ocfs2: Flush drive's caches on fdatasync
[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_sync_inode(struct inode *inode)
68 {
69         filemap_fdatawrite(inode->i_mapping);
70         return sync_mapping_buffers(inode->i_mapping);
71 }
72
73 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
74 {
75         struct ocfs2_file_private *fp;
76
77         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
78         if (!fp)
79                 return -ENOMEM;
80
81         fp->fp_file = file;
82         mutex_init(&fp->fp_mutex);
83         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
84         file->private_data = fp;
85
86         return 0;
87 }
88
89 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
90 {
91         struct ocfs2_file_private *fp = file->private_data;
92         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
93
94         if (fp) {
95                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
96                 ocfs2_lock_res_free(&fp->fp_flock);
97                 kfree(fp);
98                 file->private_data = NULL;
99         }
100 }
101
102 static int ocfs2_file_open(struct inode *inode, struct file *file)
103 {
104         int status;
105         int mode = file->f_flags;
106         struct ocfs2_inode_info *oi = OCFS2_I(inode);
107
108         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
109                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
110
111         if (file->f_mode & FMODE_WRITE)
112                 dquot_initialize(inode);
113
114         spin_lock(&oi->ip_lock);
115
116         /* Check that the inode hasn't been wiped from disk by another
117          * node. If it hasn't then we're safe as long as we hold the
118          * spin lock until our increment of open count. */
119         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
120                 spin_unlock(&oi->ip_lock);
121
122                 status = -ENOENT;
123                 goto leave;
124         }
125
126         if (mode & O_DIRECT)
127                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
128
129         oi->ip_open_count++;
130         spin_unlock(&oi->ip_lock);
131
132         status = ocfs2_init_file_private(inode, file);
133         if (status) {
134                 /*
135                  * We want to set open count back if we're failing the
136                  * open.
137                  */
138                 spin_lock(&oi->ip_lock);
139                 oi->ip_open_count--;
140                 spin_unlock(&oi->ip_lock);
141         }
142
143 leave:
144         mlog_exit(status);
145         return status;
146 }
147
148 static int ocfs2_file_release(struct inode *inode, struct file *file)
149 {
150         struct ocfs2_inode_info *oi = OCFS2_I(inode);
151
152         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
153                        file->f_path.dentry->d_name.len,
154                        file->f_path.dentry->d_name.name);
155
156         spin_lock(&oi->ip_lock);
157         if (!--oi->ip_open_count)
158                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
159         spin_unlock(&oi->ip_lock);
160
161         ocfs2_free_file_private(inode, file);
162
163         mlog_exit(0);
164
165         return 0;
166 }
167
168 static int ocfs2_dir_open(struct inode *inode, struct file *file)
169 {
170         return ocfs2_init_file_private(inode, file);
171 }
172
173 static int ocfs2_dir_release(struct inode *inode, struct file *file)
174 {
175         ocfs2_free_file_private(inode, file);
176         return 0;
177 }
178
179 static int ocfs2_sync_file(struct file *file, int datasync)
180 {
181         int err = 0;
182         journal_t *journal;
183         struct dentry *dentry = file->f_path.dentry;
184         struct inode *inode = file->f_mapping->host;
185         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
186
187         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
188                    dentry->d_name.len, dentry->d_name.name);
189
190         err = ocfs2_sync_inode(dentry->d_inode);
191         if (err)
192                 goto bail;
193
194         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
195                 /*
196                  * We still have to flush drive's caches to get data to the
197                  * platter
198                  */
199                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
200                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL,
201                                            NULL, BLKDEV_IFL_WAIT);
202                 goto bail;
203         }
204
205         journal = osb->journal->j_journal;
206         err = jbd2_journal_force_commit(journal);
207
208 bail:
209         mlog_exit(err);
210
211         return (err < 0) ? -EIO : 0;
212 }
213
214 int ocfs2_should_update_atime(struct inode *inode,
215                               struct vfsmount *vfsmnt)
216 {
217         struct timespec now;
218         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
219
220         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
221                 return 0;
222
223         if ((inode->i_flags & S_NOATIME) ||
224             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
225                 return 0;
226
227         /*
228          * We can be called with no vfsmnt structure - NFSD will
229          * sometimes do this.
230          *
231          * Note that our action here is different than touch_atime() -
232          * if we can't tell whether this is a noatime mount, then we
233          * don't know whether to trust the value of s_atime_quantum.
234          */
235         if (vfsmnt == NULL)
236                 return 0;
237
238         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
239             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
240                 return 0;
241
242         if (vfsmnt->mnt_flags & MNT_RELATIME) {
243                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
244                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
245                         return 1;
246
247                 return 0;
248         }
249
250         now = CURRENT_TIME;
251         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
252                 return 0;
253         else
254                 return 1;
255 }
256
257 int ocfs2_update_inode_atime(struct inode *inode,
258                              struct buffer_head *bh)
259 {
260         int ret;
261         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
262         handle_t *handle;
263         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
264
265         mlog_entry_void();
266
267         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
268         if (IS_ERR(handle)) {
269                 ret = PTR_ERR(handle);
270                 mlog_errno(ret);
271                 goto out;
272         }
273
274         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
275                                       OCFS2_JOURNAL_ACCESS_WRITE);
276         if (ret) {
277                 mlog_errno(ret);
278                 goto out_commit;
279         }
280
281         /*
282          * Don't use ocfs2_mark_inode_dirty() here as we don't always
283          * have i_mutex to guard against concurrent changes to other
284          * inode fields.
285          */
286         inode->i_atime = CURRENT_TIME;
287         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
288         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
289         ocfs2_journal_dirty(handle, bh);
290
291 out_commit:
292         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
293 out:
294         mlog_exit(ret);
295         return ret;
296 }
297
298 static int ocfs2_set_inode_size(handle_t *handle,
299                                 struct inode *inode,
300                                 struct buffer_head *fe_bh,
301                                 u64 new_i_size)
302 {
303         int status;
304
305         mlog_entry_void();
306         i_size_write(inode, new_i_size);
307         inode->i_blocks = ocfs2_inode_sector_count(inode);
308         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
309
310         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
311         if (status < 0) {
312                 mlog_errno(status);
313                 goto bail;
314         }
315
316 bail:
317         mlog_exit(status);
318         return status;
319 }
320
321 int ocfs2_simple_size_update(struct inode *inode,
322                              struct buffer_head *di_bh,
323                              u64 new_i_size)
324 {
325         int ret;
326         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
327         handle_t *handle = NULL;
328
329         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
330         if (IS_ERR(handle)) {
331                 ret = PTR_ERR(handle);
332                 mlog_errno(ret);
333                 goto out;
334         }
335
336         ret = ocfs2_set_inode_size(handle, inode, di_bh,
337                                    new_i_size);
338         if (ret < 0)
339                 mlog_errno(ret);
340
341         ocfs2_commit_trans(osb, handle);
342 out:
343         return ret;
344 }
345
346 static int ocfs2_cow_file_pos(struct inode *inode,
347                               struct buffer_head *fe_bh,
348                               u64 offset)
349 {
350         int status;
351         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
352         unsigned int num_clusters = 0;
353         unsigned int ext_flags = 0;
354
355         /*
356          * If the new offset is aligned to the range of the cluster, there is
357          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
358          * CoW either.
359          */
360         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
361                 return 0;
362
363         status = ocfs2_get_clusters(inode, cpos, &phys,
364                                     &num_clusters, &ext_flags);
365         if (status) {
366                 mlog_errno(status);
367                 goto out;
368         }
369
370         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
371                 goto out;
372
373         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
374
375 out:
376         return status;
377 }
378
379 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
380                                      struct inode *inode,
381                                      struct buffer_head *fe_bh,
382                                      u64 new_i_size)
383 {
384         int status;
385         handle_t *handle;
386         struct ocfs2_dinode *di;
387         u64 cluster_bytes;
388
389         mlog_entry_void();
390
391         /*
392          * We need to CoW the cluster contains the offset if it is reflinked
393          * since we will call ocfs2_zero_range_for_truncate later which will
394          * write "0" from offset to the end of the cluster.
395          */
396         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
397         if (status) {
398                 mlog_errno(status);
399                 return status;
400         }
401
402         /* TODO: This needs to actually orphan the inode in this
403          * transaction. */
404
405         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
406         if (IS_ERR(handle)) {
407                 status = PTR_ERR(handle);
408                 mlog_errno(status);
409                 goto out;
410         }
411
412         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
413                                          OCFS2_JOURNAL_ACCESS_WRITE);
414         if (status < 0) {
415                 mlog_errno(status);
416                 goto out_commit;
417         }
418
419         /*
420          * Do this before setting i_size.
