22d60460195754e8fba94e55e05e97670fb7d0fd
[linux-flexiantxendom0-3.2.10.git] / fs / ocfs2 / file.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/capability.h>
27 #include <linux/fs.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
38 #include <linux/quotaops.h>
39 #include <linux/blkdev.h>
40
41 #include <cluster/masklog.h>
42
43 #include "ocfs2.h"
44
45 #include "alloc.h"
46 #include "aops.h"
47 #include "dir.h"
48 #include "dlmglue.h"
49 #include "extent_map.h"
50 #include "file.h"
51 #include "sysfile.h"
52 #include "inode.h"
53 #include "ioctl.h"
54 #include "journal.h"
55 #include "locks.h"
56 #include "mmap.h"
57 #include "suballoc.h"
58 #include "super.h"
59 #include "xattr.h"
60 #include "acl.h"
61 #include "quota.h"
62 #include "refcounttree.h"
63 #include "ocfs2_trace.h"
64
65 #include "buffer_head_io.h"
66
67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
68 {
69         struct ocfs2_file_private *fp;
70
71         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
72         if (!fp)
73                 return -ENOMEM;
74
75         fp->fp_file = file;
76         mutex_init(&fp->fp_mutex);
77         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78         file->private_data = fp;
79
80         return 0;
81 }
82
83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
84 {
85         struct ocfs2_file_private *fp = file->private_data;
86         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
87
88         if (fp) {
89                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90                 ocfs2_lock_res_free(&fp->fp_flock);
91                 kfree(fp);
92                 file->private_data = NULL;
93         }
94 }
95
96 static int ocfs2_file_open(struct inode *inode, struct file *file)
97 {
98         int status;
99         int mode = file->f_flags;
100         struct ocfs2_inode_info *oi = OCFS2_I(inode);
101
102         trace_ocfs2_file_open(inode, file, file->f_path.dentry,
103                               (unsigned long long)OCFS2_I(inode)->ip_blkno,
104                               file->f_path.dentry->d_name.len,
105                               file->f_path.dentry->d_name.name, mode);
106
107         if (file->f_mode & FMODE_WRITE)
108                 dquot_initialize(inode);
109
110         spin_lock(&oi->ip_lock);
111
112         /* Check that the inode hasn't been wiped from disk by another
113          * node. If it hasn't then we're safe as long as we hold the
114          * spin lock until our increment of open count. */
115         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
116                 spin_unlock(&oi->ip_lock);
117
118                 status = -ENOENT;
119                 goto leave;
120         }
121
122         if (mode & O_DIRECT)
123                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
124
125         oi->ip_open_count++;
126         spin_unlock(&oi->ip_lock);
127
128         status = ocfs2_init_file_private(inode, file);
129         if (status) {
130                 /*
131                  * We want to set open count back if we're failing the
132                  * open.
133                  */
134                 spin_lock(&oi->ip_lock);
135                 oi->ip_open_count--;
136                 spin_unlock(&oi->ip_lock);
137         }
138
139 leave:
140         return status;
141 }
142
143 static int ocfs2_file_release(struct inode *inode, struct file *file)
144 {
145         struct ocfs2_inode_info *oi = OCFS2_I(inode);
146
147         spin_lock(&oi->ip_lock);
148         if (!--oi->ip_open_count)
149                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
150
151         trace_ocfs2_file_release(inode, file, file->f_path.dentry,
152                                  oi->ip_blkno,
153                                  file->f_path.dentry->d_name.len,
154                                  file->f_path.dentry->d_name.name,
155                                  oi->ip_open_count);
156         spin_unlock(&oi->ip_lock);
157
158         ocfs2_free_file_private(inode, file);
159
160         return 0;
161 }
162
163 static int ocfs2_dir_open(struct inode *inode, struct file *file)
164 {
165         return ocfs2_init_file_private(inode, file);
166 }
167
168 static int ocfs2_dir_release(struct inode *inode, struct file *file)
169 {
170         ocfs2_free_file_private(inode, file);
171         return 0;
172 }
173
174 static int ocfs2_sync_file(struct file *file, int datasync)
175 {
176         int err = 0;
177         journal_t *journal;
178         struct inode *inode = file->f_mapping->host;
179         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
180
181         trace_ocfs2_sync_file(inode, file, file->f_path.dentry,
182                               OCFS2_I(inode)->ip_blkno,
183                               file->f_path.dentry->d_name.len,
184                               file->f_path.dentry->d_name.name,
185                               (unsigned long long)datasync);
186
187         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
188                 /*
189                  * We still have to flush drive's caches to get data to the
190                  * platter
191                  */
192                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
193                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
194                 goto bail;
195         }
196
197         journal = osb->journal->j_journal;
198         err = jbd2_journal_force_commit(journal);
199
200 bail:
201         if (err)
202                 mlog_errno(err);
203
204         return (err < 0) ? -EIO : 0;
205 }
206
207 int ocfs2_should_update_atime(struct inode *inode,
208                               struct vfsmount *vfsmnt)
209 {
210         struct timespec now;
211         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
212
213         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
214                 return 0;
215
216         if ((inode->i_flags & S_NOATIME) ||
217             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
218                 return 0;
219
220         /*
221          * We can be called with no vfsmnt structure - NFSD will
222          * sometimes do this.
223          *
224          * Note that our action here is different than touch_atime() -
225          * if we can't tell whether this is a noatime mount, then we
226          * don't know whether to trust the value of s_atime_quantum.
227          */
228         if (vfsmnt == NULL)
229                 return 0;
230
231         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
232             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
233                 return 0;
234
235         if (vfsmnt->mnt_flags & MNT_RELATIME) {
236                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
237                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
238                         return 1;
239
240                 return 0;
241         }
242
243         now = CURRENT_TIME;
244         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
245                 return 0;
246         else
247                 return 1;
248 }
249
250 int ocfs2_update_inode_atime(struct inode *inode,
251                              struct buffer_head *bh)
252 {
253         int ret;
254         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
255         handle_t *handle;
256         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
257
258         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
259         if (IS_ERR(handle)) {
260                 ret = PTR_ERR(handle);
261                 mlog_errno(ret);
262                 goto out;
263         }
264
265         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
266                                       OCFS2_JOURNAL_ACCESS_WRITE);
267         if (ret) {
268                 mlog_errno(ret);
269                 goto out_commit;
270         }
271
272         /*
273          * Don't use ocfs2_mark_inode_dirty() here as we don't always
274          * have i_mutex to guard against concurrent changes to other
275          * inode fields.
276          */
277         inode->i_atime = CURRENT_TIME;
278         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
279         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
280         ocfs2_journal_dirty(handle, bh);
281
282 out_commit:
283         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
284 out:
285         return ret;
286 }
287
288 static int ocfs2_set_inode_size(handle_t *handle,
289                                 struct inode *inode,
290                                 struct buffer_head *fe_bh,
291                                 u64 new_i_size)
292 {
293         int status;
294
295         i_size_write(inode, new_i_size);
296         inode->i_blocks = ocfs2_inode_sector_count(inode);
297         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
298
299         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
300         if (status < 0) {
301                 mlog_errno(status);
302                 goto bail;
303         }
304
305 bail:
306         return status;
307 }
308
309 int ocfs2_simple_size_update(struct inode *inode,
310                              struct buffer_head *di_bh,
311                              u64 new_i_size)
312 {
313         int ret;
314         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
315         handle_t *handle = NULL;
316
317         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
318         if (IS_ERR(handle)) {
319                 ret = PTR_ERR(handle);
320                 mlog_errno(ret);
321                 goto out;
322         }
323
324         ret = ocfs2_set_inode_size(handle, inode, di_bh,
325                                    new_i_size);
326         if (ret < 0)
327                 mlog_errno(ret);
328
329         ocfs2_commit_trans(osb, handle);
330 out:
331         return ret;
332 }
333
334 static int ocfs2_cow_file_pos(struct inode *inode,
335                               struct buffer_head *fe_bh,
336                               u64 offset)
337 {
338         int status;
339         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
340         unsigned int num_clusters = 0;
341         unsigned int ext_flags = 0;
342
343         /*
344          * If the new offset is aligned to the range of the cluster, there is
345          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
346          * CoW either.
347          */
348         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
349                 return 0;
350
351         status = ocfs2_get_clusters(inode, cpos, &phys,
352                                     &num_clusters, &ext_flags);
353         if (status) {
354                 mlog_errno(status);
355                 goto out;
356         }
357
358         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
359                 goto out;
360
361         return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
362
363 out:
364         return status;
365 }
366
367 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
368                                      struct inode *inode,
369                                      struct buffer_head *fe_bh,
370                                      u64 new_i_size)
371 {
372         int status;
373         handle_t *handle;
374         struct ocfs2_dinode *di;
375         u64 cluster_bytes;
376
377         /*
378          * We need to CoW the cluster contains the offset if it is reflinked
379          * since we will call ocfs2_zero_range_for_truncate later which will
380          * write "0" from offset to the end of the cluster.
381          */
382         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
383         if (status) {
384                 mlog_errno(status);
385                 return status;
386         }
387
388         /* TODO: This needs to actually orphan the inode in this
389          * transaction. */
390
391         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
392         if (IS_ERR(handle)) {
393                 status = PTR_ERR(handle);
394                 mlog_errno(status);
395                 goto out;
396         }
397
398         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
399                                          OCFS2_JOURNAL_ACCESS_WRITE);
400         if (status < 0) {
401                 mlog_errno(status);
402                 goto out_commit;
403         }
404
405         /*
406          * Do this before setting i_size.
