fs: kill block_prepare_write
[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)
1311 {
1312         int ret;
1313
1314         mlog_entry_void();
1315
1316         ret = ocfs2_inode_lock(inode, NULL, 0);
1317         if (ret) {
1318                 if (ret != -ENOENT)
1319                         mlog_errno(ret);
1320                 goto out;
1321         }
1322
1323         ret = generic_permission(inode, mask, ocfs2_check_acl);
1324
1325         ocfs2_inode_unlock(inode, 0);
1326 out:
1327         mlog_exit(ret);
1328         return ret;
1329 }
1330
1331 static int __ocfs2_write_remove_suid(struct inode *inode,
1332                                      struct buffer_head *bh)
1333 {
1334         int ret;
1335         handle_t *handle;
1336         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1337         struct ocfs2_dinode *di;
1338
1339         mlog_entry("(Inode %llu, mode 0%o)\n",
1340                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1341
1342         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1343         if (IS_ERR(handle)) {
1344                 ret = PTR_ERR(handle);
1345                 mlog_errno(ret);
1346                 goto out;
1347         }
1348
1349         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1350                                       OCFS2_JOURNAL_ACCESS_WRITE);
1351         if (ret < 0) {
1352                 mlog_errno(ret);
1353                 goto out_trans;
1354         }
1355
1356         inode->i_mode &= ~S_ISUID;
1357         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1358                 inode->i_mode &= ~S_ISGID;
1359
1360         di = (struct ocfs2_dinode *) bh->b_data;
1361         di->i_mode = cpu_to_le16(inode->i_mode);
1362
1363         ocfs2_journal_dirty(handle, bh);
1364
1365 out_trans:
1366         ocfs2_commit_trans(osb, handle);
1367 out:
1368         mlog_exit(ret);
1369         return ret;
1370 }
1371
1372 /*
1373  * Will look for holes and unwritten extents in the range starting at
1374  * pos for count bytes (inclusive).
1375  */
1376 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1377                                        size_t count)
1378 {
1379         int ret = 0;
1380         unsigned int extent_flags;
1381         u32 cpos, clusters, extent_len, phys_cpos;
1382         struct super_block *sb = inode->i_sb;
1383
1384         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1385         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1386
1387         while (clusters) {
1388                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1389                                          &extent_flags);
1390                 if (ret < 0) {
1391                         mlog_errno(ret);
1392                         goto out;
1393                 }
1394
1395                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1396                         ret = 1;
1397                         break;
1398                 }
1399
1400                 if (extent_len > clusters)
1401                         extent_len = clusters;
1402
1403                 clusters -= extent_len;
1404                 cpos += extent_len;
1405         }
1406 out:
1407         return ret;
1408 }
1409
1410 static int ocfs2_write_remove_suid(struct inode *inode)
1411 {
1412         int ret;
1413         struct buffer_head *bh = NULL;
1414
1415         ret = ocfs2_read_inode_block(inode, &bh);
1416         if (ret < 0) {
1417                 mlog_errno(ret);
1418                 goto out;
1419         }
1420
1421         ret =  __ocfs2_write_remove_suid(inode, bh);
1422 out:
1423         brelse(bh);
1424         return ret;
1425 }
1426
1427 /*
1428  * Allocate enough extents to cover the region starting at byte offset
1429  * start for len bytes. Existing extents are skipped, any extents
1430  * added are marked as "unwritten".
1431  */
1432 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1433                                             u64 start, u64 len)
1434 {
1435         int ret;
1436         u32 cpos, phys_cpos, clusters, alloc_size;
1437         u64 end = start + len;
1438         struct buffer_head *di_bh = NULL;
1439
1440         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1441                 ret = ocfs2_read_inode_block(inode, &di_bh);
1442                 if (ret) {
1443                         mlog_errno(ret);
1444                         goto out;
1445                 }
1446
1447                 /*
1448                  * Nothing to do if the requested reservation range
1449                  * fits within the inode.
1450                  */
1451                 if (ocfs2_size_fits_inline_data(di_bh, end))
1452                         goto out;
1453
1454                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1455                 if (ret) {
1456                         mlog_errno(ret);
1457                         goto out;
1458                 }
1459         }
1460
1461         /*
1462          * We consider both start and len to be inclusive.
1463          */
1464         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1465         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1466         clusters -= cpos;
1467
1468         while (clusters) {
1469                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1470                                          &alloc_size, NULL);
1471                 if (ret) {
1472                         mlog_errno(ret);
1473                         goto out;
1474                 }
1475
1476                 /*
1477                  * Hole or existing extent len can be arbitrary, so
1478                  * cap it to our own allocation request.
1479                  */
1480                 if (alloc_size > clusters)
1481                         alloc_size = clusters;
1482
1483                 if (phys_cpos) {
1484                         /*
1485                          * We already have an allocation at this
1486                          * region so we can safely skip it.
