ocfs2: drop the BLKDEV_IFL_WAIT flag
[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
800                  * force ocfs2_{prepare,commit}_write() to zero the
801                  * whole block.
802                  */
803                 ret = ocfs2_prepare_write_nolock(inode, page,
804                                                  block_start + 1,
805                                                  block_start + 1);
806                 if (ret < 0) {
807                         mlog_errno(ret);
808                         goto out_unlock;
809                 }
810
811                 if (!handle) {
812                         handle = ocfs2_zero_start_ordered_transaction(inode);
813                         if (IS_ERR(handle)) {
814                                 ret = PTR_ERR(handle);
815                                 handle = NULL;
816                                 break;
817                         }
818                 }
819
820                 /* must not update i_size! */
821                 ret = block_commit_write(page, block_start + 1,
822                                          block_start + 1);
823                 if (ret < 0)
824                         mlog_errno(ret);
825                 else
826                         ret = 0;
827         }
828
829         if (handle)
830                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
831
832 out_unlock:
833         unlock_page(page);
834         page_cache_release(page);
835 out:
836         return ret;
837 }
838
839 /*
840  * Find the next range to zero.  We do this in terms of bytes because
841  * that's what ocfs2_zero_extend() wants, and it is dealing with the
842  * pagecache.  We may return multiple extents.
843  *
844  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
845  * needs to be zeroed.  range_start and range_end return the next zeroing
846  * range.  A subsequent call should pass the previous range_end as its
847  * zero_start.  If range_end is 0, there's nothing to do.
848  *
849  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
850  */
851 static int ocfs2_zero_extend_get_range(struct inode *inode,
852                                        struct buffer_head *di_bh,
853                                        u64 zero_start, u64 zero_end,
854                                        u64 *range_start, u64 *range_end)
855 {
856         int rc = 0, needs_cow = 0;
857         u32 p_cpos, zero_clusters = 0;
858         u32 zero_cpos =
859                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
860         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
861         unsigned int num_clusters = 0;
862         unsigned int ext_flags = 0;
863
864         while (zero_cpos < last_cpos) {
865                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
866                                         &num_clusters, &ext_flags);
867                 if (rc) {
868                         mlog_errno(rc);
869                         goto out;
870                 }
871
872                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
873                         zero_clusters = num_clusters;
874                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
875                                 needs_cow = 1;
876                         break;
877                 }
878
879                 zero_cpos += num_clusters;
880         }
881         if (!zero_clusters) {
882                 *range_end = 0;
883                 goto out;
884         }
885
886         while ((zero_cpos + zero_clusters) < last_cpos) {
887                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
888                                         &p_cpos, &num_clusters,
889                                         &ext_flags);
890                 if (rc) {
891                         mlog_errno(rc);
892                         goto out;
893                 }
894
895                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
896                         break;
897                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
898                         needs_cow = 1;
899                 zero_clusters += num_clusters;
900         }
901         if ((zero_cpos + zero_clusters) > last_cpos)
902                 zero_clusters = last_cpos - zero_cpos;
903
904         if (needs_cow) {
905                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
906                                         zero_clusters, UINT_MAX);
907                 if (rc) {
908                         mlog_errno(rc);
909                         goto out;
910                 }
911         }
912
913         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
914         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
915                                              zero_cpos + zero_clusters);
916
917 out:
918         return rc;
919 }
920
921 /*
922  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
923  * has made sure that the entire range needs zeroing.
924  */
925 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
926                                    u64 range_end)
927 {
928         int rc = 0;
929         u64 next_pos;
930         u64 zero_pos = range_start;
931
932         mlog(0, "range_start = %llu, range_end = %llu\n",
933              (unsigned long long)range_start,
934              (unsigned long long)range_end);
935         BUG_ON(range_start >= range_end);
936
937         while (zero_pos < range_end) {
938                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
939                 if (next_pos > range_end)
940                         next_pos = range_end;
941                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
942                 if (rc < 0) {
943                         mlog_errno(rc);
944                         break;
945                 }
946                 zero_pos = next_pos;
947
948                 /*
949                  * Very large extends have the potential to lock up
950                  * the cpu for extended periods of time.
951                  */
952                 cond_resched();
953         }
954
955         return rc;
956 }
957
958 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
959                       loff_t zero_to_size)
960 {
961         int ret = 0;
962         u64 zero_start, range_start = 0, range_end = 0;
963         struct super_block *sb = inode->i_sb;
964
965         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
966         mlog(0, "zero_start %llu for i_size %llu\n",
967              (unsigned long long)zero_start,
968              (unsigned long long)i_size_read(inode));
969         while (zero_start < zero_to_size) {
970                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
971                                                   zero_to_size,
972                                                   &range_start,
973                                                   &range_end);
974                 if (ret) {
975                         mlog_errno(ret);
976                         break;
977                 }
978                 if (!range_end)
979                         break;
980                 /* Trim the ends */
981                 if (range_start < zero_start)
982                         range_start = zero_start;
983                 if (range_end > zero_to_size)
984                         range_end = zero_to_size;
985
986                 ret = ocfs2_zero_extend_range(inode, range_start,
987                                               range_end);
988                 if (ret) {
989                         mlog_errno(ret);
990                         break;
991                 }
992                 zero_start = range_end;
993         }
994
995         return ret;
996 }
997
998 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
999                           u64 new_i_size, u64 zero_to)
1000 {
1001         int ret;
1002         u32 clusters_to_add;
1003         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1004
1005         /*
1006          * Only quota files call this without a bh, and they can't be
1007          * refcounted.
1008          */
1009         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
1010         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1011
1012         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1013         if (clusters_to_add < oi->ip_clusters)
1014                 clusters_to_add = 0;
1015         else
1016                 clusters_to_add -= oi->ip_clusters;
1017
1018         if (clusters_to_add) {
1019                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1020                                                 clusters_to_add, 0);
1021                 if (ret) {
1022                         mlog_errno(ret);
1023                         goto out;
1024                 }
1025         }
1026
1027         /*
1028          * Call this even if we don't add any clusters to the tree. We
1029          * still need to zero the area between the old i_size and the
1030          * new i_size.
1031          */
1032         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1033         if (ret < 0)
1034                 mlog_errno(ret);
1035
1036 out:
1037         return ret;
1038 }
1039
1040 static int ocfs2_extend_file(struct inode *inode,
1041                              struct buffer_head *di_bh,
1042                              u64 new_i_size)
1043 {
1044         int ret = 0;
1045         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1046
1047         BUG_ON(!di_bh);
1048
1049         /* setattr sometimes calls us like this. */
1050         if (new_i_size == 0)
1051                 goto out;
1052
1053         if (i_size_read(inode) == new_i_size)
1054                 goto out;
1055         BUG_ON(new_i_size < i_size_read(inode));
1056
1057         /*
1058          * The alloc sem blocks people in read/write from reading our
1059          * allocation until we're done changing it. We depend on
1060          * i_mutex to block other extend/truncate calls while we're
1061          * here.  We even have to hold it for sparse files because there
1062          * might be some tail zeroing.