407          */
408         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
409         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
410                                                cluster_bytes);
411         if (status) {
412                 mlog_errno(status);
413                 goto out_commit;
414         }
415
416         i_size_write(inode, new_i_size);
417         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
418
419         di = (struct ocfs2_dinode *) fe_bh->b_data;
420         di->i_size = cpu_to_le64(new_i_size);
421         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
422         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
423
424         ocfs2_journal_dirty(handle, fe_bh);
425
426 out_commit:
427         ocfs2_commit_trans(osb, handle);
428 out:
429         return status;
430 }
431
432 static int ocfs2_truncate_file(struct inode *inode,
433                                struct buffer_head *di_bh,
434                                u64 new_i_size)
435 {
436         int status = 0;
437         struct ocfs2_dinode *fe = NULL;
438         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
439
440         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
441          * already validated it */
442         fe = (struct ocfs2_dinode *) di_bh->b_data;
443
444         trace_ocfs2_truncate_file((unsigned long long)OCFS2_I(inode)->ip_blkno,
445                                   (unsigned long long)le64_to_cpu(fe->i_size),
446                                   (unsigned long long)new_i_size);
447
448         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
449                         "Inode %llu, inode i_size = %lld != di "
450                         "i_size = %llu, i_flags = 0x%x\n",
451                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
452                         i_size_read(inode),
453                         (unsigned long long)le64_to_cpu(fe->i_size),
454                         le32_to_cpu(fe->i_flags));
455
456         if (new_i_size > le64_to_cpu(fe->i_size)) {
457                 trace_ocfs2_truncate_file_error(
458                         (unsigned long long)le64_to_cpu(fe->i_size),
459                         (unsigned long long)new_i_size);
460                 status = -EINVAL;
461                 mlog_errno(status);
462                 goto bail;
463         }
464
465         /* lets handle the simple truncate cases before doing any more
466          * cluster locking. */
467         if (new_i_size == le64_to_cpu(fe->i_size))
468                 goto bail;
469
470         down_write(&OCFS2_I(inode)->ip_alloc_sem);
471
472         ocfs2_resv_discard(&osb->osb_la_resmap,
473                            &OCFS2_I(inode)->ip_la_data_resv);
474
475         /*
476          * The inode lock forced other nodes to sync and drop their
477          * pages, which (correctly) happens even if we have a truncate
478          * without allocation change - ocfs2 cluster sizes can be much
479          * greater than page size, so we have to truncate them
480          * anyway.
481          */
482         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
483         truncate_inode_pages(inode->i_mapping, new_i_size);
484
485         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
486                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
487                                                i_size_read(inode), 1);
488                 if (status)
489                         mlog_errno(status);
490
491                 goto bail_unlock_sem;
492         }
493
494         /* alright, we're going to need to do a full blown alloc size
495          * change. Orphan the inode so that recovery can complete the
496          * truncate if necessary. This does the task of marking
497          * i_size. */
498         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
499         if (status < 0) {
500                 mlog_errno(status);
501                 goto bail_unlock_sem;
502         }
503
504         status = ocfs2_commit_truncate(osb, inode, di_bh);
505         if (status < 0) {
506                 mlog_errno(status);
507                 goto bail_unlock_sem;
508         }
509
510         /* TODO: orphan dir cleanup here. */
511 bail_unlock_sem:
512         up_write(&OCFS2_I(inode)->ip_alloc_sem);
513
514 bail:
515         if (!status && OCFS2_I(inode)->ip_clusters == 0)
516                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
517
518         return status;
519 }
520
521 /*
522  * extend file allocation only here.
523  * we'll update all the disk stuff, and oip->alloc_size
524  *
525  * expect stuff to be locked, a transaction started and enough data /
526  * metadata reservations in the contexts.
527  *
528  * Will return -EAGAIN, and a reason if a restart is needed.
529  * If passed in, *reason will always be set, even in error.
530  */
531 int ocfs2_add_inode_data(struct ocfs2_super *osb,
532                          struct inode *inode,
533                          u32 *logical_offset,
534                          u32 clusters_to_add,
535                          int mark_unwritten,
536                          struct buffer_head *fe_bh,
537                          handle_t *handle,
538                          struct ocfs2_alloc_context *data_ac,
539                          struct ocfs2_alloc_context *meta_ac,
540                          enum ocfs2_alloc_restarted *reason_ret)
541 {
542         int ret;
543         struct ocfs2_extent_tree et;
544
545         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
546         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
547                                           clusters_to_add, mark_unwritten,
548                                           data_ac, meta_ac, reason_ret);
549
550         return ret;
551 }
552
553 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
554                                      u32 clusters_to_add, int mark_unwritten)
555 {
556         int status = 0;
557         int restart_func = 0;
558         int credits;
559         u32 prev_clusters;
560         struct buffer_head *bh = NULL;
561         struct ocfs2_dinode *fe = NULL;
562         handle_t *handle = NULL;
563         struct ocfs2_alloc_context *data_ac = NULL;
564         struct ocfs2_alloc_context *meta_ac = NULL;
565         enum ocfs2_alloc_restarted why;
566         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
567         struct ocfs2_extent_tree et;
568         int did_quota = 0;
569
570         /*
571          * This function only exists for file systems which don't
572          * support holes.
573          */
574         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
575
576         status = ocfs2_read_inode_block(inode, &bh);
577         if (status < 0) {
578                 mlog_errno(status);
579                 goto leave;
580         }
581         fe = (struct ocfs2_dinode *) bh->b_data;
582
583 restart_all:
584         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
585
586         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
587         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
588                                        &data_ac, &meta_ac);
589         if (status) {
590                 mlog_errno(status);
591                 goto leave;
592         }
593
594         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
595                                             clusters_to_add);
596         handle = ocfs2_start_trans(osb, credits);
597         if (IS_ERR(handle)) {
598                 status = PTR_ERR(handle);
599                 handle = NULL;
600                 mlog_errno(status);
601                 goto leave;
602         }
603
604 restarted_transaction:
605         trace_ocfs2_extend_allocation(
606                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
607                 (unsigned long long)i_size_read(inode),
608                 le32_to_cpu(fe->i_clusters), clusters_to_add,
609                 why, restart_func);
610
611         status = dquot_alloc_space_nodirty(inode,
612                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
613         if (status)
614                 goto leave;
615         did_quota = 1;
616
617         /* reserve a write to the file entry early on - that we if we
618          * run out of credits in the allocation path, we can still
619          * update i_size. */
620         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
621                                          OCFS2_JOURNAL_ACCESS_WRITE);
622         if (status < 0) {
623                 mlog_errno(status);
624                 goto leave;
625         }
626
627         prev_clusters = OCFS2_I(inode)->ip_clusters;
628
629         status = ocfs2_add_inode_data(osb,
630                                       inode,
631                                       &logical_start,
632                                       clusters_to_add,
633                                       mark_unwritten,
634                                       bh,
635                                       handle,
636                                       data_ac,
637                                       meta_ac,
638                                       &why);
639         if ((status < 0) && (status != -EAGAIN)) {
640                 if (status != -ENOSPC)
641                         mlog_errno(status);
642                 goto leave;
643         }
644
645         ocfs2_journal_dirty(handle, bh);
646
647         spin_lock(&OCFS2_I(inode)->ip_lock);
648         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
649         spin_unlock(&OCFS2_I(inode)->ip_lock);
650         /* Release unused quota reservation */
651         dquot_free_space(inode,
652                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
653         did_quota = 0;
654
655         if (why != RESTART_NONE && clusters_to_add) {
656                 if (why == RESTART_META) {
657                         restart_func = 1;
658                         status = 0;
659                 } else {
660                         BUG_ON(why != RESTART_TRANS);
661
662                         /* TODO: This can be more intelligent. */
663                         credits = ocfs2_calc_extend_credits(osb->sb,
664                                                             &fe->id2.i_list,
665                                                             clusters_to_add);
666                         status = ocfs2_extend_trans(handle, credits);
667                         if (status < 0) {
668                                 /* handle still has to be committed at
669                                  * this point. */
670                                 status = -ENOMEM;
671                                 mlog_errno(status);
672                                 goto leave;
673                         }
674                         goto restarted_transaction;
675                 }
676         }
677
678         trace_ocfs2_extend_allocation_end(OCFS2_I(inode)->ip_blkno,
679              le32_to_cpu(fe->i_clusters),
680              (unsigned long long)le64_to_cpu(fe->i_size),
681              OCFS2_I(inode)->ip_clusters,
682              (unsigned long long)i_size_read(inode));
683
684 leave:
685         if (status < 0 && did_quota)
686                 dquot_free_space(inode,
687                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
688         if (handle) {
689                 ocfs2_commit_trans(osb, handle);
690                 handle = NULL;
691         }
692         if (data_ac) {
693                 ocfs2_free_alloc_context(data_ac);
694                 data_ac = NULL;
695         }
696         if (meta_ac) {
697                 ocfs2_free_alloc_context(meta_ac);
698                 meta_ac = NULL;
699         }
700         if ((!status) && restart_func) {
701                 restart_func = 0;
702                 goto restart_all;
703         }
704         brelse(bh);
705         bh = NULL;
706
707         return status;
708 }
709
710 /*
711  * While a write will already be ordering the data, a truncate will not.
712  * Thus, we need to explicitly order the zeroed pages.
