ocfs2: Remove EXIT from masklog.
[linux-2.6.git] / fs / ocfs2 / file.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/capability.h>
27 #include <linux/fs.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
38 #include <linux/quotaops.h>
39 #include <linux/blkdev.h>
40
41 #define MLOG_MASK_PREFIX ML_INODE
42 #include <cluster/masklog.h>
43
44 #include "ocfs2.h"
45
46 #include "alloc.h"
47 #include "aops.h"
48 #include "dir.h"
49 #include "dlmglue.h"
50 #include "extent_map.h"
51 #include "file.h"
52 #include "sysfile.h"
53 #include "inode.h"
54 #include "ioctl.h"
55 #include "journal.h"
56 #include "locks.h"
57 #include "mmap.h"
58 #include "suballoc.h"
59 #include "super.h"
60 #include "xattr.h"
61 #include "acl.h"
62 #include "quota.h"
63 #include "refcounttree.h"
64
65 #include "buffer_head_io.h"
66
67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
68 {
69         struct ocfs2_file_private *fp;
70
71         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
72         if (!fp)
73                 return -ENOMEM;
74
75         fp->fp_file = file;
76         mutex_init(&fp->fp_mutex);
77         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78         file->private_data = fp;
79
80         return 0;
81 }
82
83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
84 {
85         struct ocfs2_file_private *fp = file->private_data;
86         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
87
88         if (fp) {
89                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90                 ocfs2_lock_res_free(&fp->fp_flock);
91                 kfree(fp);
92                 file->private_data = NULL;
93         }
94 }
95
96 static int ocfs2_file_open(struct inode *inode, struct file *file)
97 {
98         int status;
99         int mode = file->f_flags;
100         struct ocfs2_inode_info *oi = OCFS2_I(inode);
101
102         mlog(0, "(0x%p, 0x%p, '%.*s')\n", inode, file,
103              file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
104
105         if (file->f_mode & FMODE_WRITE)
106                 dquot_initialize(inode);
107
108         spin_lock(&oi->ip_lock);
109
110         /* Check that the inode hasn't been wiped from disk by another
111          * node. If it hasn't then we're safe as long as we hold the
112          * spin lock until our increment of open count. */
113         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
114                 spin_unlock(&oi->ip_lock);
115
116                 status = -ENOENT;
117                 goto leave;
118         }
119
120         if (mode & O_DIRECT)
121                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
122
123         oi->ip_open_count++;
124         spin_unlock(&oi->ip_lock);
125
126         status = ocfs2_init_file_private(inode, file);
127         if (status) {
128                 /*
129                  * We want to set open count back if we're failing the
130                  * open.
131                  */
132                 spin_lock(&oi->ip_lock);
133                 oi->ip_open_count--;
134                 spin_unlock(&oi->ip_lock);
135         }
136
137 leave:
138         return status;
139 }
140
141 static int ocfs2_file_release(struct inode *inode, struct file *file)
142 {
143         struct ocfs2_inode_info *oi = OCFS2_I(inode);
144
145         mlog(0, "(0x%p, 0x%p, '%.*s')\n", inode, file,
146              file->f_path.dentry->d_name.len,
147              file->f_path.dentry->d_name.name);
148
149         spin_lock(&oi->ip_lock);
150         if (!--oi->ip_open_count)
151                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
152         spin_unlock(&oi->ip_lock);
153
154         ocfs2_free_file_private(inode, file);
155
156         return 0;
157 }
158
159 static int ocfs2_dir_open(struct inode *inode, struct file *file)
160 {
161         return ocfs2_init_file_private(inode, file);
162 }
163
164 static int ocfs2_dir_release(struct inode *inode, struct file *file)
165 {
166         ocfs2_free_file_private(inode, file);
167         return 0;
168 }
169
170 static int ocfs2_sync_file(struct file *file, int datasync)
171 {
172         int err = 0;
173         journal_t *journal;
174         struct inode *inode = file->f_mapping->host;
175         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
176
177         mlog(0, "(0x%p, %d, 0x%p, '%.*s')\n", file, datasync,
178              file->f_path.dentry, file->f_path.dentry->d_name.len,
179              file->f_path.dentry->d_name.name);
180
181         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC)) {
182                 /*
183                  * We still have to flush drive's caches to get data to the
184                  * platter
185                  */
186                 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
187                         blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
188                 goto bail;
189         }
190
191         journal = osb->journal->j_journal;
192         err = jbd2_journal_force_commit(journal);
193
194 bail:
195         if (err)
196                 mlog_errno(err);
197
198         return (err < 0) ? -EIO : 0;
199 }
200
201 int ocfs2_should_update_atime(struct inode *inode,
202                               struct vfsmount *vfsmnt)
203 {
204         struct timespec now;
205         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
206
207         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
208                 return 0;
209
210         if ((inode->i_flags & S_NOATIME) ||
211             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
212                 return 0;
213
214         /*
215          * We can be called with no vfsmnt structure - NFSD will
216          * sometimes do this.