421          */
422         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
423         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
424                                                cluster_bytes);
425         if (status) {
426                 mlog_errno(status);
427                 goto out_commit;
428         }
429
430         i_size_write(inode, new_i_size);
431         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
432
433         di = (struct ocfs2_dinode *) fe_bh->b_data;
434         di->i_size = cpu_to_le64(new_i_size);
435         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
436         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
437
438         ocfs2_journal_dirty(handle, fe_bh);
439
440 out_commit:
441         ocfs2_commit_trans(osb, handle);
442 out:
443
444         mlog_exit(status);
445         return status;
446 }
447
448 static int ocfs2_truncate_file(struct inode *inode,
449                                struct buffer_head *di_bh,
450                                u64 new_i_size)
451 {
452         int status = 0;
453         struct ocfs2_dinode *fe = NULL;
454         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
455
456         mlog_entry("(inode = %llu, new_i_size = %llu\n",
457                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
458                    (unsigned long long)new_i_size);
459
460         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
461          * already validated it */
462         fe = (struct ocfs2_dinode *) di_bh->b_data;
463
464         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
465                         "Inode %llu, inode i_size = %lld != di "
466                         "i_size = %llu, i_flags = 0x%x\n",
467                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
468                         i_size_read(inode),
469                         (unsigned long long)le64_to_cpu(fe->i_size),
470                         le32_to_cpu(fe->i_flags));
471
472         if (new_i_size > le64_to_cpu(fe->i_size)) {
473                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
474                      (unsigned long long)le64_to_cpu(fe->i_size),
475                      (unsigned long long)new_i_size);
476                 status = -EINVAL;
477                 mlog_errno(status);
478                 goto bail;
479         }
480
481         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
482              (unsigned long long)le64_to_cpu(fe->i_blkno),
483              (unsigned long long)le64_to_cpu(fe->i_size),
484              (unsigned long long)new_i_size);
485
486         /* lets handle the simple truncate cases before doing any more
487          * cluster locking. */
488         if (new_i_size == le64_to_cpu(fe->i_size))
489                 goto bail;
490
491         down_write(&OCFS2_I(inode)->ip_alloc_sem);
492
493         ocfs2_resv_discard(&osb->osb_la_resmap,
494                            &OCFS2_I(inode)->ip_la_data_resv);
495
496         /*
497          * The inode lock forced other nodes to sync and drop their
498          * pages, which (correctly) happens even if we have a truncate
499          * without allocation change - ocfs2 cluster sizes can be much
500          * greater than page size, so we have to truncate them
501          * anyway.
502          */
503         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
504         truncate_inode_pages(inode->i_mapping, new_i_size);
505
506         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
507                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
508                                                i_size_read(inode), 1);
509                 if (status)
510                         mlog_errno(status);
511
512                 goto bail_unlock_sem;
513         }
514
515         /* alright, we're going to need to do a full blown alloc size
516          * change. Orphan the inode so that recovery can complete the
517          * truncate if necessary. This does the task of marking
518          * i_size. */
519         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
520         if (status < 0) {
521                 mlog_errno(status);
522                 goto bail_unlock_sem;
523         }
524
525         status = ocfs2_commit_truncate(osb, inode, di_bh);
526         if (status < 0) {
527                 mlog_errno(status);
528                 goto bail_unlock_sem;
529         }
530
531         /* TODO: orphan dir cleanup here. */
532 bail_unlock_sem:
533         up_write(&OCFS2_I(inode)->ip_alloc_sem);
534
535 bail:
536         if (!status && OCFS2_I(inode)->ip_clusters == 0)
537                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
538
539         mlog_exit(status);
540         return status;
541 }
542
543 /*
544  * extend file allocation only here.
545  * we'll update all the disk stuff, and oip->alloc_size
546  *
547  * expect stuff to be locked, a transaction started and enough data /
548  * metadata reservations in the contexts.
549  *
550  * Will return -EAGAIN, and a reason if a restart is needed.
551  * If passed in, *reason will always be set, even in error.
552  */
553 int ocfs2_add_inode_data(struct ocfs2_super *osb,
554                          struct inode *inode,
555                          u32 *logical_offset,
556                          u32 clusters_to_add,
557                          int mark_unwritten,
558                          struct buffer_head *fe_bh,
559                          handle_t *handle,
560                          struct ocfs2_alloc_context *data_ac,
561                          struct ocfs2_alloc_context *meta_ac,
562                          enum ocfs2_alloc_restarted *reason_ret)
563 {
564         int ret;
565         struct ocfs2_extent_tree et;
566
567         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
568         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
569                                           clusters_to_add, mark_unwritten,
570                                           data_ac, meta_ac, reason_ret);
571
572         return ret;
573 }
574
575 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
576                                      u32 clusters_to_add, int mark_unwritten)
577 {
578         int status = 0;
579         int restart_func = 0;
580         int credits;
581         u32 prev_clusters;
582         struct buffer_head *bh = NULL;
583         struct ocfs2_dinode *fe = NULL;
584         handle_t *handle = NULL;
585         struct ocfs2_alloc_context *data_ac = NULL;
586         struct ocfs2_alloc_context *meta_ac = NULL;
587         enum ocfs2_alloc_restarted why;
588         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
589         struct ocfs2_extent_tree et;
590         int did_quota = 0;
591
592         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
593
594         /*
595          * This function only exists for file systems which don't
596          * support holes.
597          */
598         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
599
600         status = ocfs2_read_inode_block(inode, &bh);
601         if (status < 0) {
602                 mlog_errno(status);
603                 goto leave;
604         }
605         fe = (struct ocfs2_dinode *) bh->b_data;
606
607 restart_all:
608         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
609
610         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
611              "clusters_to_add = %u\n",
612              (unsigned long long)OCFS2_I(inode)->ip_blkno,
613              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
614              clusters_to_add);
615         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
616         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
617                                        &data_ac, &meta_ac);
618         if (status) {
619                 mlog_errno(status);
620                 goto leave;
621         }
622
623         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
624                                             clusters_to_add);
625         handle = ocfs2_start_trans(osb, credits);
626         if (IS_ERR(handle)) {
627                 status = PTR_ERR(handle);
628                 handle = NULL;
629                 mlog_errno(status);
630                 goto leave;
631         }
632
633 restarted_transaction:
634         status = dquot_alloc_space_nodirty(inode,
635                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
636         if (status)
637                 goto leave;
638         did_quota = 1;
639
640         /* reserve a write to the file entry early on - that we if we
641          * run out of credits in the allocation path, we can still
642          * update i_size. */
643         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
644                                          OCFS2_JOURNAL_ACCESS_WRITE);
645         if (status < 0) {
646                 mlog_errno(status);
647                 goto leave;
648         }
649
650         prev_clusters = OCFS2_I(inode)->ip_clusters;
651
652         status = ocfs2_add_inode_data(osb,
653                                       inode,
654                                       &logical_start,
655                                       clusters_to_add,
656                                       mark_unwritten,
657                                       bh,
658                                       handle,
659                                       data_ac,
660                                       meta_ac,
661                                       &why);
662         if ((status < 0) && (status != -EAGAIN)) {
663                 if (status != -ENOSPC)
664                         mlog_errno(status);
665                 goto leave;
666         }
667
668         ocfs2_journal_dirty(handle, bh);
669
670         spin_lock(&OCFS2_I(inode)->ip_lock);
671         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
672         spin_unlock(&OCFS2_I(inode)->ip_lock);
673         /* Release unused quota reservation */
674         dquot_free_space(inode,
675                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
676         did_quota = 0;
677
678         if (why != RESTART_NONE && clusters_to_add) {
679                 if (why == RESTART_META) {
680                         mlog(0, "restarting function.\n");
681                         restart_func = 1;
682                         status = 0;
683                 } else {
684                         BUG_ON(why != RESTART_TRANS);
685
686                         mlog(0, "restarting transaction.\n");
687                         /* TODO: This can be more intelligent. */
688                         credits = ocfs2_calc_extend_credits(osb->sb,
689                                                             &fe->id2.i_list,
690                                                             clusters_to_add);
691                         status = ocfs2_extend_trans(handle, credits);
692                         if (status < 0) {
693                                 /* handle still has to be committed at
694                                  * this point. */
695                                 status = -ENOMEM;
696                                 mlog_errno(status);
697                                 goto leave;
698                         }
699                         goto restarted_transaction;
700                 }
701         }
702
703         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
704              le32_to_cpu(fe->i_clusters),
705              (unsigned long long)le64_to_cpu(fe->i_size));
706         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
707              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
708
709 leave:
710         if (status < 0 && did_quota)
711                 dquot_free_space(inode,
712                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
713         if (handle) {
714                 ocfs2_commit_trans(osb, handle);
715                 handle = NULL;
716         }
717         if (data_ac) {
718                 ocfs2_free_alloc_context(data_ac);
719                 data_ac = NULL;
720         }
721         if (meta_ac) {
722                 ocfs2_free_alloc_context(meta_ac);
723                 meta_ac = NULL;
724         }
725         if ((!status) && restart_func) {
726                 restart_func = 0;
727                 goto restart_all;
728         }
729         brelse(bh);
730         bh = NULL;
731
732         mlog_exit(status);
733         return status;
734 }
735
736 /*
737  * While a write will already be ordering the data, a truncate will not.
738  * Thus, we need to explicitly order the zeroed pages.