713  */
714 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
715 {
716         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
717         handle_t *handle = NULL;
718         int ret = 0;
719
720         if (!ocfs2_should_order_data(inode))
721                 goto out;
722
723         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
724         if (IS_ERR(handle)) {
725                 ret = -ENOMEM;
726                 mlog_errno(ret);
727                 goto out;
728         }
729
730         ret = ocfs2_jbd2_file_inode(handle, inode);
731         if (ret < 0)
732                 mlog_errno(ret);
733
734 out:
735         if (ret) {
736                 if (!IS_ERR(handle))
737                         ocfs2_commit_trans(osb, handle);
738                 handle = ERR_PTR(ret);
739         }
740         return handle;
741 }
742
743 /* Some parts of this taken from generic_cont_expand, which turned out
744  * to be too fragile to do exactly what we need without us having to
745  * worry about recursive locking in ->write_begin() and ->write_end(). */
746 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
747                                  u64 abs_to)
748 {
749         struct address_space *mapping = inode->i_mapping;
750         struct page *page;
751         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
752         handle_t *handle = NULL;
753         int ret = 0;
754         unsigned zero_from, zero_to, block_start, block_end;
755
756         BUG_ON(abs_from >= abs_to);
757         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
758         BUG_ON(abs_from & (inode->i_blkbits - 1));
759
760         page = find_or_create_page(mapping, index, GFP_NOFS);
761         if (!page) {
762                 ret = -ENOMEM;
763                 mlog_errno(ret);
764                 goto out;
765         }
766
767         /* Get the offsets within the page that we want to zero */
768         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
769         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
770         if (!zero_to)
771                 zero_to = PAGE_CACHE_SIZE;
772
773         trace_ocfs2_write_zero_page(
774                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
775                         (unsigned long long)abs_from,
776                         (unsigned long long)abs_to,
777                         index, zero_from, zero_to);
778
779         /* We know that zero_from is block aligned */
780         for (block_start = zero_from; block_start < zero_to;
781              block_start = block_end) {
782                 block_end = block_start + (1 << inode->i_blkbits);
783
784                 /*
785                  * block_start is block-aligned.  Bump it by one to force
786                  * __block_write_begin and block_commit_write to zero the
787                  * whole block.
788                  */
789                 ret = __block_write_begin(page, block_start + 1, 0,
790                                           ocfs2_get_block);
791                 if (ret < 0) {
792                         mlog_errno(ret);
793                         goto out_unlock;
794                 }
795
796                 if (!handle) {
797                         handle = ocfs2_zero_start_ordered_transaction(inode);
798                         if (IS_ERR(handle)) {
799                                 ret = PTR_ERR(handle);
800                                 handle = NULL;
801                                 break;
802                         }
803                 }
804
805                 /* must not update i_size! */
806                 ret = block_commit_write(page, block_start + 1,
807                                          block_start + 1);
808                 if (ret < 0)
809                         mlog_errno(ret);
810                 else
811                         ret = 0;
812         }
813
814         if (handle)
815                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
816
817 out_unlock:
818         unlock_page(page);
819         page_cache_release(page);
820 out:
821         return ret;
822 }
823
824 /*
825  * Find the next range to zero.  We do this in terms of bytes because
826  * that's what ocfs2_zero_extend() wants, and it is dealing with the
827  * pagecache.  We may return multiple extents.
828  *
829  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
830  * needs to be zeroed.  range_start and range_end return the next zeroing
831  * range.  A subsequent call should pass the previous range_end as its
832  * zero_start.  If range_end is 0, there's nothing to do.
833  *
834  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
835  */
836 static int ocfs2_zero_extend_get_range(struct inode *inode,
837                                        struct buffer_head *di_bh,
838                                        u64 zero_start, u64 zero_end,
839                                        u64 *range_start, u64 *range_end)
840 {
841         int rc = 0, needs_cow = 0;
842         u32 p_cpos, zero_clusters = 0;
843         u32 zero_cpos =
844                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
845         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
846         unsigned int num_clusters = 0;
847         unsigned int ext_flags = 0;
848
849         while (zero_cpos < last_cpos) {
850                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
851                                         &num_clusters, &ext_flags);
852                 if (rc) {
853                         mlog_errno(rc);
854                         goto out;
855                 }
856
857                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
858                         zero_clusters = num_clusters;
859                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
860                                 needs_cow = 1;
861                         break;
862                 }
863
864                 zero_cpos += num_clusters;
865         }
866         if (!zero_clusters) {
867                 *range_end = 0;
868                 goto out;
869         }
870
871         while ((zero_cpos + zero_clusters) < last_cpos) {
872                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
873                                         &p_cpos, &num_clusters,
874                                         &ext_flags);
875                 if (rc) {
876                         mlog_errno(rc);
877                         goto out;
878                 }
879
880                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
881                         break;
882                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
883                         needs_cow = 1;
884                 zero_clusters += num_clusters;
885         }
886         if ((zero_cpos + zero_clusters) > last_cpos)
887                 zero_clusters = last_cpos - zero_cpos;
888
889         if (needs_cow) {
890                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
891                                         zero_clusters, UINT_MAX);
892                 if (rc) {
893                         mlog_errno(rc);
894                         goto out;
895                 }
896         }
897
898         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
899         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
900                                              zero_cpos + zero_clusters);
901
902 out:
903         return rc;
904 }
905
906 /*
907  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
908  * has made sure that the entire range needs zeroing.
909  */
910 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
911                                    u64 range_end)
912 {
913         int rc = 0;
914         u64 next_pos;
915         u64 zero_pos = range_start;
916
917         trace_ocfs2_zero_extend_range(
918                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
919                         (unsigned long long)range_start,
920                         (unsigned long long)range_end);
921         BUG_ON(range_start >= range_end);
922
923         while (zero_pos < range_end) {
924                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
925                 if (next_pos > range_end)
926                         next_pos = range_end;
927                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
928                 if (rc < 0) {
929                         mlog_errno(rc);
930                         break;
931                 }
932                 zero_pos = next_pos;
933
934                 /*
935                  * Very large extends have the potential to lock up
936                  * the cpu for extended periods of time.
937                  */
938                 cond_resched();
939         }
940
941         return rc;
942 }
943
944 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
945                       loff_t zero_to_size)
946 {
947         int ret = 0;
948         u64 zero_start, range_start = 0, range_end = 0;
949         struct super_block *sb = inode->i_sb;
950
951         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
952         trace_ocfs2_zero_extend((unsigned long long)OCFS2_I(inode)->ip_blkno,
953                                 (unsigned long long)zero_start,
954                                 (unsigned long long)i_size_read(inode));
955         while (zero_start < zero_to_size) {
956                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
957                                                   zero_to_size,
958                                                   &range_start,
959                                                   &range_end);
960                 if (ret) {
961                         mlog_errno(ret);
962                         break;
963                 }
964                 if (!range_end)
965                         break;
966                 /* Trim the ends */
967                 if (range_start < zero_start)
968                         range_start = zero_start;
969                 if (range_end > zero_to_size)
970                         range_end = zero_to_size;
971
972                 ret = ocfs2_zero_extend_range(inode, range_start,
973                                               range_end);
974                 if (ret) {
975                         mlog_errno(ret);
976                         break;
977                 }
978                 zero_start = range_end;
979         }
980
981         return ret;
982 }
983
984 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
985                           u64 new_i_size, u64 zero_to)
986 {
987         int ret;
988         u32 clusters_to_add;
989         struct ocfs2_inode_info *oi = OCFS2_I(inode);
990
991         /*
992          * Only quota files call this without a bh, and they can't be
993          * refcounted.
994          */
995         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
996         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
997
998         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
999         if (clusters_to_add < oi->ip_clusters)
1000                 clusters_to_add = 0;
1001         else
1002                 clusters_to_add -= oi->ip_clusters;
1003
1004         if (clusters_to_add) {
1005                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1006                                                 clusters_to_add, 0);
1007                 if (ret) {
1008                         mlog_errno(ret);
1009                         goto out;
1010                 }
1011         }
1012
1013         /*
1014          * Call this even if we don't add any clusters to the tree. We
1015          * still need to zero the area between the old i_size and the
1016          * new i_size.
1017          */
1018         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1019         if (ret < 0)
1020                 mlog_errno(ret);
1021
1022 out:
1023         return ret;
1024 }
1025
1026 static int ocfs2_extend_file(struct inode *inode,
1027                              struct buffer_head *di_bh,
1028                              u64 new_i_size)
1029 {
1030         int ret = 0;
1031         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1032
1033         BUG_ON(!di_bh);
1034
1035         /* setattr sometimes calls us like this. */
1036         if (new_i_size == 0)
1037                 goto out;
1038
1039         if (i_size_read(inode) == new_i_size)
1040                 goto out;
1041         BUG_ON(new_i_size < i_size_read(inode));
1042
1043         /*
1044          * The alloc sem blocks people in read/write from reading our
1045          * allocation until we're done changing it. We depend on
1046          * i_mutex to block other extend/truncate calls while we're
1047          * here.  We even have to hold it for sparse files because there
1048          * might be some tail zeroing.
1049          */
1050         down_write(&oi->ip_alloc_sem);
1051
1052         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1053                 /*
1054                  * We can optimize small extends by keeping the inodes
1055                  * inline data.