217          *
218          * Note that our action here is different than touch_atime() -
219          * if we can't tell whether this is a noatime mount, then we
220          * don't know whether to trust the value of s_atime_quantum.
221          */
222         if (vfsmnt == NULL)
223                 return 0;
224
225         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
226             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
227                 return 0;
228
229         if (vfsmnt->mnt_flags & MNT_RELATIME) {
230                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
231                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
232                         return 1;
233
234                 return 0;
235         }
236
237         now = CURRENT_TIME;
238         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
239                 return 0;
240         else
241                 return 1;
242 }
243
244 int ocfs2_update_inode_atime(struct inode *inode,
245                              struct buffer_head *bh)
246 {
247         int ret;
248         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
249         handle_t *handle;
250         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
251
252         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
253         if (IS_ERR(handle)) {
254                 ret = PTR_ERR(handle);
255                 mlog_errno(ret);
256                 goto out;
257         }
258
259         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
260                                       OCFS2_JOURNAL_ACCESS_WRITE);
261         if (ret) {
262                 mlog_errno(ret);
263                 goto out_commit;
264         }
265
266         /*
267          * Don't use ocfs2_mark_inode_dirty() here as we don't always
268          * have i_mutex to guard against concurrent changes to other
269          * inode fields.
270          */
271         inode->i_atime = CURRENT_TIME;
272         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
273         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
274         ocfs2_journal_dirty(handle, bh);
275
276 out_commit:
277         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
278 out:
279         return ret;
280 }
281
282 static int ocfs2_set_inode_size(handle_t *handle,
283                                 struct inode *inode,
284                                 struct buffer_head *fe_bh,
285                                 u64 new_i_size)
286 {
287         int status;
288
289         i_size_write(inode, new_i_size);
290         inode->i_blocks = ocfs2_inode_sector_count(inode);
291         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
292
293         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
294         if (status < 0) {
295                 mlog_errno(status);
296                 goto bail;
297         }
298
299 bail:
300         return status;
301 }
302
303 int ocfs2_simple_size_update(struct inode *inode,
304                              struct buffer_head *di_bh,
305                              u64 new_i_size)
306 {
307         int ret;
308         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
309         handle_t *handle = NULL;
310
311         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
312         if (IS_ERR(handle)) {
313                 ret = PTR_ERR(handle);
314                 mlog_errno(ret);
315                 goto out;
316         }
317
318         ret = ocfs2_set_inode_size(handle, inode, di_bh,
319                                    new_i_size);
320         if (ret < 0)
321                 mlog_errno(ret);
322
323         ocfs2_commit_trans(osb, handle);
324 out:
325         return ret;
326 }
327
328 static int ocfs2_cow_file_pos(struct inode *inode,
329                               struct buffer_head *fe_bh,
330                               u64 offset)
331 {
332         int status;
333         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
334         unsigned int num_clusters = 0;
335         unsigned int ext_flags = 0;
336
337         /*
338          * If the new offset is aligned to the range of the cluster, there is
339          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
340          * CoW either.