739  */
740 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
741 {
742         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
743         handle_t *handle = NULL;
744         int ret = 0;
745
746         if (!ocfs2_should_order_data(inode))
747                 goto out;
748
749         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
750         if (IS_ERR(handle)) {
751                 ret = -ENOMEM;
752                 mlog_errno(ret);
753                 goto out;
754         }
755
756         ret = ocfs2_jbd2_file_inode(handle, inode);
757         if (ret < 0)
758                 mlog_errno(ret);
759
760 out:
761         if (ret) {
762                 if (!IS_ERR(handle))
763                         ocfs2_commit_trans(osb, handle);
764                 handle = ERR_PTR(ret);
765         }
766         return handle;
767 }
768
769 /* Some parts of this taken from generic_cont_expand, which turned out
770  * to be too fragile to do exactly what we need without us having to
771  * worry about recursive locking in ->write_begin() and ->write_end(). */
772 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
773                                  u64 abs_to)
774 {
775         struct address_space *mapping = inode->i_mapping;
776         struct page *page;
777         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
778         handle_t *handle = NULL;
779         int ret = 0;
780         unsigned zero_from, zero_to, block_start, block_end;
781
782         BUG_ON(abs_from >= abs_to);
783         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
784         BUG_ON(abs_from & (inode->i_blkbits - 1));
785
786         page = grab_cache_page(mapping, index);
787         if (!page) {
788                 ret = -ENOMEM;
789                 mlog_errno(ret);
790                 goto out;
791         }
792
793         /* Get the offsets within the page that we want to zero */
794         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
795         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
796         if (!zero_to)
797                 zero_to = PAGE_CACHE_SIZE;
798
799         mlog(0,
800              "abs_from = %llu, abs_to = %llu, index = %lu, zero_from = %u, zero_to = %u\n",
801              (unsigned long long)abs_from, (unsigned long long)abs_to,
802              index, zero_from, zero_to);
803
804         /* We know that zero_from is block aligned */
805         for (block_start = zero_from; block_start < zero_to;
806              block_start = block_end) {
807                 block_end = block_start + (1 << inode->i_blkbits);
808
809                 /*
810                  * block_start is block-aligned.  Bump it by one to
811                  * force ocfs2_{prepare,commit}_write() to zero the
812                  * whole block.
813                  */
814                 ret = ocfs2_prepare_write_nolock(inode, page,
815                                                  block_start + 1,
816                                                  block_start + 1);
817                 if (ret < 0) {
818                         mlog_errno(ret);
819                         goto out_unlock;
820                 }
821
822                 if (!handle) {
823                         handle = ocfs2_zero_start_ordered_transaction(inode);
824                         if (IS_ERR(handle)) {
825                                 ret = PTR_ERR(handle);
826                                 handle = NULL;
827                                 break;
828                         }
829                 }
830
831                 /* must not update i_size! */
832                 ret = block_commit_write(page, block_start + 1,
833                                          block_start + 1);
834                 if (ret < 0)
835                         mlog_errno(ret);
836                 else
837                         ret = 0;
838         }
839
840         if (handle)
841                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
842
843 out_unlock:
844         unlock_page(page);
845         page_cache_release(page);
846 out:
847         return ret;
848 }
849
850 /*
851  * Find the next range to zero.  We do this in terms of bytes because
852  * that's what ocfs2_zero_extend() wants, and it is dealing with the
853  * pagecache.  We may return multiple extents.
854  *
855  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
856  * needs to be zeroed.  range_start and range_end return the next zeroing
857  * range.  A subsequent call should pass the previous range_end as its
858  * zero_start.  If range_end is 0, there's nothing to do.
859  *
860  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
861  */
862 static int ocfs2_zero_extend_get_range(struct inode *inode,
863                                        struct buffer_head *di_bh,
864                                        u64 zero_start, u64 zero_end,
865                                        u64 *range_start, u64 *range_end)
866 {
867         int rc = 0, needs_cow = 0;
868         u32 p_cpos, zero_clusters = 0;
869         u32 zero_cpos =
870                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
871         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
872         unsigned int num_clusters = 0;
873         unsigned int ext_flags = 0;
874
875         while (zero_cpos < last_cpos) {
876                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
877                                         &num_clusters, &ext_flags);
878                 if (rc) {
879                         mlog_errno(rc);
880                         goto out;
881                 }
882
883                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
884                         zero_clusters = num_clusters;
885                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
886                                 needs_cow = 1;
887                         break;
888                 }
889
890                 zero_cpos += num_clusters;
891         }
892         if (!zero_clusters) {
893                 *range_end = 0;
894                 goto out;
895         }
896
897         while ((zero_cpos + zero_clusters) < last_cpos) {
898                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
899                                         &p_cpos, &num_clusters,
900                                         &ext_flags);
901                 if (rc) {
902                         mlog_errno(rc);
903                         goto out;
904                 }
905
906                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
907                         break;
908                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
909                         needs_cow = 1;
910                 zero_clusters += num_clusters;
911         }
912         if ((zero_cpos + zero_clusters) > last_cpos)
913                 zero_clusters = last_cpos - zero_cpos;
914
915         if (needs_cow) {
916                 rc = ocfs2_refcount_cow(inode, di_bh, zero_cpos, zero_clusters,
917                                         UINT_MAX);
918                 if (rc) {
919                         mlog_errno(rc);
920                         goto out;
921                 }
922         }
923
924         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
925         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
926                                              zero_cpos + zero_clusters);
927
928 out:
929         return rc;
930 }
931
932 /*
933  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
934  * has made sure that the entire range needs zeroing.
935  */
936 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
937                                    u64 range_end)
938 {
939         int rc = 0;
940         u64 next_pos;
941         u64 zero_pos = range_start;
942
943         mlog(0, "range_start = %llu, range_end = %llu\n",
944              (unsigned long long)range_start,
945              (unsigned long long)range_end);
946         BUG_ON(range_start >= range_end);
947
948         while (zero_pos < range_end) {
949                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
950                 if (next_pos > range_end)
951                         next_pos = range_end;
952                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
953                 if (rc < 0) {
954                         mlog_errno(rc);
955                         break;
956                 }
957                 zero_pos = next_pos;
958
959                 /*
960                  * Very large extends have the potential to lock up
961                  * the cpu for extended periods of time.
962                  */
963                 cond_resched();
964         }
965
966         return rc;
967 }
968
969 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
970                       loff_t zero_to_size)
971 {
972         int ret = 0;
973         u64 zero_start, range_start = 0, range_end = 0;
974         struct super_block *sb = inode->i_sb;
975
976         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
977         mlog(0, "zero_start %llu for i_size %llu\n",
978              (unsigned long long)zero_start,
979              (unsigned long long)i_size_read(inode));
980         while (zero_start < zero_to_size) {
981                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
982                                                   zero_to_size,
983                                                   &range_start,
984                                                   &range_end);
985                 if (ret) {
986                         mlog_errno(ret);
987                         break;
988                 }
989                 if (!range_end)
990                         break;
991                 /* Trim the ends */
992                 if (range_start < zero_start)
993                         range_start = zero_start;
994                 if (range_end > zero_to_size)
995                         range_end = zero_to_size;
996
997                 ret = ocfs2_zero_extend_range(inode, range_start,
998                                               range_end);
999                 if (ret) {
1000                         mlog_errno(ret);
1001                         break;
1002                 }
1003                 zero_start = range_end;
1004         }
1005
1006         return ret;
1007 }
1008
1009 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
1010                           u64 new_i_size, u64 zero_to)
1011 {
1012         int ret;
1013         u32 clusters_to_add;
1014         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1015
1016         /*
1017          * Only quota files call this without a bh, and they can't be
1018          * refcounted.
1019          */
1020         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
1021         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1022
1023         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1024         if (clusters_to_add < oi->ip_clusters)
1025                 clusters_to_add = 0;
1026         else
1027                 clusters_to_add -= oi->ip_clusters;
1028
1029         if (clusters_to_add) {
1030                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1031                                                 clusters_to_add, 0);
1032                 if (ret) {
1033                         mlog_errno(ret);
1034                         goto out;
1035                 }
1036         }
1037
1038         /*
1039          * Call this even if we don't add any clusters to the tree. We
1040          * still need to zero the area between the old i_size and the
1041          * new i_size.
1042          */
1043         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1044         if (ret < 0)
1045                 mlog_errno(ret);
1046
1047 out:
1048         return ret;
1049 }
1050
1051 static int ocfs2_extend_file(struct inode *inode,
1052                              struct buffer_head *di_bh,
1053                              u64 new_i_size)
1054 {
1055         int ret = 0;
1056         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1057
1058         BUG_ON(!di_bh);
1059
1060         /* setattr sometimes calls us like this. */
1061         if (new_i_size == 0)
1062                 goto out;
1063
1064         if (i_size_read(inode) == new_i_size)
1065                 goto out;
1066         BUG_ON(new_i_size < i_size_read(inode));
1067
1068         /*
1069          * The alloc sem blocks people in read/write from reading our
1070          * allocation until we're done changing it. We depend on
1071          * i_mutex to block other extend/truncate calls while we're
1072          * here.  We even have to hold it for sparse files because there
1073          * might be some tail zeroing.
1074          */
1075         down_write(&oi->ip_alloc_sem);
1076
1077         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1078                 /*
1079                  * We can optimize small extends by keeping the inodes
1080                  * inline data.
