fs: provide rcu-walk aware permission i_ops
[linux-2.6.git] / fs / ocfs2 / file.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/capability.h>
27 #include <linux/fs.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
38 #include <linux/quotaops.h>
39 #include <linux/blkdev.h>
40
41 #define MLOG_MASK_PREFIX ML_INODE
42 #include <cluster/masklog.h>
43
44 #include "ocfs2.h"
45
46 #include "alloc.h"
47 #include "aops.h"
48 #include "dir.h"
49 #include "dlmglue.h"
50 #include "extent_map.h"
51 #include "file.h"
52 #include "sysfile.h"
53 #include "inode.h"
54 #include "ioctl.h"
55 #include "journal.h"
56 #include "locks.h"
57 #include "mmap.h"
58 #include "suballoc.h"
59 #include "super.h"
60 #include "xattr.h"
61 #include "acl.h"
62 #include "quota.h"
63 #include "refcounttree.h"
64
65 #include "buffer_head_io.h"
66
67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
68 {
69         struct ocfs2_file_private *fp;
70
71         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
72         if (!fp)
73                 return -ENOMEM;
74
75         fp->fp_file = file;
76         mutex_init(&fp->fp_mutex);
77         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78         file->private_data = fp;
79
80         return 0;
81 }
82
83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
84 {
85         struct ocfs2_file_private *fp = file->private_data;
86         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
87
88         if (fp) {
89                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90                 ocfs2_lock_res_free(&fp->fp_flock);
91                 kfree(fp);
92                 file->private_data = NULL;
93         }
94 }
95
96 static int ocfs2_file_open(struct inode *inode, struct file *file)
97 {
98         int status;
99         int mode = file->f_flags;
100         struct ocfs2_inode_info *oi = OCFS2_I(inode);
101
102         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
103                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
104
105         if (file->f_mode & FMODE_WRITE)
106                 dquot_initialize(inode);
107
108         spin_lock(&oi->ip_lock);
109
110         /* Check that the inode hasn't been wiped from disk by another
111          * node. If it hasn't then we're safe as long as we hold the
112          * spin lock until our increment of open count. */
113         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
114                 spin_unlock(&oi->ip_lock);
115
116                 status = -ENOENT;
117                 goto leave;
118         }
119
120         if (mode & O_DIRECT)
121                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
122
123         oi->ip_open_count++;
124         spin_unlock(&oi->ip_lock);
125
126         status = ocfs2_init_file_private(inode, file);
127         if (status) {
128                 /*
129                  * We want to set open count back if we're failing the
130                  * open.
131                  */
132                 spin_lock(&oi->ip_lock);
133                 oi->ip_open_count--;
134                 spin_unlock(&oi->ip_lock);
135         }
136
137 leave:
138         mlog_exit(status);
139         return status;
140 }
141
142 static int ocfs2_file_release(struct inode *inode, struct file *file)
143 {
144         struct ocfs2_inode_info *oi = OCFS2_I(inode);
145
146         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
147                        file->f_path.dentry->d_name.len,
148                        file->f_path.dentry->d_name.name);
149
150         spin_lock(&oi->ip_lock);
151         if (!--oi->ip_open_count)
152                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
153         spin_unlock(&oi->ip_lock);
154
155         ocfs2_free_file_private(inode, file);
156
157         mlog_exit(0);
158
159         return 0;
160 }
161
162 static int ocfs2_dir_open(struct inode *inode, struct file *file)
163 {
164         return ocfs2_init_file_private(inode, file);
165 }
166
167 static int ocfs2_dir_release(struct inode *inode, struct file *file)
168 {
169         ocfs2_free_file_private(inode, file);
170         return 0;
171 }
172
173 static int ocfs2_sync_file(struct file *file, int datasync)
174 {
175         int err = 0;
176         journal_t *journal;
177         struct inode *inode = file->f_mapping->host;
178         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
179
180         mlog_entry("(0x%p, %d, 0x%p, '%.*s')\n", file, datasync,
181                    file->f_path.dentry, file->f_path.dentry->d_name.len,
182                    file->f_path.dentry->d_name.name);
183
184         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
185                 /*
186                  * We still have to flush drive's caches to get data to the
187                  * platter
188                  */
189                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
190                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
191                 goto bail;
192         }
193
194         journal = osb->journal->j_journal;
195         err = jbd2_journal_force_commit(journal);
196
197 bail:
198         mlog_exit(err);
199
200         return (err < 0) ? -EIO : 0;
201 }
202
203 int ocfs2_should_update_atime(struct inode *inode,
204                               struct vfsmount *vfsmnt)
205 {
206         struct timespec now;
207         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
208
209         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
210                 return 0;
211
212         if ((inode->i_flags & S_NOATIME) ||
213             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
214                 return 0;
215
216         /*
217          * We can be called with no vfsmnt structure - NFSD will
218          * sometimes do this.
219          *
220          * Note that our action here is different than touch_atime() -
221          * if we can't tell whether this is a noatime mount, then we
222          * don't know whether to trust the value of s_atime_quantum.
223          */
224         if (vfsmnt == NULL)
225                 return 0;
226
227         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
228             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
229                 return 0;
230
231         if (vfsmnt->mnt_flags & MNT_RELATIME) {
232                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
233                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
234                         return 1;
235
236                 return 0;
237         }
238
239         now = CURRENT_TIME;
240         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
241                 return 0;
242         else
243                 return 1;
244 }
245
246 int ocfs2_update_inode_atime(struct inode *inode,
247                              struct buffer_head *bh)
248 {
249         int ret;
250         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
251         handle_t *handle;
252         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
253
254         mlog_entry_void();
255
256         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
257         if (IS_ERR(handle)) {
258                 ret = PTR_ERR(handle);
259                 mlog_errno(ret);
260                 goto out;
261         }
262
263         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
264                                       OCFS2_JOURNAL_ACCESS_WRITE);
265         if (ret) {
266                 mlog_errno(ret);
267                 goto out_commit;
268         }
269
270         /*
271          * Don't use ocfs2_mark_inode_dirty() here as we don't always
272          * have i_mutex to guard against concurrent changes to other
273          * inode fields.
274          */
275         inode->i_atime = CURRENT_TIME;
276         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
277         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
278         ocfs2_journal_dirty(handle, bh);
279
280 out_commit:
281         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
282 out:
283         mlog_exit(ret);
284         return ret;
285 }
286
287 static int ocfs2_set_inode_size(handle_t *handle,
288                                 struct inode *inode,
289                                 struct buffer_head *fe_bh,
290                                 u64 new_i_size)
291 {
292         int status;
293
294         mlog_entry_void();
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         mlog_exit(status);
307         return status;
308 }
309
310 int ocfs2_simple_size_update(struct inode *inode,
311                              struct buffer_head *di_bh,
312                              u64 new_i_size)
313 {
314         int ret;
315         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
316         handle_t *handle = NULL;
317
318         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
319         if (IS_ERR(handle)) {
320                 ret = PTR_ERR(handle);
321                 mlog_errno(ret);
322                 goto out;
323         }
324
325         ret = ocfs2_set_inode_size(handle, inode, di_bh,
326                                    new_i_size);
327         if (ret < 0)
328                 mlog_errno(ret);
329
330         ocfs2_commit_trans(osb, handle);
331 out:
332         return ret;
333 }
334
335 static int ocfs2_cow_file_pos(struct inode *inode,
336                               struct buffer_head *fe_bh,
337                               u64 offset)
338 {
339         int status;
340         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
341         unsigned int num_clusters = 0;
342         unsigned int ext_flags = 0;
343
344         /*
345          * If the new offset is aligned to the range of the cluster, there is
346          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
347          * CoW either.
348          */
349         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
350                 return 0;
351
352         status = ocfs2_get_clusters(inode, cpos, &phys,
353                                     &num_clusters, &ext_flags);
354         if (status) {
355                 mlog_errno(status);
356                 goto out;
357         }
358
359         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
360                 goto out;
361
362         return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
363
364 out:
365         return status;
366 }
367
368 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
369                                      struct inode *inode,
370                                      struct buffer_head *fe_bh,
371                                      u64 new_i_size)
372 {
373         int status;
374         handle_t *handle;
375         struct ocfs2_dinode *di;
376         u64 cluster_bytes;
377
378         mlog_entry_void();
379
380         /*
381          * We need to CoW the cluster contains the offset if it is reflinked
382          * since we will call ocfs2_zero_range_for_truncate later which will
383          * write "0" from offset to the end of the cluster.
384          */
385         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
386         if (status) {
387                 mlog_errno(status);
388                 return status;
389         }
390
391         /* TODO: This needs to actually orphan the inode in this
392          * transaction. */
393
394         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
395         if (IS_ERR(handle)) {
396                 status = PTR_ERR(handle);
397                 mlog_errno(status);
398                 goto out;
399         }
400
401         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
402                                          OCFS2_JOURNAL_ACCESS_WRITE);
403         if (status < 0) {
404                 mlog_errno(status);
405                 goto out_commit;
406         }
407
408         /*
409          * Do this before setting i_size.