1487                          */
1488                         goto next;
1489                 }
1490
1491                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1492                 if (ret) {
1493                         if (ret != -ENOSPC)
1494                                 mlog_errno(ret);
1495                         goto out;
1496                 }
1497
1498 next:
1499                 cpos += alloc_size;
1500                 clusters -= alloc_size;
1501         }
1502
1503         ret = 0;
1504 out:
1505
1506         brelse(di_bh);
1507         return ret;
1508 }
1509
1510 /*
1511  * Truncate a byte range, avoiding pages within partial clusters. This
1512  * preserves those pages for the zeroing code to write to.
1513  */
1514 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1515                                          u64 byte_len)
1516 {
1517         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1518         loff_t start, end;
1519         struct address_space *mapping = inode->i_mapping;
1520
1521         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1522         end = byte_start + byte_len;
1523         end = end & ~(osb->s_clustersize - 1);
1524
1525         if (start < end) {
1526                 unmap_mapping_range(mapping, start, end - start, 0);
1527                 truncate_inode_pages_range(mapping, start, end - 1);
1528         }
1529 }
1530
1531 static int ocfs2_zero_partial_clusters(struct inode *inode,
1532                                        u64 start, u64 len)
1533 {
1534         int ret = 0;
1535         u64 tmpend, end = start + len;
1536         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1537         unsigned int csize = osb->s_clustersize;
1538         handle_t *handle;
1539
1540         /*
1541          * The "start" and "end" values are NOT necessarily part of
1542          * the range whose allocation is being deleted. Rather, this
1543          * is what the user passed in with the request. We must zero
1544          * partial clusters here. There's no need to worry about
1545          * physical allocation - the zeroing code knows to skip holes.
1546          */
1547         mlog(0, "byte start: %llu, end: %llu\n",
1548              (unsigned long long)start, (unsigned long long)end);
1549
1550         /*
1551          * If both edges are on a cluster boundary then there's no
1552          * zeroing required as the region is part of the allocation to
1553          * be truncated.
1554          */
1555         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1556                 goto out;
1557
1558         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1559         if (IS_ERR(handle)) {
1560                 ret = PTR_ERR(handle);
1561                 mlog_errno(ret);
1562                 goto out;
1563         }
1564
1565         /*
1566          * We want to get the byte offset of the end of the 1st cluster.
1567          */
1568         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1569         if (tmpend > end)
1570                 tmpend = end;
1571
1572         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1573              (unsigned long long)start, (unsigned long long)tmpend);
1574
1575         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1576         if (ret)
1577                 mlog_errno(ret);
1578
1579         if (tmpend < end) {
1580                 /*
1581                  * This may make start and end equal, but the zeroing
1582                  * code will skip any work in that case so there's no
1583                  * need to catch it up here.
1584                  */
1585                 start = end & ~(osb->s_clustersize - 1);
1586
1587                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1588                      (unsigned long long)start, (unsigned long long)end);
1589
1590                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1591                 if (ret)
1592                         mlog_errno(ret);
1593         }
1594
1595         ocfs2_commit_trans(osb, handle);
1596 out:
1597         return ret;
1598 }
1599
1600 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1601 {
1602         int i;
1603         struct ocfs2_extent_rec *rec = NULL;
1604
1605         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1606
1607                 rec = &el->l_recs[i];
1608
1609                 if (le32_to_cpu(rec->e_cpos) < pos)
1610                         break;
1611         }
1612
1613         return i;
1614 }
1615
1616 /*
1617  * Helper to calculate the punching pos and length in one run, we handle the
1618  * following three cases in order:
1619  *
1620  * - remove the entire record
1621  * - remove a partial record
1622  * - no record needs to be removed (hole-punching completed)
1623 */
1624 static void ocfs2_calc_trunc_pos(struct inode *inode,
1625                                  struct ocfs2_extent_list *el,
1626                                  struct ocfs2_extent_rec *rec,
1627                                  u32 trunc_start, u32 *trunc_cpos,
1628                                  u32 *trunc_len, u32 *trunc_end,
1629                                  u64 *blkno, int *done)
1630 {
1631         int ret = 0;
1632         u32 coff, range;
1633
1634         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1635
1636         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1637                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1638                 /*
1639                  * Skip holes if any.
1640                  */
1641                 if (range < *trunc_end)
1642                         *trunc_end = range;
1643                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1644                 *blkno = le64_to_cpu(rec->e_blkno);
1645                 *trunc_end = le32_to_cpu(rec->e_cpos);
1646         } else if (range > trunc_start) {
1647                 *trunc_cpos = trunc_start;
1648                 *trunc_len = *trunc_end - trunc_start;
1649                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1650                 *blkno = le64_to_cpu(rec->e_blkno) +
1651                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1652                 *trunc_end = trunc_start;
1653         } else {
1654                 /*
1655                  * It may have two following possibilities:
1656                  *
1657                  * - last record has been removed
1658                  * - trunc_start was within a hole
1659                  *
1660                  * both two cases mean the completion of hole punching.