1063          */
1064         down_write(&oi->ip_alloc_sem);
1065
1066         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1067                 /*
1068                  * We can optimize small extends by keeping the inodes
1069                  * inline data.
1070                  */
1071                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1072                         up_write(&oi->ip_alloc_sem);
1073                         goto out_update_size;
1074                 }
1075
1076                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1077                 if (ret) {
1078                         up_write(&oi->ip_alloc_sem);
1079                         mlog_errno(ret);
1080                         goto out;
1081                 }
1082         }
1083
1084         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1085                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1086         else
1087                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1088                                             new_i_size);
1089
1090         up_write(&oi->ip_alloc_sem);
1091
1092         if (ret < 0) {
1093                 mlog_errno(ret);
1094                 goto out;
1095         }
1096
1097 out_update_size:
1098         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1099         if (ret < 0)
1100                 mlog_errno(ret);
1101
1102 out:
1103         return ret;
1104 }
1105
1106 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1107 {
1108         int status = 0, size_change;
1109         struct inode *inode = dentry->d_inode;
1110         struct super_block *sb = inode->i_sb;
1111         struct ocfs2_super *osb = OCFS2_SB(sb);
1112         struct buffer_head *bh = NULL;
1113         handle_t *handle = NULL;
1114         struct dquot *transfer_to[MAXQUOTAS] = { };
1115         int qtype;
1116
1117         mlog_entry("(0x%p, '%.*s')\n", dentry,
1118                    dentry->d_name.len, dentry->d_name.name);
1119
1120         /* ensuring we don't even attempt to truncate a symlink */
1121         if (S_ISLNK(inode->i_mode))
1122                 attr->ia_valid &= ~ATTR_SIZE;
1123
1124         if (attr->ia_valid & ATTR_MODE)
1125                 mlog(0, "mode change: %d\n", attr->ia_mode);
1126         if (attr->ia_valid & ATTR_UID)
1127                 mlog(0, "uid change: %d\n", attr->ia_uid);
1128         if (attr->ia_valid & ATTR_GID)
1129                 mlog(0, "gid change: %d\n", attr->ia_gid);
1130         if (attr->ia_valid & ATTR_SIZE)
1131                 mlog(0, "size change...\n");
1132         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1133                 mlog(0, "time change...\n");
1134
1135 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1136                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1137         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1138                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1139                 return 0;
1140         }
1141
1142         status = inode_change_ok(inode, attr);
1143         if (status)
1144                 return status;
1145
1146         if (is_quota_modification(inode, attr))
1147                 dquot_initialize(inode);
1148         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1149         if (size_change) {
1150                 status = ocfs2_rw_lock(inode, 1);
1151                 if (status < 0) {
1152                         mlog_errno(status);
1153                         goto bail;
1154                 }
1155         }
1156
1157         status = ocfs2_inode_lock(inode, &bh, 1);
1158         if (status < 0) {
1159                 if (status != -ENOENT)
1160                         mlog_errno(status);
1161                 goto bail_unlock_rw;
1162         }
1163
1164         if (size_change && attr->ia_size != i_size_read(inode)) {
1165                 status = inode_newsize_ok(inode, attr->ia_size);
1166                 if (status)
1167                         goto bail_unlock;
1168
1169                 if (i_size_read(inode) > attr->ia_size) {
1170                         if (ocfs2_should_order_data(inode)) {
1171                                 status = ocfs2_begin_ordered_truncate(inode,
1172                                                                       attr->ia_size);
1173                                 if (status)
1174                                         goto bail_unlock;
1175                         }
1176                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1177                 } else
1178                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1179                 if (status < 0) {
1180                         if (status != -ENOSPC)
1181                                 mlog_errno(status);
1182                         status = -ENOSPC;
1183                         goto bail_unlock;
1184                 }
1185         }
1186
1187         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1188             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1189                 /*
1190                  * Gather pointers to quota structures so that allocation /
1191                  * freeing of quota structures happens here and not inside
1192                  * dquot_transfer() where we have problems with lock ordering
1193                  */
1194                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1195                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1196                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1197                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1198                                                       USRQUOTA);
1199                         if (!transfer_to[USRQUOTA]) {
1200                                 status = -ESRCH;
1201                                 goto bail_unlock;
1202                         }
1203                 }
1204                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1205                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1206                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1207                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1208                                                       GRPQUOTA);
1209                         if (!transfer_to[GRPQUOTA]) {
1210                                 status = -ESRCH;
1211                                 goto bail_unlock;
1212                         }
1213                 }
1214                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1215                                            2 * ocfs2_quota_trans_credits(sb));
1216                 if (IS_ERR(handle)) {
1217                         status = PTR_ERR(handle);
1218                         mlog_errno(status);
1219                         goto bail_unlock;
1220                 }
1221                 status = __dquot_transfer(inode, transfer_to);
1222                 if (status < 0)
1223                         goto bail_commit;
1224         } else {
1225                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1226                 if (IS_ERR(handle)) {
1227                         status = PTR_ERR(handle);
1228                         mlog_errno(status);
1229                         goto bail_unlock;
1230                 }
1231         }
1232
1233         /*
1234          * This will intentionally not wind up calling truncate_setsize(),
1235          * since all the work for a size change has been done above.
1236          * Otherwise, we could get into problems with truncate as
1237          * ip_alloc_sem is used there to protect against i_size
1238          * changes.
1239          *
1240          * XXX: this means the conditional below can probably be removed.