410          */
411         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
412         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
413                                                cluster_bytes);
414         if (status) {
415                 mlog_errno(status);
416                 goto out_commit;
417         }
418
419         i_size_write(inode, new_i_size);
420         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
421
422         di = (struct ocfs2_dinode *) fe_bh->b_data;
423         di->i_size = cpu_to_le64(new_i_size);
424         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
425         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
426
427         ocfs2_journal_dirty(handle, fe_bh);
428
429 out_commit:
430         ocfs2_commit_trans(osb, handle);
431 out:
432
433         mlog_exit(status);
434         return status;
435 }
436
437 static int ocfs2_truncate_file(struct inode *inode,
438                                struct buffer_head *di_bh,
439                                u64 new_i_size)
440 {
441         int status = 0;
442         struct ocfs2_dinode *fe = NULL;
443         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
444
445         mlog_entry("(inode = %llu, new_i_size = %llu\n",
446                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
447                    (unsigned long long)new_i_size);
448
449         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
450          * already validated it */
451         fe = (struct ocfs2_dinode *) di_bh->b_data;
452
453         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
454                         "Inode %llu, inode i_size = %lld != di "
455                         "i_size = %llu, i_flags = 0x%x\n",
456                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
457                         i_size_read(inode),
458                         (unsigned long long)le64_to_cpu(fe->i_size),
459                         le32_to_cpu(fe->i_flags));
460
461         if (new_i_size > le64_to_cpu(fe->i_size)) {
462                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
463                      (unsigned long long)le64_to_cpu(fe->i_size),
464                      (unsigned long long)new_i_size);
465                 status = -EINVAL;
466                 mlog_errno(status);
467                 goto bail;
468         }
469
470         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
471              (unsigned long long)le64_to_cpu(fe->i_blkno),
472              (unsigned long long)le64_to_cpu(fe->i_size),
473              (unsigned long long)new_i_size);
474
475         /* lets handle the simple truncate cases before doing any more
476          * cluster locking. */
477         if (new_i_size == le64_to_cpu(fe->i_size))
478                 goto bail;
479
480         down_write(&OCFS2_I(inode)->ip_alloc_sem);
481
482         ocfs2_resv_discard(&osb->osb_la_resmap,
483                            &OCFS2_I(inode)->ip_la_data_resv);
484
485         /*
486          * The inode lock forced other nodes to sync and drop their
487          * pages, which (correctly) happens even if we have a truncate
488          * without allocation change - ocfs2 cluster sizes can be much
489          * greater than page size, so we have to truncate them
490          * anyway.
491          */
492         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
493         truncate_inode_pages(inode->i_mapping, new_i_size);
494
495         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
496                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
497                                                i_size_read(inode), 1);
498                 if (status)
499                         mlog_errno(status);
500
501                 goto bail_unlock_sem;
502         }
503
504         /* alright, we're going to need to do a full blown alloc size
505          * change. Orphan the inode so that recovery can complete the
506          * truncate if necessary. This does the task of marking
507          * i_size. */
508         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
509         if (status < 0) {
510                 mlog_errno(status);
511                 goto bail_unlock_sem;
512         }
513
514         status = ocfs2_commit_truncate(osb, inode, di_bh);
515         if (status < 0) {
516                 mlog_errno(status);
517                 goto bail_unlock_sem;
518         }
519
520         /* TODO: orphan dir cleanup here. */
521 bail_unlock_sem:
522         up_write(&OCFS2_I(inode)->ip_alloc_sem);
523
524 bail:
525         if (!status && OCFS2_I(inode)->ip_clusters == 0)
526                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
527
528         mlog_exit(status);
529         return status;
530 }
531
532 /*
533  * extend file allocation only here.
534  * we'll update all the disk stuff, and oip->alloc_size
535  *
536  * expect stuff to be locked, a transaction started and enough data /
537  * metadata reservations in the contexts.
538  *
539  * Will return -EAGAIN, and a reason if a restart is needed.
540  * If passed in, *reason will always be set, even in error.
541  */
542 int ocfs2_add_inode_data(struct ocfs2_super *osb,
543                          struct inode *inode,
544                          u32 *logical_offset,
545                          u32 clusters_to_add,
546                          int mark_unwritten,
547                          struct buffer_head *fe_bh,
548                          handle_t *handle,
549                          struct ocfs2_alloc_context *data_ac,
550                          struct ocfs2_alloc_context *meta_ac,
551                          enum ocfs2_alloc_restarted *reason_ret)
552 {
553         int ret;
554         struct ocfs2_extent_tree et;
555
556         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
557         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
558                                           clusters_to_add, mark_unwritten,
559                                           data_ac, meta_ac, reason_ret);
560
561         return ret;
562 }
563
564 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
565                                      u32 clusters_to_add, int mark_unwritten)
566 {
567         int status = 0;
568         int restart_func = 0;
569         int credits;
570         u32 prev_clusters;
571         struct buffer_head *bh = NULL;
572         struct ocfs2_dinode *fe = NULL;
573         handle_t *handle = NULL;
574         struct ocfs2_alloc_context *data_ac = NULL;
575         struct ocfs2_alloc_context *meta_ac = NULL;
576         enum ocfs2_alloc_restarted why;
577         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
578         struct ocfs2_extent_tree et;
579         int did_quota = 0;
580
581         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
582
583         /*
584          * This function only exists for file systems which don't
585          * support holes.
586          */
587         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
588
589         status = ocfs2_read_inode_block(inode, &bh);
590         if (status < 0) {
591                 mlog_errno(status);
592                 goto leave;
593         }
594         fe = (struct ocfs2_dinode *) bh->b_data;
595
596 restart_all:
597         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
598
599         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
600              "clusters_to_add = %u\n",
601              (unsigned long long)OCFS2_I(inode)->ip_blkno,
602              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
603              clusters_to_add);
604         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
605         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
606                                        &data_ac, &meta_ac);
607         if (status) {
608                 mlog_errno(status);
609                 goto leave;
610         }
611
612         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
613                                             clusters_to_add);
614         handle = ocfs2_start_trans(osb, credits);
615         if (IS_ERR(handle)) {
616                 status = PTR_ERR(handle);
617                 handle = NULL;
618                 mlog_errno(status);
619                 goto leave;
620         }
621
622 restarted_transaction:
623         status = dquot_alloc_space_nodirty(inode,
624                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
625         if (status)
626                 goto leave;
627         did_quota = 1;
628
629         /* reserve a write to the file entry early on - that we if we
630          * run out of credits in the allocation path, we can still
631          * update i_size. */
632         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
633                                          OCFS2_JOURNAL_ACCESS_WRITE);
634         if (status < 0) {
635                 mlog_errno(status);
636                 goto leave;
637         }
638
639         prev_clusters = OCFS2_I(inode)->ip_clusters;
640
641         status = ocfs2_add_inode_data(osb,
642                                       inode,
643                                       &logical_start,
644                                       clusters_to_add,
645                                       mark_unwritten,
646                                       bh,
647                                       handle,
648                                       data_ac,
649                                       meta_ac,
650                                       &why);
651         if ((status < 0) && (status != -EAGAIN)) {
652                 if (status != -ENOSPC)
653                         mlog_errno(status);
654                 goto leave;
655         }
656
657         ocfs2_journal_dirty(handle, bh);
658
659         spin_lock(&OCFS2_I(inode)->ip_lock);
660         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
661         spin_unlock(&OCFS2_I(inode)->ip_lock);
662         /* Release unused quota reservation */
663         dquot_free_space(inode,
664                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
665         did_quota = 0;
666
667         if (why != RESTART_NONE && clusters_to_add) {
668                 if (why == RESTART_META) {
669                         mlog(0, "restarting function.\n");
670                         restart_func = 1;
671                         status = 0;
672                 } else {
673                         BUG_ON(why != RESTART_TRANS);
674
675                         mlog(0, "restarting transaction.\n");
676                         /* TODO: This can be more intelligent. */
677                         credits = ocfs2_calc_extend_credits(osb->sb,
678                                                             &fe->id2.i_list,
679                                                             clusters_to_add);
680                         status = ocfs2_extend_trans(handle, credits);
681                         if (status < 0) {
682                                 /* handle still has to be committed at
683                                  * this point. */
684                                 status = -ENOMEM;
685                                 mlog_errno(status);
686                                 goto leave;
687                         }
688                         goto restarted_transaction;
689                 }
690         }
691
692         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
693              le32_to_cpu(fe->i_clusters),
694              (unsigned long long)le64_to_cpu(fe->i_size));
695         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
696              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
697
698 leave:
699         if (status < 0 && did_quota)
700                 dquot_free_space(inode,
701                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
702         if (handle) {
703                 ocfs2_commit_trans(osb, handle);
704                 handle = NULL;
705         }
706         if (data_ac) {
707                 ocfs2_free_alloc_context(data_ac);
708                 data_ac = NULL;
709         }
710         if (meta_ac) {
711                 ocfs2_free_alloc_context(meta_ac);
712                 meta_ac = NULL;
713         }
714         if ((!status) && restart_func) {
715                 restart_func = 0;
716                 goto restart_all;
717         }
718         brelse(bh);
719         bh = NULL;
720
721         mlog_exit(status);
722         return status;
723 }
724
725 /*
726  * While a write will already be ordering the data, a truncate will not.
727  * Thus, we need to explicitly order the zeroed pages.