1081                  */
1082                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1083                         up_write(&oi->ip_alloc_sem);
1084                         goto out_update_size;
1085                 }
1086
1087                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1088                 if (ret) {
1089                         up_write(&oi->ip_alloc_sem);
1090                         mlog_errno(ret);
1091                         goto out;
1092                 }
1093         }
1094
1095         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1096                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1097         else
1098                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1099                                             new_i_size);
1100
1101         up_write(&oi->ip_alloc_sem);
1102
1103         if (ret < 0) {
1104                 mlog_errno(ret);
1105                 goto out;
1106         }
1107
1108 out_update_size:
1109         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1110         if (ret < 0)
1111                 mlog_errno(ret);
1112
1113 out:
1114         return ret;
1115 }
1116
1117 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1118 {
1119         int status = 0, size_change;
1120         struct inode *inode = dentry->d_inode;
1121         struct super_block *sb = inode->i_sb;
1122         struct ocfs2_super *osb = OCFS2_SB(sb);
1123         struct buffer_head *bh = NULL;
1124         handle_t *handle = NULL;
1125         struct dquot *transfer_to[MAXQUOTAS] = { };
1126         int qtype;
1127
1128         mlog_entry("(0x%p, '%.*s')\n", dentry,
1129                    dentry->d_name.len, dentry->d_name.name);
1130
1131         /* ensuring we don't even attempt to truncate a symlink */
1132         if (S_ISLNK(inode->i_mode))
1133                 attr->ia_valid &= ~ATTR_SIZE;
1134
1135         if (attr->ia_valid & ATTR_MODE)
1136                 mlog(0, "mode change: %d\n", attr->ia_mode);
1137         if (attr->ia_valid & ATTR_UID)
1138                 mlog(0, "uid change: %d\n", attr->ia_uid);
1139         if (attr->ia_valid & ATTR_GID)
1140                 mlog(0, "gid change: %d\n", attr->ia_gid);
1141         if (attr->ia_valid & ATTR_SIZE)
1142                 mlog(0, "size change...\n");
1143         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1144                 mlog(0, "time change...\n");
1145
1146 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1147                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1148         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1149                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1150                 return 0;
1151         }
1152
1153         status = inode_change_ok(inode, attr);
1154         if (status)
1155                 return status;
1156
1157         if (is_quota_modification(inode, attr))
1158                 dquot_initialize(inode);
1159         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1160         if (size_change) {
1161                 status = ocfs2_rw_lock(inode, 1);
1162                 if (status < 0) {
1163                         mlog_errno(status);
1164                         goto bail;
1165                 }
1166         }
1167
1168         status = ocfs2_inode_lock(inode, &bh, 1);
1169         if (status < 0) {
1170                 if (status != -ENOENT)
1171                         mlog_errno(status);
1172                 goto bail_unlock_rw;
1173         }
1174
1175         if (size_change && attr->ia_size != i_size_read(inode)) {
1176                 status = inode_newsize_ok(inode, attr->ia_size);
1177                 if (status)
1178                         goto bail_unlock;
1179
1180                 if (i_size_read(inode) > attr->ia_size) {
1181                         if (ocfs2_should_order_data(inode)) {
1182                                 status = ocfs2_begin_ordered_truncate(inode,
1183                                                                       attr->ia_size);
1184                                 if (status)
1185                                         goto bail_unlock;
1186                         }
1187                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1188                 } else
1189                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1190                 if (status < 0) {
1191                         if (status != -ENOSPC)
1192                                 mlog_errno(status);
1193                         status = -ENOSPC;
1194                         goto bail_unlock;
1195                 }
1196         }
1197
1198         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1199             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1200                 /*
1201                  * Gather pointers to quota structures so that allocation /
1202                  * freeing of quota structures happens here and not inside
1203                  * dquot_transfer() where we have problems with lock ordering
1204                  */
1205                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1206                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1207                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1208                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1209                                                       USRQUOTA);
1210                         if (!transfer_to[USRQUOTA]) {
1211                                 status = -ESRCH;
1212                                 goto bail_unlock;
1213                         }
1214                 }
1215                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1216                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1217                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1218                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1219                                                       GRPQUOTA);
1220                         if (!transfer_to[GRPQUOTA]) {
1221                                 status = -ESRCH;
1222                                 goto bail_unlock;
1223                         }
1224                 }
1225                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1226                                            2 * ocfs2_quota_trans_credits(sb));
1227                 if (IS_ERR(handle)) {
1228                         status = PTR_ERR(handle);
1229                         mlog_errno(status);
1230                         goto bail_unlock;
1231                 }
1232                 status = __dquot_transfer(inode, transfer_to);
1233                 if (status < 0)
1234                         goto bail_commit;
1235         } else {
1236                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1237                 if (IS_ERR(handle)) {
1238                         status = PTR_ERR(handle);
1239                         mlog_errno(status);
1240                         goto bail_unlock;
1241                 }
1242         }
1243
1244         /*
1245          * This will intentionally not wind up calling truncate_setsize(),
1246          * since all the work for a size change has been done above.
1247          * Otherwise, we could get into problems with truncate as
1248          * ip_alloc_sem is used there to protect against i_size
1249          * changes.
1250          *
1251          * XXX: this means the conditional below can probably be removed.
1252          */
1253         if ((attr->ia_valid & ATTR_SIZE) &&
1254             attr->ia_size != i_size_read(inode)) {
1255                 status = vmtruncate(inode, attr->ia_size);
1256                 if (status) {
1257                         mlog_errno(status);
1258                         goto bail_commit;
1259                 }
1260         }
1261
1262         setattr_copy(inode, attr);
1263         mark_inode_dirty(inode);
1264
1265         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1266         if (status < 0)
1267                 mlog_errno(status);
1268
1269 bail_commit:
1270         ocfs2_commit_trans(osb, handle);
1271 bail_unlock:
1272         ocfs2_inode_unlock(inode, 1);
1273 bail_unlock_rw:
1274         if (size_change)
1275                 ocfs2_rw_unlock(inode, 1);
1276 bail:
1277         brelse(bh);
1278
1279         /* Release quota pointers in case we acquired them */
1280         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1281                 dqput(transfer_to[qtype]);
1282
1283         if (!status && attr->ia_valid & ATTR_MODE) {
1284                 status = ocfs2_acl_chmod(inode);
1285                 if (status < 0)
1286                         mlog_errno(status);
1287         }
1288
1289         mlog_exit(status);
1290         return status;
1291 }
1292
1293 int ocfs2_getattr(struct vfsmount *mnt,
1294                   struct dentry *dentry,
1295                   struct kstat *stat)
1296 {
1297         struct inode *inode = dentry->d_inode;
1298         struct super_block *sb = dentry->d_inode->i_sb;
1299         struct ocfs2_super *osb = sb->s_fs_info;
1300         int err;
1301
1302         mlog_entry_void();
1303
1304         err = ocfs2_inode_revalidate(dentry);
1305         if (err) {
1306                 if (err != -ENOENT)
1307                         mlog_errno(err);
1308                 goto bail;
1309         }
1310
1311         generic_fillattr(inode, stat);
1312
1313         /* We set the blksize from the cluster size for performance */
1314         stat->blksize = osb->s_clustersize;
1315
1316 bail:
1317         mlog_exit(err);
1318
1319         return err;
1320 }
1321
1322 int ocfs2_permission(struct inode *inode, int mask)
1323 {
1324         int ret;
1325
1326         mlog_entry_void();
1327
1328         ret = ocfs2_inode_lock(inode, NULL, 0);
1329         if (ret) {
1330                 if (ret != -ENOENT)
1331                         mlog_errno(ret);
1332                 goto out;
1333         }
1334
1335         ret = generic_permission(inode, mask, ocfs2_check_acl);
1336
1337         ocfs2_inode_unlock(inode, 0);
1338 out:
1339         mlog_exit(ret);
1340         return ret;
1341 }
1342
1343 static int __ocfs2_write_remove_suid(struct inode *inode,
1344                                      struct buffer_head *bh)
1345 {
1346         int ret;
1347         handle_t *handle;
1348         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1349         struct ocfs2_dinode *di;
1350
1351         mlog_entry("(Inode %llu, mode 0%o)\n",
1352                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1353
1354         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1355         if (IS_ERR(handle)) {
1356                 ret = PTR_ERR(handle);
1357                 mlog_errno(ret);
1358                 goto out;
1359         }
1360
1361         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1362                                       OCFS2_JOURNAL_ACCESS_WRITE);
1363         if (ret < 0) {
1364                 mlog_errno(ret);
1365                 goto out_trans;
1366         }
1367
1368         inode->i_mode &= ~S_ISUID;
1369         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1370                 inode->i_mode &= ~S_ISGID;
1371
1372         di = (struct ocfs2_dinode *) bh->b_data;
1373         di->i_mode = cpu_to_le16(inode->i_mode);
1374
1375         ocfs2_journal_dirty(handle, bh);
1376
1377 out_trans:
1378         ocfs2_commit_trans(osb, handle);
1379 out:
1380         mlog_exit(ret);
1381         return ret;
1382 }
1383
1384 /*
1385  * Will look for holes and unwritten extents in the range starting at
1386  * pos for count bytes (inclusive).
