direct_io && full_coherency) {
2252                 /*
2253                  * We need to take and drop the inode lock to force
2254                  * other nodes to drop their caches.  Buffered I/O
2255                  * already does this in write_begin().
2256                  */
2257                 ret = ocfs2_inode_lock(inode, NULL, 1);
2258                 if (ret < 0) {
2259                         mlog_errno(ret);
2260                         goto out_sems;
2261                 }
2262
2263                 ocfs2_inode_unlock(inode, 1);
2264         }
2265
2266         can_do_direct = direct_io;
2267         ret = ocfs2_prepare_inode_for_write(file, ppos,
2268                                             iocb->ki_left, appending,
2269                                             &can_do_direct, &has_refcount);
2270         if (ret < 0) {
2271                 mlog_errno(ret);
2272                 goto out;
2273         }
2274
2275         /*
2276          * We can't complete the direct I/O as requested, fall back to
2277          * buffered I/O.
2278          */
2279         if (direct_io && !can_do_direct) {
2280                 ocfs2_rw_unlock(inode, rw_level);
2281                 up_read(&inode->i_alloc_sem);
2282
2283                 have_alloc_sem = 0;
2284                 rw_level = -1;
2285
2286                 direct_io = 0;
2287                 goto relock;
2288         }
2289
2290         /*
2291          * To later detect whether a journal commit for sync writes is
2292          * necessary, we sample i_size, and cluster count here.
2293          */
2294         old_size = i_size_read(inode);
2295         old_clusters = OCFS2_I(inode)->ip_clusters;
2296
2297         /* communicate with ocfs2_dio_end_io */
2298         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2299
2300         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2301                                      VERIFY_READ);
2302         if (ret)
2303                 goto out_dio;
2304
2305         count = ocount;
2306         ret = generic_write_checks(file, ppos, &count,
2307                                    S_ISBLK(inode->i_mode));
2308         if (ret)
2309                 goto out_dio;
2310
2311         if (direct_io) {
2312                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2313                                                     ppos, count, ocount);
2314                 if (written < 0) {
2315                         ret = written;
2316                         goto out_dio;
2317                 }
2318         } else {
2319                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2320                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2321                                                       ppos, count, 0);
2322                 current->backing_dev_info = NULL;
2323         }
2324
2325 out_dio:
2326         /* buffered aio wouldn't have proper lock coverage today */
2327         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2328
2329         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2330             ((file->f_flags & O_DIRECT) && !direct_io)) {
2331                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2332                                                pos + count - 1);
2333                 if (ret < 0)
2334                         written = ret;
2335
2336                 if (!ret && ((old_size != i_size_read(inode)) ||
2337                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2338                              has_refcount)) {
2339                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2340                         if (ret < 0)
2341                                 written = ret;
2342                 }
2343
2344                 if (!ret)
2345                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2346                                                       pos + count - 1);
2347         }
2348
2349         /*
2350          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2351          * function pointer which is called when o_direct io completes so that
2352          * it can unlock our rw lock.  (it's the clustered equivalent of
2353          * i_alloc_sem; protects truncate from racing with pending ios).
2354          * Unfortunately there are error cases which call end_io and others
2355          * that don't.  so we don't have to unlock the rw_lock if either an
2356          * async dio is going to do it in the future or an end_io after an
2357          * error has already done it.