728  */
729 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
730 {
731         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
732         handle_t *handle = NULL;
733         int ret = 0;
734
735         if (!ocfs2_should_order_data(inode))
736                 goto out;
737
738         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
739         if (IS_ERR(handle)) {
740                 ret = -ENOMEM;
741                 mlog_errno(ret);
742                 goto out;
743         }
744
745         ret = ocfs2_jbd2_file_inode(handle, inode);
746         if (ret < 0)
747                 mlog_errno(ret);
748
749 out:
750         if (ret) {
751                 if (!IS_ERR(handle))
752                         ocfs2_commit_trans(osb, handle);
753                 handle = ERR_PTR(ret);
754         }
755         return handle;
756 }
757
758 /* Some parts of this taken from generic_cont_expand, which turned out
759  * to be too fragile to do exactly what we need without us having to
760  * worry about recursive locking in ->write_begin() and ->write_end(). */
761 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
762                                  u64 abs_to)
763 {
764         struct address_space *mapping = inode->i_mapping;
765         struct page *page;
766         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
767         handle_t *handle = NULL;
768         int ret = 0;
769         unsigned zero_from, zero_to, block_start, block_end;
770
771         BUG_ON(abs_from >= abs_to);
772         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
773         BUG_ON(abs_from & (inode->i_blkbits - 1));
774
775         page = find_or_create_page(mapping, index, GFP_NOFS);
776         if (!page) {
777                 ret = -ENOMEM;
778                 mlog_errno(ret);
779                 goto out;
780         }
781
782         /* Get the offsets within the page that we want to zero */
783         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
784         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
785         if (!zero_to)
786                 zero_to = PAGE_CACHE_SIZE;
787
788         mlog(0,
789              "abs_from = %llu, abs_to = %llu, index = %lu, zero_from = %u, zero_to = %u\n",
790              (unsigned long long)abs_from, (unsigned long long)abs_to,
791              index, zero_from, zero_to);
792
793         /* We know that zero_from is block aligned */
794         for (block_start = zero_from; block_start < zero_to;
795              block_start = block_end) {
796                 block_end = block_start + (1 << inode->i_blkbits);
797
798                 /*
799                  * block_start is block-aligned.  Bump it by one to force
800                  * __block_write_begin and block_commit_write to zero the
801                  * whole block.
802                  */
803                 ret = __block_write_begin(page, block_start + 1, 0,
804                                           ocfs2_get_block);
805                 if (ret < 0) {
806                         mlog_errno(ret);
807                         goto out_unlock;
808                 }
809
810                 if (!handle) {
811                         handle = ocfs2_zero_start_ordered_transaction(inode);
812                         if (IS_ERR(handle)) {
813                                 ret = PTR_ERR(handle);
814                                 handle = NULL;
815                                 break;
816                         }
817                 }
818
819                 /* must not update i_size! */
820                 ret = block_commit_write(page, block_start + 1,
821                                          block_start + 1);
822                 if (ret < 0)
823                         mlog_errno(ret);
824                 else
825                         ret = 0;
826         }
827
828         if (handle)
829                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
830
831 out_unlock:
832         unlock_page(page);
833         page_cache_release(page);
834 out:
835         return ret;
836 }
837
838 /*
839  * Find the next range to zero.  We do this in terms of bytes because
840  * that's what ocfs2_zero_extend() wants, and it is dealing with the
841  * pagecache.  We may return multiple extents.
842  *
843  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
844  * needs to be zeroed.  range_start and range_end return the next zeroing
845  * range.  A subsequent call should pass the previous range_end as its
846  * zero_start.  If range_end is 0, there's nothing to do.
847  *
848  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
849  */
850 static int ocfs2_zero_extend_get_range(struct inode *inode,
851                                        struct buffer_head *di_bh,
852                                        u64 zero_start, u64 zero_end,
853                                        u64 *range_start, u64 *range_end)
854 {
855         int rc = 0, needs_cow = 0;
856         u32 p_cpos, zero_clusters = 0;
857         u32 zero_cpos =
858                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
859         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
860         unsigned int num_clusters = 0;
861         unsigned int ext_flags = 0;
862
863         while (zero_cpos < last_cpos) {
864                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
865                                         &num_clusters, &ext_flags);
866                 if (rc) {
867                         mlog_errno(rc);
868                         goto out;
869                 }
870
871                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
872                         zero_clusters = num_clusters;
873                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
874                                 needs_cow = 1;
875                         break;
876                 }
877
878                 zero_cpos += num_clusters;
879         }
880         if (!zero_clusters) {
881                 *range_end = 0;
882                 goto out;
883         }
884
885         while ((zero_cpos + zero_clusters) < last_cpos) {
886                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
887                                         &p_cpos, &num_clusters,
888                                         &ext_flags);
889                 if (rc) {
890                         mlog_errno(rc);
891                         goto out;
892                 }
893
894                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
895                         break;
896                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
897                         needs_cow = 1;
898                 zero_clusters += num_clusters;
899         }
900         if ((zero_cpos + zero_clusters) > last_cpos)
901                 zero_clusters = last_cpos - zero_cpos;
902
903         if (needs_cow) {
904                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
905                                         zero_clusters, UINT_MAX);
906                 if (rc) {
907                         mlog_errno(rc);
908                         goto out;
909                 }
910         }
911
912         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
913         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
914                                              zero_cpos + zero_clusters);
915
916 out:
917         return rc;
918 }
919
920 /*
921  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
922  * has made sure that the entire range needs zeroing.
923  */
924 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
925                                    u64 range_end)
926 {
927         int rc = 0;
928         u64 next_pos;
929         u64 zero_pos = range_start;
930
931         mlog(0, "range_start = %llu, range_end = %llu\n",
932              (unsigned long long)range_start,
933              (unsigned long long)range_end);
934         BUG_ON(range_start >= range_end);
935
936         while (zero_pos < range_end) {
937                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
938                 if (next_pos > range_end)
939                         next_pos = range_end;
940                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
941                 if (rc < 0) {
942                         mlog_errno(rc);
943                         break;
944                 }
945                 zero_pos = next_pos;
946
947                 /*
948                  * Very large extends have the potential to lock up
949                  * the cpu for extended periods of time.
950                  */
951                 cond_resched();
952         }
953
954         return rc;
955 }
956
957 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
958                       loff_t zero_to_size)
959 {
960         int ret = 0;
961         u64 zero_start, range_start = 0, range_end = 0;
962         struct super_block *sb = inode->i_sb;
963
964         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
965         mlog(0, "zero_start %llu for i_size %llu\n",
966              (unsigned long long)zero_start,
967              (unsigned long long)i_size_read(inode));
968         while (zero_start < zero_to_size) {
969                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
970                                                   zero_to_size,
971                                                   &range_start,
972                                                   &range_end);
973                 if (ret) {
974                         mlog_errno(ret);
975                         break;
976                 }
977                 if (!range_end)
978                         break;
979                 /* Trim the ends */
980                 if (range_start < zero_start)
981                         range_start = zero_start;
982                 if (range_end > zero_to_size)
983                         range_end = zero_to_size;
984
985                 ret = ocfs2_zero_extend_range(inode, range_start,
986                                               range_end);
987                 if (ret) {
988                         mlog_errno(ret);
989                         break;
990                 }
991                 zero_start = range_end;
992         }
993
994         return ret;
995 }
996
997 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
998                           u64 new_i_size, u64 zero_to)
999 {
1000         int ret;
1001         u32 clusters_to_add;
1002         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1003
1004         /*
1005          * Only quota files call this without a bh, and they can't be
1006          * refcounted.
1007          */
1008         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
1009         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1010
1011         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1012         if (clusters_to_add < oi->ip_clusters)
1013                 clusters_to_add = 0;
1014         else
1015                 clusters_to_add -= oi->ip_clusters;
1016
1017         if (clusters_to_add) {
1018                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1019                                                 clusters_to_add, 0);
1020                 if (ret) {
1021                         mlog_errno(ret);
1022                         goto out;
1023                 }
1024         }
1025
1026         /*
1027          * Call this even if we don't add any clusters to the tree. We
1028          * still need to zero the area between the old i_size and the
1029          * new i_size.
1030          */
1031         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1032         if (ret < 0)
1033                 mlog_errno(ret);
1034
1035 out:
1036         return ret;
1037 }
1038
1039 static int ocfs2_extend_file(struct inode *inode,
1040                              struct buffer_head *di_bh,
1041                              u64 new_i_size)
1042 {
1043         int ret = 0;
1044         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1045
1046         BUG_ON(!di_bh);
1047
1048         /* setattr sometimes calls us like this. */
1049         if (new_i_size == 0)
1050                 goto out;
1051
1052         if (i_size_read(inode) == new_i_size)
1053                 goto out;
1054         BUG_ON(new_i_size < i_size_read(inode));
1055
1056         /*
1057          * The alloc sem blocks people in read/write from reading our
1058          * allocation until we're done changing it. We depend on
1059          * i_mutex to block other extend/truncate calls while we're
1060          * here.  We even have to hold it for sparse files because there
1061          * might be some tail zeroing.
1062          */
1063         down_write(&oi->ip_alloc_sem);
1064
1065         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1066                 /*
1067                  * We can optimize small extends by keeping the inodes
1068                  * inline data.