341          */
342         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
343                 return 0;
344
345         status = ocfs2_get_clusters(inode, cpos, &phys,
346                                     &num_clusters, &ext_flags);
347         if (status) {
348                 mlog_errno(status);
349                 goto out;
350         }
351
352         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
353                 goto out;
354
355         return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
356
357 out:
358         return status;
359 }
360
361 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
362                                      struct inode *inode,
363                                      struct buffer_head *fe_bh,
364                                      u64 new_i_size)
365 {
366         int status;
367         handle_t *handle;
368         struct ocfs2_dinode *di;
369         u64 cluster_bytes;
370
371         /*
372          * We need to CoW the cluster contains the offset if it is reflinked
373          * since we will call ocfs2_zero_range_for_truncate later which will
374          * write "0" from offset to the end of the cluster.
375          */
376         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
377         if (status) {
378                 mlog_errno(status);
379                 return status;
380         }
381
382         /* TODO: This needs to actually orphan the inode in this
383          * transaction. */
384
385         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
386         if (IS_ERR(handle)) {
387                 status = PTR_ERR(handle);
388                 mlog_errno(status);
389                 goto out;
390         }
391
392         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
393                                          OCFS2_JOURNAL_ACCESS_WRITE);
394         if (status < 0) {
395                 mlog_errno(status);
396                 goto out_commit;
397         }
398
399         /*
400          * Do this before setting i_size.
401          */
402         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
403         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
404                                                cluster_bytes);
405         if (status) {
406                 mlog_errno(status);
407                 goto out_commit;
408         }
409
410         i_size_write(inode, new_i_size);
411         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
412
413         di = (struct ocfs2_dinode *) fe_bh->b_data;
414         di->i_size = cpu_to_le64(new_i_size);
415         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
416         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
417
418         ocfs2_journal_dirty(handle, fe_bh);
419
420 out_commit:
421         ocfs2_commit_trans(osb, handle);
422 out:
423         return status;
424 }
425
426 static int ocfs2_truncate_file(struct inode *inode,
427                                struct buffer_head *di_bh,
428                                u64 new_i_size)
429 {
430         int status = 0;
431         struct ocfs2_dinode *fe = NULL;
432         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
433
434         mlog(0, "(inode = %llu, new_i_size = %llu\n",
435              (unsigned long long)OCFS2_I(inode)->ip_blkno,
436              (unsigned long long)new_i_size);
437
438         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
439          * already validated it */
440         fe = (struct ocfs2_dinode *) di_bh->b_data;
441
442         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
443                         "Inode %llu, inode i_size = %lld != di "
444                         "i_size = %llu, i_flags = 0x%x\n",
445                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
446                         i_size_read(inode),
447                         (unsigned long long)le64_to_cpu(fe->i_size),
448                         le32_to_cpu(fe->i_flags));
449
450         if (new_i_size > le64_to_cpu(fe->i_size)) {
451                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
452                      (unsigned long long)le64_to_cpu(fe->i_size),
453                      (unsigned long long)new_i_size);
454                 status = -EINVAL;
455                 mlog_errno(status);
456                 goto bail;
457         }
458
459         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
460              (unsigned long long)le64_to_cpu(fe->i_blkno),
461              (unsigned long long)le64_to_cpu(fe->i_size),
462              (unsigned long long)new_i_size);
463
464         /* lets handle the simple truncate cases before doing any more
465          * cluster locking. */
466         if (new_i_size == le64_to_cpu(fe->i_size))
467                 goto bail;
468
469         down_write(&OCFS2_I(inode)->ip_alloc_sem);
470
471         ocfs2_resv_discard(&osb->osb_la_resmap,
472                            &OCFS2_I(inode)->ip_la_data_resv);
473
474         /*
475          * The inode lock forced other nodes to sync and drop their
476          * pages, which (correctly) happens even if we have a truncate
477          * without allocation change - ocfs2 cluster sizes can be much
478          * greater than page size, so we have to truncate them
479          * anyway.