1387  */
1388 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1389                                        size_t count)
1390 {
1391         int ret = 0;
1392         unsigned int extent_flags;
1393         u32 cpos, clusters, extent_len, phys_cpos;
1394         struct super_block *sb = inode->i_sb;
1395
1396         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1397         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1398
1399         while (clusters) {
1400                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1401                                          &extent_flags);
1402                 if (ret < 0) {
1403                         mlog_errno(ret);
1404                         goto out;
1405                 }
1406
1407                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1408                         ret = 1;
1409                         break;
1410                 }
1411
1412                 if (extent_len > clusters)
1413                         extent_len = clusters;
1414
1415                 clusters -= extent_len;
1416                 cpos += extent_len;
1417         }
1418 out:
1419         return ret;
1420 }
1421
1422 static int ocfs2_write_remove_suid(struct inode *inode)
1423 {
1424         int ret;
1425         struct buffer_head *bh = NULL;
1426
1427         ret = ocfs2_read_inode_block(inode, &bh);
1428         if (ret < 0) {
1429                 mlog_errno(ret);
1430                 goto out;
1431         }
1432
1433         ret =  __ocfs2_write_remove_suid(inode, bh);
1434 out:
1435         brelse(bh);
1436         return ret;
1437 }
1438
1439 /*
1440  * Allocate enough extents to cover the region starting at byte offset
1441  * start for len bytes. Existing extents are skipped, any extents
1442  * added are marked as "unwritten".
1443  */
1444 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1445                                             u64 start, u64 len)
1446 {
1447         int ret;
1448         u32 cpos, phys_cpos, clusters, alloc_size;
1449         u64 end = start + len;
1450         struct buffer_head *di_bh = NULL;
1451
1452         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1453                 ret = ocfs2_read_inode_block(inode, &di_bh);
1454                 if (ret) {
1455                         mlog_errno(ret);
1456                         goto out;
1457                 }
1458
1459                 /*
1460                  * Nothing to do if the requested reservation range
1461                  * fits within the inode.
1462                  */
1463                 if (ocfs2_size_fits_inline_data(di_bh, end))
1464                         goto out;
1465
1466                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1467                 if (ret) {
1468                         mlog_errno(ret);
1469                         goto out;
1470                 }
1471         }
1472
1473         /*
1474          * We consider both start and len to be inclusive.
1475          */
1476         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1477         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1478         clusters -= cpos;
1479
1480         while (clusters) {
1481                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1482                                          &alloc_size, NULL);
1483                 if (ret) {
1484                         mlog_errno(ret);
1485                         goto out;
1486                 }
1487
1488                 /*
1489                  * Hole or existing extent len can be arbitrary, so
1490                  * cap it to our own allocation request.
1491                  */
1492                 if (alloc_size > clusters)
1493                         alloc_size = clusters;
1494
1495                 if (phys_cpos) {
1496                         /*
1497                          * We already have an allocation at this
1498                          * region so we can safely skip it.
1499                          */
1500                         goto next;
1501                 }
1502
1503                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1504                 if (ret) {
1505                         if (ret != -ENOSPC)
1506                                 mlog_errno(ret);
1507                         goto out;
1508                 }
1509
1510 next:
1511                 cpos += alloc_size;
1512                 clusters -= alloc_size;
1513         }
1514
1515         ret = 0;
1516 out:
1517
1518         brelse(di_bh);
1519         return ret;
1520 }
1521
1522 /*
1523  * Truncate a byte range, avoiding pages within partial clusters. This
1524  * preserves those pages for the zeroing code to write to.
1525  */
1526 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1527                                          u64 byte_len)
1528 {
1529         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1530         loff_t start, end;
1531         struct address_space *mapping = inode->i_mapping;
1532
1533         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1534         end = byte_start + byte_len;
1535         end = end & ~(osb->s_clustersize - 1);
1536
1537         if (start < end) {
1538                 unmap_mapping_range(mapping, start, end - start, 0);
1539                 truncate_inode_pages_range(mapping, start, end - 1);
1540         }
1541 }
1542
1543 static int ocfs2_zero_partial_clusters(struct inode *inode,
1544                                        u64 start, u64 len)
1545 {
1546         int ret = 0;
1547         u64 tmpend, end = start + len;
1548         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1549         unsigned int csize = osb->s_clustersize;
1550         handle_t *handle;
1551
1552         /*
1553          * The "start" and "end" values are NOT necessarily part of
1554          * the range whose allocation is being deleted. Rather, this
1555          * is what the user passed in with the request. We must zero
1556          * partial clusters here. There's no need to worry about
1557          * physical allocation - the zeroing code knows to skip holes.
1558          */
1559         mlog(0, "byte start: %llu, end: %llu\n",
1560              (unsigned long long)start, (unsigned long long)end);
1561
1562         /*
1563          * If both edges are on a cluster boundary then there's no
1564          * zeroing required as the region is part of the allocation to
1565          * be truncated.
1566          */
1567         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1568                 goto out;
1569
1570         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1571         if (IS_ERR(handle)) {
1572                 ret = PTR_ERR(handle);
1573                 mlog_errno(ret);
1574                 goto out;
1575         }
1576
1577         /*
1578          * We want to get the byte offset of the end of the 1st cluster.
1579          */
1580         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1581         if (tmpend > end)
1582                 tmpend = end;
1583
1584         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1585              (unsigned long long)start, (unsigned long long)tmpend);
1586
1587         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1588         if (ret)
1589                 mlog_errno(ret);
1590
1591         if (tmpend < end) {
1592                 /*
1593                  * This may make start and end equal, but the zeroing
1594                  * code will skip any work in that case so there's no
1595                  * need to catch it up here.
1596                  */
1597                 start = end & ~(osb->s_clustersize - 1);
1598
1599                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1600                      (unsigned long long)start, (unsigned long long)end);
1601
1602                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1603                 if (ret)
1604                         mlog_errno(ret);
1605         }
1606
1607         ocfs2_commit_trans(osb, handle);
1608 out:
1609         return ret;
1610 }
1611
1612 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1613 {
1614         int i;
1615         struct ocfs2_extent_rec *rec = NULL;
1616
1617         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1618
1619                 rec = &el->l_recs[i];
1620
1621                 if (le32_to_cpu(rec->e_cpos) < pos)
1622                         break;
1623         }
1624
1625         return i;
1626 }
1627
1628 /*
1629  * Helper to calculate the punching pos and length in one run, we handle the
1630  * following three cases in order:
1631  *
1632  * - remove the entire record
1633  * - remove a partial record
1634  * - no record needs to be removed (hole-punching completed)
1635 */
1636 static void ocfs2_calc_trunc_pos(struct inode *inode,
1637                                  struct ocfs2_extent_list *el,
1638                                  struct ocfs2_extent_rec *rec,
1639                                  u32 trunc_start, u32 *trunc_cpos,
1640                                  u32 *trunc_len, u32 *trunc_end,
1641                                  u64 *blkno, int *done)
1642 {
1643         int ret = 0;
1644         u32 coff, range;
1645
1646         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1647
1648         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1649                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1650                 /*
1651                  * Skip holes if any.
1652                  */
1653                 if (range < *trunc_end)
1654                         *trunc_end = range;
1655                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1656                 *blkno = le64_to_cpu(rec->e_blkno);
1657                 *trunc_end = le32_to_cpu(rec->e_cpos);
1658         } else if (range > trunc_start) {
1659                 *trunc_cpos = trunc_start;
1660                 *trunc_len = *trunc_end - trunc_start;
1661                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1662                 *blkno = le64_to_cpu(rec->e_blkno) +
1663                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1664                 *trunc_end = trunc_start;
1665         } else {
1666                 /*
1667                  * It may have two following possibilities:
1668                  *
1669                  * - last record has been removed
1670                  * - trunc_start was within a hole
1671                  *
1672                  * both two cases mean the completion of hole punching.
1673                  */
1674                 ret = 1;
1675         }
1676
1677         *done = ret;
1678 }
1679
1680 static int ocfs2_remove_inode_range(struct inode *inode,
1681                                     struct buffer_head *di_bh, u64 byte_start,
1682                                     u64 byte_len)
1683 {
1684         int ret = 0, flags = 0, done = 0, i;
1685         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1686         u32 cluster_in_el;
1687         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1688         struct ocfs2_cached_dealloc_ctxt dealloc;
1689         struct address_space *mapping = inode->i_mapping;
1690         struct ocfs2_extent_tree et;
1691         struct ocfs2_path *path = NULL;
1692         struct ocfs2_extent_list *el = NULL;
1693         struct ocfs2_extent_rec *rec = NULL;
1694         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1695         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1696
1697         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1698         ocfs2_init_dealloc_ctxt(&dealloc);
1699
1700         if (byte_len == 0)
1701                 return 0;
1702
1703         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1704                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1705                                             byte_start + byte_len, 0);
1706                 if (ret) {
1707                         mlog_errno(ret);
1708                         goto out;
1709                 }
1710                 /*
1711                  * There's no need to get fancy with the page cache
1712                  * truncate of an inline-data inode. We're talking
1713                  * about less than a page here, which will be cached
1714                  * in the dinode buffer anyway.
1715                  */
1716                 unmap_mapping_range(mapping, 0, 0, 0);
1717                 truncate_inode_pages(mapping, 0);
1718                 goto out;
1719         }
1720
1721         /*
1722          * For reflinks, we may need to CoW 2 clusters which might be
1723          * partially zero'd later, if hole's start and end offset were
1724          * within one cluster(means is not exactly aligned to clustersize).