1241          */
1242         if ((attr->ia_valid & ATTR_SIZE) &&
1243             attr->ia_size != i_size_read(inode)) {
1244                 status = vmtruncate(inode, attr->ia_size);
1245                 if (status) {
1246                         mlog_errno(status);
1247                         goto bail_commit;
1248                 }
1249         }
1250
1251         setattr_copy(inode, attr);
1252         mark_inode_dirty(inode);
1253
1254         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1255         if (status < 0)
1256                 mlog_errno(status);
1257
1258 bail_commit:
1259         ocfs2_commit_trans(osb, handle);
1260 bail_unlock:
1261         ocfs2_inode_unlock(inode, 1);
1262 bail_unlock_rw:
1263         if (size_change)
1264                 ocfs2_rw_unlock(inode, 1);
1265 bail:
1266         brelse(bh);
1267
1268         /* Release quota pointers in case we acquired them */
1269         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1270                 dqput(transfer_to[qtype]);
1271
1272         if (!status && attr->ia_valid & ATTR_MODE) {
1273                 status = ocfs2_acl_chmod(inode);
1274                 if (status < 0)
1275                         mlog_errno(status);
1276         }
1277
1278         mlog_exit(status);
1279         return status;
1280 }
1281
1282 int ocfs2_getattr(struct vfsmount *mnt,
1283                   struct dentry *dentry,
1284                   struct kstat *stat)
1285 {
1286         struct inode *inode = dentry->d_inode;
1287         struct super_block *sb = dentry->d_inode->i_sb;
1288         struct ocfs2_super *osb = sb->s_fs_info;
1289         int err;
1290
1291         mlog_entry_void();
1292
1293         err = ocfs2_inode_revalidate(dentry);
1294         if (err) {
1295                 if (err != -ENOENT)
1296                         mlog_errno(err);
1297                 goto bail;
1298         }
1299
1300         generic_fillattr(inode, stat);
1301
1302         /* We set the blksize from the cluster size for performance */
1303         stat->blksize = osb->s_clustersize;
1304
1305 bail:
1306         mlog_exit(err);
1307
1308         return err;
1309 }
1310
1311 int ocfs2_permission(struct inode *inode, int mask)
1312 {
1313         int ret;
1314
1315         mlog_entry_void();
1316
1317         ret = ocfs2_inode_lock(inode, NULL, 0);
1318         if (ret) {
1319                 if (ret != -ENOENT)
1320                         mlog_errno(ret);
1321                 goto out;
1322         }
1323
1324         ret = generic_permission(inode, mask, ocfs2_check_acl);
1325
1326         ocfs2_inode_unlock(inode, 0);
1327 out:
1328         mlog_exit(ret);
1329         return ret;
1330 }
1331
1332 static int __ocfs2_write_remove_suid(struct inode *inode,
1333                                      struct buffer_head *bh)
1334 {
1335         int ret;
1336         handle_t *handle;
1337         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1338         struct ocfs2_dinode *di;
1339
1340         mlog_entry("(Inode %llu, mode 0%o)\n",
1341                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1342
1343         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1344         if (IS_ERR(handle)) {
1345                 ret = PTR_ERR(handle);
1346                 mlog_errno(ret);
1347                 goto out;
1348         }
1349
1350         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1351                                       OCFS2_JOURNAL_ACCESS_WRITE);
1352         if (ret < 0) {
1353                 mlog_errno(ret);
1354                 goto out_trans;
1355         }
1356
1357         inode->i_mode &= ~S_ISUID;
1358         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1359                 inode->i_mode &= ~S_ISGID;
1360
1361         di = (struct ocfs2_dinode *) bh->b_data;
1362         di->i_mode = cpu_to_le16(inode->i_mode);
1363
1364         ocfs2_journal_dirty(handle, bh);
1365
1366 out_trans:
1367         ocfs2_commit_trans(osb, handle);
1368 out:
1369         mlog_exit(ret);
1370         return ret;
1371 }
1372
1373 /*
1374  * Will look for holes and unwritten extents in the range starting at
1375  * pos for count bytes (inclusive).
1376  */
1377 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1378                                        size_t count)
1379 {
1380         int ret = 0;
1381         unsigned int extent_flags;
1382         u32 cpos, clusters, extent_len, phys_cpos;
1383         struct super_block *sb = inode->i_sb;
1384
1385         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1386         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1387
1388         while (clusters) {
1389                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1390                                          &extent_flags);
1391                 if (ret < 0) {
1392                         mlog_errno(ret);
1393                         goto out;
1394                 }
1395
1396                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1397                         ret = 1;
1398                         break;
1399                 }
1400
1401                 if (extent_len > clusters)
1402                         extent_len = clusters;
1403
1404                 clusters -= extent_len;
1405                 cpos += extent_len;
1406         }
1407 out:
1408         return ret;
1409 }
1410
1411 static int ocfs2_write_remove_suid(struct inode *inode)
1412 {
1413         int ret;
1414         struct buffer_head *bh = NULL;
1415
1416         ret = ocfs2_read_inode_block(inode, &bh);
1417         if (ret < 0) {
1418                 mlog_errno(ret);
1419                 goto out;
1420         }
1421
1422         ret =  __ocfs2_write_remove_suid(inode, bh);
1423 out:
1424         brelse(bh);
1425         return ret;
1426 }
1427
1428 /*
1429  * Allocate enough extents to cover the region starting at byte offset
1430  * start for len bytes. Existing extents are skipped, any extents
1431  * added are marked as "unwritten".
1432  */
1433 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1434                                             u64 start, u64 len)
1435 {
1436         int ret;
1437         u32 cpos, phys_cpos, clusters, alloc_size;
1438         u64 end = start + len;
1439         struct buffer_head *di_bh = NULL;
1440
1441         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1442                 ret = ocfs2_read_inode_block(inode, &di_bh);
1443                 if (ret) {
1444                         mlog_errno(ret);
1445                         goto out;
1446                 }
1447
1448                 /*
1449                  * Nothing to do if the requested reservation range
1450                  * fits within the inode.
1451                  */
1452                 if (ocfs2_size_fits_inline_data(di_bh, end))
1453                         goto out;
1454
1455                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1456                 if (ret) {
1457                         mlog_errno(ret);
1458                         goto out;
1459                 }
1460         }
1461
1462         /*
1463          * We consider both start and len to be inclusive.
1464          */
1465         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1466         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1467         clusters -= cpos;
1468
1469         while (clusters) {
1470                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1471                                          &alloc_size, NULL);
1472                 if (ret) {
1473                         mlog_errno(ret);
1474                         goto out;
1475                 }
1476
1477                 /*
1478                  * Hole or existing extent len can be arbitrary, so
1479                  * cap it to our own allocation request.
1480                  */
1481                 if (alloc_size > clusters)
1482                         alloc_size = clusters;
1483
1484                 if (phys_cpos) {
1485                         /*
1486                          * We already have an allocation at this
1487                          * region so we can safely skip it.
1488                          */
1489                         goto next;
1490                 }
1491
1492                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1493                 if (ret) {
1494                         if (ret != -ENOSPC)
1495                                 mlog_errno(ret);
1496                         goto out;
1497                 }
1498
1499 next:
1500                 cpos += alloc_size;
1501                 clusters -= alloc_size;
1502         }
1503
1504         ret = 0;
1505 out:
1506
1507         brelse(di_bh);
1508         return ret;
1509 }
1510
1511 /*
1512  * Truncate a byte range, avoiding pages within partial clusters. This
1513  * preserves those pages for the zeroing code to write to.