480          */
481         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
482         truncate_inode_pages(inode->i_mapping, new_i_size);
483
484         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
485                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
486                                                i_size_read(inode), 1);
487                 if (status)
488                         mlog_errno(status);
489
490                 goto bail_unlock_sem;
491         }
492
493         /* alright, we're going to need to do a full blown alloc size
494          * change. Orphan the inode so that recovery can complete the
495          * truncate if necessary. This does the task of marking
496          * i_size. */
497         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
498         if (status < 0) {
499                 mlog_errno(status);
500                 goto bail_unlock_sem;
501         }
502
503         status = ocfs2_commit_truncate(osb, inode, di_bh);
504         if (status < 0) {
505                 mlog_errno(status);
506                 goto bail_unlock_sem;
507         }
508
509         /* TODO: orphan dir cleanup here. */
510 bail_unlock_sem:
511         up_write(&OCFS2_I(inode)->ip_alloc_sem);
512
513 bail:
514         if (!status && OCFS2_I(inode)->ip_clusters == 0)
515                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
516
517         return status;
518 }
519
520 /*
521  * extend file allocation only here.
522  * we'll update all the disk stuff, and oip->alloc_size
523  *
524  * expect stuff to be locked, a transaction started and enough data /
525  * metadata reservations in the contexts.
526  *
527  * Will return -EAGAIN, and a reason if a restart is needed.
528  * If passed in, *reason will always be set, even in error.
529  */
530 int ocfs2_add_inode_data(struct ocfs2_super *osb,
531                          struct inode *inode,
532                          u32 *logical_offset,
533                          u32 clusters_to_add,
534                          int mark_unwritten,
535                          struct buffer_head *fe_bh,
536                          handle_t *handle,
537                          struct ocfs2_alloc_context *data_ac,
538                          struct ocfs2_alloc_context *meta_ac,
539                          enum ocfs2_alloc_restarted *reason_ret)
540 {
541         int ret;
542         struct ocfs2_extent_tree et;
543
544         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
545         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
546                                           clusters_to_add, mark_unwritten,
547                                           data_ac, meta_ac, reason_ret);
548
549         return ret;
550 }
551
552 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
553                                      u32 clusters_to_add, int mark_unwritten)
554 {
555         int status = 0;
556         int restart_func = 0;
557         int credits;
558         u32 prev_clusters;
559         struct buffer_head *bh = NULL;
560         struct ocfs2_dinode *fe = NULL;
561         handle_t *handle = NULL;
562         struct ocfs2_alloc_context *data_ac = NULL;
563         struct ocfs2_alloc_context *meta_ac = NULL;
564         enum ocfs2_alloc_restarted why;
565         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
566         struct ocfs2_extent_tree et;
567         int did_quota = 0;
568
569         mlog(0, "(clusters_to_add = %u)\n", clusters_to_add);
570
571         /*
572          * This function only exists for file systems which don't
573          * support holes.