1725          */
1726
1727         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1728
1729                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1730                 if (ret) {
1731                         mlog_errno(ret);
1732                         goto out;
1733                 }
1734
1735                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1736                 if (ret) {
1737                         mlog_errno(ret);
1738                         goto out;
1739                 }
1740         }
1741
1742         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1743         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1744         cluster_in_el = trunc_end;
1745
1746         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1747              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1748              (unsigned long long)byte_start,
1749              (unsigned long long)byte_len, trunc_start, trunc_end);
1750
1751         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1752         if (ret) {
1753                 mlog_errno(ret);
1754                 goto out;
1755         }
1756
1757         path = ocfs2_new_path_from_et(&et);
1758         if (!path) {
1759                 ret = -ENOMEM;
1760                 mlog_errno(ret);
1761                 goto out;
1762         }
1763
1764         while (trunc_end > trunc_start) {
1765
1766                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1767                                       cluster_in_el);
1768                 if (ret) {
1769                         mlog_errno(ret);
1770                         goto out;
1771                 }
1772
1773                 el = path_leaf_el(path);
1774
1775                 i = ocfs2_find_rec(el, trunc_end);
1776                 /*
1777                  * Need to go to previous extent block.
1778                  */
1779                 if (i < 0) {
1780                         if (path->p_tree_depth == 0)
1781                                 break;
1782
1783                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1784                                                             path,
1785                                                             &cluster_in_el);
1786                         if (ret) {
1787                                 mlog_errno(ret);
1788                                 goto out;
1789                         }
1790
1791                         /*
1792                          * We've reached the leftmost extent block,
1793                          * it's safe to leave.
1794                          */
1795                         if (cluster_in_el == 0)
1796                                 break;
1797
1798                         /*
1799                          * The 'pos' searched for previous extent block is
1800                          * always one cluster less than actual trunc_end.
1801                          */
1802                         trunc_end = cluster_in_el + 1;
1803
1804                         ocfs2_reinit_path(path, 1);
1805
1806                         continue;
1807
1808                 } else
1809                         rec = &el->l_recs[i];
1810
1811                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1812                                      &trunc_len, &trunc_end, &blkno, &done);
1813                 if (done)
1814                         break;
1815
1816                 flags = rec->e_flags;
1817                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1818
1819                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1820                                                phys_cpos, trunc_len, flags,
1821                                                &dealloc, refcount_loc);
1822                 if (ret < 0) {
1823                         mlog_errno(ret);
1824                         goto out;
1825                 }
1826
1827                 cluster_in_el = trunc_end;
1828
1829                 ocfs2_reinit_path(path, 1);
1830         }
1831
1832         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1833
1834 out:
1835         ocfs2_schedule_truncate_log_flush(osb, 1);
1836         ocfs2_run_deallocs(osb, &dealloc);
1837
1838         return ret;
1839 }
1840
1841 /*
1842  * Parts of this function taken from xfs_change_file_space()
1843  */
1844 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1845                                      loff_t f_pos, unsigned int cmd,
1846                                      struct ocfs2_space_resv *sr,
1847                                      int change_size)
1848 {
1849         int ret;
1850         s64 llen;
1851         loff_t size;
1852         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1853         struct buffer_head *di_bh = NULL;
1854         handle_t *handle;
1855         unsigned long long max_off = inode->i_sb->s_maxbytes;
1856
1857         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1858                 return -EROFS;
1859
1860         mutex_lock(&inode->i_mutex);
1861
1862         /*
1863          * This prevents concurrent writes on other nodes
1864          */
1865         ret = ocfs2_rw_lock(inode, 1);
1866         if (ret) {
1867                 mlog_errno(ret);
1868                 goto out;
1869         }
1870
1871         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1872         if (ret) {
1873                 mlog_errno(ret);
1874                 goto out_rw_unlock;
1875         }
1876
1877         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1878                 ret = -EPERM;
1879                 goto out_inode_unlock;
1880         }
1881
1882         switch (sr->l_whence) {
1883         case 0: /*SEEK_SET*/
1884                 break;
1885         case 1: /*SEEK_CUR*/
1886                 sr->l_start += f_pos;
1887                 break;
1888         case 2: /*SEEK_END*/
1889                 sr->l_start += i_size_read(inode);
1890                 break;
1891         default:
1892                 ret = -EINVAL;
1893                 goto out_inode_unlock;
1894         }
1895         sr->l_whence = 0;
1896
1897         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1898
1899         if (sr->l_start < 0
1900             || sr->l_start > max_off
1901             || (sr->l_start + llen) < 0
1902             || (sr->l_start + llen) > max_off) {
1903                 ret = -EINVAL;
1904                 goto out_inode_unlock;
1905         }
1906         size = sr->l_start + sr->l_len;
1907
1908         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1909                 if (sr->l_len <= 0) {
1910                         ret = -EINVAL;
1911                         goto out_inode_unlock;
1912                 }
1913         }
1914
1915         if (file && should_remove_suid(file->f_path.dentry)) {
1916                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1917                 if (ret) {
1918                         mlog_errno(ret);
1919                         goto out_inode_unlock;
1920                 }
1921         }
1922
1923         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1924         switch (cmd) {
1925         case OCFS2_IOC_RESVSP:
1926         case OCFS2_IOC_RESVSP64:
1927                 /*
1928                  * This takes unsigned offsets, but the signed ones we
1929                  * pass have been checked against overflow above.
1930                  */
1931                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1932                                                        sr->l_len);
1933                 break;
1934         case OCFS2_IOC_UNRESVSP:
1935         case OCFS2_IOC_UNRESVSP64:
1936                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1937                                                sr->l_len);
1938                 break;
1939         default:
1940                 ret = -EINVAL;
1941         }
1942         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1943         if (ret) {
1944                 mlog_errno(ret);
1945                 goto out_inode_unlock;
1946         }
1947
1948         /*
1949          * We update c/mtime for these changes
1950          */
1951         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1952         if (IS_ERR(handle)) {
1953                 ret = PTR_ERR(handle);
1954                 mlog_errno(ret);
1955                 goto out_inode_unlock;
1956         }
1957
1958         if (change_size && i_size_read(inode) < size)
1959                 i_size_write(inode, size);
1960
1961         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1962         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1963         if (ret < 0)
1964                 mlog_errno(ret);
1965
1966         ocfs2_commit_trans(osb, handle);
1967
1968 out_inode_unlock:
1969         brelse(di_bh);
1970         ocfs2_inode_unlock(inode, 1);
1971 out_rw_unlock:
1972         ocfs2_rw_unlock(inode, 1);
1973
1974 out:
1975         mutex_unlock(&inode->i_mutex);
1976         return ret;
1977 }
1978
1979 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1980                             struct ocfs2_space_resv *sr)
1981 {
1982         struct inode *inode = file->f_path.dentry->d_inode;
1983         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1984
1985         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1986             !ocfs2_writes_unwritten_extents(osb))
1987                 return -ENOTTY;
1988         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1989                  !ocfs2_sparse_alloc(osb))
1990                 return -ENOTTY;
1991
1992         if (!S_ISREG(inode->i_mode))
1993                 return -EINVAL;
1994
1995         if (!(file->f_mode & FMODE_WRITE))
1996                 return -EBADF;
1997
1998         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1999 }
2000
2001 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
2002                             loff_t len)
2003 {
2004         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2005         struct ocfs2_space_resv sr;
2006         int change_size = 1;
2007
2008         if (!ocfs2_writes_unwritten_extents(osb))
2009                 return -EOPNOTSUPP;
2010
2011         if (S_ISDIR(inode->i_mode))
2012                 return -ENODEV;
2013
2014         if (mode & FALLOC_FL_KEEP_SIZE)
2015                 change_size = 0;
2016
2017         sr.l_whence = 0;
2018         sr.l_start = (s64)offset;
2019         sr.l_len = (s64)len;
2020
2021         return __ocfs2_change_file_space(NULL, inode, offset,
2022                                          OCFS2_IOC_RESVSP64, &sr, change_size);
2023 }
2024
2025 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2026                                    size_t count)
2027 {
2028         int ret = 0;
2029         unsigned int extent_flags;
2030         u32 cpos, clusters, extent_len, phys_cpos;
2031         struct super_block *sb = inode->i_sb;
2032
2033         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2034             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2035             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2036                 return 0;
2037
2038         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2039         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2040
2041         while (clusters) {
2042                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2043                                          &extent_flags);
2044                 if (ret < 0) {
2045                         mlog_errno(ret);
2046                         goto out;
2047                 }
2048
2049                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2050                         ret = 1;
2051                         break;
2052                 }
2053
2054                 if (extent_len > clusters)
2055                         extent_len = clusters;
2056
2057                 clusters -= extent_len;
2058                 cpos += extent_len;
2059         }
2060 out:
2061         return ret;
2062 }
2063
2064 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2065                                             loff_t pos, size_t count,
2066                                             int *meta_level)
2067 {
2068         int ret;
2069         struct buffer_head *di_bh = NULL;
2070         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2071         u32 clusters =
2072                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2073
2074         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2075         if (ret) {
2076                 mlog_errno(ret);
2077                 goto out;
2078         }
2079
2080         *meta_level = 1;
2081
2082         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
2083         if (ret)
2084                 mlog_errno(ret);
2085 out:
2086         brelse(di_bh);
2087         return ret;
2088 }
2089
2090 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
2091                                          loff_t *ppos,
2092                                          size_t count,
2093                                          int appending,
2094                                          int *direct_io,
2095                                          int *has_refcount)
2096 {
2097         int ret = 0, meta_level = 0;
2098         struct inode *inode = dentry->d_inode;
2099         loff_t saved_pos, end;
2100
2101         /*
2102          * We start with a read level meta lock and only jump to an ex
2103          * if we need to make modifications here.
