ocfs2: Add a mount option "coherency=*" to handle cluster coherency for O_DIRECT...
[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 #define MLOG_MASK_PREFIX ML_INODE
42 #include <cluster/masklog.h>
43
44 #include "ocfs2.h"
45
46 #include "alloc.h"
47 #include "aops.h"
48 #include "dir.h"
49 #include "dlmglue.h"
50 #include "extent_map.h"
51 #include "file.h"
52 #include "sysfile.h"
53 #include "inode.h"
54 #include "ioctl.h"
55 #include "journal.h"
56 #include "locks.h"
57 #include "mmap.h"
58 #include "suballoc.h"
59 #include "super.h"
60 #include "xattr.h"
61 #include "acl.h"
62 #include "quota.h"
63 #include "refcounttree.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         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
103                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
104
105         if (file->f_mode & FMODE_WRITE)
106                 dquot_initialize(inode);
107
108         spin_lock(&oi->ip_lock);
109
110         /* Check that the inode hasn't been wiped from disk by another
111          * node. If it hasn't then we're safe as long as we hold the
112          * spin lock until our increment of open count. */
113         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
114                 spin_unlock(&oi->ip_lock);
115
116                 status = -ENOENT;
117                 goto leave;
118         }
119
120         if (mode & O_DIRECT)
121                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
122
123         oi->ip_open_count++;
124         spin_unlock(&oi->ip_lock);
125
126         status = ocfs2_init_file_private(inode, file);
127         if (status) {
128                 /*
129                  * We want to set open count back if we're failing the
130                  * open.
131                  */
132                 spin_lock(&oi->ip_lock);
133                 oi->ip_open_count--;
134                 spin_unlock(&oi->ip_lock);
135         }
136
137 leave:
138         mlog_exit(status);
139         return status;
140 }
141
142 static int ocfs2_file_release(struct inode *inode, struct file *file)
143 {
144         struct ocfs2_inode_info *oi = OCFS2_I(inode);
145
146         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
147                        file->f_path.dentry->d_name.len,
148                        file->f_path.dentry->d_name.name);
149
150         spin_lock(&oi->ip_lock);
151         if (!--oi->ip_open_count)
152                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
153         spin_unlock(&oi->ip_lock);
154
155         ocfs2_free_file_private(inode, file);
156
157         mlog_exit(0);
158
159         return 0;
160 }
161
162 static int ocfs2_dir_open(struct inode *inode, struct file *file)
163 {
164         return ocfs2_init_file_private(inode, file);
165 }
166
167 static int ocfs2_dir_release(struct inode *inode, struct file *file)
168 {
169         ocfs2_free_file_private(inode, file);
170         return 0;
171 }
172
173 static int ocfs2_sync_file(struct file *file, int datasync)
174 {
175         int err = 0;
176         journal_t *journal;
177         struct inode *inode = file->f_mapping->host;
178         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
179
180         mlog_entry("(0x%p, %d, 0x%p, '%.*s')\n", file, datasync,
181                    file->f_path.dentry, file->f_path.dentry->d_name.len,
182                    file->f_path.dentry->d_name.name);
183
184         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
185                 /*
186                  * We still have to flush drive's caches to get data to the
187                  * platter
188                  */
189                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
190                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL,
191                                            NULL, BLKDEV_IFL_WAIT);
192                 goto bail;
193         }
194
195         journal = osb->journal->j_journal;
196         err = jbd2_journal_force_commit(journal);
197
198 bail:
199         mlog_exit(err);
200
201         return (err < 0) ? -EIO : 0;
202 }
203
204 int ocfs2_should_update_atime(struct inode *inode,
205                               struct vfsmount *vfsmnt)
206 {
207         struct timespec now;
208         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
209
210         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
211                 return 0;
212
213         if ((inode->i_flags & S_NOATIME) ||
214             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
215                 return 0;
216
217         /*
218          * We can be called with no vfsmnt structure - NFSD will
219          * sometimes do this.
220          *
221          * Note that our action here is different than touch_atime() -
222          * if we can't tell whether this is a noatime mount, then we
223          * don't know whether to trust the value of s_atime_quantum.
224          */
225         if (vfsmnt == NULL)
226                 return 0;
227
228         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
229             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
230                 return 0;
231
232         if (vfsmnt->mnt_flags & MNT_RELATIME) {
233                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
234                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
235                         return 1;
236
237                 return 0;
238         }
239
240         now = CURRENT_TIME;
241         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
242                 return 0;
243         else
244                 return 1;
245 }
246
247 int ocfs2_update_inode_atime(struct inode *inode,
248                              struct buffer_head *bh)
249 {
250         int ret;
251         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
252         handle_t *handle;
253         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
254
255         mlog_entry_void();
256
257         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
258         if (IS_ERR(handle)) {
259                 ret = PTR_ERR(handle);
260                 mlog_errno(ret);
261                 goto out;
262         }
263
264         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
265                                       OCFS2_JOURNAL_ACCESS_WRITE);
266         if (ret) {
267                 mlog_errno(ret);
268                 goto out_commit;
269         }
270
271         /*
272          * Don't use ocfs2_mark_inode_dirty() here as we don't always
273          * have i_mutex to guard against concurrent changes to other
274          * inode fields.
275          */
276         inode->i_atime = CURRENT_TIME;
277         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
278         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
279         ocfs2_journal_dirty(handle, bh);
280
281 out_commit:
282         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
283 out:
284         mlog_exit(ret);
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         mlog_entry_void();
296         i_size_write(inode, new_i_size);
297         inode->i_blocks = ocfs2_inode_sector_count(inode);
298         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
299
300         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
301         if (status < 0) {
302                 mlog_errno(status);
303                 goto bail;
304         }
305
306 bail:
307         mlog_exit(status);
308         return status;
309 }
310
311 int ocfs2_simple_size_update(struct inode *inode,
312                              struct buffer_head *di_bh,
313                              u64 new_i_size)
314 {
315         int ret;
316         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
317         handle_t *handle = NULL;
318
319         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
320         if (IS_ERR(handle)) {
321                 ret = PTR_ERR(handle);
322                 mlog_errno(ret);
323                 goto out;
324         }
325
326         ret = ocfs2_set_inode_size(handle, inode, di_bh,
327                                    new_i_size);
328         if (ret < 0)
329                 mlog_errno(ret);
330
331         ocfs2_commit_trans(osb, handle);
332 out:
333         return ret;
334 }
335
336 static int ocfs2_cow_file_pos(struct inode *inode,
337                               struct buffer_head *fe_bh,
338                               u64 offset)
339 {
340         int status;
341         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
342         unsigned int num_clusters = 0;
343         unsigned int ext_flags = 0;
344
345         /*
346          * If the new offset is aligned to the range of the cluster, there is
347          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
348          * CoW either.
349          */
350         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
351                 return 0;
352
353         status = ocfs2_get_clusters(inode, cpos, &phys,
354                                     &num_clusters, &ext_flags);
355         if (status) {
356                 mlog_errno(status);
357                 goto out;
358         }
359
360         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
361                 goto out;
362
363         return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
364
365 out:
366         return status;
367 }
368
369 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
370                                      struct inode *inode,
371                                      struct buffer_head *fe_bh,
372                                      u64 new_i_size)
373 {
374         int status;
375         handle_t *handle;
376         struct ocfs2_dinode *di;
377         u64 cluster_bytes;
378
379         mlog_entry_void();
380
381         /*
382          * We need to CoW the cluster contains the offset if it is reflinked
383          * since we will call ocfs2_zero_range_for_truncate later which will
384          * write "0" from offset to the end of the cluster.
385          */
386         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
387         if (status) {
388                 mlog_errno(status);
389                 return status;
390         }
391
392         /* TODO: This needs to actually orphan the inode in this
393          * transaction. */
394
395         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
396         if (IS_ERR(handle)) {
397                 status = PTR_ERR(handle);
398                 mlog_errno(status);
399                 goto out;
400         }
401
402         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
403                                          OCFS2_JOURNAL_ACCESS_WRITE);
404         if (status < 0) {
405                 mlog_errno(status);
406                 goto out_commit;
407         }
408
409         /*
410          * Do this before setting i_size.