574          */
575         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
576
577         status = ocfs2_read_inode_block(inode, &bh);
578         if (status < 0) {
579                 mlog_errno(status);
580                 goto leave;
581         }
582         fe = (struct ocfs2_dinode *) bh->b_data;
583
584 restart_all:
585         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
586
587         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
588              "clusters_to_add = %u\n",
589              (unsigned long long)OCFS2_I(inode)->ip_blkno,
590              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
591              clusters_to_add);
592         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
593         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
594                                        &data_ac, &meta_ac);
595         if (status) {
596                 mlog_errno(status);
597                 goto leave;
598         }
599
600         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
601                                             clusters_to_add);
602         handle = ocfs2_start_trans(osb, credits);
603         if (IS_ERR(handle)) {
604                 status = PTR_ERR(handle);
605                 handle = NULL;
606                 mlog_errno(status);
607                 goto leave;
608         }
609
610 restarted_transaction:
611         status = dquot_alloc_space_nodirty(inode,
612                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
613         if (status)
614                 goto leave;
615         did_quota = 1;
616
617         /* reserve a write to the file entry early on - that we if we
618          * run out of credits in the allocation path, we can still
619          * update i_size. */
620         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
621                                          OCFS2_JOURNAL_ACCESS_WRITE);
622         if (status < 0) {
623                 mlog_errno(status);
624                 goto leave;
625         }
626
627         prev_clusters = OCFS2_I(inode)->ip_clusters;
628
629         status = ocfs2_add_inode_data(osb,
630                                       inode,
631                                       &logical_start,
632                                       clusters_to_add,
633                                       mark_unwritten,
634                                       bh,
635                                       handle,
636                                       data_ac,
637                                       meta_ac,
638                                       &why);
639         if ((status < 0) && (status != -EAGAIN)) {
640                 if (status != -ENOSPC)
641                         mlog_errno(status);
642                 goto leave;
643         }
644
645         ocfs2_journal_dirty(handle, bh);
646
647         spin_lock(&OCFS2_I(inode)->ip_lock);
648         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
649         spin_unlock(&OCFS2_I(inode)->ip_lock);
650         /* Release unused quota reservation */
651         dquot_free_space(inode,
652                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
653         did_quota = 0;
654
655         if (why != RESTART_NONE && clusters_to_add) {
656                 if (why == RESTART_META) {
657                         mlog(0, "restarting function.\n");
658                         restart_func = 1;
659                         status = 0;
660                 } else {
661                         BUG_ON(why != RESTART_TRANS);
662
663                         mlog(0, "restarting transaction.\n");
664                         /* TODO: This can be more intelligent. */
665                         credits = ocfs2_calc_extend_credits(osb->sb,
666                                                             &fe->id2.i_list,
667                                                             clusters_to_add);
668                         status = ocfs2_extend_trans(handle, credits);
669                         if (status < 0) {
670                                 /* handle still has to be committed at
671                                  * this point. */
672                                 status = -ENOMEM;
673                                 mlog_errno(status);
674                                 goto leave;
675                         }
676                         goto restarted_transaction;
677                 }
678         }
679
680         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
681              le32_to_cpu(fe->i_clusters),
682              (unsigned long long)le64_to_cpu(fe->i_size));
683         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
684              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
685
686 leave:
687         if (status < 0 && did_quota)
688                 dquot_free_space(inode,
689                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
690         if (handle) {
691                 ocfs2_commit_trans(osb, handle);
692                 handle = NULL;
693         }
694         if (data_ac) {
695                 ocfs2_free_alloc_context(data_ac);
696                 data_ac = NULL;
697         }
698         if (meta_ac) {
699                 ocfs2_free_alloc_context(meta_ac);
700                 meta_ac = NULL;
701         }
702         if ((!status) && restart_func) {
703                 restart_func = 0;
704                 goto restart_all;
705         }
706         brelse(bh);
707         bh = NULL;
708
709         return status;
710 }
711
712 /*
713  * While a write will already be ordering the data, a truncate will not.
714  * Thus, we need to explicitly order the zeroed pages.