1514  */
1515 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1516                                          u64 byte_len)
1517 {
1518         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1519         loff_t start, end;
1520         struct address_space *mapping = inode->i_mapping;
1521
1522         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1523         end = byte_start + byte_len;
1524         end = end & ~(osb->s_clustersize - 1);
1525
1526         if (start < end) {
1527                 unmap_mapping_range(mapping, start, end - start, 0);
1528                 truncate_inode_pages_range(mapping, start, end - 1);
1529         }
1530 }
1531
1532 static int ocfs2_zero_partial_clusters(struct inode *inode,
1533                                        u64 start, u64 len)
1534 {
1535         int ret = 0;
1536         u64 tmpend, end = start + len;
1537         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1538         unsigned int csize = osb->s_clustersize;
1539         handle_t *handle;
1540
1541         /*
1542          * The "start" and "end" values are NOT necessarily part of
1543          * the range whose allocation is being deleted. Rather, this
1544          * is what the user passed in with the request. We must zero
1545          * partial clusters here. There's no need to worry about
1546          * physical allocation - the zeroing code knows to skip holes.
1547          */
1548         mlog(0, "byte start: %llu, end: %llu\n",
1549              (unsigned long long)start, (unsigned long long)end);
1550
1551         /*
1552          * If both edges are on a cluster boundary then there's no
1553          * zeroing required as the region is part of the allocation to
1554          * be truncated.
1555          */
1556         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1557                 goto out;
1558
1559         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1560         if (IS_ERR(handle)) {
1561                 ret = PTR_ERR(handle);
1562                 mlog_errno(ret);
1563                 goto out;
1564         }
1565
1566         /*
1567          * We want to get the byte offset of the end of the 1st cluster.
1568          */
1569         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1570         if (tmpend > end)
1571                 tmpend = end;
1572
1573         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1574              (unsigned long long)start, (unsigned long long)tmpend);
1575
1576         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1577         if (ret)
1578                 mlog_errno(ret);
1579
1580         if (tmpend < end) {
1581                 /*
1582                  * This may make start and end equal, but the zeroing
1583                  * code will skip any work in that case so there's no
1584                  * need to catch it up here.
1585                  */
1586                 start = end & ~(osb->s_clustersize - 1);
1587
1588                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1589                      (unsigned long long)start, (unsigned long long)end);
1590
1591                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1592                 if (ret)
1593                         mlog_errno(ret);
1594         }
1595
1596         ocfs2_commit_trans(osb, handle);
1597 out:
1598         return ret;
1599 }
1600
1601 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1602 {
1603         int i;
1604         struct ocfs2_extent_rec *rec = NULL;
1605
1606         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1607
1608                 rec = &el->l_recs[i];
1609
1610                 if (le32_to_cpu(rec->e_cpos) < pos)
1611                         break;
1612         }
1613
1614         return i;
1615 }
1616
1617 /*
1618  * Helper to calculate the punching pos and length in one run, we handle the
1619  * following three cases in order:
1620  *
1621  * - remove the entire record
1622  * - remove a partial record
1623  * - no record needs to be removed (hole-punching completed)
1624 */
1625 static void ocfs2_calc_trunc_pos(struct inode *inode,
1626                                  struct ocfs2_extent_list *el,
1627                                  struct ocfs2_extent_rec *rec,
1628                                  u32 trunc_start, u32 *trunc_cpos,
1629                                  u32 *trunc_len, u32 *trunc_end,
1630                                  u64 *blkno, int *done)
1631 {
1632         int ret = 0;
1633         u32 coff, range;
1634
1635         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1636
1637         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1638                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1639                 /*
1640                  * Skip holes if any.
1641                  */
1642                 if (range < *trunc_end)
1643                         *trunc_end = range;
1644                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1645                 *blkno = le64_to_cpu(rec->e_blkno);
1646                 *trunc_end = le32_to_cpu(rec->e_cpos);
1647         } else if (range > trunc_start) {
1648                 *trunc_cpos = trunc_start;
1649                 *trunc_len = *trunc_end - trunc_start;
1650                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1651                 *blkno = le64_to_cpu(rec->e_blkno) +
1652                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1653                 *trunc_end = trunc_start;
1654         } else {
1655                 /*
1656                  * It may have two following possibilities:
1657                  *
1658                  * - last record has been removed
1659                  * - trunc_start was within a hole
1660                  *
1661                  * both two cases mean the completion of hole punching.
1662                  */
1663                 ret = 1;
1664         }
1665
1666         *done = ret;
1667 }
1668
1669 static int ocfs2_remove_inode_range(struct inode *inode,
1670                                     struct buffer_head *di_bh, u64 byte_start,
1671                                     u64 byte_len)
1672 {
1673         int ret = 0, flags = 0, done = 0, i;
1674         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1675         u32 cluster_in_el;
1676         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1677         struct ocfs2_cached_dealloc_ctxt dealloc;
1678         struct address_space *mapping = inode->i_mapping;
1679         struct ocfs2_extent_tree et;
1680         struct ocfs2_path *path = NULL;
1681         struct ocfs2_extent_list *el = NULL;
1682         struct ocfs2_extent_rec *rec = NULL;
1683         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1684         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1685
1686         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1687         ocfs2_init_dealloc_ctxt(&dealloc);
1688
1689         if (byte_len == 0)
1690                 return 0;
1691
1692         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1693                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1694                                             byte_start + byte_len, 0);
1695                 if (ret) {
1696                         mlog_errno(ret);
1697                         goto out;
1698                 }
1699                 /*
1700                  * There's no need to get fancy with the page cache
1701                  * truncate of an inline-data inode. We're talking
1702                  * about less than a page here, which will be cached
1703                  * in the dinode buffer anyway.
1704                  */
1705                 unmap_mapping_range(mapping, 0, 0, 0);
1706                 truncate_inode_pages(mapping, 0);
1707                 goto out;
1708         }
1709
1710         /*
1711          * For reflinks, we may need to CoW 2 clusters which might be
1712          * partially zero'd later, if hole's start and end offset were
1713          * within one cluster(means is not exactly aligned to clustersize).
1714          */
1715
1716         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1717
1718                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1719                 if (ret) {
1720                         mlog_errno(ret);
1721                         goto out;
1722                 }
1723
1724                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1725                 if (ret) {
1726                         mlog_errno(ret);
1727                         goto out;
1728                 }
1729         }
1730
1731         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1732         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1733         cluster_in_el = trunc_end;
1734
1735         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1736              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1737              (unsigned long long)byte_start,
1738              (unsigned long long)byte_len, trunc_start, trunc_end);
1739
1740         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1741         if (ret) {
1742                 mlog_errno(ret);
1743                 goto out;
1744         }
1745
1746         path = ocfs2_new_path_from_et(&et);
1747         if (!path) {
1748                 ret = -ENOMEM;
1749                 mlog_errno(ret);
1750                 goto out;
1751         }
1752
1753         while (trunc_end > trunc_start) {
1754
1755                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1756                                       cluster_in_el);
1757                 if (ret) {
1758                         mlog_errno(ret);
1759                         goto out;
1760                 }
1761
1762                 el = path_leaf_el(path);
1763
1764                 i = ocfs2_find_rec(el, trunc_end);
1765                 /*
1766                  * Need to go to previous extent block.