1056                  */
1057                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1058                         up_write(&oi->ip_alloc_sem);
1059                         goto out_update_size;
1060                 }
1061
1062                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1063                 if (ret) {
1064                         up_write(&oi->ip_alloc_sem);
1065                         mlog_errno(ret);
1066                         goto out;
1067                 }
1068         }
1069
1070         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1071                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1072         else
1073                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1074                                             new_i_size);
1075
1076         up_write(&oi->ip_alloc_sem);
1077
1078         if (ret < 0) {
1079                 mlog_errno(ret);
1080                 goto out;
1081         }
1082
1083 out_update_size:
1084         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1085         if (ret < 0)
1086                 mlog_errno(ret);
1087
1088 out:
1089         return ret;
1090 }
1091
1092 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1093 {
1094         int status = 0, size_change;
1095         struct inode *inode = dentry->d_inode;
1096         struct super_block *sb = inode->i_sb;
1097         struct ocfs2_super *osb = OCFS2_SB(sb);
1098         struct buffer_head *bh = NULL;
1099         handle_t *handle = NULL;
1100         struct dquot *transfer_to[MAXQUOTAS] = { };
1101         int qtype;
1102
1103         trace_ocfs2_setattr(inode, dentry,
1104                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1105                             dentry->d_name.len, dentry->d_name.name,
1106                             attr->ia_valid, attr->ia_mode,
1107                             attr->ia_uid, attr->ia_gid);
1108
1109         /* ensuring we don't even attempt to truncate a symlink */
1110         if (S_ISLNK(inode->i_mode))
1111                 attr->ia_valid &= ~ATTR_SIZE;
1112
1113 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1114                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1115         if (!(attr->ia_valid & OCFS2_VALID_ATTRS))
1116                 return 0;
1117
1118         status = inode_change_ok(inode, attr);
1119         if (status)
1120                 return status;
1121
1122         if (is_quota_modification(inode, attr))
1123                 dquot_initialize(inode);
1124         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1125         if (size_change) {
1126                 status = ocfs2_rw_lock(inode, 1);
1127                 if (status < 0) {
1128                         mlog_errno(status);
1129                         goto bail;
1130                 }
1131         }
1132
1133         status = ocfs2_inode_lock(inode, &bh, 1);
1134         if (status < 0) {
1135                 if (status != -ENOENT)
1136                         mlog_errno(status);
1137                 goto bail_unlock_rw;
1138         }
1139
1140         if (size_change && attr->ia_size != i_size_read(inode)) {
1141                 status = inode_newsize_ok(inode, attr->ia_size);
1142                 if (status)
1143                         goto bail_unlock;
1144
1145                 inode_dio_wait(inode);
1146
1147                 if (i_size_read(inode) > attr->ia_size) {
1148                         if (ocfs2_should_order_data(inode)) {
1149                                 status = ocfs2_begin_ordered_truncate(inode,
1150                                                                       attr->ia_size);
1151                                 if (status)
1152                                         goto bail_unlock;
1153                         }
1154                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1155                 } else
1156                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1157                 if (status < 0) {
1158                         if (status != -ENOSPC)
1159                                 mlog_errno(status);
1160                         status = -ENOSPC;
1161                         goto bail_unlock;
1162                 }
1163         }
1164
1165         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1166             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1167                 /*
1168                  * Gather pointers to quota structures so that allocation /
1169                  * freeing of quota structures happens here and not inside
1170                  * dquot_transfer() where we have problems with lock ordering
1171                  */
1172                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1173                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1174                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1175                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1176                                                       USRQUOTA);
1177                         if (!transfer_to[USRQUOTA]) {
1178                                 status = -ESRCH;
1179                                 goto bail_unlock;
1180                         }
1181                 }
1182                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1183                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1184                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1185                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1186                                                       GRPQUOTA);
1187                         if (!transfer_to[GRPQUOTA]) {
1188                                 status = -ESRCH;
1189                                 goto bail_unlock;
1190                         }
1191                 }
1192                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1193                                            2 * ocfs2_quota_trans_credits(sb));
1194                 if (IS_ERR(handle)) {
1195                         status = PTR_ERR(handle);
1196                         mlog_errno(status);
1197                         goto bail_unlock;
1198                 }
1199                 status = __dquot_transfer(inode, transfer_to);
1200                 if (status < 0)
1201                         goto bail_commit;
1202         } else {
1203                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1204                 if (IS_ERR(handle)) {
1205                         status = PTR_ERR(handle);
1206                         mlog_errno(status);
1207                         goto bail_unlock;
1208                 }
1209         }
1210
1211         /*
1212          * This will intentionally not wind up calling truncate_setsize(),
1213          * since all the work for a size change has been done above.
1214          * Otherwise, we could get into problems with truncate as
1215          * ip_alloc_sem is used there to protect against i_size
1216          * changes.
1217          *
1218          * XXX: this means the conditional below can probably be removed.
1219          */
1220         if ((attr->ia_valid & ATTR_SIZE) &&
1221             attr->ia_size != i_size_read(inode)) {
1222                 status = vmtruncate(inode, attr->ia_size);
1223                 if (status) {
1224                         mlog_errno(status);
1225                         goto bail_commit;
1226                 }
1227         }
1228
1229         setattr_copy(inode, attr);
1230         mark_inode_dirty(inode);
1231
1232         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1233         if (status < 0)
1234                 mlog_errno(status);
1235
1236 bail_commit:
1237         ocfs2_commit_trans(osb, handle);
1238 bail_unlock:
1239         ocfs2_inode_unlock(inode, 1);
1240 bail_unlock_rw:
1241         if (size_change)
1242                 ocfs2_rw_unlock(inode, 1);
1243 bail:
1244         brelse(bh);
1245
1246         /* Release quota pointers in case we acquired them */
1247         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1248                 dqput(transfer_to[qtype]);
1249
1250         if (!status && attr->ia_valid & ATTR_MODE) {
1251                 status = ocfs2_acl_chmod(inode);
1252                 if (status < 0)
1253                         mlog_errno(status);
1254         }
1255
1256         return status;
1257 }
1258
1259 int ocfs2_getattr(struct vfsmount *mnt,
1260                   struct dentry *dentry,
1261                   struct kstat *stat)
1262 {
1263         struct inode *inode = dentry->d_inode;
1264         struct super_block *sb = dentry->d_inode->i_sb;
1265         struct ocfs2_super *osb = sb->s_fs_info;
1266         int err;
1267
1268         err = ocfs2_inode_revalidate(dentry);
1269         if (err) {
1270                 if (err != -ENOENT)
1271                         mlog_errno(err);
1272                 goto bail;
1273         }
1274
1275         generic_fillattr(inode, stat);
1276
1277         /* We set the blksize from the cluster size for performance */
1278         stat->blksize = osb->s_clustersize;
1279
1280 bail:
1281         return err;
1282 }
1283
1284 int ocfs2_permission(struct inode *inode, int mask)
1285 {
1286         int ret;
1287
1288         if (mask & MAY_NOT_BLOCK)
1289                 return -ECHILD;
1290
1291         ret = ocfs2_inode_lock(inode, NULL, 0);
1292         if (ret) {
1293                 if (ret != -ENOENT)
1294                         mlog_errno(ret);
1295                 goto out;
1296         }
1297
1298         ret = generic_permission(inode, mask);
1299
1300         ocfs2_inode_unlock(inode, 0);
1301 out:
1302         return ret;
1303 }
1304
1305 static int __ocfs2_write_remove_suid(struct inode *inode,
1306                                      struct buffer_head *bh)
1307 {
1308         int ret;
1309         handle_t *handle;
1310         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1311         struct ocfs2_dinode *di;
1312
1313         trace_ocfs2_write_remove_suid(
1314                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1315                         inode->i_mode);
1316
1317         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1318         if (IS_ERR(handle)) {
1319                 ret = PTR_ERR(handle);
1320                 mlog_errno(ret);
1321                 goto out;
1322         }
1323
1324         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1325                                       OCFS2_JOURNAL_ACCESS_WRITE);
1326         if (ret < 0) {
1327                 mlog_errno(ret);
1328                 goto out_trans;
1329         }
1330
1331         inode->i_mode &= ~S_ISUID;
1332         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1333                 inode->i_mode &= ~S_ISGID;
1334
1335         di = (struct ocfs2_dinode *) bh->b_data;
1336         di->i_mode = cpu_to_le16(inode->i_mode);
1337
1338         ocfs2_journal_dirty(handle, bh);
1339
1340 out_trans:
1341         ocfs2_commit_trans(osb, handle);
1342 out:
1343         return ret;
1344 }
1345
1346 /*
1347  * Will look for holes and unwritten extents in the range starting at
1348  * pos for count bytes (inclusive).
1349  */
1350 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1351                                        size_t count)
1352 {
1353         int ret = 0;
1354         unsigned int extent_flags;
1355         u32 cpos, clusters, extent_len, phys_cpos;
1356         struct super_block *sb = inode->i_sb;
1357
1358         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1359         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1360
1361         while (clusters) {
1362                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1363                                          &extent_flags);
1364                 if (ret < 0) {
1365                         mlog_errno(ret);
1366                         goto out;
1367                 }
1368
1369                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1370                         ret = 1;
1371                         break;
1372                 }
1373
1374                 if (extent_len > clusters)
1375                         extent_len = clusters;
1376
1377                 clusters -= extent_len;
1378                 cpos += extent_len;
1379         }
1380 out:
1381         return ret;
1382 }
1383
1384 static int ocfs2_write_remove_suid(struct inode *inode)
1385 {
1386         int ret;
1387         struct buffer_head *bh = NULL;
1388
1389         ret = ocfs2_read_inode_block(inode, &bh);
1390         if (ret < 0) {
1391                 mlog_errno(ret);
1392                 goto out;
1393         }
1394
1395         ret =  __ocfs2_write_remove_suid(inode, bh);
1396 out:
1397         brelse(bh);
1398         return ret;
1399 }
1400
1401 /*
1402  * Allocate enough extents to cover the region starting at byte offset
1403  * start for len bytes. Existing extents are skipped, any extents
1404  * added are marked as "unwritten".