715  */
716 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
717 {
718         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
719         handle_t *handle = NULL;
720         int ret = 0;
721
722         if (!ocfs2_should_order_data(inode))
723                 goto out;
724
725         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
726         if (IS_ERR(handle)) {
727                 ret = -ENOMEM;
728                 mlog_errno(ret);
729                 goto out;
730         }
731
732         ret = ocfs2_jbd2_file_inode(handle, inode);
733         if (ret < 0)
734                 mlog_errno(ret);
735
736 out:
737         if (ret) {
738                 if (!IS_ERR(handle))
739                         ocfs2_commit_trans(osb, handle);
740                 handle = ERR_PTR(ret);
741         }
742         return handle;
743 }
744
745 /* Some parts of this taken from generic_cont_expand, which turned out
746  * to be too fragile to do exactly what we need without us having to
747  * worry about recursive locking in ->write_begin() and ->write_end(). */
748 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
749                                  u64 abs_to)
750 {
751         struct address_space *mapping = inode->i_mapping;
752         struct page *page;
753         unsigned long index = abs_from >> PAGE_CACHE_SHIFT;
754         handle_t *handle = NULL;
755         int ret = 0;
756         unsigned zero_from, zero_to, block_start, block_end;
757
758         BUG_ON(abs_from >= abs_to);
759         BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
760         BUG_ON(abs_from & (inode->i_blkbits - 1));
761
762         page = find_or_create_page(mapping, index, GFP_NOFS);
763         if (!page) {
764                 ret = -ENOMEM;
765                 mlog_errno(ret);
766                 goto out;
767         }
768
769         /* Get the offsets within the page that we want to zero */
770         zero_from = abs_from & (PAGE_CACHE_SIZE - 1);
771         zero_to = abs_to & (PAGE_CACHE_SIZE - 1);
772         if (!zero_to)
773                 zero_to = PAGE_CACHE_SIZE;
774
775         mlog(0,
776              "abs_from = %llu, abs_to = %llu, index = %lu, zero_from = %u, zero_to = %u\n",
777              (unsigned long long)abs_from, (unsigned long long)abs_to,
778              index, zero_from, zero_to);
779
780         /* We know that zero_from is block aligned */
781         for (block_start = zero_from; block_start < zero_to;
782              block_start = block_end) {
783                 block_end = block_start + (1 << inode->i_blkbits);
784
785                 /*
786                  * block_start is block-aligned.  Bump it by one to force
787                  * __block_write_begin and block_commit_write to zero the
788                  * whole block.
789                  */
790                 ret = __block_write_begin(page, block_start + 1, 0,
791                                           ocfs2_get_block);
792                 if (ret < 0) {
793                         mlog_errno(ret);
794                         goto out_unlock;
795                 }
796
797                 if (!handle) {
798                         handle = ocfs2_zero_start_ordered_transaction(inode);
799                         if (IS_ERR(handle)) {
800                                 ret = PTR_ERR(handle);
801                                 handle = NULL;
802                                 break;
803                         }
804                 }
805
806                 /* must not update i_size! */
807                 ret = block_commit_write(page, block_start + 1,
808                                          block_start + 1);
809                 if (ret < 0)
810                         mlog_errno(ret);
811                 else
812                         ret = 0;
813         }
814
815         if (handle)
816                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
817
818 out_unlock:
819         unlock_page(page);
820         page_cache_release(page);
821 out:
822         return ret;
823 }
824
825 /*
826  * Find the next range to zero.  We do this in terms of bytes because
827  * that's what ocfs2_zero_extend() wants, and it is dealing with the
828  * pagecache.  We may return multiple extents.
829  *
830  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
831  * needs to be zeroed.  range_start and range_end return the next zeroing
832  * range.  A subsequent call should pass the previous range_end as its
833  * zero_start.  If range_end is 0, there's nothing to do.
834  *
835  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
836  */
837 static int ocfs2_zero_extend_get_range(struct inode *inode,
838                                        struct buffer_head *di_bh,
839                                        u64 zero_start, u64 zero_end,
840                                        u64 *range_start, u64 *range_end)
841 {
842         int rc = 0, needs_cow = 0;
843         u32 p_cpos, zero_clusters = 0;
844         u32 zero_cpos =
845                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
846         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
847         unsigned int num_clusters = 0;
848         unsigned int ext_flags = 0;
849
850         while (zero_cpos < last_cpos) {
851                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
852                                         &num_clusters, &ext_flags);
853                 if (rc) {
854                         mlog_errno(rc);
855                         goto out;
856                 }
857
858                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
859                         zero_clusters = num_clusters;
860                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
861                                 needs_cow = 1;
862                         break;
863                 }
864
865                 zero_cpos += num_clusters;
866         }
867         if (!zero_clusters) {
868                 *range_end = 0;
869                 goto out;
870         }
871
872         while ((zero_cpos + zero_clusters) < last_cpos) {
873                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
874                                         &p_cpos, &num_clusters,
875                                         &ext_flags);
876                 if (rc) {
877                         mlog_errno(rc);
878                         goto out;
879                 }
880
881                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
882                         break;
883                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
884                         needs_cow = 1;
885                 zero_clusters += num_clusters;
886         }
887         if ((zero_cpos + zero_clusters) > last_cpos)
888                 zero_clusters = last_cpos - zero_cpos;
889
890         if (needs_cow) {
891                 rc = ocfs2_refcount_cow(inode, NULL, di_bh, zero_cpos,
892                                         zero_clusters, UINT_MAX);
893                 if (rc) {
894                         mlog_errno(rc);
895                         goto out;
896                 }
897         }
898
899         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
900         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
901                                              zero_cpos + zero_clusters);
902
903 out:
904         return rc;
905 }
906
907 /*
908  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
909  * has made sure that the entire range needs zeroing.