1767                  */
1768                 if (i < 0) {
1769                         if (path->p_tree_depth == 0)
1770                                 break;
1771
1772                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1773                                                             path,
1774                                                             &cluster_in_el);
1775                         if (ret) {
1776                                 mlog_errno(ret);
1777                                 goto out;
1778                         }
1779
1780                         /*
1781                          * We've reached the leftmost extent block,
1782                          * it's safe to leave.
1783                          */
1784                         if (cluster_in_el == 0)
1785                                 break;
1786
1787                         /*
1788                          * The 'pos' searched for previous extent block is
1789                          * always one cluster less than actual trunc_end.
1790                          */
1791                         trunc_end = cluster_in_el + 1;
1792
1793                         ocfs2_reinit_path(path, 1);
1794
1795                         continue;
1796
1797                 } else
1798                         rec = &el->l_recs[i];
1799
1800                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1801                                      &trunc_len, &trunc_end, &blkno, &done);
1802                 if (done)
1803                         break;
1804
1805                 flags = rec->e_flags;
1806                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1807
1808                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1809                                                phys_cpos, trunc_len, flags,
1810                                                &dealloc, refcount_loc);
1811                 if (ret < 0) {
1812                         mlog_errno(ret);
1813                         goto out;
1814                 }
1815
1816                 cluster_in_el = trunc_end;
1817
1818                 ocfs2_reinit_path(path, 1);
1819         }
1820
1821         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1822
1823 out:
1824         ocfs2_schedule_truncate_log_flush(osb, 1);
1825         ocfs2_run_deallocs(osb, &dealloc);
1826
1827         return ret;
1828 }
1829
1830 /*
1831  * Parts of this function taken from xfs_change_file_space()
1832  */
1833 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1834                                      loff_t f_pos, unsigned int cmd,
1835                                      struct ocfs2_space_resv *sr,
1836                                      int change_size)
1837 {
1838         int ret;
1839         s64 llen;
1840         loff_t size;
1841         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1842         struct buffer_head *di_bh = NULL;
1843         handle_t *handle;
1844         unsigned long long max_off = inode->i_sb->s_maxbytes;
1845
1846         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1847                 return -EROFS;
1848
1849         mutex_lock(&inode->i_mutex);
1850
1851         /*
1852          * This prevents concurrent writes on other nodes
1853          */
1854         ret = ocfs2_rw_lock(inode, 1);
1855         if (ret) {
1856                 mlog_errno(ret);
1857                 goto out;
1858         }
1859
1860         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1861         if (ret) {
1862                 mlog_errno(ret);
1863                 goto out_rw_unlock;
1864         }
1865
1866         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1867                 ret = -EPERM;
1868                 goto out_inode_unlock;
1869         }
1870
1871         switch (sr->l_whence) {
1872         case 0: /*SEEK_SET*/
1873                 break;
1874         case 1: /*SEEK_CUR*/
1875                 sr->l_start += f_pos;
1876                 break;
1877         case 2: /*SEEK_END*/
1878                 sr->l_start += i_size_read(inode);
1879                 break;
1880         default:
1881                 ret = -EINVAL;
1882                 goto out_inode_unlock;
1883         }
1884         sr->l_whence = 0;
1885
1886         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1887
1888         if (sr->l_start < 0
1889             || sr->l_start > max_off
1890             || (sr->l_start + llen) < 0
1891             || (sr->l_start + llen) > max_off) {
1892                 ret = -EINVAL;
1893                 goto out_inode_unlock;
1894         }
1895         size = sr->l_start + sr->l_len;
1896
1897         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1898                 if (sr->l_len <= 0) {
1899                         ret = -EINVAL;
1900                         goto out_inode_unlock;
1901                 }
1902         }
1903
1904         if (file && should_remove_suid(file->f_path.dentry)) {
1905                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1906                 if (ret) {
1907                         mlog_errno(ret);
1908                         goto out_inode_unlock;
1909                 }
1910         }
1911
1912         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1913         switch (cmd) {
1914         case OCFS2_IOC_RESVSP:
1915         case OCFS2_IOC_RESVSP64:
1916                 /*
1917                  * This takes unsigned offsets, but the signed ones we
1918                  * pass have been checked against overflow above.
1919                  */
1920                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1921                                                        sr->l_len);
1922                 break;
1923         case OCFS2_IOC_UNRESVSP:
1924         case OCFS2_IOC_UNRESVSP64:
1925                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1926                                                sr->l_len);
1927                 break;
1928         default:
1929                 ret = -EINVAL;
1930         }
1931         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1932         if (ret) {
1933                 mlog_errno(ret);
1934                 goto out_inode_unlock;
1935         }
1936
1937         /*
1938          * We update c/mtime for these changes
1939          */
1940         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1941         if (IS_ERR(handle)) {
1942                 ret = PTR_ERR(handle);
1943                 mlog_errno(ret);
1944                 goto out_inode_unlock;
1945         }
1946
1947         if (change_size && i_size_read(inode) < size)
1948                 i_size_write(inode, size);
1949
1950         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1951         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1952         if (ret < 0)
1953                 mlog_errno(ret);
1954
1955         ocfs2_commit_trans(osb, handle);
1956
1957 out_inode_unlock:
1958         brelse(di_bh);
1959         ocfs2_inode_unlock(inode, 1);
1960 out_rw_unlock:
1961         ocfs2_rw_unlock(inode, 1);
1962
1963 out:
1964         mutex_unlock(&inode->i_mutex);
1965         return ret;
1966 }
1967
1968 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1969                             struct ocfs2_space_resv *sr)
1970 {
1971         struct inode *inode = file->f_path.dentry->d_inode;
1972         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1973
1974         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1975             !ocfs2_writes_unwritten_extents(osb))
1976                 return -ENOTTY;
1977         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1978                  !ocfs2_sparse_alloc(osb))
1979                 return -ENOTTY;
1980
1981         if (!S_ISREG(inode->i_mode))
1982                 return -EINVAL;
1983
1984         if (!(file->f_mode & FMODE_WRITE))
1985                 return -EBADF;
1986
1987         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1988 }
1989
1990 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1991                             loff_t len)
1992 {
1993         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1994         struct ocfs2_space_resv sr;
1995         int change_size = 1;
1996
1997         if (!ocfs2_writes_unwritten_extents(osb))
1998                 return -EOPNOTSUPP;
1999
2000         if (S_ISDIR(inode->i_mode))
2001                 return -ENODEV;
2002
2003         if (mode & FALLOC_FL_KEEP_SIZE)
2004                 change_size = 0;
2005
2006         sr.l_whence = 0;
2007         sr.l_start = (s64)offset;
2008         sr.l_len = (s64)len;
2009
2010         return __ocfs2_change_file_space(NULL, inode, offset,
2011                                          OCFS2_IOC_RESVSP64, &sr, change_size);
2012 }
2013
2014 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2015                                    size_t count)
2016 {
2017         int ret = 0;
2018         unsigned int extent_flags;
2019         u32 cpos, clusters, extent_len, phys_cpos;
2020         struct super_block *sb = inode->i_sb;
2021
2022         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2023             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2024             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2025                 return 0;
2026
2027         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2028         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2029
2030         while (clusters) {
2031                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2032                                          &extent_flags);
2033                 if (ret < 0) {
2034                         mlog_errno(ret);
2035                         goto out;
2036                 }
2037
2038                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2039                         ret = 1;
2040                         break;
2041                 }
2042
2043                 if (extent_len > clusters)
2044                         extent_len = clusters;
2045
2046                 clusters -= extent_len;
2047                 cpos += extent_len;
2048         }
2049 out:
2050         return ret;
2051 }
2052
2053 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2054                                             struct file *file,
2055                                             loff_t pos, size_t count,
2056                                             int *meta_level)
2057 {
2058         int ret;
2059         struct buffer_head *di_bh = NULL;
2060         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2061         u32 clusters =
2062                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2063
2064         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2065         if (ret) {
2066                 mlog_errno(ret);
2067                 goto out;
2068         }
2069
2070         *meta_level = 1;
2071
2072         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2073         if (ret)
2074                 mlog_errno(ret);
2075 out:
2076         brelse(di_bh);
2077         return ret;
2078 }
2079
2080 static int ocfs2_prepare_inode_for_write(struct file *file,
2081                                          loff_t *ppos,
2082                                          size_t count,
2083                                          int appending,
2084                                          int *direct_io,
2085                                          int *has_refcount)
2086 {
2087         int ret = 0, meta_level = 0;
2088         struct dentry *dentry = file->f_path.dentry;
2089         struct inode *inode = dentry->d_inode;
2090         loff_t saved_pos, end;
2091
2092         /*
2093          * We start with a read level meta lock and only jump to an ex
2094          * if we need to make modifications here.