1405  */
1406 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1407                                             u64 start, u64 len)
1408 {
1409         int ret;
1410         u32 cpos, phys_cpos, clusters, alloc_size;
1411         u64 end = start + len;
1412         struct buffer_head *di_bh = NULL;
1413
1414         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1415                 ret = ocfs2_read_inode_block(inode, &di_bh);
1416                 if (ret) {
1417                         mlog_errno(ret);
1418                         goto out;
1419                 }
1420
1421                 /*
1422                  * Nothing to do if the requested reservation range
1423                  * fits within the inode.
1424                  */
1425                 if (ocfs2_size_fits_inline_data(di_bh, end))
1426                         goto out;
1427
1428                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1429                 if (ret) {
1430                         mlog_errno(ret);
1431                         goto out;
1432                 }
1433         }
1434
1435         /*
1436          * We consider both start and len to be inclusive.
1437          */
1438         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1439         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1440         clusters -= cpos;
1441
1442         while (clusters) {
1443                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1444                                          &alloc_size, NULL);
1445                 if (ret) {
1446                         mlog_errno(ret);
1447                         goto out;
1448                 }
1449
1450                 /*
1451                  * Hole or existing extent len can be arbitrary, so
1452                  * cap it to our own allocation request.
1453                  */
1454                 if (alloc_size > clusters)
1455                         alloc_size = clusters;
1456
1457                 if (phys_cpos) {
1458                         /*
1459                          * We already have an allocation at this
1460                          * region so we can safely skip it.
1461                          */
1462                         goto next;
1463                 }
1464
1465                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1466                 if (ret) {
1467                         if (ret != -ENOSPC)
1468                                 mlog_errno(ret);
1469                         goto out;
1470                 }
1471
1472 next:
1473                 cpos += alloc_size;
1474                 clusters -= alloc_size;
1475         }
1476
1477         ret = 0;
1478 out:
1479
1480         brelse(di_bh);
1481         return ret;
1482 }
1483
1484 /*
1485  * Truncate a byte range, avoiding pages within partial clusters. This
1486  * preserves those pages for the zeroing code to write to.
1487  */
1488 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1489                                          u64 byte_len)
1490 {
1491         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1492         loff_t start, end;
1493         struct address_space *mapping = inode->i_mapping;
1494
1495         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1496         end = byte_start + byte_len;
1497         end = end & ~(osb->s_clustersize - 1);
1498
1499         if (start < end) {
1500                 unmap_mapping_range(mapping, start, end - start, 0);
1501                 truncate_inode_pages_range(mapping, start, end - 1);
1502         }
1503 }
1504
1505 static int ocfs2_zero_partial_clusters(struct inode *inode,
1506                                        u64 start, u64 len)
1507 {
1508         int ret = 0;
1509         u64 tmpend, end = start + len;
1510         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1511         unsigned int csize = osb->s_clustersize;
1512         handle_t *handle;
1513
1514         /*
1515          * The "start" and "end" values are NOT necessarily part of
1516          * the range whose allocation is being deleted. Rather, this
1517          * is what the user passed in with the request. We must zero
1518          * partial clusters here. There's no need to worry about
1519          * physical allocation - the zeroing code knows to skip holes.
1520          */
1521         trace_ocfs2_zero_partial_clusters(
1522                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1523                 (unsigned long long)start, (unsigned long long)end);
1524
1525         /*
1526          * If both edges are on a cluster boundary then there's no
1527          * zeroing required as the region is part of the allocation to
1528          * be truncated.
1529          */
1530         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1531                 goto out;
1532
1533         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1534         if (IS_ERR(handle)) {
1535                 ret = PTR_ERR(handle);
1536                 mlog_errno(ret);
1537                 goto out;
1538         }
1539
1540         /*
1541          * We want to get the byte offset of the end of the 1st cluster.
1542          */
1543         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1544         if (tmpend > end)
1545                 tmpend = end;
1546
1547         trace_ocfs2_zero_partial_clusters_range1((unsigned long long)start,
1548                                                  (unsigned long long)tmpend);
1549
1550         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1551         if (ret)
1552                 mlog_errno(ret);
1553
1554         if (tmpend < end) {
1555                 /*
1556                  * This may make start and end equal, but the zeroing
1557                  * code will skip any work in that case so there's no
1558                  * need to catch it up here.
1559                  */
1560                 start = end & ~(osb->s_clustersize - 1);
1561
1562                 trace_ocfs2_zero_partial_clusters_range2(
1563                         (unsigned long long)start, (unsigned long long)end);
1564
1565                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1566                 if (ret)
1567                         mlog_errno(ret);
1568         }
1569
1570         ocfs2_commit_trans(osb, handle);
1571 out:
1572         return ret;
1573 }
1574
1575 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1576 {
1577         int i;
1578         struct ocfs2_extent_rec *rec = NULL;
1579
1580         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1581
1582                 rec = &el->l_recs[i];
1583
1584                 if (le32_to_cpu(rec->e_cpos) < pos)
1585                         break;
1586         }
1587
1588         return i;
1589 }
1590
1591 /*
1592  * Helper to calculate the punching pos and length in one run, we handle the
1593  * following three cases in order:
1594  *
1595  * - remove the entire record
1596  * - remove a partial record
1597  * - no record needs to be removed (hole-punching completed)
1598 */
1599 static void ocfs2_calc_trunc_pos(struct inode *inode,
1600                                  struct ocfs2_extent_list *el,
1601                                  struct ocfs2_extent_rec *rec,
1602                                  u32 trunc_start, u32 *trunc_cpos,
1603                                  u32 *trunc_len, u32 *trunc_end,
1604                                  u64 *blkno, int *done)
1605 {
1606         int ret = 0;
1607         u32 coff, range;
1608
1609         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1610
1611         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1612                 /*
1613                  * remove an entire extent record.
1614                  */
1615                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1616                 /*
1617                  * Skip holes if any.
1618                  */
1619                 if (range < *trunc_end)
1620                         *trunc_end = range;
1621                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1622                 *blkno = le64_to_cpu(rec->e_blkno);
1623                 *trunc_end = le32_to_cpu(rec->e_cpos);
1624         } else if (range > trunc_start) {
1625                 /*
1626                  * remove a partial extent record, which means we're
1627                  * removing the last extent record.
1628                  */
1629                 *trunc_cpos = trunc_start;
1630                 /*
1631                  * skip hole if any.
1632                  */
1633                 if (range < *trunc_end)
1634                         *trunc_end = range;
1635                 *trunc_len = *trunc_end - trunc_start;
1636                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1637                 *blkno = le64_to_cpu(rec->e_blkno) +
1638                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1639                 *trunc_end = trunc_start;
1640         } else {
1641                 /*
1642                  * It may have two following possibilities:
1643                  *
1644                  * - last record has been removed
1645                  * - trunc_start was within a hole
1646                  *
1647                  * both two cases mean the completion of hole punching.
1648                  */
1649                 ret = 1;
1650         }
1651
1652         *done = ret;
1653 }
1654
1655 static int ocfs2_remove_inode_range(struct inode *inode,
1656                                     struct buffer_head *di_bh, u64 byte_start,
1657                                     u64 byte_len)
1658 {
1659         int ret = 0, flags = 0, done = 0, i;
1660         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1661         u32 cluster_in_el;
1662         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1663         struct ocfs2_cached_dealloc_ctxt dealloc;
1664         struct address_space *mapping = inode->i_mapping;
1665         struct ocfs2_extent_tree et;
1666         struct ocfs2_path *path = NULL;
1667         struct ocfs2_extent_list *el = NULL;
1668         struct ocfs2_extent_rec *rec = NULL;
1669         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1670         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1671
1672         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1673         ocfs2_init_dealloc_ctxt(&dealloc);
1674
1675         trace_ocfs2_remove_inode_range(
1676                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1677                         (unsigned long long)byte_start,
1678                         (unsigned long long)byte_len);
1679
1680         if (byte_len == 0)
1681                 return 0;
1682
1683         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1684                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1685                                             byte_start + byte_len, 0);
1686                 if (ret) {
1687                         mlog_errno(ret);
1688                         goto out;
1689                 }
1690                 /*
1691                  * There's no need to get fancy with the page cache
1692                  * truncate of an inline-data inode. We're talking
1693                  * about less than a page here, which will be cached
1694                  * in the dinode buffer anyway.
1695                  */
1696                 unmap_mapping_range(mapping, 0, 0, 0);
1697                 truncate_inode_pages(mapping, 0);
1698                 goto out;
1699         }
1700
1701         /*
1702          * For reflinks, we may need to CoW 2 clusters which might be
1703          * partially zero'd later, if hole's start and end offset were
1704          * within one cluster(means is not exactly aligned to clustersize).