1661                  */
1662                 ret = 1;
1663         }
1664
1665         *done = ret;
1666 }
1667
1668 static int ocfs2_remove_inode_range(struct inode *inode,
1669                                     struct buffer_head *di_bh, u64 byte_start,
1670                                     u64 byte_len)
1671 {
1672         int ret = 0, flags = 0, done = 0, i;
1673         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1674         u32 cluster_in_el;
1675         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1676         struct ocfs2_cached_dealloc_ctxt dealloc;
1677         struct address_space *mapping = inode->i_mapping;
1678         struct ocfs2_extent_tree et;
1679         struct ocfs2_path *path = NULL;
1680         struct ocfs2_extent_list *el = NULL;
1681         struct ocfs2_extent_rec *rec = NULL;
1682         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1683         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1684
1685         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1686         ocfs2_init_dealloc_ctxt(&dealloc);
1687
1688         if (byte_len == 0)
1689                 return 0;
1690
1691         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1692                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1693                                             byte_start + byte_len, 0);
1694                 if (ret) {
1695                         mlog_errno(ret);
1696                         goto out;
1697                 }
1698                 /*
1699                  * There's no need to get fancy with the page cache
1700                  * truncate of an inline-data inode. We're talking
1701                  * about less than a page here, which will be cached
1702                  * in the dinode buffer anyway.
1703                  */
1704                 unmap_mapping_range(mapping, 0, 0, 0);
1705                 truncate_inode_pages(mapping, 0);
1706                 goto out;
1707         }
1708
1709         /*
1710          * For reflinks, we may need to CoW 2 clusters which might be
1711          * partially zero'd later, if hole's start and end offset were
1712          * within one cluster(means is not exactly aligned to clustersize).
1713          */
1714
1715         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1716
1717                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1718                 if (ret) {
1719                         mlog_errno(ret);
1720                         goto out;
1721                 }
1722
1723                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1724                 if (ret) {
1725                         mlog_errno(ret);
1726                         goto out;
1727                 }
1728         }
1729
1730         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1731         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1732         cluster_in_el = trunc_end;
1733
1734         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1735              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1736              (unsigned long long)byte_start,
1737              (unsigned long long)byte_len, trunc_start, trunc_end);
1738
1739         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1740         if (ret) {
1741                 mlog_errno(ret);
1742                 goto out;
1743         }
1744
1745         path = ocfs2_new_path_from_et(&et);
1746         if (!path) {
1747                 ret = -ENOMEM;
1748                 mlog_errno(ret);
1749                 goto out;
1750         }
1751
1752         while (trunc_end > trunc_start) {
1753
1754                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1755                                       cluster_in_el);
1756                 if (ret) {
1757                         mlog_errno(ret);
1758                         goto out;
1759                 }
1760
1761                 el = path_leaf_el(path);
1762
1763                 i = ocfs2_find_rec(el, trunc_end);
1764                 /*
1765                  * Need to go to previous extent block.
1766                  */
1767                 if (i < 0) {
1768                         if (path->p_tree_depth == 0)
1769                                 break;
1770
1771                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1772                                                             path,
1773                                                             &cluster_in_el);
1774                         if (ret) {
1775                                 mlog_errno(ret);
1776                                 goto out;
1777                         }
1778
1779                         /*
1780                          * We've reached the leftmost extent block,
1781                          * it's safe to leave.
1782                          */
1783                         if (cluster_in_el == 0)
1784                                 break;
1785
1786                         /*
1787                          * The 'pos' searched for previous extent block is
1788                          * always one cluster less than actual trunc_end.
1789                          */
1790                         trunc_end = cluster_in_el + 1;
1791
1792                         ocfs2_reinit_path(path, 1);
1793
1794                         continue;
1795
1796                 } else
1797                         rec = &el->l_recs[i];
1798
1799                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1800                                      &trunc_len, &trunc_end, &blkno, &done);
1801                 if (done)
1802                         break;
1803
1804                 flags = rec->e_flags;
1805                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1806
1807                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1808                                                phys_cpos, trunc_len, flags,
1809                                                &dealloc, refcount_loc);
1810                 if (ret < 0) {
1811                         mlog_errno(ret);
1812                         goto out;
1813                 }
1814
1815                 cluster_in_el = trunc_end;
1816
1817                 ocfs2_reinit_path(path, 1);
1818         }
1819
1820         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1821
1822 out:
1823         ocfs2_schedule_truncate_log_flush(osb, 1);
1824         ocfs2_run_deallocs(osb, &dealloc);
1825
1826         return ret;
1827 }
1828
1829 /*
1830  * Parts of this function taken from xfs_change_file_space()
1831  */
1832 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1833                                      loff_t f_pos, unsigned int cmd,
1834                                      struct ocfs2_space_resv *sr,
1835                                      int change_size)
1836 {
1837         int ret;
1838         s64 llen;
1839         loff_t size;
1840         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1841         struct buffer_head *di_bh = NULL;
1842         handle_t *handle;
1843         unsigned long long max_off = inode->i_sb->s_maxbytes;
1844
1845         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1846                 return -EROFS;
1847
1848         mutex_lock(&inode->i_mutex);
1849
1850         /*
1851          * This prevents concurrent writes on other nodes
1852          */
1853         ret = ocfs2_rw_lock(inode, 1);
1854         if (ret) {
1855                 mlog_errno(ret);
1856                 goto out;
1857         }
1858
1859         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1860         if (ret) {
1861                 mlog_errno(ret);
1862                 goto out_rw_unlock;
1863         }
1864
1865         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1866                 ret = -EPERM;
1867                 goto out_inode_unlock;
1868         }
1869
1870         switch (sr->l_whence) {
1871         case 0: /*SEEK_SET*/