2358          */
2359         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2360                 rw_level = -1;
2361                 have_alloc_sem = 0;
2362         }
2363
2364 out:
2365         if (rw_level != -1)
2366                 ocfs2_rw_unlock(inode, rw_level);
2367
2368 out_sems:
2369         if (have_alloc_sem) {
2370                 up_read(&inode->i_alloc_sem);
2371                 ocfs2_iocb_clear_sem_locked(iocb);
2372         }
2373
2374         mutex_unlock(&inode->i_mutex);
2375
2376         if (written)
2377                 ret = written;
2378         return ret;
2379 }
2380
2381 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2382                                 struct file *out,
2383                                 struct splice_desc *sd)
2384 {
2385         int ret;
2386
2387         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2388                                             sd->total_len, 0, NULL, NULL);
2389         if (ret < 0) {
2390                 mlog_errno(ret);
2391                 return ret;
2392         }
2393
2394         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2395 }
2396
2397 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2398                                        struct file *out,
2399                                        loff_t *ppos,
2400                                        size_t len,
2401                                        unsigned int flags)
2402 {
2403         int ret;
2404         struct address_space *mapping = out->f_mapping;
2405         struct inode *inode = mapping->host;
2406         struct splice_desc sd = {
2407                 .total_len = len,
2408                 .flags = flags,
2409                 .pos = *ppos,
2410                 .u.file = out,
2411         };
2412
2413
2414         trace_ocfs2_file_splice_write(inode, out, out->f_path.dentry,
2415                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2416                         out->f_path.dentry->d_name.len,
2417                         out->f_path.dentry->d_name.name, len);
2418
2419         if (pipe->inode)
2420                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2421
2422         splice_from_pipe_begin(&sd);
2423         do {
2424                 ret = splice_from_pipe_next(pipe, &sd);
2425                 if (ret <= 0)
2426                         break;
2427
2428                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2429                 ret = ocfs2_rw_lock(inode, 1);
2430                 if (ret < 0)
2431                         mlog_errno(ret);
2432                 else {
2433                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2434                         ocfs2_rw_unlock(inode, 1);
2435                 }
2436                 mutex_unlock(&inode->i_mutex);
2437         } while (ret > 0);
2438         splice_from_pipe_end(pipe, &sd);
2439
2440         if (pipe->inode)
2441                 mutex_unlock(&pipe->inode->i_mutex);
2442
2443         if (sd.num_spliced)
2444                 ret = sd.num_spliced;
2445
2446         if (ret > 0) {
2447                 unsigned long nr_pages;
2448                 int err;
2449
2450                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2451
2452                 err = generic_write_sync(out, *ppos, ret);
2453                 if (err)
2454                         ret = err;
2455                 else
2456                         *ppos += ret;
2457
2458                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2459         }
2460
2461         return ret;
2462 }
2463
2464 static ssize_t ocfs2_file_splice_read(struct file *in,
2465                                       loff_t *ppos,
2466                                       struct pipe_inode_info *pipe,
2467                                       size_t len,
2468                                       unsigned int flags)
2469 {
2470         int ret = 0, lock_level = 0;
2471         struct inode *inode = in->f_path.dentry->d_inode;
2472
2473         trace_ocfs2_file_splice_read(inode, in, in->f_path.dentry,
2474                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2475                         in->f_path.dentry->d_name.len,
2476                         in->f_path.dentry->d_name.name, len);
2477
2478         /*
2479          * See the comment in ocfs2_file_aio_read()
2480          */
2481         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2482         if (ret < 0) {
2483                 mlog_errno(ret);
2484                 goto bail;
2485         }
2486         ocfs2_inode_unlock(inode, lock_level);
2487
2488         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2489
2490 bail:
2491         return ret;
2492 }
2493
2494 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2495                                    const struct iovec *iov,
2496                                    unsigned long nr_segs,
2497                                    loff_t pos)
2498 {
2499         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2500         struct file *filp = iocb->ki_filp;
2501         struct inode *inode = filp->f_path.dentry->d_inode;
2502
2503         trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2504                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2505                         filp->f_path.dentry->d_name.len,
2506                         filp->f_path.dentry->d_name.name, nr_segs);
2507
2508
2509         if (!inode) {
2510                 ret = -EINVAL;
2511                 mlog_errno(ret);
2512                 goto bail;
2513         }
2514
2515         ocfs2_iocb_clear_sem_locked(iocb);
2516
2517         /*
2518          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2519          * need locks to protect pending reads from racing with truncate.
2520          */
2521         if (filp->f_flags & O_DIRECT) {
2522                 down_read(&inode->i_alloc_sem);
2523                 have_alloc_sem = 1;
2524                 ocfs2_iocb_set_sem_locked(iocb);
2525
2526                 ret = ocfs2_rw_lock(inode, 0);
2527                 if (ret < 0) {
2528                         mlog_errno(ret);
2529                         goto bail;
2530                 }
2531                 rw_level = 0;
2532                 /* communicate with ocfs2_dio_end_io */
2533                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2534         }
2535
2536         /*
2537          * We're fine letting folks race truncates and extending
2538          * writes with read across the cluster, just like they can
2539          * locally. Hence no rw_lock during read.
2540          *
2541          * Take and drop the meta data lock to update inode fields
2542          * like i_size. This allows the checks down below
2543          * generic_file_aio_read() a chance of actually working.