2095          */
2096         for(;;) {
2097                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2098                 if (ret < 0) {
2099                         meta_level = -1;
2100                         mlog_errno(ret);
2101                         goto out;
2102                 }
2103
2104                 /* Clear suid / sgid if necessary. We do this here
2105                  * instead of later in the write path because
2106                  * remove_suid() calls ->setattr without any hint that
2107                  * we may have already done our cluster locking. Since
2108                  * ocfs2_setattr() *must* take cluster locks to
2109                  * proceeed, this will lead us to recursively lock the
2110                  * inode. There's also the dinode i_size state which
2111                  * can be lost via setattr during extending writes (we
2112                  * set inode->i_size at the end of a write. */
2113                 if (should_remove_suid(dentry)) {
2114                         if (meta_level == 0) {
2115                                 ocfs2_inode_unlock(inode, meta_level);
2116                                 meta_level = 1;
2117                                 continue;
2118                         }
2119
2120                         ret = ocfs2_write_remove_suid(inode);
2121                         if (ret < 0) {
2122                                 mlog_errno(ret);
2123                                 goto out_unlock;
2124                         }
2125                 }
2126
2127                 /* work on a copy of ppos until we're sure that we won't have
2128                  * to recalculate it due to relocking. */
2129                 if (appending) {
2130                         saved_pos = i_size_read(inode);
2131                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2132                 } else {
2133                         saved_pos = *ppos;
2134                 }
2135
2136                 end = saved_pos + count;
2137
2138                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2139                 if (ret == 1) {
2140                         ocfs2_inode_unlock(inode, meta_level);
2141                         meta_level = -1;
2142
2143                         ret = ocfs2_prepare_inode_for_refcount(inode,
2144                                                                file,
2145                                                                saved_pos,
2146                                                                count,
2147                                                                &meta_level);
2148                         if (has_refcount)
2149                                 *has_refcount = 1;
2150                         if (direct_io)
2151                                 *direct_io = 0;
2152                 }
2153
2154                 if (ret < 0) {
2155                         mlog_errno(ret);
2156                         goto out_unlock;
2157                 }
2158
2159                 /*
2160                  * Skip the O_DIRECT checks if we don't need
2161                  * them.
2162                  */
2163                 if (!direct_io || !(*direct_io))
2164                         break;
2165
2166                 /*
2167                  * There's no sane way to do direct writes to an inode
2168                  * with inline data.
2169                  */
2170                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2171                         *direct_io = 0;
2172                         break;
2173                 }
2174
2175                 /*
2176                  * Allowing concurrent direct writes means
2177                  * i_size changes wouldn't be synchronized, so
2178                  * one node could wind up truncating another
2179                  * nodes writes.
2180                  */
2181                 if (end > i_size_read(inode)) {
2182                         *direct_io = 0;
2183                         break;
2184                 }
2185
2186                 /*
2187                  * We don't fill holes during direct io, so
2188                  * check for them here. If any are found, the
2189                  * caller will have to retake some cluster
2190                  * locks and initiate the io as buffered.
2191                  */
2192                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2193                 if (ret == 1) {
2194                         *direct_io = 0;
2195                         ret = 0;
2196                 } else if (ret < 0)
2197                         mlog_errno(ret);
2198                 break;
2199         }
2200
2201         if (appending)
2202                 *ppos = saved_pos;
2203
2204 out_unlock:
2205         if (meta_level >= 0)
2206                 ocfs2_inode_unlock(inode, meta_level);
2207
2208 out:
2209         return ret;
2210 }
2211
2212 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2213                                     const struct iovec *iov,
2214                                     unsigned long nr_segs,
2215                                     loff_t pos)
2216 {
2217         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2218         int can_do_direct, has_refcount = 0;
2219         ssize_t written = 0;
2220         size_t ocount;          /* original count */
2221         size_t count;           /* after file limit checks */
2222         loff_t old_size, *ppos = &iocb->ki_pos;
2223         u32 old_clusters;
2224         struct file *file = iocb->ki_filp;
2225         struct inode *inode = file->f_path.dentry->d_inode;
2226         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2227         int full_coherency = !(osb->s_mount_opt &
2228                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2229
2230         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2231                    (unsigned int)nr_segs,
2232                    file->f_path.dentry->d_name.len,
2233                    file->f_path.dentry->d_name.name);
2234
2235         if (iocb->ki_left == 0)
2236                 return 0;
2237
2238         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2239
2240         appending = file->f_flags & O_APPEND ? 1 : 0;
2241         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2242
2243         mutex_lock(&inode->i_mutex);
2244
2245 relock:
2246         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2247         if (direct_io) {
2248                 down_read(&inode->i_alloc_sem);
2249                 have_alloc_sem = 1;
2250         }
2251
2252         /*
2253          * Concurrent O_DIRECT writes are allowed with
2254          * mount_option "coherency=buffered".