1872                 break;
1873         case 1: /*SEEK_CUR*/
1874                 sr->l_start += f_pos;
1875                 break;
1876         case 2: /*SEEK_END*/
1877                 sr->l_start += i_size_read(inode);
1878                 break;
1879         default:
1880                 ret = -EINVAL;
1881                 goto out_inode_unlock;
1882         }
1883         sr->l_whence = 0;
1884
1885         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1886
1887         if (sr->l_start < 0
1888             || sr->l_start > max_off
1889             || (sr->l_start + llen) < 0
1890             || (sr->l_start + llen) > max_off) {
1891                 ret = -EINVAL;
1892                 goto out_inode_unlock;
1893         }
1894         size = sr->l_start + sr->l_len;
1895
1896         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1897                 if (sr->l_len <= 0) {
1898                         ret = -EINVAL;
1899                         goto out_inode_unlock;
1900                 }
1901         }
1902
1903         if (file && should_remove_suid(file->f_path.dentry)) {
1904                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1905                 if (ret) {
1906                         mlog_errno(ret);
1907                         goto out_inode_unlock;
1908                 }
1909         }
1910
1911         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1912         switch (cmd) {
1913         case OCFS2_IOC_RESVSP:
1914         case OCFS2_IOC_RESVSP64:
1915                 /*
1916                  * This takes unsigned offsets, but the signed ones we
1917                  * pass have been checked against overflow above.
1918                  */
1919                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1920                                                        sr->l_len);
1921                 break;
1922         case OCFS2_IOC_UNRESVSP:
1923         case OCFS2_IOC_UNRESVSP64:
1924                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1925                                                sr->l_len);
1926                 break;
1927         default:
1928                 ret = -EINVAL;
1929         }
1930         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1931         if (ret) {
1932                 mlog_errno(ret);
1933                 goto out_inode_unlock;
1934         }
1935
1936         /*
1937          * We update c/mtime for these changes
1938          */
1939         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1940         if (IS_ERR(handle)) {
1941                 ret = PTR_ERR(handle);
1942                 mlog_errno(ret);
1943                 goto out_inode_unlock;
1944         }
1945
1946         if (change_size && i_size_read(inode) < size)
1947                 i_size_write(inode, size);
1948
1949         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1950         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1951         if (ret < 0)
1952                 mlog_errno(ret);
1953
1954         ocfs2_commit_trans(osb, handle);
1955
1956 out_inode_unlock:
1957         brelse(di_bh);
1958         ocfs2_inode_unlock(inode, 1);
1959 out_rw_unlock:
1960         ocfs2_rw_unlock(inode, 1);
1961
1962 out:
1963         mutex_unlock(&inode->i_mutex);
1964         return ret;
1965 }
1966
1967 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1968                             struct ocfs2_space_resv *sr)
1969 {
1970         struct inode *inode = file->f_path.dentry->d_inode;
1971         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1972
1973         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1974             !ocfs2_writes_unwritten_extents(osb))
1975                 return -ENOTTY;
1976         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1977                  !ocfs2_sparse_alloc(osb))
1978                 return -ENOTTY;
1979
1980         if (!S_ISREG(inode->i_mode))
1981                 return -EINVAL;
1982
1983         if (!(file->f_mode & FMODE_WRITE))
1984                 return -EBADF;
1985
1986         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1987 }
1988
1989 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1990                             loff_t len)
1991 {
1992         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1993         struct ocfs2_space_resv sr;
1994         int change_size = 1;
1995
1996         if (!ocfs2_writes_unwritten_extents(osb))
1997                 return -EOPNOTSUPP;
1998
1999         if (S_ISDIR(inode->i_mode))
2000                 return -ENODEV;
2001
2002         if (mode & FALLOC_FL_KEEP_SIZE)
2003                 change_size = 0;
2004
2005         sr.l_whence = 0;
2006         sr.l_start = (s64)offset;
2007         sr.l_len = (s64)len;
2008
2009         return __ocfs2_change_file_space(NULL, inode, offset,
2010                                          OCFS2_IOC_RESVSP64, &sr, change_size);
2011 }
2012
2013 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2014                                    size_t count)
2015 {
2016         int ret = 0;
2017         unsigned int extent_flags;
2018         u32 cpos, clusters, extent_len, phys_cpos;
2019         struct super_block *sb = inode->i_sb;
2020
2021         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2022             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2023             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2024                 return 0;
2025
2026         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2027         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2028
2029         while (clusters) {
2030                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2031                                          &extent_flags);
2032                 if (ret < 0) {
2033                         mlog_errno(ret);
2034                         goto out;
2035                 }
2036
2037                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2038                         ret = 1;
2039                         break;
2040                 }
2041
2042                 if (extent_len > clusters)
2043                         extent_len = clusters;
2044
2045                 clusters -= extent_len;
2046                 cpos += extent_len;
2047         }
2048 out:
2049         return ret;
2050 }
2051
2052 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2053                                             struct file *file,
2054                                             loff_t pos, size_t count,
2055                                             int *meta_level)
2056 {
2057         int ret;
2058         struct buffer_head *di_bh = NULL;
2059         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2060         u32 clusters =
2061                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2062
2063         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2064         if (ret) {
2065                 mlog_errno(ret);
2066                 goto out;
2067         }
2068
2069         *meta_level = 1;
2070
2071         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2072         if (ret)
2073                 mlog_errno(ret);
2074 out:
2075         brelse(di_bh);
2076         return ret;
2077 }
2078
2079 static int ocfs2_prepare_inode_for_write(struct file *file,
2080                                          loff_t *ppos,
2081                                          size_t count,
2082                                          int appending,
2083                                          int *direct_io,
2084                                          int *has_refcount)
2085 {
2086         int ret = 0, meta_level = 0;
2087         struct dentry *dentry = file->f_path.dentry;
2088         struct inode *inode = dentry->d_inode;
2089         loff_t saved_pos, end;
2090
2091         /*
2092          * We start with a read level meta lock and only jump to an ex
2093          * if we need to make modifications here.