1069                  */
1070                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1071                         up_write(&oi->ip_alloc_sem);
1072                         goto out_update_size;
1073                 }
1074
1075                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1076                 if (ret) {
1077                         up_write(&oi->ip_alloc_sem);
1078                         mlog_errno(ret);
1079                         goto out;
1080                 }
1081         }
1082
1083         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1084                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1085         else
1086                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1087                                             new_i_size);
1088
1089         up_write(&oi->ip_alloc_sem);
1090
1091         if (ret < 0) {
1092                 mlog_errno(ret);
1093                 goto out;
1094         }
1095
1096 out_update_size:
1097         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1098         if (ret < 0)
1099                 mlog_errno(ret);
1100
1101 out:
1102         return ret;
1103 }
1104
1105 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1106 {
1107         int status = 0, size_change;
1108         struct inode *inode = dentry->d_inode;
1109         struct super_block *sb = inode->i_sb;
1110         struct ocfs2_super *osb = OCFS2_SB(sb);
1111         struct buffer_head *bh = NULL;
1112         handle_t *handle = NULL;
1113         struct dquot *transfer_to[MAXQUOTAS] = { };
1114         int qtype;
1115
1116         mlog_entry("(0x%p, '%.*s')\n", dentry,
1117                    dentry->d_name.len, dentry->d_name.name);
1118
1119         /* ensuring we don't even attempt to truncate a symlink */
1120         if (S_ISLNK(inode->i_mode))
1121                 attr->ia_valid &= ~ATTR_SIZE;
1122
1123         if (attr->ia_valid & ATTR_MODE)
1124                 mlog(0, "mode change: %d\n", attr->ia_mode);
1125         if (attr->ia_valid & ATTR_UID)
1126                 mlog(0, "uid change: %d\n", attr->ia_uid);
1127         if (attr->ia_valid & ATTR_GID)
1128                 mlog(0, "gid change: %d\n", attr->ia_gid);
1129         if (attr->ia_valid & ATTR_SIZE)
1130                 mlog(0, "size change...\n");
1131         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1132                 mlog(0, "time change...\n");
1133
1134 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1135                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1136         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1137                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1138                 return 0;
1139         }
1140
1141         status = inode_change_ok(inode, attr);
1142         if (status)
1143                 return status;
1144
1145         if (is_quota_modification(inode, attr))
1146                 dquot_initialize(inode);
1147         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1148         if (size_change) {
1149                 status = ocfs2_rw_lock(inode, 1);
1150                 if (status < 0) {
1151                         mlog_errno(status);
1152                         goto bail;
1153                 }
1154         }
1155
1156         status = ocfs2_inode_lock(inode, &bh, 1);
1157         if (status < 0) {
1158                 if (status != -ENOENT)
1159                         mlog_errno(status);
1160                 goto bail_unlock_rw;
1161         }
1162
1163         if (size_change && attr->ia_size != i_size_read(inode)) {
1164                 status = inode_newsize_ok(inode, attr->ia_size);
1165                 if (status)
1166                         goto bail_unlock;
1167
1168                 if (i_size_read(inode) > attr->ia_size) {
1169                         if (ocfs2_should_order_data(inode)) {
1170                                 status = ocfs2_begin_ordered_truncate(inode,
1171                                                                       attr->ia_size);
1172                                 if (status)
1173                                         goto bail_unlock;
1174                         }
1175                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1176                 } else
1177                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1178                 if (status < 0) {
1179                         if (status != -ENOSPC)
1180                                 mlog_errno(status);
1181                         status = -ENOSPC;
1182                         goto bail_unlock;
1183                 }
1184         }
1185
1186         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1187             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1188                 /*
1189                  * Gather pointers to quota structures so that allocation /
1190                  * freeing of quota structures happens here and not inside
1191                  * dquot_transfer() where we have problems with lock ordering
1192                  */
1193                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1194                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1195                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1196                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1197                                                       USRQUOTA);
1198                         if (!transfer_to[USRQUOTA]) {
1199                                 status = -ESRCH;
1200                                 goto bail_unlock;
1201                         }
1202                 }
1203                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1204                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1205                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1206                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1207                                                       GRPQUOTA);
1208                         if (!transfer_to[GRPQUOTA]) {
1209                                 status = -ESRCH;
1210                                 goto bail_unlock;
1211                         }
1212                 }
1213                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1214                                            2 * ocfs2_quota_trans_credits(sb));
1215                 if (IS_ERR(handle)) {
1216                         status = PTR_ERR(handle);
1217                         mlog_errno(status);
1218                         goto bail_unlock;
1219                 }
1220                 status = __dquot_transfer(inode, transfer_to);
1221                 if (status < 0)
1222                         goto bail_commit;
1223         } else {
1224                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1225                 if (IS_ERR(handle)) {
1226                         status = PTR_ERR(handle);
1227                         mlog_errno(status);
1228                         goto bail_unlock;
1229                 }
1230         }
1231
1232         /*
1233          * This will intentionally not wind up calling truncate_setsize(),
1234          * since all the work for a size change has been done above.
1235          * Otherwise, we could get into problems with truncate as
1236          * ip_alloc_sem is used there to protect against i_size
1237          * changes.
1238          *
1239          * XXX: this means the conditional below can probably be removed.
1240          */
1241         if ((attr->ia_valid & ATTR_SIZE) &&
1242             attr->ia_size != i_size_read(inode)) {
1243                 status = vmtruncate(inode, attr->ia_size);
1244                 if (status) {
1245                         mlog_errno(status);
1246                         goto bail_commit;
1247                 }
1248         }
1249
1250         setattr_copy(inode, attr);
1251         mark_inode_dirty(inode);
1252
1253         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1254         if (status < 0)
1255                 mlog_errno(status);
1256
1257 bail_commit:
1258         ocfs2_commit_trans(osb, handle);
1259 bail_unlock:
1260         ocfs2_inode_unlock(inode, 1);
1261 bail_unlock_rw:
1262         if (size_change)
1263                 ocfs2_rw_unlock(inode, 1);
1264 bail:
1265         brelse(bh);
1266
1267         /* Release quota pointers in case we acquired them */
1268         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1269                 dqput(transfer_to[qtype]);
1270
1271         if (!status && attr->ia_valid & ATTR_MODE) {
1272                 status = ocfs2_acl_chmod(inode);
1273                 if (status < 0)
1274                         mlog_errno(status);
1275         }
1276
1277         mlog_exit(status);
1278         return status;
1279 }
1280
1281 int ocfs2_getattr(struct vfsmount *mnt,
1282                   struct dentry *dentry,
1283                   struct kstat *stat)
1284 {
1285         struct inode *inode = dentry->d_inode;
1286         struct super_block *sb = dentry->d_inode->i_sb;
1287         struct ocfs2_super *osb = sb->s_fs_info;
1288         int err;
1289
1290         mlog_entry_void();
1291
1292         err = ocfs2_inode_revalidate(dentry);
1293         if (err) {
1294                 if (err != -ENOENT)
1295                         mlog_errno(err);
1296                 goto bail;
1297         }
1298
1299         generic_fillattr(inode, stat);
1300
1301         /* We set the blksize from the cluster size for performance */
1302         stat->blksize = osb->s_clustersize;
1303
1304 bail:
1305         mlog_exit(err);
1306
1307         return err;
1308 }
1309
1310 int ocfs2_permission(struct inode *inode, int mask, unsigned int flags)
1311 {
1312         int ret;
1313
1314         if (flags & IPERM_FLAG_RCU)
1315                 return -ECHILD;
1316
1317         mlog_entry_void();
1318
1319         ret = ocfs2_inode_lock(inode, NULL, 0);
1320         if (ret) {
1321                 if (ret != -ENOENT)
1322                         mlog_errno(ret);
1323                 goto out;
1324         }
1325
1326         ret = generic_permission(inode, mask, flags, ocfs2_check_acl);
1327
1328         ocfs2_inode_unlock(inode, 0);
1329 out:
1330         mlog_exit(ret);
1331         return ret;
1332 }
1333
1334 static int __ocfs2_write_remove_suid(struct inode *inode,
1335                                      struct buffer_head *bh)
1336 {
1337         int ret;
1338         handle_t *handle;
1339         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1340         struct ocfs2_dinode *di;
1341
1342         mlog_entry("(Inode %llu, mode 0%o)\n",
1343                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1344
1345         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1346         if (IS_ERR(handle)) {
1347                 ret = PTR_ERR(handle);
1348                 mlog_errno(ret);
1349                 goto out;
1350         }
1351
1352         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1353                                       OCFS2_JOURNAL_ACCESS_WRITE);
1354         if (ret < 0) {
1355                 mlog_errno(ret);
1356                 goto out_trans;
1357         }
1358
1359         inode->i_mode &= ~S_ISUID;
1360         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1361                 inode->i_mode &= ~S_ISGID;
1362
1363         di = (struct ocfs2_dinode *) bh->b_data;
1364         di->i_mode = cpu_to_le16(inode->i_mode);
1365
1366         ocfs2_journal_dirty(handle, bh);
1367
1368 out_trans:
1369         ocfs2_commit_trans(osb, handle);
1370 out:
1371         mlog_exit(ret);
1372         return ret;
1373 }
1374
1375 /*
1376  * Will look for holes and unwritten extents in the range starting at
1377  * pos for count bytes (inclusive).