2255          */
2256         rw_level = (!direct_io || full_coherency);
2257
2258         ret = ocfs2_rw_lock(inode, rw_level);
2259         if (ret < 0) {
2260                 mlog_errno(ret);
2261                 goto out_sems;
2262         }
2263
2264         /*
2265          * O_DIRECT writes with "coherency=full" need to take EX cluster
2266          * inode_lock to guarantee coherency.
2267          */
2268         if (direct_io && full_coherency) {
2269                 /*
2270                  * We need to take and drop the inode lock to force
2271                  * other nodes to drop their caches.  Buffered I/O
2272                  * already does this in write_begin().
2273                  */
2274                 ret = ocfs2_inode_lock(inode, NULL, 1);
2275                 if (ret < 0) {
2276                         mlog_errno(ret);
2277                         goto out_sems;
2278                 }
2279
2280                 ocfs2_inode_unlock(inode, 1);
2281         }
2282
2283         can_do_direct = direct_io;
2284         ret = ocfs2_prepare_inode_for_write(file, ppos,
2285                                             iocb->ki_left, appending,
2286                                             &can_do_direct, &has_refcount);
2287         if (ret < 0) {
2288                 mlog_errno(ret);
2289                 goto out;
2290         }
2291
2292         /*
2293          * We can't complete the direct I/O as requested, fall back to
2294          * buffered I/O.
2295          */
2296         if (direct_io && !can_do_direct) {
2297                 ocfs2_rw_unlock(inode, rw_level);
2298                 up_read(&inode->i_alloc_sem);
2299
2300                 have_alloc_sem = 0;
2301                 rw_level = -1;
2302
2303                 direct_io = 0;
2304                 goto relock;
2305         }
2306
2307         /*
2308          * To later detect whether a journal commit for sync writes is
2309          * necessary, we sample i_size, and cluster count here.
2310          */
2311         old_size = i_size_read(inode);
2312         old_clusters = OCFS2_I(inode)->ip_clusters;
2313
2314         /* communicate with ocfs2_dio_end_io */
2315         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2316
2317         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2318                                      VERIFY_READ);
2319         if (ret)
2320                 goto out_dio;
2321
2322         count = ocount;
2323         ret = generic_write_checks(file, ppos, &count,
2324                                    S_ISBLK(inode->i_mode));
2325         if (ret)
2326                 goto out_dio;
2327
2328         if (direct_io) {
2329                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2330                                                     ppos, count, ocount);
2331                 if (written < 0) {
2332                         ret = written;
2333                         goto out_dio;
2334                 }
2335         } else {
2336                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2337                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2338                                                       ppos, count, 0);
2339                 current->backing_dev_info = NULL;
2340         }
2341
2342 out_dio:
2343         /* buffered aio wouldn't have proper lock coverage today */
2344         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2345
2346         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2347             ((file->f_flags & O_DIRECT) && !direct_io)) {
2348                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2349                                                pos + count - 1);
2350                 if (ret < 0)
2351                         written = ret;
2352
2353                 if (!ret && ((old_size != i_size_read(inode)) ||
2354                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2355                              has_refcount)) {
2356                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2357                         if (ret < 0)
2358                                 written = ret;
2359                 }
2360
2361                 if (!ret)
2362                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2363                                                       pos + count - 1);
2364         }
2365
2366         /*
2367          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2368          * function pointer which is called when o_direct io completes so that
2369          * it can unlock our rw lock.  (it's the clustered equivalent of
2370          * i_alloc_sem; protects truncate from racing with pending ios).
2371          * Unfortunately there are error cases which call end_io and others
2372          * that don't.  so we don't have to unlock the rw_lock if either an
2373          * async dio is going to do it in the future or an end_io after an
2374          * error has already done it.
2375          */
2376         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2377                 rw_level = -1;
2378                 have_alloc_sem = 0;
2379         }
2380
2381 out:
2382         if (rw_level != -1)
2383                 ocfs2_rw_unlock(inode, rw_level);
2384
2385 out_sems:
2386         if (have_alloc_sem)
2387                 up_read(&inode->i_alloc_sem);
2388
2389         mutex_unlock(&inode->i_mutex);
2390
2391         if (written)
2392                 ret = written;
2393         mlog_exit(ret);
2394         return ret;
2395 }
2396
2397 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2398                                 struct file *out,
2399                                 struct splice_desc *sd)
2400 {
2401         int ret;
2402
2403         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2404                                             sd->total_len, 0, NULL, NULL);
2405         if (ret < 0) {
2406                 mlog_errno(ret);
2407                 return ret;
2408         }
2409
2410         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2411 }
2412
2413 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2414                                        struct file *out,
2415                                        loff_t *ppos,
2416                                        size_t len,
2417                                        unsigned int flags)
2418 {
2419         int ret;
2420         struct address_space *mapping = out->f_mapping;
2421         struct inode *inode = mapping->host;
2422         struct splice_desc sd = {
2423                 .total_len = len,
2424                 .flags = flags,
2425                 .pos = *ppos,
2426                 .u.file = out,
2427         };
2428
2429         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2430                    (unsigned int)len,
2431                    out->f_path.dentry->d_name.len,
2432                    out->f_path.dentry->d_name.name);
2433
2434         if (pipe->inode)
2435                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2436
2437         splice_from_pipe_begin(&sd);
2438         do {
2439                 ret = splice_from_pipe_next(pipe, &sd);
2440                 if (ret <= 0)
2441                         break;
2442
2443                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2444                 ret = ocfs2_rw_lock(inode, 1);
2445                 if (ret < 0)
2446                         mlog_errno(ret);
2447                 else {
2448                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2449                         ocfs2_rw_unlock(inode, 1);
2450                 }
2451                 mutex_unlock(&inode->i_mutex);
2452         } while (ret > 0);
2453         splice_from_pipe_end(pipe, &sd);
2454
2455         if (pipe->inode)
2456                 mutex_unlock(&pipe->inode->i_mutex);
2457
2458         if (sd.num_spliced)
2459                 ret = sd.num_spliced;
2460
2461         if (ret > 0) {
2462                 unsigned long nr_pages;
2463                 int err;
2464
2465                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2466
2467                 err = generic_write_sync(out, *ppos, ret);
2468                 if (err)
2469                         ret = err;
2470                 else
2471                         *ppos += ret;
2472
2473                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2474         }
2475
2476         mlog_exit(ret);
2477         return ret;
2478 }
2479
2480 static ssize_t ocfs2_file_splice_read(struct file *in,
2481                                       loff_t *ppos,
2482                                       struct pipe_inode_info *pipe,
2483                                       size_t len,
2484                                       unsigned int flags)
2485 {
2486         int ret = 0, lock_level = 0;
2487         struct inode *inode = in->f_path.dentry->d_inode;
2488
2489         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2490                    (unsigned int)len,
2491                    in->f_path.dentry->d_name.len,
2492                    in->f_path.dentry->d_name.name);
2493
2494         /*
2495          * See the comment in ocfs2_file_aio_read()
2496          */
2497         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2498         if (ret < 0) {
2499                 mlog_errno(ret);
2500                 goto bail;
2501         }
2502         ocfs2_inode_unlock(inode, lock_level);
2503
2504         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2505
2506 bail:
2507         mlog_exit(ret);
2508         return ret;
2509 }
2510
2511 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2512                                    const struct iovec *iov,
2513                                    unsigned long nr_segs,
2514                                    loff_t pos)
2515 {
2516         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2517         struct file *filp = iocb->ki_filp;
2518         struct inode *inode = filp->f_path.dentry->d_inode;
2519
2520         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2521                    (unsigned int)nr_segs,
2522                    filp->f_path.dentry->d_name.len,
2523                    filp->f_path.dentry->d_name.name);
2524
2525         if (!inode) {
2526                 ret = -EINVAL;
2527                 mlog_errno(ret);
2528                 goto bail;
2529         }
2530
2531         /*
2532          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2533          * need locks to protect pending reads from racing with truncate.