910  */
911 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
912                                    u64 range_end)
913 {
914         int rc = 0;
915         u64 next_pos;
916         u64 zero_pos = range_start;
917
918         mlog(0, "range_start = %llu, range_end = %llu\n",
919              (unsigned long long)range_start,
920              (unsigned long long)range_end);
921         BUG_ON(range_start >= range_end);
922
923         while (zero_pos < range_end) {
924                 next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
925                 if (next_pos > range_end)
926                         next_pos = range_end;
927                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
928                 if (rc < 0) {
929                         mlog_errno(rc);
930                         break;
931                 }
932                 zero_pos = next_pos;
933
934                 /*
935                  * Very large extends have the potential to lock up
936                  * the cpu for extended periods of time.
937                  */
938                 cond_resched();
939         }
940
941         return rc;
942 }
943
944 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
945                       loff_t zero_to_size)
946 {
947         int ret = 0;
948         u64 zero_start, range_start = 0, range_end = 0;
949         struct super_block *sb = inode->i_sb;
950
951         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
952         mlog(0, "zero_start %llu for i_size %llu\n",
953              (unsigned long long)zero_start,
954              (unsigned long long)i_size_read(inode));
955         while (zero_start < zero_to_size) {
956                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
957                                                   zero_to_size,
958                                                   &range_start,
959                                                   &range_end);
960                 if (ret) {
961                         mlog_errno(ret);
962                         break;
963                 }
964                 if (!range_end)
965                         break;
966                 /* Trim the ends */
967                 if (range_start < zero_start)
968                         range_start = zero_start;
969                 if (range_end > zero_to_size)
970                         range_end = zero_to_size;
971
972                 ret = ocfs2_zero_extend_range(inode, range_start,
973                                               range_end);
974                 if (ret) {
975                         mlog_errno(ret);
976                         break;
977                 }
978                 zero_start = range_end;
979         }
980
981         return ret;
982 }
983
984 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
985                           u64 new_i_size, u64 zero_to)
986 {
987         int ret;
988         u32 clusters_to_add;
989         struct ocfs2_inode_info *oi = OCFS2_I(inode);
990
991         /*
992          * Only quota files call this without a bh, and they can't be
993          * refcounted.
994          */
995         BUG_ON(!di_bh && (oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL));
996         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
997
998         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
999         if (clusters_to_add < oi->ip_clusters)
1000                 clusters_to_add = 0;
1001         else
1002                 clusters_to_add -= oi->ip_clusters;
1003
1004         if (clusters_to_add) {
1005                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1006                                                 clusters_to_add, 0);
1007                 if (ret) {
1008                         mlog_errno(ret);
1009                         goto out;
1010                 }
1011         }
1012
1013         /*
1014          * Call this even if we don't add any clusters to the tree. We
1015          * still need to zero the area between the old i_size and the
1016          * new i_size.