411          */
412         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
413         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
414                                                cluster_bytes);
415         if (status) {
416                 mlog_errno(status);
417                 goto out_commit;
418         }
419
420         i_size_write(inode, new_i_size);
421         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
422
423         di = (struct ocfs2_dinode *) fe_bh->b_data;
424         di->i_size = cpu_to_le64(new_i_size);
425         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
426         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
427
428         ocfs2_journal_dirty(handle, fe_bh);
429
430 out_commit:
431         ocfs2_commit_trans(osb, handle);
432 out:
433
434         mlog_exit(status);
435         return status;
436 }
437
438 static int ocfs2_truncate_file(struct inode *inode,
439                                struct buffer_head *di_bh,
440                                u64 new_i_size)
441 {
442         int status = 0;
443         struct ocfs2_dinode *fe = NULL;
444         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
445
446         mlog_entry("(inode = %llu, new_i_size = %llu\n",
447                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
448                    (unsigned long long)new_i_size);
449
450         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
451          * already validated it */
452         fe = (struct ocfs2_dinode *) di_bh->b_data;
453
454         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
455                         "Inode %llu, inode i_size = %lld != di "
456                         "i_size = %llu, i_flags = 0x%x\n",
457                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
458                         i_size_read(inode),
459                         (unsigned long long)le64_to_cpu(fe->i_size),
460                         le32_to_cpu(fe->i_flags));
461
462         if (new_i_size > le64_to_cpu(fe->i_size)) {
463                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
464                      (unsigned long long)le64_to_cpu(fe->i_size),
465                      (unsigned long long)new_i_size);
466                 status = -EINVAL;
467                 mlog_errno(status);
468                 goto bail;
469         }
470
471         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
472              (unsigned long long)le64_to_cpu(fe->i_blkno),
473              (unsigned long long)le64_to_cpu(fe->i_size),
474              (unsigned long long)new_i_size);
475
476         /* lets handle the simple truncate cases before doing any more
477          * cluster locking. */
478         if (new_i_size == le64_to_cpu(fe->i_size))
479                 goto bail;
480
481         down_write(&OCFS2_I(inode)->ip_alloc_sem);
482
483         ocfs2_resv_discard(&osb->osb_la_resmap,
484                            &OCFS2_I(inode)->ip_la_data_resv);
485
486         /*
487          * The inode lock forced other nodes to sync and drop their
488          * pages, which (correctly) happens even if we have a truncate
489          * without allocation change - ocfs2 cluster sizes can be much
490          * greater than page size, so we have to truncate them
491          * anyway.
492          */
493         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
494         truncate_inode_pages(inode->i_mapping, new_i_size);
495
496         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
497                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
498                                                i_size_read(inode), 1);
499                 if (status)
500                         mlog_errno(status);
501
502                 goto bail_unlock_sem;
503         }
504
505         /* alright, we're going to need to do a full blown alloc size
506          * change. Orphan the inode so that recovery can complete the
507          * truncate if necessary. This does the task of marking
508          * i_size. */
509         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
510         if (status < 0) {
511                 mlog_errno(status);
512                 goto bail_unlock_sem;
513         }
514
515         status = ocfs2_commit_truncate(osb, inode, di_bh);
516         if (status < 0) {
517                 mlog_errno(status);
518                 goto bail_unlock_sem;
519         }
520
521         /* TODO: orphan dir cleanup here. */
522 bail_unlock_sem:
523         up_write(&OCFS2_I(inode)->ip_alloc_sem);
524
525 bail:
526         if (!status && OCFS2_I(inode)->ip_clusters == 0)
527                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
528
529         mlog_exit(status);
530         return status;
531 }
532
533 /*
534  * extend file allocation only here.
535  * we'll update all the disk stuff, and oip->alloc_size
536  *
537  * expect stuff to be locked, a transaction started and enough data /
538  * metadata reservations in the contexts.
539  *
540  * Will return -EAGAIN, and a reason if a restart is needed.
541  * If passed in, *reason will always be set, even in error.
542  */
543 int ocfs2_add_inode_data(struct ocfs2_super *osb,
544                          struct inode *inode,
545                          u32 *logical_offset,
546                          u32 clusters_to_add,
547                          int mark_unwritten,
548                          struct buffer_head *fe_bh,
549                          handle_t *handle,
550                          struct ocfs2_alloc_context *data_ac,
551                          struct ocfs2_alloc_context *meta_ac,
552                          enum ocfs2_alloc_restarted *reason_ret)
553 {
554         int ret;
555         struct ocfs2_extent_tree et;
556
557         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
558         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
559                                           clusters_to_add, mark_unwritten,
560                                           data_ac, meta_ac, reason_ret);
561
562         return ret;
563 }
564
565 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
566                                      u32 clusters_to_add, int mark_unwritten)
567 {
568         int status = 0;
569         int restart_func = 0;
570         int credits;
571         u32 prev_clusters;
572         struct buffer_head *bh = NULL;
573         struct ocfs2_dinode *fe = NULL;
574         handle_t *handle = NULL;
575         struct ocfs2_alloc_context *data_ac = NULL;
576         struct ocfs2_alloc_context *meta_ac = NULL;
577         enum ocfs2_alloc_restarted why;
578         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
579         struct ocfs2_extent_tree et;
580         int did_quota = 0;
581
582         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
583
584         /*
585          * This function only exists for file systems which don't
586          * support holes.
587          */
588         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
589
590         status = ocfs2_read_inode_block(inode, &bh);
591         if (status < 0) {
592                 mlog_errno(status);
593                 goto leave;
594         }
595         fe = (struct ocfs2_dinode *) bh->b_data;
596
597 restart_all:
598         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
599
600         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
601              "clusters_to_add = %u\n",
602              (unsigned long long)OCFS2_I(inode)->ip_blkno,
603              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
604              clusters_to_add);
605         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
606         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
607                                        &data_ac, &meta_ac);
608         if (status) {
609                 mlog_errno(status);
610                 goto leave;
611         }
612
613         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
614                                             clusters_to_add);
615         handle = ocfs2_start_trans(osb, credits);
616         if (IS_ERR(handle)) {
617                 status = PTR_ERR(handle);
618                 handle = NULL;
619                 mlog_errno(status);
620                 goto leave;
621         }
622
623 restarted_transaction:
624         status = dquot_alloc_space_nodirty(inode,
625                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
626         if (status)
627                 goto leave;
628         did_quota = 1;
629
630         /* reserve a write to the file entry early on - that we if we
631          * run out of credits in the allocation path, we can still
632          * update i_size. */
633         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
634                                          OCFS2_JOURNAL_ACCESS_WRITE);
635         if (status < 0) {
636                 mlog_errno(status);
637                 goto leave;
638         }
639
640         prev_clusters = OCFS2_I(inode)->ip_clusters;
641
642         status = ocfs2_add_inode_data(osb,
643                                       inode,
644                                       &logical_start,
645                                       clusters_to_add,
646                                       mark_unwritten,
647                                       bh,
648                                       handle,
649                                       data_ac,
650                                       meta_ac,
651                                       &why);
652         if ((status < 0) && (status != -EAGAIN)) {
653                 if (status != -ENOSPC)
654                         mlog_errno(status);
655                 goto leave;
656         }
657
658         ocfs2_journal_dirty(handle, bh);
659
660         spin_lock(&OCFS2_I(inode)->ip_lock);
661         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
662         spin_unlock(&OCFS2_I(inode)->ip_lock);
663         /* Release unused quota reservation */
664         dquot_free_space(inode,
665                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
666         did_quota = 0;
667
668         if (why != RESTART_NONE && clusters_to_add) {
669                 if (why == RESTART_META) {
670                         mlog(0, "restarting function.\n");
671                         restart_func = 1;
672                         status = 0;
673                 } else {
674                         BUG_ON(why != RESTART_TRANS);
675
676                         mlog(0, "restarting transaction.\n");
677                         /* TODO: This can be more intelligent. */
678                         credits = ocfs2_calc_extend_credits(osb->sb,
679                                                             &fe->id2.i_list,
680                                                             clusters_to_add);
681                         status = ocfs2_extend_trans(handle, credits);
682                         if (status < 0) {
683                                 /* handle still has to be committed at
684                                  * this point. */
685                                 status = -ENOMEM;
686                                 mlog_errno(status);
687                                 goto leave;
688                         }
689                         goto restarted_transaction;
690                 }
691         }
692
693         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
694              le32_to_cpu(fe->i_clusters),
695              (unsigned long long)le64_to_cpu(fe->i_size));
696         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
697              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
698
699 leave:
700         if (status < 0 && did_quota)
701                 dquot_free_space(inode,
702                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
703         if (handle) {
704                 ocfs2_commit_trans(osb, handle);
705                 handle = NULL;
706         }
707         if (data_ac) {
708                 ocfs2_free_alloc_context(data_ac);
709                 data_ac = NULL;
710         }
711         if (meta_ac) {
712                 ocfs2_free_alloc_context(meta_ac);
713                 meta_ac = NULL;
714         }
715         if ((!status) && restart_func) {
716                 restart_func = 0;
717                 goto restart_all;
718         }
719         brelse(bh);
720         bh = NULL;
721
722         mlog_exit(status);
723         return status;
724 }
725
726 /*
727  * While a write will already be ordering the data, a truncate will not.
728  * Thus, we need to explicitly order the zeroed pages.