2094          */
2095         for(;;) {
2096                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2097                 if (ret < 0) {
2098                         meta_level = -1;
2099                         mlog_errno(ret);
2100                         goto out;
2101                 }
2102
2103                 /* Clear suid / sgid if necessary. We do this here
2104                  * instead of later in the write path because
2105                  * remove_suid() calls ->setattr without any hint that
2106                  * we may have already done our cluster locking. Since
2107                  * ocfs2_setattr() *must* take cluster locks to
2108                  * proceeed, this will lead us to recursively lock the
2109                  * inode. There's also the dinode i_size state which
2110                  * can be lost via setattr during extending writes (we
2111                  * set inode->i_size at the end of a write. */
2112                 if (should_remove_suid(dentry)) {
2113                         if (meta_level == 0) {
2114                                 ocfs2_inode_unlock(inode, meta_level);
2115                                 meta_level = 1;
2116                                 continue;
2117                         }
2118
2119                         ret = ocfs2_write_remove_suid(inode);
2120                         if (ret < 0) {
2121                                 mlog_errno(ret);
2122                                 goto out_unlock;
2123                         }
2124                 }
2125
2126                 /* work on a copy of ppos until we're sure that we won't have
2127                  * to recalculate it due to relocking. */
2128                 if (appending) {
2129                         saved_pos = i_size_read(inode);
2130                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2131                 } else {
2132                         saved_pos = *ppos;
2133                 }
2134
2135                 end = saved_pos + count;
2136
2137                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2138                 if (ret == 1) {
2139                         ocfs2_inode_unlock(inode, meta_level);
2140                         meta_level = -1;
2141
2142                         ret = ocfs2_prepare_inode_for_refcount(inode,
2143                                                                file,
2144                                                                saved_pos,
2145                                                                count,
2146                                                                &meta_level);
2147                         if (has_refcount)
2148                                 *has_refcount = 1;
2149                         if (direct_io)
2150                                 *direct_io = 0;
2151                 }
2152
2153                 if (ret < 0) {
2154                         mlog_errno(ret);
2155                         goto out_unlock;
2156                 }
2157
2158                 /*
2159                  * Skip the O_DIRECT checks if we don't need
2160                  * them.
2161                  */
2162                 if (!direct_io || !(*direct_io))
2163                         break;
2164
2165                 /*
2166                  * There's no sane way to do direct writes to an inode
2167                  * with inline data.
2168                  */
2169                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2170                         *direct_io = 0;
2171                         break;
2172                 }
2173
2174                 /*
2175                  * Allowing concurrent direct writes means
2176                  * i_size changes wouldn't be synchronized, so
2177                  * one node could wind up truncating another
2178                  * nodes writes.
2179                  */
2180                 if (end > i_size_read(inode)) {
2181                         *direct_io = 0;
2182                         break;
2183                 }
2184
2185                 /*
2186                  * We don't fill holes during direct io, so
2187                  * check for them here. If any are found, the
2188                  * caller will have to retake some cluster
2189                  * locks and initiate the io as buffered.