1705          */
1706
1707         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1708
1709                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1710                 if (ret) {
1711                         mlog_errno(ret);
1712                         goto out;
1713                 }
1714
1715                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1716                 if (ret) {
1717                         mlog_errno(ret);
1718                         goto out;
1719                 }
1720         }
1721
1722         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1723         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1724         cluster_in_el = trunc_end;
1725
1726         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1727         if (ret) {
1728                 mlog_errno(ret);
1729                 goto out;
1730         }
1731
1732         path = ocfs2_new_path_from_et(&et);
1733         if (!path) {
1734                 ret = -ENOMEM;
1735                 mlog_errno(ret);
1736                 goto out;
1737         }
1738
1739         while (trunc_end > trunc_start) {
1740
1741                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1742                                       cluster_in_el);
1743                 if (ret) {
1744                         mlog_errno(ret);
1745                         goto out;
1746                 }
1747
1748                 el = path_leaf_el(path);
1749
1750                 i = ocfs2_find_rec(el, trunc_end);
1751                 /*
1752                  * Need to go to previous extent block.
1753                  */
1754                 if (i < 0) {
1755                         if (path->p_tree_depth == 0)
1756                                 break;
1757
1758                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1759                                                             path,
1760                                                             &cluster_in_el);
1761                         if (ret) {
1762                                 mlog_errno(ret);
1763                                 goto out;
1764                         }
1765
1766                         /*
1767                          * We've reached the leftmost extent block,
1768                          * it's safe to leave.
1769                          */
1770                         if (cluster_in_el == 0)
1771                                 break;
1772
1773                         /*
1774                          * The 'pos' searched for previous extent block is
1775                          * always one cluster less than actual trunc_end.
1776                          */
1777                         trunc_end = cluster_in_el + 1;
1778
1779                         ocfs2_reinit_path(path, 1);
1780
1781                         continue;
1782
1783                 } else
1784                         rec = &el->l_recs[i];
1785
1786                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1787                                      &trunc_len, &trunc_end, &blkno, &done);
1788                 if (done)
1789                         break;
1790
1791                 flags = rec->e_flags;
1792                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1793
1794                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1795                                                phys_cpos, trunc_len, flags,
1796                                                &dealloc, refcount_loc);
1797                 if (ret < 0) {
1798                         mlog_errno(ret);
1799                         goto out;
1800                 }
1801
1802                 cluster_in_el = trunc_end;
1803
1804                 ocfs2_reinit_path(path, 1);
1805         }
1806
1807         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1808
1809 out:
1810         ocfs2_schedule_truncate_log_flush(osb, 1);
1811         ocfs2_run_deallocs(osb, &dealloc);
1812
1813         return ret;
1814 }
1815
1816 /*
1817  * Parts of this function taken from xfs_change_file_space()
1818  */
1819 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1820                                      loff_t f_pos, unsigned int cmd,
1821                                      struct ocfs2_space_resv *sr,
1822                                      int change_size)
1823 {
1824         int ret;
1825         s64 llen;
1826         loff_t size;
1827         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1828         struct buffer_head *di_bh = NULL;
1829         handle_t *handle;
1830         unsigned long long max_off = inode->i_sb->s_maxbytes;
1831
1832         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1833                 return -EROFS;
1834
1835         mutex_lock(&inode->i_mutex);
1836
1837         /*
1838          * This prevents concurrent writes on other nodes
1839          */
1840         ret = ocfs2_rw_lock(inode, 1);
1841         if (ret) {
1842                 mlog_errno(ret);
1843                 goto out;
1844         }
1845
1846         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1847         if (ret) {
1848                 mlog_errno(ret);
1849                 goto out_rw_unlock;
1850         }
1851
1852         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1853                 ret = -EPERM;
1854                 goto out_inode_unlock;
1855         }
1856
1857         switch (sr->l_whence) {
1858         case 0: /*SEEK_SET*/
1859                 break;
1860         case 1: /*SEEK_CUR*/
1861                 sr->l_start += f_pos;
1862                 break;
1863         case 2: /*SEEK_END*/
1864                 sr->l_start += i_size_read(inode);
1865                 break;
1866         default:
1867                 ret = -EINVAL;
1868                 goto out_inode_unlock;
1869         }
1870         sr->l_whence = 0;
1871
1872         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1873
1874         if (sr->l_start < 0
1875             || sr->l_start > max_off
1876             || (sr->l_start + llen) < 0
1877             || (sr->l_start + llen) > max_off) {
1878                 ret = -EINVAL;
1879                 goto out_inode_unlock;
1880         }
1881         size = sr->l_start + sr->l_len;
1882
1883         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1884                 if (sr->l_len <= 0) {
1885                         ret = -EINVAL;
1886                         goto out_inode_unlock;
1887                 }
1888         }
1889
1890         if (file && should_remove_suid(file->f_path.dentry)) {
1891                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1892                 if (ret) {
1893                         mlog_errno(ret);
1894                         goto out_inode_unlock;
1895                 }
1896         }
1897
1898         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1899         switch (cmd) {
1900         case OCFS2_IOC_RESVSP:
1901         case OCFS2_IOC_RESVSP64:
1902                 /*
1903                  * This takes unsigned offsets, but the signed ones we
1904                  * pass have been checked against overflow above.
1905                  */
1906                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1907                                                        sr->l_len);
1908                 break;
1909         case OCFS2_IOC_UNRESVSP:
1910         case OCFS2_IOC_UNRESVSP64:
1911                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1912                                                sr->l_len);
1913                 break;
1914         default:
1915                 ret = -EINVAL;
1916         }
1917         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1918         if (ret) {
1919                 mlog_errno(ret);
1920                 goto out_inode_unlock;
1921         }
1922
1923         /*
1924          * We update c/mtime for these changes
1925          */
1926         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1927         if (IS_ERR(handle)) {
1928                 ret = PTR_ERR(handle);
1929                 mlog_errno(ret);
1930                 goto out_inode_unlock;
1931         }
1932
1933         if (change_size && i_size_read(inode) < size)
1934                 i_size_write(inode, size);
1935
1936         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1937         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1938         if (ret < 0)
1939                 mlog_errno(ret);
1940
1941         ocfs2_commit_trans(osb, handle);
1942
1943 out_inode_unlock:
1944         brelse(di_bh);
1945         ocfs2_inode_unlock(inode, 1);
1946 out_rw_unlock:
1947         ocfs2_rw_unlock(inode, 1);
1948
1949 out:
1950         mutex_unlock(&inode->i_mutex);
1951         return ret;
1952 }
1953
1954 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1955                             struct ocfs2_space_resv *sr)
1956 {
1957         struct inode *inode = file->f_path.dentry->d_inode;
1958         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1959
1960         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1961             !ocfs2_writes_unwritten_extents(osb))
1962                 return -ENOTTY;
1963         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1964                  !ocfs2_sparse_alloc(osb))
1965                 return -ENOTTY;
1966
1967         if (!S_ISREG(inode->i_mode))
1968                 return -EINVAL;
1969
1970         if (!(file->f_mode & FMODE_WRITE))
1971                 return -EBADF;
1972
1973         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1974 }
1975
1976 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
1977                             loff_t len)
1978 {
1979         struct inode *inode = file->f_path.dentry->d_inode;
1980         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1981         struct ocfs2_space_resv sr;
1982         int change_size = 1;
1983         int cmd = OCFS2_IOC_RESVSP64;
1984
1985         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1986                 return -EOPNOTSUPP;
1987         if (!ocfs2_writes_unwritten_extents(osb))
1988                 return -EOPNOTSUPP;
1989
1990         if (mode & FALLOC_FL_KEEP_SIZE)
1991                 change_size = 0;
1992
1993         if (mode & FALLOC_FL_PUNCH_HOLE)
1994                 cmd = OCFS2_IOC_UNRESVSP64;
1995
1996         sr.l_whence = 0;
1997         sr.l_start = (s64)offset;
1998         sr.l_len = (s64)len;
1999
2000         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
2001                                          change_size);
2002 }
2003
2004 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2005                                    size_t count)
2006 {
2007         int ret = 0;
2008         unsigned int extent_flags;
2009         u32 cpos, clusters, extent_len, phys_cpos;
2010         struct super_block *sb = inode->i_sb;
2011
2012         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2013             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2014             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2015                 return 0;
2016
2017         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2018         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2019
2020         while (clusters) {
2021                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2022                                          &extent_flags);
2023                 if (ret < 0) {
2024                         mlog_errno(ret);
2025                         goto out;
2026                 }
2027
2028                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2029                         ret = 1;
2030                         break;
2031                 }
2032
2033                 if (extent_len > clusters)
2034                         extent_len = clusters;
2035
2036                 clusters -= extent_len;
2037                 cpos += extent_len;
2038         }
2039 out:
2040         return ret;
2041 }
2042
2043 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2044                                             struct file *file,
2045                                             loff_t pos, size_t count,
2046                                             int *meta_level)
2047 {
2048         int ret;
2049         struct buffer_head *di_bh = NULL;
2050         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2051         u32 clusters =
2052                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2053
2054         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2055         if (ret) {
2056                 mlog_errno(ret);
2057                 goto out;
2058         }
2059
2060         *meta_level = 1;
2061
2062         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2063         if (ret)
2064                 mlog_errno(ret);
2065 out:
2066         brelse(di_bh);
2067         return ret;
2068 }
2069
2070 static int ocfs2_prepare_inode_for_write(struct file *file,
2071                                          loff_t *ppos,
2072                                          size_t count,
2073                                          int appending,
2074                                          int *direct_io,
2075                                          int *has_refcount)
2076 {
2077         int ret = 0, meta_level = 0;
2078         struct dentry *dentry = file->f_path.dentry;
2079         struct inode *inode = dentry->d_inode;
2080         loff_t saved_pos = 0, end;
2081
2082         /*
2083          * We start with a read level meta lock and only jump to an ex
2084          * if we need to make modifications here.