729  */
730 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
731 {
732         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
733         handle_t *handle = NULL;
734         int ret = 0;
735
736         if (!ocfs2_should_order_data(inode))
737                 goto out;
738
739         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
740         if (IS_ERR(handle)) {
741                 ret = -ENOMEM;
742                 mlog_errno(ret);
743                 goto out;
744         }
745
746         ret = ocfs2_jbd2_file_inode(handle, inode);
747         if (ret < 0)
748                 mlog_errno(ret);
749
750 out:
751         if (ret) {
752                 if (!IS_ERR(handle))
753                         ocfs2_commit_trans(osb, handle);
754                 handle = ERR_PTR(ret);
755         }
756         return handle;
757 }
758
759 /* Some parts of this taken from generic_cont_expand, which turned out
760  * to be too fragile to do exactly what we need without us having to
761  * worry about recursive locking in ->write_begin() and ->write_end(). */
762 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
763                                  u64 abs_to)
764 {
765         struct address_space *mapping = inode->i_mapping;
766         struct page *page;
767         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
768         handle_t *handle = NULL;
769         int ret = 0;
770         unsigned zero_from, zero_to, block_start, block_end;
771
772         BUG_ON(abs_from >= abs_to);
773         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
774         BUG_ON(abs_from & (inode->i_blkbits - 1));
775
776         page = find_or_create_page(mapping, index, GFP_NOFS);
777         if (!page) {
778                 ret = -ENOMEM;
779                 mlog_errno(ret);
780                 goto out;
781         }
782
783         /* Get the offsets within the page that we want to zero */
784         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
785         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
786         if (!zero_to)
787                 zero_to = PAGE_CACHE_SIZE;
788
789         mlog(0,
790              "abs_from = %llu, abs_to = %llu, index = %lu, zero_from = %u, zero_to = %u\n",
791              (unsigned long long)abs_from, (unsigned long long)abs_to,
792              index, zero_from, zero_to);
793
794         /* We know that zero_from is block aligned */
795         for (block_start = zero_from; block_start < zero_to;
796              block_start = block_end) {
797                 block_end = block_start + (1 << inode->i_blkbits);
798
799                 /*
800                  * block_start is block-aligned.  Bump it by one to
801                  * force ocfs2_{prepare,commit}_write() to zero the
802                  * whole block.
803                  */
804                 ret = ocfs2_prepare_write_nolock(inode, page,
805                                                  block_start + 1,
806                                                  block_start + 1);
807                 if (ret < 0) {
808                         mlog_errno(ret);
809                         goto out_unlock;
810                 }
811
812                 if (!handle) {
813                         handle = ocfs2_zero_start_ordered_transaction(inode);
814                         if (IS_ERR(handle)) {
815                                 ret = PTR_ERR(handle);
816                                 handle = NULL;
817                                 break;
818                         }
819                 }
820
821                 /* must not update i_size! */
822                 ret = block_commit_write(page, block_start + 1,
823                                          block_start + 1);
824                 if (ret < 0)
825                         mlog_errno(ret);
826                 else
827                         ret = 0;
828         }
829
830         if (handle)
831                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
832
833 out_unlock:
834         unlock_page(page);
835         page_cache_release(page);
836 out:
837         return ret;
838 }
839
840 /*
841  * Find the next range to zero.  We do this in terms of bytes because
842  * that's what ocfs2_zero_extend() wants, and it is dealing with the
843  * pagecache.  We may return multiple extents.
844  *
845  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
846  * needs to be zeroed.  range_start and range_end return the next zeroing
847  * range.  A subsequent call should pass the previous range_end as its
848  * zero_start.  If range_end is 0, there's nothing to do.
849  *
850  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
851  */
852 static int ocfs2_zero_extend_get_range(struct inode *inode,
853                                        struct buffer_head *di_bh,
854                                        u64 zero_start, u64 zero_end,
855                                        u64 *range_start, u64 *range_end)
856 {
857         int rc = 0, needs_cow = 0;
858         u32 p_cpos, zero_clusters = 0;
859         u32 zero_cpos =
860                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
861         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
862         unsigned int num_clusters = 0;
863         unsigned int ext_flags = 0;
864
865         while (zero_cpos < last_cpos) {
866                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
867                                         &num_clusters, &ext_flags);
868                 if (rc) {
869                         mlog_errno(rc);
870                         goto out;
871                 }
872
873                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
874                         zero_clusters = num_clusters;
875                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
876                                 needs_cow = 1;
877                         break;
878                 }
879
880                 zero_cpos += num_clusters;
881         }
882         if (!zero_clusters) {
883                 *range_end = 0;
884                 goto out;
885         }
886
887         while ((zero_cpos + zero_clusters) < last_cpos) {
888                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
889                                         &p_cpos, &num_clusters,
890                                         &ext_flags);
891                 if (rc) {
892                         mlog_errno(rc);
893                         goto out;
894                 }
895
896                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
897                         break;
898                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
899                         needs_cow = 1;
900                 zero_clusters += num_clusters;
901         }
902         if ((zero_cpos + zero_clusters) > last_cpos)
903                 zero_clusters = last_cpos - zero_cpos;
904
905         if (needs_cow) {
906                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
907                                         zero_clusters, UINT_MAX);
908                 if (rc) {
909                         mlog_errno(rc);
910                         goto out;
911                 }
912         }
913
914         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
915         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
916                                              zero_cpos + zero_clusters);
917
918 out:
919         return rc;
920 }
921
922 /*
923  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
924  * has made sure that the entire range needs zeroing.
925  */
926 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
927                                    u64 range_end)
928 {
929         int rc = 0;
930         u64 next_pos;
931         u64 zero_pos = range_start;
932
933         mlog(0, "range_start = %llu, range_end = %llu\n",
934              (unsigned long long)range_start,
935              (unsigned long long)range_end);
936         BUG_ON(range_start >= range_end);
937
938         while (zero_pos < range_end) {
939                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
940                 if (next_pos > range_end)
941                         next_pos = range_end;
942                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
943                 if (rc < 0) {
944                         mlog_errno(rc);
945                         break;
946                 }
947                 zero_pos = next_pos;
948
949                 /*
950                  * Very large extends have the potential to lock up
951                  * the cpu for extended periods of time.
952                  */
953                 cond_resched();
954         }
955
956         return rc;
957 }
958
959 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
960                       loff_t zero_to_size)
961 {
962         int ret = 0;
963         u64 zero_start, range_start = 0, range_end = 0;
964         struct super_block *sb = inode->i_sb;
965
966         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
967         mlog(0, "zero_start %llu for i_size %llu\n",
968              (unsigned long long)zero_start,
969              (unsigned long long)i_size_read(inode));
970         while (zero_start < zero_to_size) {
971                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
972                                                   zero_to_size,
973                                                   &range_start,
974                                                   &range_end);
975                 if (ret) {
976                         mlog_errno(ret);
977                         break;
978                 }
979                 if (!range_end)
980                         break;
981                 /* Trim the ends */
982                 if (range_start < zero_start)
983                         range_start = zero_start;
984                 if (range_end > zero_to_size)
985                         range_end = zero_to_size;
986
987                 ret = ocfs2_zero_extend_range(inode, range_start,
988                                               range_end);
989                 if (ret) {
990                         mlog_errno(ret);
991                         break;
992                 }
993                 zero_start = range_end;
994         }
995
996         return ret;
997 }
998
999 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
1000                           u64 new_i_size, u64 zero_to)
1001 {
1002         int ret;
1003         u32 clusters_to_add;
1004         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1005
1006         /*
1007          * Only quota files call this without a bh, and they can't be
1008          * refcounted.
1009          */
1010         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
1011         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1012
1013         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1014         if (clusters_to_add < oi->ip_clusters)
1015                 clusters_to_add = 0;
1016         else
1017                 clusters_to_add -= oi->ip_clusters;
1018
1019         if (clusters_to_add) {
1020                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1021                                                 clusters_to_add, 0);
1022                 if (ret) {
1023                         mlog_errno(ret);
1024                         goto out;
1025                 }
1026         }
1027
1028         /*
1029          * Call this even if we don't add any clusters to the tree. We
1030          * still need to zero the area between the old i_size and the
1031          * new i_size.
1032          */
1033         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1034         if (ret < 0)
1035                 mlog_errno(ret);
1036
1037 out:
1038         return ret;
1039 }
1040
1041 static int ocfs2_extend_file(struct inode *inode,
1042                              struct buffer_head *di_bh,
1043                              u64 new_i_size)
1044 {
1045         int ret = 0;
1046         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1047
1048         BUG_ON(!di_bh);
1049
1050         /* setattr sometimes calls us like this. */
1051         if (new_i_size == 0)
1052                 goto out;
1053
1054         if (i_size_read(inode) == new_i_size)
1055                 goto out;
1056         BUG_ON(new_i_size < i_size_read(inode));
1057
1058         /*
1059          * The alloc sem blocks people in read/write from reading our
1060          * allocation until we're done changing it. We depend on
1061          * i_mutex to block other extend/truncate calls while we're
1062          * here.  We even have to hold it for sparse files because there
1063          * might be some tail zeroing.
1064          */
1065         down_write(&oi->ip_alloc_sem);
1066
1067         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1068                 /*
1069                  * We can optimize small extends by keeping the inodes
1070                  * inline data.