2190                  */
2191                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2192                 if (ret == 1) {
2193                         *direct_io = 0;
2194                         ret = 0;
2195                 } else if (ret < 0)
2196                         mlog_errno(ret);
2197                 break;
2198         }
2199
2200         if (appending)
2201                 *ppos = saved_pos;
2202
2203 out_unlock:
2204         if (meta_level >= 0)
2205                 ocfs2_inode_unlock(inode, meta_level);
2206
2207 out:
2208         return ret;
2209 }
2210
2211 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2212                                     const struct iovec *iov,
2213                                     unsigned long nr_segs,
2214                                     loff_t pos)
2215 {
2216         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2217         int can_do_direct, has_refcount = 0;
2218         ssize_t written = 0;
2219         size_t ocount;          /* original count */
2220         size_t count;           /* after file limit checks */
2221         loff_t old_size, *ppos = &iocb->ki_pos;
2222         u32 old_clusters;
2223         struct file *file = iocb->ki_filp;
2224         struct inode *inode = file->f_path.dentry->d_inode;
2225         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2226         int full_coherency = !(osb->s_mount_opt &
2227                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2228
2229         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2230                    (unsigned int)nr_segs,
2231                    file->f_path.dentry->d_name.len,
2232                    file->f_path.dentry->d_name.name);
2233
2234         if (iocb->ki_left == 0)
2235                 return 0;
2236
2237         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2238
2239         appending = file->f_flags & O_APPEND ? 1 : 0;
2240         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2241
2242         mutex_lock(&inode->i_mutex);
2243
2244 relock:
2245         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2246         if (direct_io) {
2247                 down_read(&inode->i_alloc_sem);
2248                 have_alloc_sem = 1;
2249         }
2250
2251         /*
2252          * Concurrent O_DIRECT writes are allowed with
2253          * mount_option "coherency=buffered".
2254          */
2255         rw_level = (!direct_io || full_coherency);
2256
2257         ret = ocfs2_rw_lock(inode, rw_level);
2258         if (ret < 0) {
2259                 mlog_errno(ret);
2260                 goto out_sems;
2261         }
2262
2263         /*
2264          * O_DIRECT writes with "coherency=full" need to take EX cluster
2265          * inode_lock to guarantee coherency.
2266          */
2267         if (direct_io && full_coherency) {
2268                 /*
2269                  * We need to take and drop the inode lock to force
2270                  * other nodes to drop their caches.  Buffered I/O
2271                  * already does this in write_begin().
2272                  */
2273                 ret = ocfs2_inode_lock(inode, NULL, 1);
2274                 if (ret < 0) {
2275                         mlog_errno(ret);
2276                         goto out_sems;
2277                 }
2278
2279                 ocfs2_inode_unlock(inode, 1);
2280         }
2281
2282         can_do_direct = direct_io;
2283         ret = ocfs2_prepare_inode_for_write(file, ppos,
2284                                             iocb->ki_left, appending,
2285                                             &can_do_direct, &has_refcount);
2286         if (ret < 0) {
2287                 mlog_errno(ret);
2288                 goto out;
2289         }
2290
2291         /*
2292          * We can't complete the direct I/O as requested, fall back to
2293          * buffered I/O.
2294          */
2295         if (direct_io && !can_do_direct) {
2296                 ocfs2_rw_unlock(inode, rw_level);
2297                 up_read(&inode->i_alloc_sem);
2298
2299                 have_alloc_sem = 0;
2300                 rw_level = -1;
2301
2302                 direct_io = 0;
2303                 goto relock;
2304         }
2305
2306         /*
2307          * To later detect whether a journal commit for sync writes is
2308          * necessary, we sample i_size, and cluster count here.
2309          */
2310         old_size = i_size_read(inode);
2311         old_clusters = OCFS2_I(inode)->ip_clusters;
2312
2313         /* communicate with ocfs2_dio_end_io */
2314         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2315
2316         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2317                                      VERIFY_READ);
2318         if (ret)
2319                 goto out_dio;
2320
2321         count = ocount;
2322         ret = generic_write_checks(file, ppos, &count,
2323                                    S_ISBLK(inode->i_mode));
2324         if (ret)
2325                 goto out_dio;
2326
2327         if (direct_io) {
2328                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2329                                                     ppos, count, ocount);
2330                 if (written < 0) {
2331                         ret = written;
2332                         goto out_dio;
2333                 }
2334         } else {
2335                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2336                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2337                                                       ppos, count, 0);
2338                 current->backing_dev_info = NULL;
2339         }
2340
2341 out_dio:
2342         /* buffered aio wouldn't have proper lock coverage today */
2343         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2344
2345         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2346             ((file->f_flags & O_DIRECT) && !direct_io)) {
2347                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2348                                                pos + count - 1);
2349                 if (ret < 0)
2350                         written = ret;
2351
2352                 if (!ret && ((old_size != i_size_read(inode)) ||
2353                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2354                              has_refcount)) {
2355                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2356                         if (ret < 0)
2357                                 written = ret;
2358                 }
2359
2360                 if (!ret)
2361                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2362                                                       pos + count - 1);
2363         }
2364
2365         /*
2366          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2367          * function pointer which is called when o_direct io completes so that
2368          * it can unlock our rw lock.  (it's the clustered equivalent of
2369          * i_alloc_sem; protects truncate from racing with pending ios).
2370          * Unfortunately there are error cases which call end_io and others
2371          * that don't.  so we don't have to unlock the rw_lock if either an
2372          * async dio is going to do it in the future or an end_io after an
2373          * error has already done it.