1378  */
1379 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1380                                        size_t count)
1381 {
1382         int ret = 0;
1383         unsigned int extent_flags;
1384         u32 cpos, clusters, extent_len, phys_cpos;
1385         struct super_block *sb = inode->i_sb;
1386
1387         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1388         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1389
1390         while (clusters) {
1391                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1392                                          &extent_flags);
1393                 if (ret < 0) {
1394                         mlog_errno(ret);
1395                         goto out;
1396                 }
1397
1398                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1399                         ret = 1;
1400                         break;
1401                 }
1402
1403                 if (extent_len > clusters)
1404                         extent_len = clusters;
1405
1406                 clusters -= extent_len;
1407                 cpos += extent_len;
1408         }
1409 out:
1410         return ret;
1411 }
1412
1413 static int ocfs2_write_remove_suid(struct inode *inode)
1414 {
1415         int ret;
1416         struct buffer_head *bh = NULL;
1417
1418         ret = ocfs2_read_inode_block(inode, &bh);
1419         if (ret < 0) {
1420                 mlog_errno(ret);
1421                 goto out;
1422         }
1423
1424         ret =  __ocfs2_write_remove_suid(inode, bh);
1425 out:
1426         brelse(bh);
1427         return ret;
1428 }
1429
1430 /*
1431  * Allocate enough extents to cover the region starting at byte offset
1432  * start for len bytes. Existing extents are skipped, any extents
1433  * added are marked as "unwritten".
1434  */
1435 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1436                                             u64 start, u64 len)
1437 {
1438         int ret;
1439         u32 cpos, phys_cpos, clusters, alloc_size;
1440         u64 end = start + len;
1441         struct buffer_head *di_bh = NULL;
1442
1443         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1444                 ret = ocfs2_read_inode_block(inode, &di_bh);
1445                 if (ret) {
1446                         mlog_errno(ret);
1447                         goto out;
1448                 }
1449
1450                 /*
1451                  * Nothing to do if the requested reservation range
1452                  * fits within the inode.
1453                  */
1454                 if (ocfs2_size_fits_inline_data(di_bh, end))
1455                         goto out;
1456
1457                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1458                 if (ret) {
1459                         mlog_errno(ret);
1460                         goto out;
1461                 }
1462         }
1463
1464         /*
1465          * We consider both start and len to be inclusive.
1466          */
1467         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1468         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1469         clusters -= cpos;
1470
1471         while (clusters) {
1472                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1473                                          &alloc_size, NULL);
1474                 if (ret) {
1475                         mlog_errno(ret);
1476                         goto out;
1477                 }
1478
1479                 /*
1480                  * Hole or existing extent len can be arbitrary, so
1481                  * cap it to our own allocation request.
1482                  */
1483                 if (alloc_size > clusters)
1484                         alloc_size = clusters;
1485
1486                 if (phys_cpos) {
1487                         /*
1488                          * We already have an allocation at this
1489                          * region so we can safely skip it.
1490                          */
1491                         goto next;
1492                 }
1493
1494                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1495                 if (ret) {
1496                         if (ret != -ENOSPC)
1497                                 mlog_errno(ret);
1498                         goto out;
1499                 }
1500
1501 next:
1502                 cpos += alloc_size;
1503                 clusters -= alloc_size;
1504         }
1505
1506         ret = 0;
1507 out:
1508
1509         brelse(di_bh);
1510         return ret;
1511 }
1512
1513 /*
1514  * Truncate a byte range, avoiding pages within partial clusters. This
1515  * preserves those pages for the zeroing code to write to.
1516  */
1517 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1518                                          u64 byte_len)
1519 {
1520         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1521         loff_t start, end;
1522         struct address_space *mapping = inode->i_mapping;
1523
1524         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1525         end = byte_start + byte_len;
1526         end = end & ~(osb->s_clustersize - 1);
1527
1528         if (start < end) {
1529                 unmap_mapping_range(mapping, start, end - start, 0);
1530                 truncate_inode_pages_range(mapping, start, end - 1);
1531         }
1532 }
1533
1534 static int ocfs2_zero_partial_clusters(struct inode *inode,
1535                                        u64 start, u64 len)
1536 {
1537         int ret = 0;
1538         u64 tmpend, end = start + len;
1539         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1540         unsigned int csize = osb->s_clustersize;
1541         handle_t *handle;
1542
1543         /*
1544          * The "start" and "end" values are NOT necessarily part of
1545          * the range whose allocation is being deleted. Rather, this
1546          * is what the user passed in with the request. We must zero
1547          * partial clusters here. There's no need to worry about
1548          * physical allocation - the zeroing code knows to skip holes.
1549          */
1550         mlog(0, "byte start: %llu, end: %llu\n",
1551              (unsigned long long)start, (unsigned long long)end);
1552
1553         /*
1554          * If both edges are on a cluster boundary then there's no
1555          * zeroing required as the region is part of the allocation to
1556          * be truncated.
1557          */
1558         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1559                 goto out;
1560
1561         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1562         if (IS_ERR(handle)) {
1563                 ret = PTR_ERR(handle);
1564                 mlog_errno(ret);
1565                 goto out;
1566         }
1567
1568         /*
1569          * We want to get the byte offset of the end of the 1st cluster.
1570          */
1571         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1572         if (tmpend > end)
1573                 tmpend = end;
1574
1575         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1576              (unsigned long long)start, (unsigned long long)tmpend);
1577
1578         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1579         if (ret)
1580                 mlog_errno(ret);
1581
1582         if (tmpend < end) {
1583                 /*
1584                  * This may make start and end equal, but the zeroing
1585                  * code will skip any work in that case so there's no
1586                  * need to catch it up here.
1587                  */
1588                 start = end & ~(osb->s_clustersize - 1);
1589
1590                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1591                      (unsigned long long)start, (unsigned long long)end);
1592
1593                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1594                 if (ret)
1595                         mlog_errno(ret);
1596         }
1597
1598         ocfs2_commit_trans(osb, handle);
1599 out:
1600         return ret;
1601 }
1602
1603 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1604 {
1605         int i;
1606         struct ocfs2_extent_rec *rec = NULL;
1607
1608         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1609
1610                 rec = &el->l_recs[i];
1611
1612                 if (le32_to_cpu(rec->e_cpos) < pos)
1613                         break;
1614         }
1615
1616         return i;
1617 }
1618
1619 /*
1620  * Helper to calculate the punching pos and length in one run, we handle the
1621  * following three cases in order:
1622  *
1623  * - remove the entire record
1624  * - remove a partial record
1625  * - no record needs to be removed (hole-punching completed)
1626 */
1627 static void ocfs2_calc_trunc_pos(struct inode *inode,
1628                                  struct ocfs2_extent_list *el,
1629                                  struct ocfs2_extent_rec *rec,
1630                                  u32 trunc_start, u32 *trunc_cpos,
1631                                  u32 *trunc_len, u32 *trunc_end,
1632                                  u64 *blkno, int *done)
1633 {
1634         int ret = 0;
1635         u32 coff, range;
1636
1637         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1638
1639         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1640                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1641                 /*
1642                  * Skip holes if any.
1643                  */
1644                 if (range < *trunc_end)
1645                         *trunc_end = range;
1646                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1647                 *blkno = le64_to_cpu(rec->e_blkno);
1648                 *trunc_end = le32_to_cpu(rec->e_cpos);
1649         } else if (range > trunc_start) {
1650                 *trunc_cpos = trunc_start;
1651                 *trunc_len = *trunc_end - trunc_start;
1652                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1653                 *blkno = le64_to_cpu(rec->e_blkno) +
1654                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1655                 *trunc_end = trunc_start;
1656         } else {
1657                 /*
1658                  * It may have two following possibilities:
1659                  *
1660                  * - last record has been removed
1661                  * - trunc_start was within a hole
1662                  *
1663                  * both two cases mean the completion of hole punching.
1664                  */
1665                 ret = 1;
1666         }
1667
1668         *done = ret;
1669 }
1670
1671 static int ocfs2_remove_inode_range(struct inode *inode,
1672                                     struct buffer_head *di_bh, u64 byte_start,
1673                                     u64 byte_len)
1674 {
1675         int ret = 0, flags = 0, done = 0, i;
1676         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1677         u32 cluster_in_el;
1678         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1679         struct ocfs2_cached_dealloc_ctxt dealloc;
1680         struct address_space *mapping = inode->i_mapping;
1681         struct ocfs2_extent_tree et;
1682         struct ocfs2_path *path = NULL;
1683         struct ocfs2_extent_list *el = NULL;
1684         struct ocfs2_extent_rec *rec = NULL;
1685         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1686         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1687
1688         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1689         ocfs2_init_dealloc_ctxt(&dealloc);
1690
1691         if (byte_len == 0)
1692                 return 0;
1693
1694         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1695                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1696                                             byte_start + byte_len, 0);
1697                 if (ret) {
1698                         mlog_errno(ret);
1699                         goto out;
1700                 }
1701                 /*
1702                  * There's no need to get fancy with the page cache
1703                  * truncate of an inline-data inode. We're talking
1704                  * about less than a page here, which will be cached
1705                  * in the dinode buffer anyway.
1706                  */
1707                 unmap_mapping_range(mapping, 0, 0, 0);
1708                 truncate_inode_pages(mapping, 0);
1709                 goto out;
1710         }
1711
1712         /*
1713          * For reflinks, we may need to CoW 2 clusters which might be
1714          * partially zero'd later, if hole's start and end offset were
1715          * within one cluster(means is not exactly aligned to clustersize).