2104          */
2105         for(;;) {
2106                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2107                 if (ret < 0) {
2108                         meta_level = -1;
2109                         mlog_errno(ret);
2110                         goto out;
2111                 }
2112
2113                 /* Clear suid / sgid if necessary. We do this here
2114                  * instead of later in the write path because
2115                  * remove_suid() calls ->setattr without any hint that
2116                  * we may have already done our cluster locking. Since
2117                  * ocfs2_setattr() *must* take cluster locks to
2118                  * proceeed, this will lead us to recursively lock the
2119                  * inode. There's also the dinode i_size state which
2120                  * can be lost via setattr during extending writes (we
2121                  * set inode->i_size at the end of a write. */
2122                 if (should_remove_suid(dentry)) {
2123                         if (meta_level == 0) {
2124                                 ocfs2_inode_unlock(inode, meta_level);
2125                                 meta_level = 1;
2126                                 continue;
2127                         }
2128
2129                         ret = ocfs2_write_remove_suid(inode);
2130                         if (ret < 0) {
2131                                 mlog_errno(ret);
2132                                 goto out_unlock;
2133                         }
2134                 }
2135
2136                 /* work on a copy of ppos until we're sure that we won't have
2137                  * to recalculate it due to relocking. */
2138                 if (appending) {
2139                         saved_pos = i_size_read(inode);
2140                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2141                 } else {
2142                         saved_pos = *ppos;
2143                 }
2144
2145                 end = saved_pos + count;
2146
2147                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2148                 if (ret == 1) {
2149                         ocfs2_inode_unlock(inode, meta_level);
2150                         meta_level = -1;
2151
2152                         ret = ocfs2_prepare_inode_for_refcount(inode,
2153                                                                saved_pos,
2154                                                                count,
2155                                                                &meta_level);
2156                         if (has_refcount)
2157                                 *has_refcount = 1;
2158                         if (direct_io)
2159                                 *direct_io = 0;
2160                 }
2161
2162                 if (ret < 0) {
2163                         mlog_errno(ret);
2164                         goto out_unlock;
2165                 }
2166
2167                 /*
2168                  * Skip the O_DIRECT checks if we don't need
2169                  * them.
2170                  */
2171                 if (!direct_io || !(*direct_io))
2172                         break;
2173
2174                 /*
2175                  * There's no sane way to do direct writes to an inode
2176                  * with inline data.
2177                  */
2178                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2179                         *direct_io = 0;
2180                         break;
2181                 }
2182
2183                 /*
2184                  * Allowing concurrent direct writes means
2185                  * i_size changes wouldn't be synchronized, so
2186                  * one node could wind up truncating another
2187                  * nodes writes.
2188                  */
2189                 if (end > i_size_read(inode)) {
2190                         *direct_io = 0;
2191                         break;
2192                 }
2193
2194                 /*
2195                  * We don't fill holes during direct io, so
2196                  * check for them here. If any are found, the
2197                  * caller will have to retake some cluster
2198                  * locks and initiate the io as buffered.
2199                  */
2200                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2201                 if (ret == 1) {
2202                         *direct_io = 0;
2203                         ret = 0;
2204                 } else if (ret < 0)
2205                         mlog_errno(ret);
2206                 break;
2207         }
2208
2209         if (appending)
2210                 *ppos = saved_pos;
2211
2212 out_unlock:
2213         if (meta_level >= 0)
2214                 ocfs2_inode_unlock(inode, meta_level);
2215
2216 out:
2217         return ret;
2218 }
2219
2220 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2221                                     const struct iovec *iov,
2222                                     unsigned long nr_segs,
2223                                     loff_t pos)
2224 {
2225         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2226         int can_do_direct, has_refcount = 0;
2227         ssize_t written = 0;
2228         size_t ocount;          /* original count */
2229         size_t count;           /* after file limit checks */
2230         loff_t old_size, *ppos = &iocb->ki_pos;
2231         u32 old_clusters;
2232         struct file *file = iocb->ki_filp;
2233         struct inode *inode = file->f_path.dentry->d_inode;
2234         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2235
2236         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2237                    (unsigned int)nr_segs,
2238                    file->f_path.dentry->d_name.len,
2239                    file->f_path.dentry->d_name.name);
2240
2241         if (iocb->ki_left == 0)
2242                 return 0;
2243
2244         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2245
2246         appending = file->f_flags & O_APPEND ? 1 : 0;
2247         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2248
2249         mutex_lock(&inode->i_mutex);
2250
2251 relock:
2252         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2253         if (direct_io) {
2254                 down_read(&inode->i_alloc_sem);
2255                 have_alloc_sem = 1;
2256         }
2257
2258         /* concurrent O_DIRECT writes are allowed */
2259         rw_level = !direct_io;
2260         ret = ocfs2_rw_lock(inode, rw_level);
2261         if (ret < 0) {
2262                 mlog_errno(ret);
2263                 goto out_sems;
2264         }
2265
2266         can_do_direct = direct_io;
2267         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2268                                             iocb->ki_left, appending,
2269                                             &can_do_direct, &has_refcount);
2270         if (ret < 0) {
2271                 mlog_errno(ret);
2272                 goto out;
2273         }
2274
2275         /*
2276          * We can't complete the direct I/O as requested, fall back to
2277          * buffered I/O.
2278          */
2279         if (direct_io && !can_do_direct) {
2280                 ocfs2_rw_unlock(inode, rw_level);
2281                 up_read(&inode->i_alloc_sem);
2282
2283                 have_alloc_sem = 0;
2284                 rw_level = -1;
2285
2286                 direct_io = 0;
2287                 goto relock;
2288         }
2289
2290         /*
2291          * To later detect whether a journal commit for sync writes is
2292          * necessary, we sample i_size, and cluster count here.
2293          */
2294         old_size = i_size_read(inode);
2295         old_clusters = OCFS2_I(inode)->ip_clusters;
2296
2297         /* communicate with ocfs2_dio_end_io */
2298         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2299
2300         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2301                                      VERIFY_READ);
2302         if (ret)
2303                 goto out_dio;
2304
2305         count = ocount;
2306         ret = generic_write_checks(file, ppos, &count,
2307                                    S_ISBLK(inode->i_mode));
2308         if (ret)
2309                 goto out_dio;
2310
2311         if (direct_io) {
2312                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2313                                                     ppos, count, ocount);
2314                 if (written < 0) {
2315                         /*
2316                          * direct write may have instantiated a few
2317                          * blocks outside i_size. Trim these off again.
2318                          * Don't need i_size_read because we hold i_mutex.
2319                          *
2320                          * XXX(truncate): this looks buggy because ocfs2 did not
2321                          * actually implement ->truncate.  Take a look at
2322                          * the new truncate sequence and update this accordingly
2323                          */
2324                         if (*ppos + count > inode->i_size)
2325                                 truncate_setsize(inode, inode->i_size);
2326                         ret = written;
2327                         goto out_dio;
2328                 }
2329         } else {
2330                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2331                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2332                                                       ppos, count, 0);
2333                 current->backing_dev_info = NULL;
2334         }
2335
2336 out_dio:
2337         /* buffered aio wouldn't have proper lock coverage today */
2338         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2339
2340         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2341             ((file->f_flags & O_DIRECT) && has_refcount)) {
2342                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2343                                                pos + count - 1);
2344                 if (ret < 0)
2345                         written = ret;
2346
2347                 if (!ret && ((old_size != i_size_read(inode)) ||
2348                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2349                              has_refcount)) {
2350                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2351                         if (ret < 0)
2352                                 written = ret;
2353                 }
2354
2355                 if (!ret)
2356                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2357                                                       pos + count - 1);
2358         }
2359
2360         /*
2361          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2362          * function pointer which is called when o_direct io completes so that
2363          * it can unlock our rw lock.  (it's the clustered equivalent of
2364          * i_alloc_sem; protects truncate from racing with pending ios).
2365          * Unfortunately there are error cases which call end_io and others
2366          * that don't.  so we don't have to unlock the rw_lock if either an
2367          * async dio is going to do it in the future or an end_io after an
2368          * error has already done it.