2374          */
2375         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2376                 rw_level = -1;
2377                 have_alloc_sem = 0;
2378         }
2379
2380 out:
2381         if (rw_level != -1)
2382                 ocfs2_rw_unlock(inode, rw_level);
2383
2384 out_sems:
2385         if (have_alloc_sem)
2386                 up_read(&inode->i_alloc_sem);
2387
2388         mutex_unlock(&inode->i_mutex);
2389
2390         if (written)
2391                 ret = written;
2392         mlog_exit(ret);
2393         return ret;
2394 }
2395
2396 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2397                                 struct file *out,
2398                                 struct splice_desc *sd)
2399 {
2400         int ret;
2401
2402         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2403                                             sd->total_len, 0, NULL, NULL);
2404         if (ret < 0) {
2405                 mlog_errno(ret);
2406                 return ret;
2407         }
2408
2409         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2410 }
2411
2412 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2413                                        struct file *out,
2414                                        loff_t *ppos,
2415                                        size_t len,
2416                                        unsigned int flags)
2417 {
2418         int ret;
2419         struct address_space *mapping = out->f_mapping;
2420         struct inode *inode = mapping->host;
2421         struct splice_desc sd = {
2422                 .total_len = len,
2423                 .flags = flags,
2424                 .pos = *ppos,
2425                 .u.file = out,
2426         };
2427
2428         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2429                    (unsigned int)len,
2430                    out->f_path.dentry->d_name.len,
2431                    out->f_path.dentry->d_name.name);
2432
2433         if (pipe->inode)
2434                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2435
2436         splice_from_pipe_begin(&sd);
2437         do {
2438                 ret = splice_from_pipe_next(pipe, &sd);
2439                 if (ret <= 0)
2440                         break;
2441
2442                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2443                 ret = ocfs2_rw_lock(inode, 1);
2444                 if (ret < 0)
2445                         mlog_errno(ret);
2446                 else {
2447                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2448                         ocfs2_rw_unlock(inode, 1);
2449                 }
2450                 mutex_unlock(&inode->i_mutex);
2451         } while (ret > 0);
2452         splice_from_pipe_end(pipe, &sd);
2453
2454         if (pipe->inode)
2455                 mutex_unlock(&pipe->inode->i_mutex);
2456
2457         if (sd.num_spliced)
2458                 ret = sd.num_spliced;
2459
2460         if (ret > 0) {
2461                 unsigned long nr_pages;
2462                 int err;
2463
2464                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2465
2466                 err = generic_write_sync(out, *ppos, ret);
2467                 if (err)
2468                         ret = err;
2469                 else
2470                         *ppos += ret;
2471
2472                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2473         }
2474
2475         mlog_exit(ret);
2476         return ret;
2477 }
2478
2479 static ssize_t ocfs2_file_splice_read(struct file *in,
2480                                       loff_t *ppos,
2481                                       struct pipe_inode_info *pipe,
2482                                       size_t len,
2483                                       unsigned int flags)
2484 {
2485         int ret = 0, lock_level = 0;
2486         struct inode *inode = in->f_path.dentry->d_inode;
2487
2488         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2489                    (unsigned int)len,
2490                    in->f_path.dentry->d_name.len,
2491                    in->f_path.dentry->d_name.name);
2492
2493         /*
2494          * See the comment in ocfs2_file_aio_read()
2495          */
2496         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2497         if (ret < 0) {
2498                 mlog_errno(ret);
2499                 goto bail;
2500         }
2501         ocfs2_inode_unlock(inode, lock_level);
2502
2503         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2504
2505 bail:
2506         mlog_exit(ret);
2507         return ret;
2508 }
2509
2510 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2511                                    const struct iovec *iov,
2512                                    unsigned long nr_segs,
2513                                    loff_t pos)
2514 {
2515         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2516         struct file *filp = iocb->ki_filp;
2517         struct inode *inode = filp->f_path.dentry->d_inode;
2518
2519         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2520                    (unsigned int)nr_segs,
2521                    filp->f_path.dentry->d_name.len,
2522                    filp->f_path.dentry->d_name.name);
2523
2524         if (!inode) {
2525                 ret = -EINVAL;
2526                 mlog_errno(ret);
2527                 goto bail;
2528         }
2529
2530         /*
2531          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2532          * need locks to protect pending reads from racing with truncate.
2533          */
2534         if (filp->f_flags & O_DIRECT) {
2535                 down_read(&inode->i_alloc_sem);
2536                 have_alloc_sem = 1;
2537
2538                 ret = ocfs2_rw_lock(inode, 0);
2539                 if (ret < 0) {
2540                         mlog_errno(ret);
2541                         goto bail;
2542                 }
2543                 rw_level = 0;
2544                 /* communicate with ocfs2_dio_end_io */
2545                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2546         }
2547
2548         /*
2549          * We're fine letting folks race truncates and extending
2550          * writes with read across the cluster, just like they can
2551          * locally. Hence no rw_lock during read.
2552          *
2553          * Take and drop the meta data lock to update inode fields
2554          * like i_size. This allows the checks down below
2555          * generic_file_aio_read() a chance of actually working.