2085          */
2086         for(;;) {
2087                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2088                 if (ret < 0) {
2089                         meta_level = -1;
2090                         mlog_errno(ret);
2091                         goto out;
2092                 }
2093
2094                 /* Clear suid / sgid if necessary. We do this here
2095                  * instead of later in the write path because
2096                  * remove_suid() calls ->setattr without any hint that
2097                  * we may have already done our cluster locking. Since
2098                  * ocfs2_setattr() *must* take cluster locks to
2099                  * proceeed, this will lead us to recursively lock the
2100                  * inode. There's also the dinode i_size state which
2101                  * can be lost via setattr during extending writes (we
2102                  * set inode->i_size at the end of a write. */
2103                 if (should_remove_suid(dentry)) {
2104                         if (meta_level == 0) {
2105                                 ocfs2_inode_unlock(inode, meta_level);
2106                                 meta_level = 1;
2107                                 continue;
2108                         }
2109
2110                         ret = ocfs2_write_remove_suid(inode);
2111                         if (ret < 0) {
2112                                 mlog_errno(ret);
2113                                 goto out_unlock;
2114                         }
2115                 }
2116
2117                 /* work on a copy of ppos until we're sure that we won't have
2118                  * to recalculate it due to relocking. */
2119                 if (appending)
2120                         saved_pos = i_size_read(inode);
2121                 else
2122                         saved_pos = *ppos;
2123
2124                 end = saved_pos + count;
2125
2126                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2127                 if (ret == 1) {
2128                         ocfs2_inode_unlock(inode, meta_level);
2129                         meta_level = -1;
2130
2131                         ret = ocfs2_prepare_inode_for_refcount(inode,
2132                                                                file,
2133                                                                saved_pos,
2134                                                                count,
2135                                                                &meta_level);
2136                         if (has_refcount)
2137                                 *has_refcount = 1;
2138                         if (direct_io)
2139                                 *direct_io = 0;
2140                 }
2141
2142                 if (ret < 0) {
2143                         mlog_errno(ret);
2144                         goto out_unlock;
2145                 }
2146
2147                 /*
2148                  * Skip the O_DIRECT checks if we don't need
2149                  * them.
2150                  */
2151                 if (!direct_io || !(*direct_io))
2152                         break;
2153
2154                 /*
2155                  * There's no sane way to do direct writes to an inode
2156                  * with inline data.
2157                  */
2158                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2159                         *direct_io = 0;
2160                         break;
2161                 }
2162
2163                 /*
2164                  * Allowing concurrent direct writes means
2165                  * i_size changes wouldn't be synchronized, so
2166                  * one node could wind up truncating another
2167                  * nodes writes.
2168                  */
2169                 if (end > i_size_read(inode)) {
2170                         *direct_io = 0;
2171                         break;
2172                 }
2173
2174                 /*
2175                  * We don't fill holes during direct io, so
2176                  * check for them here. If any are found, the
2177                  * caller will have to retake some cluster
2178                  * locks and initiate the io as buffered.
2179                  */
2180                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2181                 if (ret == 1) {
2182                         *direct_io = 0;
2183                         ret = 0;
2184                 } else if (ret < 0)
2185                         mlog_errno(ret);
2186                 break;
2187         }
2188
2189         if (appending)
2190                 *ppos = saved_pos;
2191
2192 out_unlock:
2193         trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2194                                             saved_pos, appending, count,
2195                                             direct_io, has_refcount);
2196
2197         if (meta_level >= 0)
2198                 ocfs2_inode_unlock(inode, meta_level);
2199
2200 out:
2201         return ret;
2202 }
2203
2204 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2205                                     const struct iovec *iov,
2206                                     unsigned long nr_segs,
2207                                     loff_t pos)
2208 {
2209         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2210         int can_do_direct, has_refcount = 0;
2211         ssize_t written = 0;
2212         size_t ocount;          /* original count */
2213         size_t count;           /* after file limit checks */
2214         loff_t old_size, *ppos = &iocb->ki_pos;
2215         u32 old_clusters;
2216         struct file *file = iocb->ki_filp;
2217         struct inode *inode = file->f_path.dentry->d_inode;
2218         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2219         int full_coherency = !(osb->s_mount_opt &
2220                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2221
2222         trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2223                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2224                 file->f_path.dentry->d_name.len,
2225                 file->f_path.dentry->d_name.name,
2226                 (unsigned int)nr_segs);
2227
2228         if (iocb->ki_left == 0)
2229                 return 0;
2230
2231         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2232
2233         appending = file->f_flags & O_APPEND ? 1 : 0;
2234         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2235
2236         mutex_lock(&inode->i_mutex);
2237
2238         ocfs2_iocb_clear_sem_locked(iocb);
2239
2240 relock:
2241         /* to match setattr's i_mutex -> rw_lock ordering */
2242         if (direct_io) {
2243                 have_alloc_sem = 1;
2244                 /* communicate with ocfs2_dio_end_io */
2245                 ocfs2_iocb_set_sem_locked(iocb);
2246         }
2247
2248         /*
2249          * Concurrent O_DIRECT writes are allowed with
2250          * mount_option "coherency=buffered".
2251          */
2252         rw_level = (!direct_io || full_coherency);
2253
2254         ret = ocfs2_rw_lock(inode, rw_level);
2255         if (ret < 0) {
2256                 mlog_errno(ret);
2257                 goto out_sems;
2258         }
2259
2260         /*
2261          * O_DIRECT writes with "coherency=full" need to take EX cluster
2262          * inode_lock to guarantee coherency.
2263          */
2264         if (direct_io && full_coherency) {
2265                 /*
2266                  * We need to take and drop the inode lock to force
2267                  * other nodes to drop their caches.  Buffered I/O
2268                  * already does this in write_begin().
2269                  */
2270                 ret = ocfs2_inode_lock(inode, NULL, 1);
2271                 if (ret < 0) {
2272                         mlog_errno(ret);
2273                         goto out_sems;
2274                 }
2275
2276                 ocfs2_inode_unlock(inode, 1);
2277         }
2278
2279         can_do_direct = direct_io;
2280         ret = ocfs2_prepare_inode_for_write(file, ppos,
2281                                             iocb->ki_left, appending,
2282                                             &can_do_direct, &has_refcount);
2283         if (ret < 0) {
2284                 mlog_errno(ret);
2285                 goto out;
2286         }
2287
2288         /*
2289          * We can't complete the direct I/O as requested, fall back to
2290          * buffered I/O.
2291          */
2292         if (direct_io && !can_do_direct) {
2293                 ocfs2_rw_unlock(inode, rw_level);
2294
2295                 have_alloc_sem = 0;
2296                 rw_level = -1;
2297
2298                 direct_io = 0;
2299                 goto relock;
2300         }
2301
2302         /*
2303          * To later detect whether a journal commit for sync writes is
2304          * necessary, we sample i_size, and cluster count here.
2305          */
2306         old_size = i_size_read(inode);
2307         old_clusters = OCFS2_I(inode)->ip_clusters;
2308
2309         /* communicate with ocfs2_dio_end_io */
2310         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2311
2312         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2313                                      VERIFY_READ);
2314         if (ret)
2315                 goto out_dio;
2316
2317         count = ocount;
2318         ret = generic_write_checks(file, ppos, &count,
2319                                    S_ISBLK(inode->i_mode));
2320         if (ret)
2321                 goto out_dio;
2322
2323         if (direct_io) {
2324                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2325                                                     ppos, count, ocount);
2326                 if (written < 0) {
2327                         ret = written;
2328                         goto out_dio;
2329                 }
2330         } else {
2331                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2332                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2333                                                       ppos, count, 0);
2334                 current->backing_dev_info = NULL;
2335         }
2336
2337 out_dio:
2338         /* buffered aio wouldn't have proper lock coverage today */
2339         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2340
2341         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2342             ((file->f_flags & O_DIRECT) && !direct_io)) {
2343                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2344                                                pos + count - 1);
2345                 if (ret < 0)
2346                         written = ret;
2347
2348                 if (!ret && ((old_size != i_size_read(inode)) ||
2349                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2350                              has_refcount)) {
2351                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2352                         if (ret < 0)
2353                                 written = ret;
2354                 }
2355
2356                 if (!ret)
2357                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2358                                                       pos + count - 1);
2359         }
2360
2361         /*
2362          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2363          * function pointer which is called when o_direct io completes so that
2364          * it can unlock our rw lock.
2365          * Unfortunately there are error cases which call end_io and others
2366          * that don't.  so we don't have to unlock the rw_lock if either an
2367          * async dio is going to do it in the future or an end_io after an
2368          * error has already done it.