1071                  */
1072                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1073                         up_write(&oi->ip_alloc_sem);
1074                         goto out_update_size;
1075                 }
1076
1077                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1078                 if (ret) {
1079                         up_write(&oi->ip_alloc_sem);
1080                         mlog_errno(ret);
1081                         goto out;
1082                 }
1083         }
1084
1085         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1086                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1087         else
1088                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1089                                             new_i_size);
1090
1091         up_write(&oi->ip_alloc_sem);
1092
1093         if (ret < 0) {
1094                 mlog_errno(ret);
1095                 goto out;
1096         }
1097
1098 out_update_size:
1099         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1100         if (ret < 0)
1101                 mlog_errno(ret);
1102
1103 out:
1104         return ret;
1105 }
1106
1107 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1108 {
1109         int status = 0, size_change;
1110         struct inode *inode = dentry->d_inode;
1111         struct super_block *sb = inode->i_sb;
1112         struct ocfs2_super *osb = OCFS2_SB(sb);
1113         struct buffer_head *bh = NULL;
1114         handle_t *handle = NULL;
1115         struct dquot *transfer_to[MAXQUOTAS] = { };
1116         int qtype;
1117
1118         mlog_entry("(0x%p, '%.*s')\n", dentry,
1119                    dentry->d_name.len, dentry->d_name.name);
1120
1121         /* ensuring we don't even attempt to truncate a symlink */
1122         if (S_ISLNK(inode->i_mode))
1123                 attr->ia_valid &= ~ATTR_SIZE;
1124
1125         if (attr->ia_valid & ATTR_MODE)
1126                 mlog(0, "mode change: %d\n", attr->ia_mode);
1127         if (attr->ia_valid & ATTR_UID)
1128                 mlog(0, "uid change: %d\n", attr->ia_uid);
1129         if (attr->ia_valid & ATTR_GID)
1130                 mlog(0, "gid change: %d\n", attr->ia_gid);
1131         if (attr->ia_valid & ATTR_SIZE)
1132                 mlog(0, "size change...\n");
1133         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1134                 mlog(0, "time change...\n");
1135
1136 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1137                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1138         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1139                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1140                 return 0;
1141         }
1142
1143         status = inode_change_ok(inode, attr);
1144         if (status)
1145                 return status;
1146
1147         if (is_quota_modification(inode, attr))
1148                 dquot_initialize(inode);
1149         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1150         if (size_change) {
1151                 status = ocfs2_rw_lock(inode, 1);
1152                 if (status < 0) {
1153                         mlog_errno(status);
1154                         goto bail;
1155                 }
1156         }
1157
1158         status = ocfs2_inode_lock(inode, &bh, 1);
1159         if (status < 0) {
1160                 if (status != -ENOENT)
1161                         mlog_errno(status);
1162                 goto bail_unlock_rw;
1163         }
1164
1165         if (size_change && attr->ia_size != i_size_read(inode)) {
1166                 status = inode_newsize_ok(inode, attr->ia_size);
1167                 if (status)
1168                         goto bail_unlock;
1169
1170                 if (i_size_read(inode) > attr->ia_size) {
1171                         if (ocfs2_should_order_data(inode)) {
1172                                 status = ocfs2_begin_ordered_truncate(inode,
1173                                                                       attr->ia_size);
1174                                 if (status)
1175                                         goto bail_unlock;
1176                         }
1177                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1178                 } else
1179                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1180                 if (status < 0) {
1181                         if (status != -ENOSPC)
1182                                 mlog_errno(status);
1183                         status = -ENOSPC;
1184                         goto bail_unlock;
1185                 }
1186         }
1187
1188         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1189             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1190                 /*
1191                  * Gather pointers to quota structures so that allocation /
1192                  * freeing of quota structures happens here and not inside
1193                  * dquot_transfer() where we have problems with lock ordering
1194                  */
1195                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1196                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1197                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1198                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1199                                                       USRQUOTA);
1200                         if (!transfer_to[USRQUOTA]) {
1201                                 status = -ESRCH;
1202                                 goto bail_unlock;
1203                         }
1204                 }
1205                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1206                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1207                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1208                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1209                                                       GRPQUOTA);
1210                         if (!transfer_to[GRPQUOTA]) {
1211                                 status = -ESRCH;
1212                                 goto bail_unlock;
1213                         }
1214                 }
1215                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1216                                            2 * ocfs2_quota_trans_credits(sb));
1217                 if (IS_ERR(handle)) {
1218                         status = PTR_ERR(handle);
1219                         mlog_errno(status);
1220                         goto bail_unlock;
1221                 }
1222                 status = __dquot_transfer(inode, transfer_to);
1223                 if (status < 0)
1224                         goto bail_commit;
1225         } else {
1226                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1227                 if (IS_ERR(handle)) {
1228                         status = PTR_ERR(handle);
1229                         mlog_errno(status);
1230                         goto bail_unlock;
1231                 }
1232         }
1233
1234         /*
1235          * This will intentionally not wind up calling truncate_setsize(),
1236          * since all the work for a size change has been done above.
1237          * Otherwise, we could get into problems with truncate as
1238          * ip_alloc_sem is used there to protect against i_size
1239          * changes.
1240          *
1241          * XXX: this means the conditional below can probably be removed.
1242          */
1243         if ((attr->ia_valid & ATTR_SIZE) &&
1244             attr->ia_size != i_size_read(inode)) {
1245                 status = vmtruncate(inode, attr->ia_size);
1246                 if (status) {
1247                         mlog_errno(status);
1248                         goto bail_commit;
1249                 }
1250         }
1251
1252         setattr_copy(inode, attr);
1253         mark_inode_dirty(inode);
1254
1255         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1256         if (status < 0)
1257                 mlog_errno(status);
1258
1259 bail_commit:
1260         ocfs2_commit_trans(osb, handle);
1261 bail_unlock:
1262         ocfs2_inode_unlock(inode, 1);
1263 bail_unlock_rw:
1264         if (size_change)
1265                 ocfs2_rw_unlock(inode, 1);
1266 bail:
1267         brelse(bh);
1268
1269         /* Release quota pointers in case we acquired them */
1270         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1271                 dqput(transfer_to[qtype]);
1272
1273         if (!status && attr->ia_valid & ATTR_MODE) {
1274                 status = ocfs2_acl_chmod(inode);
1275                 if (status < 0)
1276                         mlog_errno(status);
1277         }
1278
1279         mlog_exit(status);
1280         return status;
1281 }
1282
1283 int ocfs2_getattr(struct vfsmount *mnt,
1284                   struct dentry *dentry,
1285                   struct kstat *stat)
1286 {
1287         struct inode *inode = dentry->d_inode;
1288         struct super_block *sb = dentry->d_inode->i_sb;
1289         struct ocfs2_super *osb = sb->s_fs_info;
1290         int err;
1291
1292         mlog_entry_void();
1293
1294         err = ocfs2_inode_revalidate(dentry);
1295         if (err) {
1296                 if (err != -ENOENT)
1297                         mlog_errno(err);
1298                 goto bail;
1299         }
1300
1301         generic_fillattr(inode, stat);
1302
1303         /* We set the blksize from the cluster size for performance */
1304         stat->blksize = osb->s_clustersize;
1305
1306 bail:
1307         mlog_exit(err);
1308
1309         return err;
1310 }
1311
1312 int ocfs2_permission(struct inode *inode, int mask)
1313 {
1314         int ret;
1315
1316         mlog_entry_void();
1317
1318         ret = ocfs2_inode_lock(inode, NULL, 0);
1319         if (ret) {
1320                 if (ret != -ENOENT)
1321                         mlog_errno(ret);
1322                 goto out;
1323         }
1324
1325         ret = generic_permission(inode, mask, ocfs2_check_acl);
1326
1327         ocfs2_inode_unlock(inode, 0);
1328 out:
1329         mlog_exit(ret);
1330         return ret;
1331 }
1332
1333 static int __ocfs2_write_remove_suid(struct inode *inode,
1334                                      struct buffer_head *bh)
1335 {
1336         int ret;
1337         handle_t *handle;
1338         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1339         struct ocfs2_dinode *di;
1340
1341         mlog_entry("(Inode %llu, mode 0%o)\n",
1342                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1343
1344         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1345         if (IS_ERR(handle)) {
1346                 ret = PTR_ERR(handle);
1347                 mlog_errno(ret);
1348                 goto out;
1349         }
1350
1351         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1352                                       OCFS2_JOURNAL_ACCESS_WRITE);
1353         if (ret < 0) {
1354                 mlog_errno(ret);
1355                 goto out_trans;
1356         }
1357
1358         inode->i_mode &= ~S_ISUID;
1359         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1360                 inode->i_mode &= ~S_ISGID;
1361
1362         di = (struct ocfs2_dinode *) bh->b_data;
1363         di->i_mode = cpu_to_le16(inode->i_mode);
1364
1365         ocfs2_journal_dirty(handle, bh);
1366
1367 out_trans:
1368         ocfs2_commit_trans(osb, handle);
1369 out:
1370         mlog_exit(ret);
1371         return ret;
1372 }
1373
1374 /*
1375  * Will look for holes and unwritten extents in the range starting at
1376  * pos for count bytes (inclusive).
1377  */
1378 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1379                                        size_t count)
1380 {
1381         int ret = 0;
1382         unsigned int extent_flags;
1383         u32 cpos, clusters, extent_len, phys_cpos;
1384         struct super_block *sb = inode->i_sb;
1385
1386         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1387         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1388
1389         while (clusters) {
1390                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1391                                          &extent_flags);
1392                 if (ret < 0) {
1393                         mlog_errno(ret);
1394                         goto out;
1395                 }
1396
1397                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1398                         ret = 1;
1399                         break;
1400                 }
1401
1402                 if (extent_len > clusters)
1403                         extent_len = clusters;
1404
1405                 clusters -= extent_len;
1406                 cpos += extent_len;
1407         }
1408 out:
1409         return ret;
1410 }
1411
1412 static int ocfs2_write_remove_suid(struct inode *inode)
1413 {
1414         int ret;
1415         struct buffer_head *bh = NULL;
1416
1417         ret = ocfs2_read_inode_block(inode, &bh);
1418         if (ret < 0) {
1419                 mlog_errno(ret);
1420                 goto out;
1421         }
1422
1423         ret =  __ocfs2_write_remove_suid(inode, bh);
1424 out:
1425         brelse(bh);
1426         return ret;
1427 }
1428
1429 /*
1430  * Allocate enough extents to cover the region starting at byte offset
1431  * start for len bytes. Existing extents are skipped, any extents
1432  * added are marked as "unwritten".