1716          */
1717
1718         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1719
1720                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1721                 if (ret) {
1722                         mlog_errno(ret);
1723                         goto out;
1724                 }
1725
1726                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1727                 if (ret) {
1728                         mlog_errno(ret);
1729                         goto out;
1730                 }
1731         }
1732
1733         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1734         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1735         cluster_in_el = trunc_end;
1736
1737         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1738              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1739              (unsigned long long)byte_start,
1740              (unsigned long long)byte_len, trunc_start, trunc_end);
1741
1742         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1743         if (ret) {
1744                 mlog_errno(ret);
1745                 goto out;
1746         }
1747
1748         path = ocfs2_new_path_from_et(&et);
1749         if (!path) {
1750                 ret = -ENOMEM;
1751                 mlog_errno(ret);
1752                 goto out;
1753         }
1754
1755         while (trunc_end > trunc_start) {
1756
1757                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1758                                       cluster_in_el);
1759                 if (ret) {
1760                         mlog_errno(ret);
1761                         goto out;
1762                 }
1763
1764                 el = path_leaf_el(path);
1765
1766                 i = ocfs2_find_rec(el, trunc_end);
1767                 /*
1768                  * Need to go to previous extent block.
1769                  */
1770                 if (i < 0) {
1771                         if (path->p_tree_depth == 0)
1772                                 break;
1773
1774                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1775                                                             path,
1776                                                             &cluster_in_el);
1777                         if (ret) {
1778                                 mlog_errno(ret);
1779                                 goto out;
1780                         }
1781
1782                         /*
1783                          * We've reached the leftmost extent block,
1784                          * it's safe to leave.
1785                          */
1786                         if (cluster_in_el == 0)
1787                                 break;
1788
1789                         /*
1790                          * The 'pos' searched for previous extent block is
1791                          * always one cluster less than actual trunc_end.
1792                          */
1793                         trunc_end = cluster_in_el + 1;
1794
1795                         ocfs2_reinit_path(path, 1);
1796
1797                         continue;
1798
1799                 } else
1800                         rec = &el->l_recs[i];
1801
1802                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1803                                      &trunc_len, &trunc_end, &blkno, &done);
1804                 if (done)
1805                         break;
1806
1807                 flags = rec->e_flags;
1808                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1809
1810                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1811                                                phys_cpos, trunc_len, flags,
1812                                                &dealloc, refcount_loc);
1813                 if (ret < 0) {
1814                         mlog_errno(ret);
1815                         goto out;
1816                 }
1817
1818                 cluster_in_el = trunc_end;
1819
1820                 ocfs2_reinit_path(path, 1);
1821         }
1822
1823         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1824
1825 out:
1826         ocfs2_schedule_truncate_log_flush(osb, 1);
1827         ocfs2_run_deallocs(osb, &dealloc);
1828
1829         return ret;
1830 }
1831
1832 /*
1833  * Parts of this function taken from xfs_change_file_space()
1834  */
1835 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1836                                      loff_t f_pos, unsigned int cmd,
1837                                      struct ocfs2_space_resv *sr,
1838                                      int change_size)
1839 {
1840         int ret;
1841         s64 llen;
1842         loff_t size;
1843         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1844         struct buffer_head *di_bh = NULL;
1845         handle_t *handle;
1846         unsigned long long max_off = inode->i_sb->s_maxbytes;
1847
1848         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1849                 return -EROFS;
1850
1851         mutex_lock(&inode->i_mutex);
1852
1853         /*
1854          * This prevents concurrent writes on other nodes
1855          */
1856         ret = ocfs2_rw_lock(inode, 1);
1857         if (ret) {
1858                 mlog_errno(ret);
1859                 goto out;
1860         }
1861
1862         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1863         if (ret) {
1864                 mlog_errno(ret);
1865                 goto out_rw_unlock;
1866         }
1867
1868         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1869                 ret = -EPERM;
1870                 goto out_inode_unlock;
1871         }
1872
1873         switch (sr->l_whence) {
1874         case 0: /*SEEK_SET*/
1875                 break;
1876         case 1: /*SEEK_CUR*/
1877                 sr->l_start += f_pos;
1878                 break;
1879         case 2: /*SEEK_END*/
1880                 sr->l_start += i_size_read(inode);
1881                 break;
1882         default:
1883                 ret = -EINVAL;
1884                 goto out_inode_unlock;
1885         }
1886         sr->l_whence = 0;
1887
1888         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1889
1890         if (sr->l_start < 0
1891             || sr->l_start > max_off
1892             || (sr->l_start + llen) < 0
1893             || (sr->l_start + llen) > max_off) {
1894                 ret = -EINVAL;
1895                 goto out_inode_unlock;
1896         }
1897         size = sr->l_start + sr->l_len;
1898
1899         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1900                 if (sr->l_len <= 0) {
1901                         ret = -EINVAL;
1902                         goto out_inode_unlock;
1903                 }
1904         }
1905
1906         if (file && should_remove_suid(file->f_path.dentry)) {
1907                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1908                 if (ret) {
1909                         mlog_errno(ret);
1910                         goto out_inode_unlock;
1911                 }
1912         }
1913
1914         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1915         switch (cmd) {
1916         case OCFS2_IOC_RESVSP:
1917         case OCFS2_IOC_RESVSP64:
1918                 /*
1919                  * This takes unsigned offsets, but the signed ones we
1920                  * pass have been checked against overflow above.
1921                  */
1922                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1923                                                        sr->l_len);
1924                 break;
1925         case OCFS2_IOC_UNRESVSP:
1926         case OCFS2_IOC_UNRESVSP64:
1927                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1928                                                sr->l_len);
1929                 break;
1930         default:
1931                 ret = -EINVAL;
1932         }
1933         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1934         if (ret) {
1935                 mlog_errno(ret);
1936                 goto out_inode_unlock;
1937         }
1938
1939         /*
1940          * We update c/mtime for these changes
1941          */
1942         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1943         if (IS_ERR(handle)) {
1944                 ret = PTR_ERR(handle);
1945                 mlog_errno(ret);
1946                 goto out_inode_unlock;
1947         }
1948
1949         if (change_size && i_size_read(inode) < size)
1950                 i_size_write(inode, size);
1951
1952         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1953         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1954         if (ret < 0)
1955                 mlog_errno(ret);
1956
1957         ocfs2_commit_trans(osb, handle);
1958
1959 out_inode_unlock:
1960         brelse(di_bh);
1961         ocfs2_inode_unlock(inode, 1);
1962 out_rw_unlock:
1963         ocfs2_rw_unlock(inode, 1);
1964
1965 out:
1966         mutex_unlock(&inode->i_mutex);
1967         return ret;
1968 }
1969
1970 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1971                             struct ocfs2_space_resv *sr)
1972 {
1973         struct inode *inode = file->f_path.dentry->d_inode;
1974         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1975
1976         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1977             !ocfs2_writes_unwritten_extents(osb))
1978                 return -ENOTTY;
1979         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1980                  !ocfs2_sparse_alloc(osb))
1981                 return -ENOTTY;
1982
1983         if (!S_ISREG(inode->i_mode))
1984                 return -EINVAL;
1985
1986         if (!(file->f_mode & FMODE_WRITE))
1987                 return -EBADF;
1988
1989         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1990 }
1991
1992 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1993                             loff_t len)
1994 {
1995         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1996         struct ocfs2_space_resv sr;
1997         int change_size = 1;
1998
1999         if (!ocfs2_writes_unwritten_extents(osb))
2000                 return -EOPNOTSUPP;
2001
2002         if (S_ISDIR(inode->i_mode))
2003                 return -ENODEV;
2004
2005         if (mode & FALLOC_FL_KEEP_SIZE)
2006                 change_size = 0;
2007
2008         sr.l_whence = 0;
2009         sr.l_start = (s64)offset;
2010         sr.l_len = (s64)len;
2011
2012         return __ocfs2_change_file_space(NULL, inode, offset,
2013                                          OCFS2_IOC_RESVSP64, &sr, change_size);
2014 }
2015
2016 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2017                                    size_t count)
2018 {
2019         int ret = 0;
2020         unsigned int extent_flags;
2021         u32 cpos, clusters, extent_len, phys_cpos;
2022         struct super_block *sb = inode->i_sb;
2023
2024         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2025             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2026             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2027                 return 0;
2028
2029         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2030         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2031
2032         while (clusters) {
2033                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2034                                          &extent_flags);
2035                 if (ret < 0) {
2036                         mlog_errno(ret);
2037                         goto out;
2038                 }
2039
2040                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2041                         ret = 1;
2042                         break;
2043                 }
2044
2045                 if (extent_len > clusters)
2046                         extent_len = clusters;
2047
2048                 clusters -= extent_len;
2049                 cpos += extent_len;
2050         }
2051 out:
2052         return ret;
2053 }
2054
2055 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2056                                             struct file *file,
2057                                             loff_t pos, size_t count,
2058                                             int *meta_level)
2059 {
2060         int ret;
2061         struct buffer_head *di_bh = NULL;
2062         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2063         u32 clusters =
2064                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2065
2066         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2067         if (ret) {
2068                 mlog_errno(ret);
2069                 goto out;
2070         }
2071
2072         *meta_level = 1;
2073
2074         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2075         if (ret)
2076                 mlog_errno(ret);
2077 out:
2078         brelse(di_bh);
2079         return ret;
2080 }
2081
2082 static int ocfs2_prepare_inode_for_write(struct file *file,
2083                                          loff_t *ppos,
2084                                          size_t count,
2085                                          int appending,
2086                                          int *direct_io,
2087                                          int *has_refcount)
2088 {
2089         int ret = 0, meta_level = 0;
2090         struct dentry *dentry = file->f_path.dentry;
2091         struct inode *inode = dentry->d_inode;
2092         loff_t saved_pos, end;
2093
2094         /*
2095          * We start with a read level meta lock and only jump to an ex
2096          * if we need to make modifications here.