2369          */
2370         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2371                 rw_level = -1;
2372                 have_alloc_sem = 0;
2373         }
2374
2375 out:
2376         if (rw_level != -1)
2377                 ocfs2_rw_unlock(inode, rw_level);
2378
2379 out_sems:
2380         if (have_alloc_sem)
2381                 up_read(&inode->i_alloc_sem);
2382
2383         mutex_unlock(&inode->i_mutex);
2384
2385         if (written)
2386                 ret = written;
2387         mlog_exit(ret);
2388         return ret;
2389 }
2390
2391 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2392                                 struct file *out,
2393                                 struct splice_desc *sd)
2394 {
2395         int ret;
2396
2397         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2398                                             sd->total_len, 0, NULL, NULL);
2399         if (ret < 0) {
2400                 mlog_errno(ret);
2401                 return ret;
2402         }
2403
2404         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2405 }
2406
2407 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2408                                        struct file *out,
2409                                        loff_t *ppos,
2410                                        size_t len,
2411                                        unsigned int flags)
2412 {
2413         int ret;
2414         struct address_space *mapping = out->f_mapping;
2415         struct inode *inode = mapping->host;
2416         struct splice_desc sd = {
2417                 .total_len = len,
2418                 .flags = flags,
2419                 .pos = *ppos,
2420                 .u.file = out,
2421         };
2422
2423         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2424                    (unsigned int)len,
2425                    out->f_path.dentry->d_name.len,
2426                    out->f_path.dentry->d_name.name);
2427
2428         if (pipe->inode)
2429                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2430
2431         splice_from_pipe_begin(&sd);
2432         do {
2433                 ret = splice_from_pipe_next(pipe, &sd);
2434                 if (ret <= 0)
2435                         break;
2436
2437                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2438                 ret = ocfs2_rw_lock(inode, 1);
2439                 if (ret < 0)
2440                         mlog_errno(ret);
2441                 else {
2442                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2443                         ocfs2_rw_unlock(inode, 1);
2444                 }
2445                 mutex_unlock(&inode->i_mutex);
2446         } while (ret > 0);
2447         splice_from_pipe_end(pipe, &sd);
2448
2449         if (pipe->inode)
2450                 mutex_unlock(&pipe->inode->i_mutex);
2451
2452         if (sd.num_spliced)
2453                 ret = sd.num_spliced;
2454
2455         if (ret > 0) {
2456                 unsigned long nr_pages;
2457                 int err;
2458
2459                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2460
2461                 err = generic_write_sync(out, *ppos, ret);
2462                 if (err)
2463                         ret = err;
2464                 else
2465                         *ppos += ret;
2466
2467                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2468         }
2469
2470         mlog_exit(ret);
2471         return ret;
2472 }
2473
2474 static ssize_t ocfs2_file_splice_read(struct file *in,
2475                                       loff_t *ppos,
2476                                       struct pipe_inode_info *pipe,
2477                                       size_t len,
2478                                       unsigned int flags)
2479 {
2480         int ret = 0, lock_level = 0;
2481         struct inode *inode = in->f_path.dentry->d_inode;
2482
2483         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2484                    (unsigned int)len,
2485                    in->f_path.dentry->d_name.len,
2486                    in->f_path.dentry->d_name.name);
2487
2488         /*
2489          * See the comment in ocfs2_file_aio_read()
2490          */
2491         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2492         if (ret < 0) {
2493                 mlog_errno(ret);
2494                 goto bail;
2495         }
2496         ocfs2_inode_unlock(inode, lock_level);
2497
2498         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2499
2500 bail:
2501         mlog_exit(ret);
2502         return ret;
2503 }
2504
2505 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2506                                    const struct iovec *iov,
2507                                    unsigned long nr_segs,
2508                                    loff_t pos)
2509 {
2510         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2511         struct file *filp = iocb->ki_filp;
2512         struct inode *inode = filp->f_path.dentry->d_inode;
2513
2514         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2515                    (unsigned int)nr_segs,
2516                    filp->f_path.dentry->d_name.len,
2517                    filp->f_path.dentry->d_name.name);
2518
2519         if (!inode) {
2520                 ret = -EINVAL;
2521                 mlog_errno(ret);
2522                 goto bail;
2523         }
2524
2525         /*
2526          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2527          * need locks to protect pending reads from racing with truncate.
2528          */
2529         if (filp->f_flags & O_DIRECT) {
2530                 down_read(&inode->i_alloc_sem);
2531                 have_alloc_sem = 1;
2532
2533                 ret = ocfs2_rw_lock(inode, 0);
2534                 if (ret < 0) {
2535                         mlog_errno(ret);
2536                         goto bail;
2537                 }
2538                 rw_level = 0;
2539                 /* communicate with ocfs2_dio_end_io */
2540                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2541         }
2542
2543         /*
2544          * We're fine letting folks race truncates and extending
2545          * writes with read across the cluster, just like they can
2546          * locally. Hence no rw_lock during read.
2547          *
2548          * Take and drop the meta data lock to update inode fields
2549          * like i_size. This allows the checks down below
2550          * generic_file_aio_read() a chance of actually working.
2551          */
2552         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2553         if (ret < 0) {
2554                 mlog_errno(ret);
2555                 goto bail;
2556         }
2557         ocfs2_inode_unlock(inode, lock_level);
2558
2559         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2560         if (ret == -EINVAL)
2561                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2562
2563         /* buffered aio wouldn't have proper lock coverage today */
2564         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2565
2566         /* see ocfs2_file_aio_write */
2567         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2568                 rw_level = -1;
2569                 have_alloc_sem = 0;
2570         }
2571
2572 bail:
2573         if (have_alloc_sem)
2574                 up_read(&inode->i_alloc_sem);
2575         if (rw_level != -1)
2576                 ocfs2_rw_unlock(inode, rw_level);
2577         mlog_exit(ret);
2578
2579         return ret;
2580 }
2581
2582 const struct inode_operations ocfs2_file_iops = {
2583         .setattr        = ocfs2_setattr,
2584         .getattr        = ocfs2_getattr,
2585         .permission     = ocfs2_permission,
2586         .setxattr       = generic_setxattr,
2587         .getxattr       = generic_getxattr,
2588         .listxattr      = ocfs2_listxattr,
2589         .removexattr    = generic_removexattr,
2590         .fallocate      = ocfs2_fallocate,
2591         .fiemap         = ocfs2_fiemap,
2592 };
2593
2594 const struct inode_operations ocfs2_special_file_iops = {
2595         .setattr        = ocfs2_setattr,
2596         .getattr        = ocfs2_getattr,
2597         .permission     = ocfs2_permission,
2598 };
2599
2600 /*
2601  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2602  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2603  */
2604 const struct file_operations ocfs2_fops = {
2605         .llseek         = generic_file_llseek,
2606         .read           = do_sync_read,
2607         .write          = do_sync_write,
2608         .mmap           = ocfs2_mmap,
2609         .fsync          = ocfs2_sync_file,
2610         .release        = ocfs2_file_release,
2611         .open           = ocfs2_file_open,
2612         .aio_read       = ocfs2_file_aio_read,
2613         .aio_write      = ocfs2_file_aio_write,
2614         .unlocked_ioctl = ocfs2_ioctl,
2615 #ifdef CONFIG_COMPAT
2616         .compat_ioctl   = ocfs2_compat_ioctl,
2617 #endif
2618         .lock           = ocfs2_lock,
2619         .flock          = ocfs2_flock,
2620         .splice_read    = ocfs2_file_splice_read,
2621         .splice_write   = ocfs2_file_splice_write,
2622 };
2623
2624 const struct file_operations ocfs2_dops = {
2625         .llseek         = generic_file_llseek,
2626         .read           = generic_read_dir,
2627         .readdir        = ocfs2_readdir,
2628         .fsync          = ocfs2_sync_file,
2629         .release        = ocfs2_dir_release,
2630         .open           = ocfs2_dir_open,
2631         .unlocked_ioctl = ocfs2_ioctl,
2632 #ifdef CONFIG_COMPAT
2633         .compat_ioctl   = ocfs2_compat_ioctl,
2634 #endif
2635         .lock           = ocfs2_lock,
2636         .flock          = ocfs2_flock,
2637 };
2638
2639 /*
2640  * POSIX-lockless variants of our file_operations.
2641  *
2642  * These will be used if the underlying cluster stack does not support
2643  * posix file locking, if the user passes the "localflocks" mount
2644  * option, or if we have a local-only fs.
2645  *
2646  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2647  * so we still want it in the case of no stack support for
2648  * plocks. Internally, it will do the right thing when asked to ignore
2649  * the cluster.
2650  */
2651 const struct file_operations ocfs2_fops_no_plocks = {
2652         .llseek         = generic_file_llseek,
2653         .read           = do_sync_read,
2654         .write          = do_sync_write,
2655         .mmap           = ocfs2_mmap,
2656         .fsync          = ocfs2_sync_file,
2657         .release        = ocfs2_file_release,
2658         .open           = ocfs2_file_open,
2659         .aio_read       = ocfs2_file_aio_read,
2660         .aio_write      = ocfs2_file_aio_write,
2661         .unlocked_ioctl = ocfs2_ioctl,
2662 #ifdef CONFIG_COMPAT
2663         .compat_ioctl   = ocfs2_compat_ioctl,
2664 #endif
2665         .flock          = ocfs2_flock,
2666         .splice_read    = ocfs2_file_splice_read,
2667         .splice_write   = ocfs2_file_splice_write,
2668 };
2669
2670 const struct file_operations ocfs2_dops_no_plocks = {
2671         .llseek         = generic_file_llseek,
2672         .read           = generic_read_dir,
2673         .readdir        = ocfs2_readdir,
2674         .fsync          = ocfs2_sync_file,
2675         .release        = ocfs2_dir_release,
2676         .open           = ocfs2_dir_open,
2677         .unlocked_ioctl = ocfs2_ioctl,
2678 #ifdef CONFIG_COMPAT
2679         .compat_ioctl   = ocfs2_compat_ioctl,
2680 #endif
2681         .flock          = ocfs2_flock,
2682 };