1433  */
1434 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1435                                             u64 start, u64 len)
1436 {
1437         int ret;
1438         u32 cpos, phys_cpos, clusters, alloc_size;
1439         u64 end = start + len;
1440         struct buffer_head *di_bh = NULL;
1441
1442         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1443                 ret = ocfs2_read_inode_block(inode, &di_bh);
1444                 if (ret) {
1445                         mlog_errno(ret);
1446                         goto out;
1447                 }
1448
1449                 /*
1450                  * Nothing to do if the requested reservation range
1451                  * fits within the inode.
1452                  */
1453                 if (ocfs2_size_fits_inline_data(di_bh, end))
1454                         goto out;
1455
1456                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1457                 if (ret) {
1458                         mlog_errno(ret);
1459                         goto out;
1460                 }
1461         }
1462
1463         /*
1464          * We consider both start and len to be inclusive.
1465          */
1466         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1467         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1468         clusters -= cpos;
1469
1470         while (clusters) {
1471                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1472                                          &alloc_size, NULL);
1473                 if (ret) {
1474                         mlog_errno(ret);
1475                         goto out;
1476                 }
1477
1478                 /*
1479                  * Hole or existing extent len can be arbitrary, so
1480                  * cap it to our own allocation request.
1481                  */
1482                 if (alloc_size > clusters)
1483                         alloc_size = clusters;
1484
1485                 if (phys_cpos) {
1486                         /*
1487                          * We already have an allocation at this
1488                          * region so we can safely skip it.
1489                          */
1490                         goto next;
1491                 }
1492
1493                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1494                 if (ret) {
1495                         if (ret != -ENOSPC)
1496                                 mlog_errno(ret);
1497                         goto out;
1498                 }
1499
1500 next:
1501                 cpos += alloc_size;
1502                 clusters -= alloc_size;
1503         }
1504
1505         ret = 0;
1506 out:
1507
1508         brelse(di_bh);
1509         return ret;
1510 }
1511
1512 /*
1513  * Truncate a byte range, avoiding pages within partial clusters. This
1514  * preserves those pages for the zeroing code to write to.
1515  */
1516 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1517                                          u64 byte_len)
1518 {
1519         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1520         loff_t start, end;
1521         struct address_space *mapping = inode->i_mapping;
1522
1523         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1524         end = byte_start + byte_len;
1525         end = end & ~(osb->s_clustersize - 1);
1526
1527         if (start < end) {
1528                 unmap_mapping_range(mapping, start, end - start, 0);
1529                 truncate_inode_pages_range(mapping, start, end - 1);
1530         }
1531 }
1532
1533 static int ocfs2_zero_partial_clusters(struct inode *inode,
1534                                        u64 start, u64 len)
1535 {
1536         int ret = 0;
1537         u64 tmpend, end = start + len;
1538         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1539         unsigned int csize = osb->s_clustersize;
1540         handle_t *handle;
1541
1542         /*
1543          * The "start" and "end" values are NOT necessarily part of
1544          * the range whose allocation is being deleted. Rather, this
1545          * is what the user passed in with the request. We must zero
1546          * partial clusters here. There's no need to worry about
1547          * physical allocation - the zeroing code knows to skip holes.
1548          */
1549         mlog(0, "byte start: %llu, end: %llu\n",
1550              (unsigned long long)start, (unsigned long long)end);
1551
1552         /*
1553          * If both edges are on a cluster boundary then there's no
1554          * zeroing required as the region is part of the allocation to
1555          * be truncated.
1556          */
1557         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1558                 goto out;
1559
1560         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1561         if (IS_ERR(handle)) {
1562                 ret = PTR_ERR(handle);
1563                 mlog_errno(ret);
1564                 goto out;
1565         }
1566
1567         /*
1568          * We want to get the byte offset of the end of the 1st cluster.
1569          */
1570         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1571         if (tmpend > end)
1572                 tmpend = end;
1573
1574         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1575              (unsigned long long)start, (unsigned long long)tmpend);
1576
1577         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1578         if (ret)
1579                 mlog_errno(ret);
1580
1581         if (tmpend < end) {
1582                 /*
1583                  * This may make start and end equal, but the zeroing
1584                  * code will skip any work in that case so there's no
1585                  * need to catch it up here.
1586                  */
1587                 start = end & ~(osb->s_clustersize - 1);
1588
1589                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1590                      (unsigned long long)start, (unsigned long long)end);
1591
1592                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1593                 if (ret)
1594                         mlog_errno(ret);
1595         }
1596
1597         ocfs2_commit_trans(osb, handle);
1598 out:
1599         return ret;
1600 }
1601
1602 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1603 {
1604         int i;
1605         struct ocfs2_extent_rec *rec = NULL;
1606
1607         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1608
1609                 rec = &el->l_recs[i];
1610
1611                 if (le32_to_cpu(rec->e_cpos) < pos)
1612                         break;
1613         }
1614
1615         return i;
1616 }
1617
1618 /*
1619  * Helper to calculate the punching pos and length in one run, we handle the
1620  * following three cases in order:
1621  *
1622  * - remove the entire record
1623  * - remove a partial record
1624  * - no record needs to be removed (hole-punching completed)
1625 */
1626 static void ocfs2_calc_trunc_pos(struct inode *inode,
1627                                  struct ocfs2_extent_list *el,
1628                                  struct ocfs2_extent_rec *rec,
1629                                  u32 trunc_start, u32 *trunc_cpos,
1630                                  u32 *trunc_len, u32 *trunc_end,
1631                                  u64 *blkno, int *done)
1632 {
1633         int ret = 0;
1634         u32 coff, range;
1635
1636         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1637
1638         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1639                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1640                 /*
1641                  * Skip holes if any.
1642                  */
1643                 if (range < *trunc_end)
1644                         *trunc_end = range;
1645                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1646                 *blkno = le64_to_cpu(rec->e_blkno);
1647                 *trunc_end = le32_to_cpu(rec->e_cpos);
1648         } else if (range > trunc_start) {
1649                 *trunc_cpos = trunc_start;
1650                 *trunc_len = *trunc_end - trunc_start;
1651                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1652                 *blkno = le64_to_cpu(rec->e_blkno) +
1653                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1654                 *trunc_end = trunc_start;
1655         } else {
1656                 /*
1657                  * It may have two following possibilities:
1658                  *
1659                  * - last record has been removed
1660                  * - trunc_start was within a hole
1661                  *
1662                  * both two cases mean the completion of hole punching.
1663                  */
1664                 ret = 1;
1665         }
1666
1667         *done = ret;
1668 }
1669
1670 static int ocfs2_remove_inode_range(struct inode *inode,
1671                                     struct buffer_head *di_bh, u64 byte_start,
1672                                     u64 byte_len)
1673 {
1674         int ret = 0, flags = 0, done = 0, i;
1675         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1676         u32 cluster_in_el;
1677         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1678         struct ocfs2_cached_dealloc_ctxt dealloc;
1679         struct address_space *mapping = inode->i_mapping;
1680         struct ocfs2_extent_tree et;
1681         struct ocfs2_path *path = NULL;
1682         struct ocfs2_extent_list *el = NULL;
1683         struct ocfs2_extent_rec *rec = NULL;
1684         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1685         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1686
1687         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1688         ocfs2_init_dealloc_ctxt(&dealloc);
1689
1690         if (byte_len == 0)
1691                 return 0;
1692
1693         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1694                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1695                                             byte_start + byte_len, 0);
1696                 if (ret) {
1697                         mlog_errno(ret);
1698                         goto out;
1699                 }
1700                 /*
1701                  * There's no need to get fancy with the page cache
1702                  * truncate of an inline-data inode. We're talking
1703                  * about less than a page here, which will be cached
1704                  * in the dinode buffer anyway.
1705                  */
1706                 unmap_mapping_range(mapping, 0, 0, 0);
1707                 truncate_inode_pages(mapping, 0);
1708                 goto out;
1709         }
1710
1711         /*
1712          * For reflinks, we may need to CoW 2 clusters which might be
1713          * partially zero'd later, if hole's start and end offset were
1714          * within one cluster(means is not exactly aligned to clustersize).
1715          */
1716
1717         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1718
1719                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1720                 if (ret) {
1721                         mlog_errno(ret);
1722                         goto out;
1723                 }
1724
1725                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1726                 if (ret) {
1727                         mlog_errno(ret);
1728                         goto out;
1729                 }
1730         }
1731
1732         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1733         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1734         cluster_in_el = trunc_end;
1735
1736         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1737              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1738              (unsigned long long)byte_start,
1739              (unsigned long long)byte_len, trunc_start, trunc_end);
1740
1741         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1742         if (ret) {
1743                 mlog_errno(ret);
1744                 goto out;
1745         }
1746
1747         path = ocfs2_new_path_from_et(&et);
1748         if (!path) {
1749                 ret = -ENOMEM;
1750                 mlog_errno(ret);
1751                 goto out;
1752         }
1753
1754         while (trunc_end > trunc_start) {
1755
1756                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1757                                       cluster_in_el);
1758                 if (ret) {
1759                         mlog_errno(ret);
1760                         goto out;
1761                 }
1762
1763                 el = path_leaf_el(path);
1764
1765                 i = ocfs2_find_rec(el, trunc_end);
1766                 /*
1767                  * Need to go to previous extent block.