2544          */
2545         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2546         if (ret < 0) {
2547                 mlog_errno(ret);
2548                 goto bail;
2549         }
2550         ocfs2_inode_unlock(inode, lock_level);
2551
2552         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2553         trace_generic_file_aio_read_ret(ret);
2554
2555         /* buffered aio wouldn't have proper lock coverage today */
2556         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2557
2558         /* see ocfs2_file_aio_write */
2559         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2560                 rw_level = -1;
2561                 have_alloc_sem = 0;
2562         }
2563
2564 bail:
2565         if (have_alloc_sem) {
2566                 up_read(&inode->i_alloc_sem);
2567                 ocfs2_iocb_clear_sem_locked(iocb);
2568         }
2569         if (rw_level != -1)
2570                 ocfs2_rw_unlock(inode, rw_level);
2571
2572         return ret;
2573 }
2574
2575 const struct inode_operations ocfs2_file_iops = {
2576         .setattr        = ocfs2_setattr,
2577         .getattr        = ocfs2_getattr,
2578         .permission     = ocfs2_permission,
2579         .setxattr       = generic_setxattr,
2580         .getxattr       = generic_getxattr,
2581         .listxattr      = ocfs2_listxattr,
2582         .removexattr    = generic_removexattr,
2583         .fiemap         = ocfs2_fiemap,
2584 };
2585
2586 const struct inode_operations ocfs2_special_file_iops = {
2587         .setattr        = ocfs2_setattr,
2588         .getattr        = ocfs2_getattr,
2589         .permission     = ocfs2_permission,
2590 };
2591
2592 /*
2593  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2594  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2595  */
2596 const struct file_operations ocfs2_fops = {
2597         .llseek         = generic_file_llseek,
2598         .read           = do_sync_read,
2599         .write          = do_sync_write,
2600         .mmap           = ocfs2_mmap,
2601         .fsync          = ocfs2_sync_file,
2602         .release        = ocfs2_file_release,
2603         .open           = ocfs2_file_open,
2604         .aio_read       = ocfs2_file_aio_read,
2605         .aio_write      = ocfs2_file_aio_write,
2606         .unlocked_ioctl = ocfs2_ioctl,
2607 #ifdef CONFIG_COMPAT
2608         .compat_ioctl   = ocfs2_compat_ioctl,
2609 #endif
2610         .lock           = ocfs2_lock,
2611         .flock          = ocfs2_flock,
2612         .splice_read    = ocfs2_file_splice_read,
2613         .splice_write   = ocfs2_file_splice_write,
2614         .fallocate      = ocfs2_fallocate,
2615 };
2616
2617 const struct file_operations ocfs2_dops = {
2618         .llseek         = generic_file_llseek,
2619         .read           = generic_read_dir,
2620         .readdir        = ocfs2_readdir,
2621         .fsync          = ocfs2_sync_file,
2622         .release        = ocfs2_dir_release,
2623         .open           = ocfs2_dir_open,
2624         .unlocked_ioctl = ocfs2_ioctl,
2625 #ifdef CONFIG_COMPAT
2626         .compat_ioctl   = ocfs2_compat_ioctl,
2627 #endif
2628         .lock           = ocfs2_lock,
2629         .flock          = ocfs2_flock,
2630 };
2631
2632 /*
2633  * POSIX-lockless variants of our file_operations.
2634  *
2635  * These will be used if the underlying cluster stack does not support
2636  * posix file locking, if the user passes the "localflocks" mount
2637  * option, or if we have a local-only fs.
2638  *
2639  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2640  * so we still want it in the case of no stack support for
2641  * plocks. Internally, it will do the right thing when asked to ignore
2642  * the cluster.
2643  */
2644 const struct file_operations ocfs2_fops_no_plocks = {
2645         .llseek         = generic_file_llseek,
2646         .read           = do_sync_read,
2647         .write          = do_sync_write,
2648         .mmap           = ocfs2_mmap,
2649         .fsync          = ocfs2_sync_file,
2650         .release        = ocfs2_file_release,
2651         .open           = ocfs2_file_open,
2652         .aio_read       = ocfs2_file_aio_read,
2653         .aio_write      = ocfs2_file_aio_write,
2654         .unlocked_ioctl = ocfs2_ioctl,
2655 #ifdef CONFIG_COMPAT
2656         .compat_ioctl   = ocfs2_compat_ioctl,
2657 #endif
2658         .flock          = ocfs2_flock,
2659         .splice_read    = ocfs2_file_splice_read,
2660         .splice_write   = ocfs2_file_splice_write,
2661 };
2662
2663 const struct file_operations ocfs2_dops_no_plocks = {
2664         .llseek         = generic_file_llseek,
2665         .read           = generic_read_dir,
2666         .readdir        = ocfs2_readdir,
2667         .fsync          = ocfs2_sync_file,
2668         .release        = ocfs2_dir_release,
2669         .open           = ocfs2_dir_open,
2670         .unlocked_ioctl = ocfs2_ioctl,
2671 #ifdef CONFIG_COMPAT
2672         .compat_ioctl   = ocfs2_compat_ioctl,
2673 #endif
2674         .flock          = ocfs2_flock,
2675 };