2097          */
2098         for(;;) {
2099                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2100                 if (ret < 0) {
2101                         meta_level = -1;
2102                         mlog_errno(ret);
2103                         goto out;
2104                 }
2105
2106                 /* Clear suid / sgid if necessary. We do this here
2107                  * instead of later in the write path because
2108                  * remove_suid() calls ->setattr without any hint that
2109                  * we may have already done our cluster locking. Since
2110                  * ocfs2_setattr() *must* take cluster locks to
2111                  * proceeed, this will lead us to recursively lock the
2112                  * inode. There's also the dinode i_size state which
2113                  * can be lost via setattr during extending writes (we
2114                  * set inode->i_size at the end of a write. */
2115                 if (should_remove_suid(dentry)) {
2116                         if (meta_level == 0) {
2117                                 ocfs2_inode_unlock(inode, meta_level);
2118                                 meta_level = 1;
2119                                 continue;
2120                         }
2121
2122                         ret = ocfs2_write_remove_suid(inode);
2123                         if (ret < 0) {
2124                                 mlog_errno(ret);
2125                                 goto out_unlock;
2126                         }
2127                 }
2128
2129                 /* work on a copy of ppos until we're sure that we won't have
2130                  * to recalculate it due to relocking. */
2131                 if (appending) {
2132                         saved_pos = i_size_read(inode);
2133                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2134                 } else {
2135                         saved_pos = *ppos;
2136                 }
2137
2138                 end = saved_pos + count;
2139
2140                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2141                 if (ret == 1) {
2142                         ocfs2_inode_unlock(inode, meta_level);
2143                         meta_level = -1;
2144
2145                         ret = ocfs2_prepare_inode_for_refcount(inode,
2146                                                                file,
2147                                                                saved_pos,
2148                                                                count,
2149                                                                &meta_level);
2150                         if (has_refcount)
2151                                 *has_refcount = 1;
2152                         if (direct_io)
2153                                 *direct_io = 0;
2154                 }
2155
2156                 if (ret < 0) {
2157                         mlog_errno(ret);
2158                         goto out_unlock;
2159                 }
2160
2161                 /*
2162                  * Skip the O_DIRECT checks if we don't need
2163                  * them.
2164                  */
2165                 if (!direct_io || !(*direct_io))
2166                         break;
2167
2168                 /*
2169                  * There's no sane way to do direct writes to an inode
2170                  * with inline data.
2171                  */
2172                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2173                         *direct_io = 0;
2174                         break;
2175                 }
2176
2177                 /*
2178                  * Allowing concurrent direct writes means
2179                  * i_size changes wouldn't be synchronized, so
2180                  * one node could wind up truncating another
2181                  * nodes writes.
2182                  */
2183                 if (end > i_size_read(inode)) {
2184                         *direct_io = 0;
2185                         break;
2186                 }
2187
2188                 /*
2189                  * We don't fill holes during direct io, so
2190                  * check for them here. If any are found, the
2191                  * caller will have to retake some cluster
2192                  * locks and initiate the io as buffered.
2193                  */
2194                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2195                 if (ret == 1) {
2196                         *direct_io = 0;
2197                         ret = 0;
2198                 } else if (ret < 0)
2199                         mlog_errno(ret);
2200                 break;
2201         }
2202
2203         if (appending)
2204                 *ppos = saved_pos;
2205
2206 out_unlock:
2207         if (meta_level >= 0)
2208                 ocfs2_inode_unlock(inode, meta_level);
2209
2210 out:
2211         return ret;
2212 }
2213
2214 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2215                                     const struct iovec *iov,
2216                                     unsigned long nr_segs,
2217                                     loff_t pos)
2218 {
2219         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2220         int can_do_direct, has_refcount = 0;
2221         ssize_t written = 0;
2222         size_t ocount;          /* original count */
2223         size_t count;           /* after file limit checks */
2224         loff_t old_size, *ppos = &iocb->ki_pos;
2225         u32 old_clusters;
2226         struct file *file = iocb->ki_filp;
2227         struct inode *inode = file->f_path.dentry->d_inode;
2228         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2229         int full_coherency = !(osb->s_mount_opt &
2230                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2231
2232         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2233                    (unsigned int)nr_segs,
2234                    file->f_path.dentry->d_name.len,
2235                    file->f_path.dentry->d_name.name);
2236
2237         if (iocb->ki_left == 0)
2238                 return 0;
2239
2240         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2241
2242         appending = file->f_flags & O_APPEND ? 1 : 0;
2243         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2244
2245         mutex_lock(&inode->i_mutex);
2246
2247         ocfs2_iocb_clear_sem_locked(iocb);
2248
2249 relock:
2250         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2251         if (direct_io) {
2252                 down_read(&inode->i_alloc_sem);
2253                 have_alloc_sem = 1;
2254                 /* communicate with ocfs2_dio_end_io */
2255                 ocfs2_iocb_set_sem_locked(iocb);
2256         }
2257
2258         /*
2259          * Concurrent O_DIRECT writes are allowed with
2260          * mount_option "coherency=buffered".
2261          */
2262         rw_level = (!direct_io || full_coherency);
2263
2264         ret = ocfs2_rw_lock(inode, rw_level);
2265         if (ret < 0) {
2266                 mlog_errno(ret);
2267                 goto out_sems;
2268         }
2269
2270         /*
2271          * O_DIRECT writes with "coherency=full" need to take EX cluster
2272          * inode_lock to guarantee coherency.
2273          */
2274         if (direct_io && full_coherency) {
2275                 /*
2276                  * We need to take and drop the inode lock to force
2277                  * other nodes to drop their caches.  Buffered I/O
2278                  * already does this in write_begin().
2279                  */
2280                 ret = ocfs2_inode_lock(inode, NULL, 1);
2281                 if (ret < 0) {
2282                         mlog_errno(ret);
2283                         goto out_sems;
2284                 }
2285
2286                 ocfs2_inode_unlock(inode, 1);
2287         }
2288
2289         can_do_direct = direct_io;
2290         ret = ocfs2_prepare_inode_for_write(file, ppos,
2291                                             iocb->ki_left, appending,
2292                                             &can_do_direct, &has_refcount);
2293         if (ret < 0) {
2294                 mlog_errno(ret);
2295                 goto out;
2296         }
2297
2298         /*
2299          * We can't complete the direct I/O as requested, fall back to
2300          * buffered I/O.
2301          */
2302         if (direct_io && !can_do_direct) {
2303                 ocfs2_rw_unlock(inode, rw_level);
2304                 up_read(&inode->i_alloc_sem);
2305
2306                 have_alloc_sem = 0;
2307                 rw_level = -1;
2308
2309                 direct_io = 0;
2310                 goto relock;
2311         }
2312
2313         /*
2314          * To later detect whether a journal commit for sync writes is
2315          * necessary, we sample i_size, and cluster count here.
2316          */
2317         old_size = i_size_read(inode);
2318         old_clusters = OCFS2_I(inode)->ip_clusters;
2319
2320         /* communicate with ocfs2_dio_end_io */
2321         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2322
2323         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2324                                      VERIFY_READ);
2325         if (ret)
2326                 goto out_dio;
2327
2328         count = ocount;
2329         ret = generic_write_checks(file, ppos, &count,
2330                                    S_ISBLK(inode->i_mode));
2331         if (ret)
2332                 goto out_dio;
2333
2334         if (direct_io) {
2335                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2336                                                     ppos, count, ocount);
2337                 if (written < 0) {
2338                         ret = written;
2339                         goto out_dio;
2340                 }
2341         } else {
2342                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2343                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2344                                                       ppos, count, 0);
2345                 current->backing_dev_info = NULL;
2346         }
2347
2348 out_dio:
2349         /* buffered aio wouldn't have proper lock coverage today */
2350         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2351
2352         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2353             ((file->f_flags & O_DIRECT) && !direct_io)) {
2354                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2355                                                pos + count - 1);
2356                 if (ret < 0)
2357                         written = ret;
2358
2359                 if (!ret && ((old_size != i_size_read(inode)) ||
2360                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2361                              has_refcount)) {
2362                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2363                         if (ret < 0)
2364                                 written = ret;
2365                 }
2366
2367                 if (!ret)
2368                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2369                                                       pos + count - 1);
2370         }
2371
2372         /*
2373          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2374          * function pointer which is called when o_direct io completes so that
2375          * it can unlock our rw lock.  (it's the clustered equivalent of
2376          * i_alloc_sem; protects truncate from racing with pending ios).
2377          * Unfortunately there are error cases which call end_io and others
2378          * that don't.  so we don't have to unlock the rw_lock if either an
2379          * async dio is going to do it in the future or an end_io after an
2380          * error has already done it.