1768                  */
1769                 if (i < 0) {
1770                         if (path->p_tree_depth == 0)
1771                                 break;
1772
1773                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1774                                                             path,
1775                                                             &cluster_in_el);
1776                         if (ret) {
1777                                 mlog_errno(ret);
1778                                 goto out;
1779                         }
1780
1781                         /*
1782                          * We've reached the leftmost extent block,
1783                          * it's safe to leave.
1784                          */
1785                         if (cluster_in_el == 0)
1786                                 break;
1787
1788                         /*
1789                          * The 'pos' searched for previous extent block is
1790                          * always one cluster less than actual trunc_end.
1791                          */
1792                         trunc_end = cluster_in_el + 1;
1793
1794                         ocfs2_reinit_path(path, 1);
1795
1796                         continue;
1797
1798                 } else
1799                         rec = &el->l_recs[i];
1800
1801                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1802                                      &trunc_len, &trunc_end, &blkno, &done);
1803                 if (done)
1804                         break;
1805
1806                 flags = rec->e_flags;
1807                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1808
1809                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1810                                                phys_cpos, trunc_len, flags,
1811                                                &dealloc, refcount_loc);
1812                 if (ret < 0) {
1813                         mlog_errno(ret);
1814                         goto out;
1815                 }
1816
1817                 cluster_in_el = trunc_end;
1818
1819                 ocfs2_reinit_path(path, 1);
1820         }
1821
1822         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1823
1824 out:
1825         ocfs2_schedule_truncate_log_flush(osb, 1);
1826         ocfs2_run_deallocs(osb, &dealloc);
1827
1828         return ret;
1829 }
1830
1831 /*
1832  * Parts of this function taken from xfs_change_file_space()
1833  */
1834 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1835                                      loff_t f_pos, unsigned int cmd,
1836                                      struct ocfs2_space_resv *sr,
1837                                      int change_size)
1838 {
1839         int ret;
1840         s64 llen;
1841         loff_t size;
1842         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1843         struct buffer_head *di_bh = NULL;
1844         handle_t *handle;
1845         unsigned long long max_off = inode->i_sb->s_maxbytes;
1846
1847         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1848                 return -EROFS;
1849
1850         mutex_lock(&inode->i_mutex);
1851
1852         /*
1853          * This prevents concurrent writes on other nodes
1854          */
1855         ret = ocfs2_rw_lock(inode, 1);
1856         if (ret) {
1857                 mlog_errno(ret);
1858                 goto out;
1859         }
1860
1861         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1862         if (ret) {
1863                 mlog_errno(ret);
1864                 goto out_rw_unlock;
1865         }
1866
1867         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1868                 ret = -EPERM;
1869                 goto out_inode_unlock;
1870         }
1871
1872         switch (sr->l_whence) {
1873         case 0: /*SEEK_SET*/
1874                 break;
1875         case 1: /*SEEK_CUR*/
1876                 sr->l_start += f_pos;
1877                 break;
1878         case 2: /*SEEK_END*/
1879                 sr->l_start += i_size_read(inode);
1880                 break;
1881         default:
1882                 ret = -EINVAL;
1883                 goto out_inode_unlock;
1884         }
1885         sr->l_whence = 0;
1886
1887         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1888
1889         if (sr->l_start < 0
1890             || sr->l_start > max_off
1891             || (sr->l_start + llen) < 0
1892             || (sr->l_start + llen) > max_off) {
1893                 ret = -EINVAL;
1894                 goto out_inode_unlock;
1895         }
1896         size = sr->l_start + sr->l_len;
1897
1898         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1899                 if (sr->l_len <= 0) {
1900                         ret = -EINVAL;
1901                         goto out_inode_unlock;
1902                 }
1903         }
1904
1905         if (file && should_remove_suid(file->f_path.dentry)) {
1906                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1907                 if (ret) {
1908                         mlog_errno(ret);
1909                         goto out_inode_unlock;
1910                 }
1911         }
1912
1913         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1914         switch (cmd) {
1915         case OCFS2_IOC_RESVSP:
1916         case OCFS2_IOC_RESVSP64:
1917                 /*
1918                  * This takes unsigned offsets, but the signed ones we
1919                  * pass have been checked against overflow above.
1920                  */
1921                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1922                                                        sr->l_len);
1923                 break;
1924         case OCFS2_IOC_UNRESVSP:
1925         case OCFS2_IOC_UNRESVSP64:
1926                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1927                                                sr->l_len);
1928                 break;
1929         default:
1930                 ret = -EINVAL;
1931         }
1932         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1933         if (ret) {
1934                 mlog_errno(ret);
1935                 goto out_inode_unlock;
1936         }
1937
1938         /*
1939          * We update c/mtime for these changes
1940          */
1941         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1942         if (IS_ERR(handle)) {
1943                 ret = PTR_ERR(handle);
1944                 mlog_errno(ret);
1945                 goto out_inode_unlock;
1946         }
1947
1948         if (change_size && i_size_read(inode) < size)
1949                 i_size_write(inode, size);
1950
1951         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1952         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1953         if (ret < 0)
1954                 mlog_errno(ret);
1955
1956         ocfs2_commit_trans(osb, handle);
1957
1958 out_inode_unlock:
1959         brelse(di_bh);
1960         ocfs2_inode_unlock(inode, 1);
1961 out_rw_unlock:
1962         ocfs2_rw_unlock(inode, 1);
1963
1964 out:
1965         mutex_unlock(&inode->i_mutex);
1966         return ret;
1967 }
1968
1969 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1970                             struct ocfs2_space_resv *sr)
1971 {
1972         struct inode *inode = file->f_path.dentry->d_inode;
1973         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1974
1975         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1976             !ocfs2_writes_unwritten_extents(osb))
1977                 return -ENOTTY;
1978         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1979                  !ocfs2_sparse_alloc(osb))
1980                 return -ENOTTY;
1981
1982         if (!S_ISREG(inode->i_mode))
1983                 return -EINVAL;
1984
1985         if (!(file->f_mode & FMODE_WRITE))
1986                 return -EBADF;
1987
1988         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1989 }
1990
1991 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1992                             loff_t len)
1993 {
1994         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1995         struct ocfs2_space_resv sr;
1996         int change_size = 1;
1997
1998         if (!ocfs2_writes_unwritten_extents(osb))
1999                 return -EOPNOTSUPP;
2000
2001         if (S_ISDIR(inode->i_mode))
2002                 return -ENODEV;
2003
2004         if (mode & FALLOC_FL_KEEP_SIZE)
2005                 change_size = 0;
2006
2007         sr.l_whence = 0;
2008         sr.l_start = (s64)offset;
2009         sr.l_len = (s64)len;
2010
2011         return __ocfs2_change_file_space(NULL, inode, offset,
2012                                          OCFS2_IOC_RESVSP64, &sr, change_size);
2013 }
2014
2015 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2016                                    size_t count)
2017 {
2018         int ret = 0;
2019         unsigned int extent_flags;
2020         u32 cpos, clusters, extent_len, phys_cpos;
2021         struct super_block *sb = inode->i_sb;
2022
2023         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2024             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2025             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2026                 return 0;
2027
2028         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2029         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2030
2031         while (clusters) {
2032                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2033                                          &extent_flags);
2034                 if (ret < 0) {
2035                         mlog_errno(ret);
2036                         goto out;
2037                 }
2038
2039                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2040                         ret = 1;
2041                         break;
2042                 }
2043
2044                 if (extent_len > clusters)
2045                         extent_len = clusters;
2046
2047                 clusters -= extent_len;
2048                 cpos += extent_len;
2049         }
2050 out:
2051         return ret;
2052 }
2053
2054 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2055                                             struct file *file,
2056                                             loff_t pos, size_t count,
2057                                             int *meta_level)
2058 {
2059         int ret;
2060         struct buffer_head *di_bh = NULL;
2061         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2062         u32 clusters =
2063                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2064
2065         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2066         if (ret) {
2067                 mlog_errno(ret);
2068                 goto out;
2069         }
2070
2071         *meta_level = 1;
2072
2073         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2074         if (ret)
2075                 mlog_errno(ret);
2076 out:
2077         brelse(di_bh);
2078         return ret;
2079 }
2080
2081 static int ocfs2_prepare_inode_for_write(struct file *file,
2082                                          loff_t *ppos,
2083                                          size_t count,
2084                                          int appending,
2085                                          int *direct_io,
2086                                          int *has_refcount)
2087 {
2088         int ret = 0, meta_level = 0;
2089         struct dentry *dentry = file->f_path.dentry;
2090         struct inode *inode = dentry->d_inode;
2091         loff_t saved_pos, end;
2092
2093         /*
2094          * We start with a read level meta lock and only jump to an ex
2095          * if we need to make modifications here.