2369          */
2370         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2371                 rw_level = -1;
2372                 have_alloc_sem = 0;
2373         }
2374
2375 out:
2376         if (rw_level != -1)
2377                 ocfs2_rw_unlock(inode, rw_level);
2378
2379 out_sems:
2380         if (have_alloc_sem)
2381                 ocfs2_iocb_clear_sem_locked(iocb);
2382
2383         mutex_unlock(&inode->i_mutex);
2384
2385         if (written)
2386                 ret = written;
2387         return ret;
2388 }
2389
2390 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2391                                 struct file *out,
2392                                 struct splice_desc *sd)
2393 {
2394         int ret;
2395
2396         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2397                                             sd->total_len, 0, NULL, NULL);
2398         if (ret < 0) {
2399                 mlog_errno(ret);
2400                 return ret;
2401         }
2402
2403         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2404 }
2405
2406 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2407                                        struct file *out,
2408                                        loff_t *ppos,
2409                                        size_t len,
2410                                        unsigned int flags)
2411 {
2412         int ret;
2413         struct address_space *mapping = out->f_mapping;
2414         struct inode *inode = mapping->host;
2415         struct splice_desc sd = {
2416                 .total_len = len,
2417                 .flags = flags,
2418                 .pos = *ppos,
2419                 .u.file = out,
2420         };
2421
2422
2423         trace_ocfs2_file_splice_write(inode, out, out->f_path.dentry,
2424                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2425                         out->f_path.dentry->d_name.len,
2426                         out->f_path.dentry->d_name.name, len);
2427
2428         if (pipe->inode)
2429                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2430
2431         splice_from_pipe_begin(&sd);
2432         do {
2433                 ret = splice_from_pipe_next(pipe, &sd);
2434                 if (ret <= 0)
2435                         break;
2436
2437                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2438                 ret = ocfs2_rw_lock(inode, 1);
2439                 if (ret < 0)
2440                         mlog_errno(ret);
2441                 else {
2442                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2443                         ocfs2_rw_unlock(inode, 1);
2444                 }
2445                 mutex_unlock(&inode->i_mutex);
2446         } while (ret > 0);
2447         splice_from_pipe_end(pipe, &sd);
2448
2449         if (pipe->inode)
2450                 mutex_unlock(&pipe->inode->i_mutex);
2451
2452         if (sd.num_spliced)
2453                 ret = sd.num_spliced;
2454
2455         if (ret > 0) {
2456                 unsigned long nr_pages;
2457                 int err;
2458
2459                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2460
2461                 err = generic_write_sync(out, *ppos, ret);
2462                 if (err)
2463                         ret = err;
2464                 else
2465                         *ppos += ret;
2466
2467                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2468         }
2469
2470         return ret;
2471 }
2472
2473 static ssize_t ocfs2_file_splice_read(struct file *in,
2474                                       loff_t *ppos,
2475                                       struct pipe_inode_info *pipe,
2476                                       size_t len,
2477                                       unsigned int flags)
2478 {
2479         int ret = 0, lock_level = 0;
2480         struct inode *inode = in->f_path.dentry->d_inode;
2481
2482         trace_ocfs2_file_splice_read(inode, in, in->f_path.dentry,
2483                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2484                         in->f_path.dentry->d_name.len,
2485                         in->f_path.dentry->d_name.name, len);
2486
2487         /*
2488          * See the comment in ocfs2_file_aio_read()
2489          */
2490         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2491         if (ret < 0) {
2492                 mlog_errno(ret);
2493                 goto bail;
2494         }
2495         ocfs2_inode_unlock(inode, lock_level);
2496
2497         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2498
2499 bail:
2500         return ret;
2501 }
2502
2503 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2504                                    const struct iovec *iov,
2505                                    unsigned long nr_segs,
2506                                    loff_t pos)
2507 {
2508         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2509         struct file *filp = iocb->ki_filp;
2510         struct inode *inode = filp->f_path.dentry->d_inode;
2511
2512         trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2513                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2514                         filp->f_path.dentry->d_name.len,
2515                         filp->f_path.dentry->d_name.name, nr_segs);
2516
2517
2518         if (!inode) {
2519                 ret = -EINVAL;
2520                 mlog_errno(ret);
2521                 goto bail;
2522         }
2523
2524         ocfs2_iocb_clear_sem_locked(iocb);
2525
2526         /*
2527          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2528          * need locks to protect pending reads from racing with truncate.
2529          */
2530         if (filp->f_flags & O_DIRECT) {
2531                 have_alloc_sem = 1;
2532                 ocfs2_iocb_set_sem_locked(iocb);
2533
2534                 ret = ocfs2_rw_lock(inode, 0);
2535                 if (ret < 0) {
2536                         mlog_errno(ret);
2537                         goto bail;
2538                 }
2539                 rw_level = 0;
2540                 /* communicate with ocfs2_dio_end_io */
2541                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2542         }
2543
2544         /*
2545          * We're fine letting folks race truncates and extending
2546          * writes with read across the cluster, just like they can
2547          * locally. Hence no rw_lock during read.
2548          *
2549          * Take and drop the meta data lock to update inode fields
2550          * like i_size. This allows the checks down below
2551          * generic_file_aio_read() a chance of actually working.
2552          */
2553         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2554         if (ret < 0) {
2555                 mlog_errno(ret);
2556                 goto bail;
2557         }
2558         ocfs2_inode_unlock(inode, lock_level);
2559
2560         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2561         trace_generic_file_aio_read_ret(ret);
2562
2563         /* buffered aio wouldn't have proper lock coverage today */
2564         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2565
2566         /* see ocfs2_file_aio_write */
2567         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2568                 rw_level = -1;
2569                 have_alloc_sem = 0;
2570         }
2571
2572 bail:
2573         if (have_alloc_sem)
2574                 ocfs2_iocb_clear_sem_locked(iocb);
2575
2576         if (rw_level != -1)
2577                 ocfs2_rw_unlock(inode, rw_level);
2578
2579         return ret;
2580 }
2581
2582 const struct inode_operations ocfs2_file_iops = {
2583         .setattr        = ocfs2_setattr,
2584         .getattr        = ocfs2_getattr,
2585         .permission     = ocfs2_permission,
2586         .setxattr       = generic_setxattr,
2587         .getxattr       = generic_getxattr,
2588         .listxattr      = ocfs2_listxattr,
2589         .removexattr    = generic_removexattr,
2590         .fiemap         = ocfs2_fiemap,
2591         .check_acl      = ocfs2_check_acl,
2592 };
2593
2594 const struct inode_operations ocfs2_special_file_iops = {
2595         .setattr        = ocfs2_setattr,
2596         .getattr        = ocfs2_getattr,
2597         .permission     = ocfs2_permission,
2598         .check_acl      = ocfs2_check_acl,
2599 };
2600
2601 /*
2602  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2603  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2604  */
2605 const struct file_operations ocfs2_fops = {
2606         .llseek         = generic_file_llseek,
2607         .read           = do_sync_read,
2608         .write          = do_sync_write,
2609         .mmap           = ocfs2_mmap,
2610         .fsync          = ocfs2_sync_file,
2611         .release        = ocfs2_file_release,
2612         .open           = ocfs2_file_open,
2613         .aio_read       = ocfs2_file_aio_read,
2614         .aio_write      = ocfs2_file_aio_write,
2615         .unlocked_ioctl = ocfs2_ioctl,
2616 #ifdef CONFIG_COMPAT
2617         .compat_ioctl   = ocfs2_compat_ioctl,
2618 #endif
2619         .lock           = ocfs2_lock,
2620         .flock          = ocfs2_flock,
2621         .splice_read    = ocfs2_file_splice_read,
2622         .splice_write   = ocfs2_file_splice_write,
2623         .fallocate      = ocfs2_fallocate,
2624 };
2625
2626 const struct file_operations ocfs2_dops = {
2627         .llseek         = generic_file_llseek,
2628         .read           = generic_read_dir,
2629         .readdir        = ocfs2_readdir,
2630         .fsync          = ocfs2_sync_file,
2631         .release        = ocfs2_dir_release,
2632         .open           = ocfs2_dir_open,
2633         .unlocked_ioctl = ocfs2_ioctl,
2634 #ifdef CONFIG_COMPAT
2635         .compat_ioctl   = ocfs2_compat_ioctl,
2636 #endif
2637         .lock           = ocfs2_lock,
2638         .flock          = ocfs2_flock,
2639 };
2640
2641 /*
2642  * POSIX-lockless variants of our file_operations.
2643  *
2644  * These will be used if the underlying cluster stack does not support
2645  * posix file locking, if the user passes the "localflocks" mount
2646  * option, or if we have a local-only fs.
2647  *
2648  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2649  * so we still want it in the case of no stack support for
2650  * plocks. Internally, it will do the right thing when asked to ignore
2651  * the cluster.
2652  */
2653 const struct file_operations ocfs2_fops_no_plocks = {
2654         .llseek         = generic_file_llseek,
2655         .read           = do_sync_read,
2656         .write          = do_sync_write,
2657         .mmap           = ocfs2_mmap,
2658         .fsync          = ocfs2_sync_file,
2659         .release        = ocfs2_file_release,
2660         .open           = ocfs2_file_open,
2661         .aio_read       = ocfs2_file_aio_read,
2662         .aio_write      = ocfs2_file_aio_write,
2663         .unlocked_ioctl = ocfs2_ioctl,
2664 #ifdef CONFIG_COMPAT
2665         .compat_ioctl   = ocfs2_compat_ioctl,
2666 #endif
2667         .flock          = ocfs2_flock,
2668         .splice_read    = ocfs2_file_splice_read,
2669         .splice_write   = ocfs2_file_splice_write,
2670         .fallocate      = ocfs2_fallocate,
2671 };
2672
2673 const struct file_operations ocfs2_dops_no_plocks = {
2674         .llseek         = generic_file_llseek,
2675         .read           = generic_read_dir,
2676         .readdir        = ocfs2_readdir,
2677         .fsync          = ocfs2_sync_file,
2678         .release        = ocfs2_dir_release,
2679         .open           = ocfs2_dir_open,
2680         .unlocked_ioctl = ocfs2_ioctl,
2681 #ifdef CONFIG_COMPAT
2682         .compat_ioctl   = ocfs2_compat_ioctl,
2683 #endif
2684         .flock          = ocfs2_flock,
2685 };