2534          */
2535         if (filp->f_flags & O_DIRECT) {
2536                 down_read(&inode->i_alloc_sem);
2537                 have_alloc_sem = 1;
2538
2539                 ret = ocfs2_rw_lock(inode, 0);
2540                 if (ret < 0) {
2541                         mlog_errno(ret);
2542                         goto bail;
2543                 }
2544                 rw_level = 0;
2545                 /* communicate with ocfs2_dio_end_io */
2546                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2547         }
2548
2549         /*
2550          * We're fine letting folks race truncates and extending
2551          * writes with read across the cluster, just like they can
2552          * locally. Hence no rw_lock during read.
2553          *
2554          * Take and drop the meta data lock to update inode fields
2555          * like i_size. This allows the checks down below
2556          * generic_file_aio_read() a chance of actually working.
2557          */
2558         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2559         if (ret < 0) {
2560                 mlog_errno(ret);
2561                 goto bail;
2562         }
2563         ocfs2_inode_unlock(inode, lock_level);
2564
2565         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2566         if (ret == -EINVAL)
2567                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2568
2569         /* buffered aio wouldn't have proper lock coverage today */
2570         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2571
2572         /* see ocfs2_file_aio_write */
2573         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2574                 rw_level = -1;
2575                 have_alloc_sem = 0;
2576         }
2577
2578 bail:
2579         if (have_alloc_sem)
2580                 up_read(&inode->i_alloc_sem);
2581         if (rw_level != -1)
2582                 ocfs2_rw_unlock(inode, rw_level);
2583         mlog_exit(ret);
2584
2585         return ret;
2586 }
2587
2588 const struct inode_operations ocfs2_file_iops = {
2589         .setattr        = ocfs2_setattr,
2590         .getattr        = ocfs2_getattr,
2591         .permission     = ocfs2_permission,
2592         .setxattr       = generic_setxattr,
2593         .getxattr       = generic_getxattr,
2594         .listxattr      = ocfs2_listxattr,
2595         .removexattr    = generic_removexattr,
2596         .fallocate      = ocfs2_fallocate,
2597         .fiemap         = ocfs2_fiemap,
2598 };
2599
2600 const struct inode_operations ocfs2_special_file_iops = {
2601         .setattr        = ocfs2_setattr,
2602         .getattr        = ocfs2_getattr,
2603         .permission     = ocfs2_permission,
2604 };
2605
2606 /*
2607  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2608  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2609  */
2610 const struct file_operations ocfs2_fops = {
2611         .llseek         = generic_file_llseek,
2612         .read           = do_sync_read,
2613         .write          = do_sync_write,
2614         .mmap           = ocfs2_mmap,
2615         .fsync          = ocfs2_sync_file,
2616         .release        = ocfs2_file_release,
2617         .open           = ocfs2_file_open,
2618         .aio_read       = ocfs2_file_aio_read,
2619         .aio_write      = ocfs2_file_aio_write,
2620         .unlocked_ioctl = ocfs2_ioctl,
2621 #ifdef CONFIG_COMPAT
2622         .compat_ioctl   = ocfs2_compat_ioctl,
2623 #endif
2624         .lock           = ocfs2_lock,
2625         .flock          = ocfs2_flock,
2626         .splice_read    = ocfs2_file_splice_read,
2627         .splice_write   = ocfs2_file_splice_write,
2628 };
2629
2630 const struct file_operations ocfs2_dops = {
2631         .llseek         = generic_file_llseek,
2632         .read           = generic_read_dir,
2633         .readdir        = ocfs2_readdir,
2634         .fsync          = ocfs2_sync_file,
2635         .release        = ocfs2_dir_release,
2636         .open           = ocfs2_dir_open,
2637         .unlocked_ioctl = ocfs2_ioctl,
2638 #ifdef CONFIG_COMPAT
2639         .compat_ioctl   = ocfs2_compat_ioctl,
2640 #endif
2641         .lock           = ocfs2_lock,
2642         .flock          = ocfs2_flock,
2643 };
2644
2645 /*
2646  * POSIX-lockless variants of our file_operations.
2647  *
2648  * These will be used if the underlying cluster stack does not support
2649  * posix file locking, if the user passes the "localflocks" mount
2650  * option, or if we have a local-only fs.
2651  *
2652  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2653  * so we still want it in the case of no stack support for
2654  * plocks. Internally, it will do the right thing when asked to ignore
2655  * the cluster.
2656  */
2657 const struct file_operations ocfs2_fops_no_plocks = {
2658         .llseek         = generic_file_llseek,
2659         .read           = do_sync_read,
2660         .write          = do_sync_write,
2661         .mmap           = ocfs2_mmap,
2662         .fsync          = ocfs2_sync_file,
2663         .release        = ocfs2_file_release,
2664         .open           = ocfs2_file_open,
2665         .aio_read       = ocfs2_file_aio_read,
2666         .aio_write      = ocfs2_file_aio_write,
2667         .unlocked_ioctl = ocfs2_ioctl,
2668 #ifdef CONFIG_COMPAT
2669         .compat_ioctl   = ocfs2_compat_ioctl,
2670 #endif
2671         .flock          = ocfs2_flock,
2672         .splice_read    = ocfs2_file_splice_read,
2673         .splice_write   = ocfs2_file_splice_write,
2674 };
2675
2676 const struct file_operations ocfs2_dops_no_plocks = {
2677         .llseek         = generic_file_llseek,
2678         .read           = generic_read_dir,
2679         .readdir        = ocfs2_readdir,
2680         .fsync          = ocfs2_sync_file,
2681         .release        = ocfs2_dir_release,
2682         .open           = ocfs2_dir_open,
2683         .unlocked_ioctl = ocfs2_ioctl,
2684 #ifdef CONFIG_COMPAT
2685         .compat_ioctl   = ocfs2_compat_ioctl,
2686 #endif
2687         .flock          = ocfs2_flock,
2688 };