2096          */
2097         for(;;) {
2098                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2099                 if (ret < 0) {
2100                         meta_level = -1;
2101                         mlog_errno(ret);
2102                         goto out;
2103                 }
2104
2105                 /* Clear suid / sgid if necessary. We do this here
2106                  * instead of later in the write path because
2107                  * remove_suid() calls ->setattr without any hint that
2108                  * we may have already done our cluster locking. Since
2109                  * ocfs2_setattr() *must* take cluster locks to
2110                  * proceeed, this will lead us to recursively lock the
2111                  * inode. There's also the dinode i_size state which
2112                  * can be lost via setattr during extending writes (we
2113                  * set inode->i_size at the end of a write. */
2114                 if (should_remove_suid(dentry)) {
2115                         if (meta_level == 0) {
2116                                 ocfs2_inode_unlock(inode, meta_level);
2117                                 meta_level = 1;
2118                                 continue;
2119                         }
2120
2121                         ret = ocfs2_write_remove_suid(inode);
2122                         if (ret < 0) {
2123                                 mlog_errno(ret);
2124                                 goto out_unlock;
2125                         }
2126                 }
2127
2128                 /* work on a copy of ppos until we're sure that we won't have
2129                  * to recalculate it due to relocking. */
2130                 if (appending) {
2131                         saved_pos = i_size_read(inode);
2132                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2133                 } else {
2134                         saved_pos = *ppos;
2135                 }
2136
2137                 end = saved_pos + count;
2138
2139                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2140                 if (ret == 1) {
2141                         ocfs2_inode_unlock(inode, meta_level);
2142                         meta_level = -1;
2143
2144                         ret = ocfs2_prepare_inode_for_refcount(inode,
2145                                                                file,
2146                                                                saved_pos,
2147                                                                count,
2148                                                                &meta_level);
2149                         if (has_refcount)
2150                                 *has_refcount = 1;
2151                         if (direct_io)
2152                                 *direct_io = 0;
2153                 }
2154
2155                 if (ret < 0) {
2156                         mlog_errno(ret);
2157                         goto out_unlock;
2158                 }
2159
2160                 /*
2161                  * Skip the O_DIRECT checks if we don't need
2162                  * them.
2163                  */
2164                 if (!direct_io || !(*direct_io))
2165                         break;
2166
2167                 /*
2168                  * There's no sane way to do direct writes to an inode
2169                  * with inline data.
2170                  */
2171                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2172                         *direct_io = 0;
2173                         break;
2174                 }
2175
2176                 /*
2177                  * Allowing concurrent direct writes means
2178                  * i_size changes wouldn't be synchronized, so
2179                  * one node could wind up truncating another
2180                  * nodes writes.
2181                  */
2182                 if (end > i_size_read(inode)) {
2183                         *direct_io = 0;
2184                         break;
2185                 }
2186
2187                 /*
2188                  * We don't fill holes during direct io, so
2189                  * check for them here. If any are found, the
2190                  * caller will have to retake some cluster
2191                  * locks and initiate the io as buffered.
2192                  */
2193                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2194                 if (ret == 1) {
2195                         *direct_io = 0;
2196                         ret = 0;
2197                 } else if (ret < 0)
2198                         mlog_errno(ret);
2199                 break;
2200         }
2201
2202         if (appending)
2203                 *ppos = saved_pos;
2204
2205 out_unlock:
2206         if (meta_level >= 0)
2207                 ocfs2_inode_unlock(inode, meta_level);
2208
2209 out:
2210         return ret;
2211 }
2212
2213 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2214                                     const struct iovec *iov,
2215                                     unsigned long nr_segs,
2216                                     loff_t pos)
2217 {
2218         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2219         int can_do_direct, has_refcount = 0;
2220         ssize_t written = 0;
2221         size_t ocount;          /* original count */
2222         size_t count;           /* after file limit checks */
2223         loff_t old_size, *ppos = &iocb->ki_pos;
2224         u32 old_clusters;
2225         struct file *file = iocb->ki_filp;
2226         struct inode *inode = file->f_path.dentry->d_inode;
2227         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2228         int full_coherency = !(osb->s_mount_opt &
2229                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2230
2231         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2232                    (unsigned int)nr_segs,
2233                    file->f_path.dentry->d_name.len,
2234                    file->f_path.dentry->d_name.name);
2235
2236         if (iocb->ki_left == 0)
2237                 return 0;
2238
2239         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2240
2241         appending = file->f_flags & O_APPEND ? 1 : 0;
2242         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2243
2244         mutex_lock(&inode->i_mutex);
2245
2246 relock:
2247         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2248         if (direct_io) {
2249                 down_read(&inode->i_alloc_sem);
2250                 have_alloc_sem = 1;
2251         }
2252
2253         /*
2254          * Concurrent O_DIRECT writes are allowed with
2255          * mount_option "coherency=buffered".
2256          */
2257         rw_level = (!direct_io || full_coherency);
2258
2259         ret = ocfs2_rw_lock(inode, rw_level);
2260         if (ret < 0) {
2261                 mlog_errno(ret);
2262                 goto out_sems;
2263         }
2264
2265         /*
2266          * O_DIRECT writes with "coherency=full" need to take EX cluster
2267          * inode_lock to guarantee coherency.
2268          */
2269         if (direct_io && full_coherency) {
2270                 /*
2271                  * We need to take and drop the inode lock to force
2272                  * other nodes to drop their caches.  Buffered I/O
2273                  * already does this in write_begin().
2274                  */
2275                 ret = ocfs2_inode_lock(inode, NULL, 1);
2276                 if (ret < 0) {
2277                         mlog_errno(ret);
2278                         goto out_sems;
2279                 }
2280
2281                 ocfs2_inode_unlock(inode, 1);
2282         }
2283
2284         can_do_direct = direct_io;
2285         ret = ocfs2_prepare_inode_for_write(file, ppos,
2286                                             iocb->ki_left, appending,
2287                                             &can_do_direct, &has_refcount);
2288         if (ret < 0) {
2289                 mlog_errno(ret);
2290                 goto out;
2291         }
2292
2293         /*
2294          * We can't complete the direct I/O as requested, fall back to
2295          * buffered I/O.
2296          */
2297         if (direct_io && !can_do_direct) {
2298                 ocfs2_rw_unlock(inode, rw_level);
2299                 up_read(&inode->i_alloc_sem);
2300
2301                 have_alloc_sem = 0;
2302                 rw_level = -1;
2303
2304                 direct_io = 0;
2305                 goto relock;
2306         }
2307
2308         /*
2309          * To later detect whether a journal commit for sync writes is
2310          * necessary, we sample i_size, and cluster count here.
2311          */
2312         old_size = i_size_read(inode);
2313         old_clusters = OCFS2_I(inode)->ip_clusters;
2314
2315         /* communicate with ocfs2_dio_end_io */
2316         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2317
2318         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2319                                      VERIFY_READ);
2320         if (ret)
2321                 goto out_dio;
2322
2323         count = ocount;
2324         ret = generic_write_checks(file, ppos, &count,
2325                                    S_ISBLK(inode->i_mode));
2326         if (ret)
2327                 goto out_dio;
2328
2329         if (direct_io) {
2330                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2331                                                     ppos, count, ocount);
2332                 if (written < 0) {
2333                         ret = written;
2334                         goto out_dio;
2335                 }
2336         } else {
2337                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2338                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2339                                                       ppos, count, 0);
2340                 current->backing_dev_info = NULL;
2341         }
2342
2343 out_dio:
2344         /* buffered aio wouldn't have proper lock coverage today */
2345         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2346
2347         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2348             ((file->f_flags & O_DIRECT) && !direct_io)) {
2349                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2350                                                pos + count - 1);
2351                 if (ret < 0)
2352                         written = ret;
2353
2354                 if (!ret && ((old_size != i_size_read(inode)) ||
2355                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2356                              has_refcount)) {
2357                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2358                         if (ret < 0)
2359                                 written = ret;
2360                 }
2361
2362                 if (!ret)
2363                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2364                                                       pos + count - 1);
2365         }
2366
2367         /*
2368          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2369          * function pointer which is called when o_direct io completes so that
2370          * it can unlock our rw lock.  (it's the clustered equivalent of
2371          * i_alloc_sem; protects truncate from racing with pending ios).
2372          * Unfortunately there are error cases which call end_io and others
2373          * that don't.  so we don't have to unlock the rw_lock if either an
2374          * async dio is going to do it in the future or an end_io after an
2375          * error has already done it.