2556          */
2557         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2558         if (ret < 0) {
2559                 mlog_errno(ret);
2560                 goto bail;
2561         }
2562         ocfs2_inode_unlock(inode, lock_level);
2563
2564         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2565         if (ret == -EINVAL)
2566                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2567
2568         /* buffered aio wouldn't have proper lock coverage today */
2569         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2570
2571         /* see ocfs2_file_aio_write */
2572         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2573                 rw_level = -1;
2574                 have_alloc_sem = 0;
2575         }
2576
2577 bail:
2578         if (have_alloc_sem)
2579                 up_read(&inode->i_alloc_sem);
2580         if (rw_level != -1)
2581                 ocfs2_rw_unlock(inode, rw_level);
2582         mlog_exit(ret);
2583
2584         return ret;
2585 }
2586
2587 const struct inode_operations ocfs2_file_iops = {
2588         .setattr        = ocfs2_setattr,
2589         .getattr        = ocfs2_getattr,
2590         .permission     = ocfs2_permission,
2591         .setxattr       = generic_setxattr,
2592         .getxattr       = generic_getxattr,
2593         .listxattr      = ocfs2_listxattr,
2594         .removexattr    = generic_removexattr,
2595         .fallocate      = ocfs2_fallocate,
2596         .fiemap         = ocfs2_fiemap,
2597 };
2598
2599 const struct inode_operations ocfs2_special_file_iops = {
2600         .setattr        = ocfs2_setattr,
2601         .getattr        = ocfs2_getattr,
2602         .permission     = ocfs2_permission,
2603 };
2604
2605 /*
2606  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2607  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2608  */
2609 const struct file_operations ocfs2_fops = {
2610         .llseek         = generic_file_llseek,
2611         .read           = do_sync_read,
2612         .write          = do_sync_write,
2613         .mmap           = ocfs2_mmap,
2614         .fsync          = ocfs2_sync_file,
2615         .release        = ocfs2_file_release,
2616         .open           = ocfs2_file_open,
2617         .aio_read       = ocfs2_file_aio_read,
2618         .aio_write      = ocfs2_file_aio_write,
2619         .unlocked_ioctl = ocfs2_ioctl,
2620 #ifdef CONFIG_COMPAT
2621         .compat_ioctl   = ocfs2_compat_ioctl,
2622 #endif
2623         .lock           = ocfs2_lock,
2624         .flock          = ocfs2_flock,
2625         .splice_read    = ocfs2_file_splice_read,
2626         .splice_write   = ocfs2_file_splice_write,
2627 };
2628
2629 const struct file_operations ocfs2_dops = {
2630         .llseek         = generic_file_llseek,
2631         .read           = generic_read_dir,
2632         .readdir        = ocfs2_readdir,
2633         .fsync          = ocfs2_sync_file,
2634         .release        = ocfs2_dir_release,
2635         .open           = ocfs2_dir_open,
2636         .unlocked_ioctl = ocfs2_ioctl,
2637 #ifdef CONFIG_COMPAT
2638         .compat_ioctl   = ocfs2_compat_ioctl,
2639 #endif
2640         .lock           = ocfs2_lock,
2641         .flock          = ocfs2_flock,
2642 };
2643
2644 /*
2645  * POSIX-lockless variants of our file_operations.
2646  *
2647  * These will be used if the underlying cluster stack does not support
2648  * posix file locking, if the user passes the "localflocks" mount
2649  * option, or if we have a local-only fs.
2650  *
2651  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2652  * so we still want it in the case of no stack support for
2653  * plocks. Internally, it will do the right thing when asked to ignore
2654  * the cluster.
2655  */
2656 const struct file_operations ocfs2_fops_no_plocks = {
2657         .llseek         = generic_file_llseek,
2658         .read           = do_sync_read,
2659         .write          = do_sync_write,
2660         .mmap           = ocfs2_mmap,
2661         .fsync          = ocfs2_sync_file,
2662         .release        = ocfs2_file_release,
2663         .open           = ocfs2_file_open,
2664         .aio_read       = ocfs2_file_aio_read,
2665         .aio_write      = ocfs2_file_aio_write,
2666         .unlocked_ioctl = ocfs2_ioctl,
2667 #ifdef CONFIG_COMPAT
2668         .compat_ioctl   = ocfs2_compat_ioctl,
2669 #endif
2670         .flock          = ocfs2_flock,
2671         .splice_read    = ocfs2_file_splice_read,
2672         .splice_write   = ocfs2_file_splice_write,
2673 };
2674
2675 const struct file_operations ocfs2_dops_no_plocks = {
2676         .llseek         = generic_file_llseek,
2677         .read           = generic_read_dir,
2678         .readdir        = ocfs2_readdir,
2679         .fsync          = ocfs2_sync_file,
2680         .release        = ocfs2_dir_release,
2681         .open           = ocfs2_dir_open,
2682         .unlocked_ioctl = ocfs2_ioctl,
2683 #ifdef CONFIG_COMPAT
2684         .compat_ioctl   = ocfs2_compat_ioctl,
2685 #endif
2686         .flock          = ocfs2_flock,
2687 };