2381          */
2382         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2383                 rw_level = -1;
2384                 have_alloc_sem = 0;
2385         }
2386
2387 out:
2388         if (rw_level != -1)
2389                 ocfs2_rw_unlock(inode, rw_level);
2390
2391 out_sems:
2392         if (have_alloc_sem) {
2393                 up_read(&inode->i_alloc_sem);
2394                 ocfs2_iocb_clear_sem_locked(iocb);
2395         }
2396
2397         mutex_unlock(&inode->i_mutex);
2398
2399         if (written)
2400                 ret = written;
2401         mlog_exit(ret);
2402         return ret;
2403 }
2404
2405 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2406                                 struct file *out,
2407                                 struct splice_desc *sd)
2408 {
2409         int ret;
2410
2411         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2412                                             sd->total_len, 0, NULL, NULL);
2413         if (ret < 0) {
2414                 mlog_errno(ret);
2415                 return ret;
2416         }
2417
2418         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2419 }
2420
2421 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2422                                        struct file *out,
2423                                        loff_t *ppos,
2424                                        size_t len,
2425                                        unsigned int flags)
2426 {
2427         int ret;
2428         struct address_space *mapping = out->f_mapping;
2429         struct inode *inode = mapping->host;
2430         struct splice_desc sd = {
2431                 .total_len = len,
2432                 .flags = flags,
2433                 .pos = *ppos,
2434                 .u.file = out,
2435         };
2436
2437         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2438                    (unsigned int)len,
2439                    out->f_path.dentry->d_name.len,
2440                    out->f_path.dentry->d_name.name);
2441
2442         if (pipe->inode)
2443                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2444
2445         splice_from_pipe_begin(&sd);
2446         do {
2447                 ret = splice_from_pipe_next(pipe, &sd);
2448                 if (ret <= 0)
2449                         break;
2450
2451                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2452                 ret = ocfs2_rw_lock(inode, 1);
2453                 if (ret < 0)
2454                         mlog_errno(ret);
2455                 else {
2456                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2457                         ocfs2_rw_unlock(inode, 1);
2458                 }
2459                 mutex_unlock(&inode->i_mutex);
2460         } while (ret > 0);
2461         splice_from_pipe_end(pipe, &sd);
2462
2463         if (pipe->inode)
2464                 mutex_unlock(&pipe->inode->i_mutex);
2465
2466         if (sd.num_spliced)
2467                 ret = sd.num_spliced;
2468
2469         if (ret > 0) {
2470                 unsigned long nr_pages;
2471                 int err;
2472
2473                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2474
2475                 err = generic_write_sync(out, *ppos, ret);
2476                 if (err)
2477                         ret = err;
2478                 else
2479                         *ppos += ret;
2480
2481                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2482         }
2483
2484         mlog_exit(ret);
2485         return ret;
2486 }
2487
2488 static ssize_t ocfs2_file_splice_read(struct file *in,
2489                                       loff_t *ppos,
2490                                       struct pipe_inode_info *pipe,
2491                                       size_t len,
2492                                       unsigned int flags)
2493 {
2494         int ret = 0, lock_level = 0;
2495         struct inode *inode = in->f_path.dentry->d_inode;
2496
2497         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2498                    (unsigned int)len,
2499                    in->f_path.dentry->d_name.len,
2500                    in->f_path.dentry->d_name.name);
2501
2502         /*
2503          * See the comment in ocfs2_file_aio_read()
2504          */
2505         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2506         if (ret < 0) {
2507                 mlog_errno(ret);
2508                 goto bail;
2509         }
2510         ocfs2_inode_unlock(inode, lock_level);
2511
2512         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2513
2514 bail:
2515         mlog_exit(ret);
2516         return ret;
2517 }
2518
2519 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2520                                    const struct iovec *iov,
2521                                    unsigned long nr_segs,
2522                                    loff_t pos)
2523 {
2524         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2525         struct file *filp = iocb->ki_filp;
2526         struct inode *inode = filp->f_path.dentry->d_inode;
2527
2528         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2529                    (unsigned int)nr_segs,
2530                    filp->f_path.dentry->d_name.len,
2531                    filp->f_path.dentry->d_name.name);
2532
2533         if (!inode) {
2534                 ret = -EINVAL;
2535                 mlog_errno(ret);
2536                 goto bail;
2537         }
2538
2539         ocfs2_iocb_clear_sem_locked(iocb);
2540
2541         /*
2542          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2543          * need locks to protect pending reads from racing with truncate.
2544          */
2545         if (filp->f_flags & O_DIRECT) {
2546                 down_read(&inode->i_alloc_sem);
2547                 have_alloc_sem = 1;
2548                 ocfs2_iocb_set_sem_locked(iocb);
2549
2550                 ret = ocfs2_rw_lock(inode, 0);
2551                 if (ret < 0) {
2552                         mlog_errno(ret);
2553                         goto bail;
2554                 }
2555                 rw_level = 0;
2556                 /* communicate with ocfs2_dio_end_io */
2557                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2558         }
2559
2560         /*
2561          * We're fine letting folks race truncates and extending
2562          * writes with read across the cluster, just like they can
2563          * locally. Hence no rw_lock during read.
2564          *
2565          * Take and drop the meta data lock to update inode fields
2566          * like i_size. This allows the checks down below
2567          * generic_file_aio_read() a chance of actually working.
2568          */
2569         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2570         if (ret < 0) {
2571                 mlog_errno(ret);
2572                 goto bail;
2573         }
2574         ocfs2_inode_unlock(inode, lock_level);
2575
2576         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2577         if (ret == -EINVAL)
2578                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2579
2580         /* buffered aio wouldn't have proper lock coverage today */
2581         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2582
2583         /* see ocfs2_file_aio_write */
2584         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2585                 rw_level = -1;
2586                 have_alloc_sem = 0;
2587         }
2588
2589 bail:
2590         if (have_alloc_sem) {
2591                 up_read(&inode->i_alloc_sem);
2592                 ocfs2_iocb_clear_sem_locked(iocb);
2593         }
2594         if (rw_level != -1)
2595                 ocfs2_rw_unlock(inode, rw_level);
2596         mlog_exit(ret);
2597
2598         return ret;
2599 }
2600
2601 const struct inode_operations ocfs2_file_iops = {
2602         .setattr        = ocfs2_setattr,
2603         .getattr        = ocfs2_getattr,
2604         .permission     = ocfs2_permission,
2605         .setxattr       = generic_setxattr,
2606         .getxattr       = generic_getxattr,
2607         .listxattr      = ocfs2_listxattr,
2608         .removexattr    = generic_removexattr,
2609         .fallocate      = ocfs2_fallocate,
2610         .fiemap         = ocfs2_fiemap,
2611 };
2612
2613 const struct inode_operations ocfs2_special_file_iops = {
2614         .setattr        = ocfs2_setattr,
2615         .getattr        = ocfs2_getattr,
2616         .permission     = ocfs2_permission,
2617 };
2618
2619 /*
2620  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2621  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2622  */
2623 const struct file_operations ocfs2_fops = {
2624         .llseek         = generic_file_llseek,
2625         .read           = do_sync_read,
2626         .write          = do_sync_write,
2627         .mmap           = ocfs2_mmap,
2628         .fsync          = ocfs2_sync_file,
2629         .release        = ocfs2_file_release,
2630         .open           = ocfs2_file_open,
2631         .aio_read       = ocfs2_file_aio_read,
2632         .aio_write      = ocfs2_file_aio_write,
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         .splice_read    = ocfs2_file_splice_read,
2640         .splice_write   = ocfs2_file_splice_write,
2641 };
2642
2643 const struct file_operations ocfs2_dops = {
2644         .llseek         = generic_file_llseek,
2645         .read           = generic_read_dir,
2646         .readdir        = ocfs2_readdir,
2647         .fsync          = ocfs2_sync_file,
2648         .release        = ocfs2_dir_release,
2649         .open           = ocfs2_dir_open,
2650         .unlocked_ioctl = ocfs2_ioctl,
2651 #ifdef CONFIG_COMPAT
2652         .compat_ioctl   = ocfs2_compat_ioctl,
2653 #endif
2654         .lock           = ocfs2_lock,
2655         .flock          = ocfs2_flock,
2656 };
2657
2658 /*
2659  * POSIX-lockless variants of our file_operations.
2660  *
2661  * These will be used if the underlying cluster stack does not support
2662  * posix file locking, if the user passes the "localflocks" mount
2663  * option, or if we have a local-only fs.
2664  *
2665  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2666  * so we still want it in the case of no stack support for
2667  * plocks. Internally, it will do the right thing when asked to ignore
2668  * the cluster.
2669  */
2670 const struct file_operations ocfs2_fops_no_plocks = {
2671         .llseek         = generic_file_llseek,
2672         .read           = do_sync_read,
2673         .write          = do_sync_write,
2674         .mmap           = ocfs2_mmap,
2675         .fsync          = ocfs2_sync_file,
2676         .release        = ocfs2_file_release,
2677         .open           = ocfs2_file_open,
2678         .aio_read       = ocfs2_file_aio_read,
2679         .aio_write      = ocfs2_file_aio_write,
2680         .unlocked_ioctl = ocfs2_ioctl,
2681 #ifdef CONFIG_COMPAT
2682         .compat_ioctl   = ocfs2_compat_ioctl,
2683 #endif
2684         .flock          = ocfs2_flock,
2685         .splice_read    = ocfs2_file_splice_read,
2686         .splice_write   = ocfs2_file_splice_write,
2687 };
2688
2689 const struct file_operations ocfs2_dops_no_plocks = {
2690         .llseek         = generic_file_llseek,
2691         .read           = generic_read_dir,
2692         .readdir        = ocfs2_readdir,
2693         .fsync          = ocfs2_sync_file,
2694         .release        = ocfs2_dir_release,
2695         .open           = ocfs2_dir_open,
2696         .unlocked_ioctl = ocfs2_ioctl,
2697 #ifdef CONFIG_COMPAT
2698         .compat_ioctl   = ocfs2_compat_ioctl,
2699 #endif
2700         .flock          = ocfs2_flock,
2701 };