2376          */
2377         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2378                 rw_level = -1;
2379                 have_alloc_sem = 0;
2380         }
2381
2382 out:
2383         if (rw_level != -1)
2384                 ocfs2_rw_unlock(inode, rw_level);
2385
2386 out_sems:
2387         if (have_alloc_sem)
2388                 up_read(&inode->i_alloc_sem);
2389
2390         mutex_unlock(&inode->i_mutex);
2391
2392         if (written)
2393                 ret = written;
2394         mlog_exit(ret);
2395         return ret;
2396 }
2397
2398 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2399                                 struct file *out,
2400                                 struct splice_desc *sd)
2401 {
2402         int ret;
2403
2404         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2405                                             sd->total_len, 0, NULL, NULL);
2406         if (ret < 0) {
2407                 mlog_errno(ret);
2408                 return ret;
2409         }
2410
2411         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2412 }
2413
2414 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2415                                        struct file *out,
2416                                        loff_t *ppos,
2417                                        size_t len,
2418                                        unsigned int flags)
2419 {
2420         int ret;
2421         struct address_space *mapping = out->f_mapping;
2422         struct inode *inode = mapping->host;
2423         struct splice_desc sd = {
2424                 .total_len = len,
2425                 .flags = flags,
2426                 .pos = *ppos,
2427                 .u.file = out,
2428         };
2429
2430         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2431                    (unsigned int)len,
2432                    out->f_path.dentry->d_name.len,
2433                    out->f_path.dentry->d_name.name);
2434
2435         if (pipe->inode)
2436                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2437
2438         splice_from_pipe_begin(&sd);
2439         do {
2440                 ret = splice_from_pipe_next(pipe, &sd);
2441                 if (ret <= 0)
2442                         break;
2443
2444                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2445                 ret = ocfs2_rw_lock(inode, 1);
2446                 if (ret < 0)
2447                         mlog_errno(ret);
2448                 else {
2449                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2450                         ocfs2_rw_unlock(inode, 1);
2451                 }
2452                 mutex_unlock(&inode->i_mutex);
2453         } while (ret > 0);
2454         splice_from_pipe_end(pipe, &sd);
2455
2456         if (pipe->inode)
2457                 mutex_unlock(&pipe->inode->i_mutex);
2458
2459         if (sd.num_spliced)
2460                 ret = sd.num_spliced;
2461
2462         if (ret > 0) {
2463                 unsigned long nr_pages;
2464                 int err;
2465
2466                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2467
2468                 err = generic_write_sync(out, *ppos, ret);
2469                 if (err)
2470                         ret = err;
2471                 else
2472                         *ppos += ret;
2473
2474                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2475         }
2476
2477         mlog_exit(ret);
2478         return ret;
2479 }
2480
2481 static ssize_t ocfs2_file_splice_read(struct file *in,
2482                                       loff_t *ppos,
2483                                       struct pipe_inode_info *pipe,
2484                                       size_t len,
2485                                       unsigned int flags)
2486 {
2487         int ret = 0, lock_level = 0;
2488         struct inode *inode = in->f_path.dentry->d_inode;
2489
2490         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2491                    (unsigned int)len,
2492                    in->f_path.dentry->d_name.len,
2493                    in->f_path.dentry->d_name.name);
2494
2495         /*
2496          * See the comment in ocfs2_file_aio_read()
2497          */
2498         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2499         if (ret < 0) {
2500                 mlog_errno(ret);
2501                 goto bail;
2502         }
2503         ocfs2_inode_unlock(inode, lock_level);
2504
2505         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2506
2507 bail:
2508         mlog_exit(ret);
2509         return ret;
2510 }
2511
2512 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2513                                    const struct iovec *iov,
2514                                    unsigned long nr_segs,
2515                                    loff_t pos)
2516 {
2517         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2518         struct file *filp = iocb->ki_filp;
2519         struct inode *inode = filp->f_path.dentry->d_inode;
2520
2521         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2522                    (unsigned int)nr_segs,
2523                    filp->f_path.dentry->d_name.len,
2524                    filp->f_path.dentry->d_name.name);
2525
2526         if (!inode) {
2527                 ret = -EINVAL;
2528                 mlog_errno(ret);
2529                 goto bail;
2530         }
2531
2532         /*
2533          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2534          * need locks to protect pending reads from racing with truncate.
2535          */
2536         if (filp->f_flags & O_DIRECT) {
2537                 down_read(&inode->i_alloc_sem);
2538                 have_alloc_sem = 1;
2539
2540                 ret = ocfs2_rw_lock(inode, 0);
2541                 if (ret < 0) {
2542                         mlog_errno(ret);
2543                         goto bail;
2544                 }
2545                 rw_level = 0;
2546                 /* communicate with ocfs2_dio_end_io */
2547                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2548         }
2549
2550         /*
2551          * We're fine letting folks race truncates and extending
2552          * writes with read across the cluster, just like they can
2553          * locally. Hence no rw_lock during read.
2554          *
2555          * Take and drop the meta data lock to update inode fields
2556          * like i_size. This allows the checks down below
2557          * generic_file_aio_read() a chance of actually working.
2558          */
2559         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2560         if (ret < 0) {
2561                 mlog_errno(ret);
2562                 goto bail;
2563         }
2564         ocfs2_inode_unlock(inode, lock_level);
2565
2566         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2567         if (ret == -EINVAL)
2568                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2569
2570         /* buffered aio wouldn't have proper lock coverage today */
2571         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2572
2573         /* see ocfs2_file_aio_write */
2574         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2575                 rw_level = -1;
2576                 have_alloc_sem = 0;
2577         }
2578
2579 bail:
2580         if (have_alloc_sem)
2581                 up_read(&inode->i_alloc_sem);
2582         if (rw_level != -1)
2583                 ocfs2_rw_unlock(inode, rw_level);
2584         mlog_exit(ret);
2585
2586         return ret;
2587 }
2588
2589 const struct inode_operations ocfs2_file_iops = {
2590         .setattr        = ocfs2_setattr,
2591         .getattr        = ocfs2_getattr,
2592         .permission     = ocfs2_permission,
2593         .setxattr       = generic_setxattr,
2594         .getxattr       = generic_getxattr,
2595         .listxattr      = ocfs2_listxattr,
2596         .removexattr    = generic_removexattr,
2597         .fallocate      = ocfs2_fallocate,
2598         .fiemap         = ocfs2_fiemap,
2599 };
2600
2601 const struct inode_operations ocfs2_special_file_iops = {
2602         .setattr        = ocfs2_setattr,
2603         .getattr        = ocfs2_getattr,
2604         .permission     = ocfs2_permission,
2605 };
2606
2607 /*
2608  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2609  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2610  */
2611 const struct file_operations ocfs2_fops = {
2612         .llseek         = generic_file_llseek,
2613         .read           = do_sync_read,
2614         .write          = do_sync_write,
2615         .mmap           = ocfs2_mmap,
2616         .fsync          = ocfs2_sync_file,
2617         .release        = ocfs2_file_release,
2618         .open           = ocfs2_file_open,
2619         .aio_read       = ocfs2_file_aio_read,
2620         .aio_write      = ocfs2_file_aio_write,
2621         .unlocked_ioctl = ocfs2_ioctl,
2622 #ifdef CONFIG_COMPAT
2623         .compat_ioctl   = ocfs2_compat_ioctl,
2624 #endif
2625         .lock           = ocfs2_lock,
2626         .flock          = ocfs2_flock,
2627         .splice_read    = ocfs2_file_splice_read,
2628         .splice_write   = ocfs2_file_splice_write,
2629 };
2630
2631 const struct file_operations ocfs2_dops = {
2632         .llseek         = generic_file_llseek,
2633         .read           = generic_read_dir,
2634         .readdir        = ocfs2_readdir,
2635         .fsync          = ocfs2_sync_file,
2636         .release        = ocfs2_dir_release,
2637         .open           = ocfs2_dir_open,
2638         .unlocked_ioctl = ocfs2_ioctl,
2639 #ifdef CONFIG_COMPAT
2640         .compat_ioctl   = ocfs2_compat_ioctl,
2641 #endif
2642         .lock           = ocfs2_lock,
2643         .flock          = ocfs2_flock,
2644 };
2645
2646 /*
2647  * POSIX-lockless variants of our file_operations.
2648  *
2649  * These will be used if the underlying cluster stack does not support
2650  * posix file locking, if the user passes the "localflocks" mount
2651  * option, or if we have a local-only fs.
2652  *
2653  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2654  * so we still want it in the case of no stack support for
2655  * plocks. Internally, it will do the right thing when asked to ignore
2656  * the cluster.
2657  */
2658 const struct file_operations ocfs2_fops_no_plocks = {
2659         .llseek         = generic_file_llseek,
2660         .read           = do_sync_read,
2661         .write          = do_sync_write,
2662         .mmap           = ocfs2_mmap,
2663         .fsync          = ocfs2_sync_file,
2664         .release        = ocfs2_file_release,
2665         .open           = ocfs2_file_open,
2666         .aio_read       = ocfs2_file_aio_read,
2667         .aio_write      = ocfs2_file_aio_write,
2668         .unlocked_ioctl = ocfs2_ioctl,
2669 #ifdef CONFIG_COMPAT
2670         .compat_ioctl   = ocfs2_compat_ioctl,
2671 #endif
2672         .flock          = ocfs2_flock,
2673         .splice_read    = ocfs2_file_splice_read,
2674         .splice_write   = ocfs2_file_splice_write,
2675 };
2676
2677 const struct file_operations ocfs2_dops_no_plocks = {
2678         .llseek         = generic_file_llseek,
2679         .read           = generic_read_dir,
2680         .readdir        = ocfs2_readdir,
2681         .fsync          = ocfs2_sync_file,
2682         .release        = ocfs2_dir_release,
2683         .open           = ocfs2_dir_open,
2684         .unlocked_ioctl = ocfs2_ioctl,
2685 #ifdef CONFIG_COMPAT
2686         .compat_ioctl   = ocfs2_compat_ioctl,
2687 #endif
2688         .flock          = ocfs2_flock,
2689 };