1017          */
1018         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1019         if (ret < 0)
1020                 mlog_errno(ret);
1021
1022 out:
1023         return ret;
1024 }
1025
1026 static int ocfs2_extend_file(struct inode *inode,
1027                              struct buffer_head *di_bh,
1028                              u64 new_i_size)
1029 {
1030         int ret = 0;
1031         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1032
1033         BUG_ON(!di_bh);
1034
1035         /* setattr sometimes calls us like this. */
1036         if (new_i_size == 0)
1037                 goto out;
1038
1039         if (i_size_read(inode) == new_i_size)
1040                 goto out;
1041         BUG_ON(new_i_size < i_size_read(inode));
1042
1043         /*
1044          * The alloc sem blocks people in read/write from reading our
1045          * allocation until we're done changing it. We depend on
1046          * i_mutex to block other extend/truncate calls while we're
1047          * here.  We even have to hold it for sparse files because there
1048          * might be some tail zeroing.
1049          */
1050         down_write(&oi->ip_alloc_sem);
1051
1052         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1053                 /*
1054                  * We can optimize small extends by keeping the inodes
1055                  * inline data.
1056                  */
1057                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1058                         up_write(&oi->ip_alloc_sem);
1059                         goto out_update_size;
1060                 }
1061
1062                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1063                 if (ret) {
1064                         up_write(&oi->ip_alloc_sem);
1065                         mlog_errno(ret);
1066                         goto out;
1067                 }
1068         }
1069
1070         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1071                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1072         else
1073                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1074                                             new_i_size);
1075
1076         up_write(&oi->ip_alloc_sem);
1077
1078         if (ret < 0) {
1079                 mlog_errno(ret);
1080                 goto out;
1081         }
1082
1083 out_update_size:
1084         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1085         if (ret < 0)
1086                 mlog_errno(ret);
1087
1088 out:
1089         return ret;
1090 }
1091
1092 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1093 {
1094         int status = 0, size_change;
1095         struct inode *inode = dentry->d_inode;
1096         struct super_block *sb = inode->i_sb;
1097         struct ocfs2_super *osb = OCFS2_SB(sb);
1098         struct buffer_head *bh = NULL;
1099         handle_t *handle = NULL;
1100         struct dquot *transfer_to[MAXQUOTAS] = { };
1101         int qtype;
1102
1103         mlog(0, "(0x%p, '%.*s')\n", dentry,
1104              dentry->d_name.len, dentry->d_name.name);
1105
1106         /* ensuring we don't even attempt to truncate a symlink */
1107         if (S_ISLNK(inode->i_mode))
1108                 attr->ia_valid &= ~ATTR_SIZE;
1109
1110         if (attr->ia_valid & ATTR_MODE)
1111                 mlog(0, "mode change: %d\n", attr->ia_mode);
1112         if (attr->ia_valid & ATTR_UID)
1113                 mlog(0, "uid change: %d\n", attr->ia_uid);
1114         if (attr->ia_valid & ATTR_GID)
1115                 mlog(0, "gid change: %d\n", attr->ia_gid);
1116         if (attr->ia_valid & ATTR_SIZE)
1117                 mlog(0, "size change...\n");
1118         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1119                 mlog(0, "time change...\n");
1120
1121 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1122                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1123         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1124                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1125                 return 0;
1126         }
1127
1128         status = inode_change_ok(inode, attr);
1129         if (status)
1130                 return status;
1131
1132         if (is_quota_modification(inode, attr))
1133                 dquot_initialize(inode);
1134         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1135         if (size_change) {
1136                 status = ocfs2_rw_lock(inode, 1);
1137                 if (status < 0) {
1138                         mlog_errno(status);
1139                         goto bail;
1140                 }
1141         }
1142
1143         status = ocfs2_inode_lock(inode, &bh, 1);
1144         if (status < 0) {
1145                 if (status != -ENOENT)
1146                         mlog_errno(status);
1147                 goto bail_unlock_rw;
1148         }
1149
1150         if (size_change && attr->ia_size != i_size_read(inode)) {
1151                 status = inode_newsize_ok(inode, attr->ia_size);
1152                 if (status)
1153                         goto bail_unlock;
1154
1155                 if (i_size_read(inode) > attr->ia_size) {
1156                         if (ocfs2_should_order_data(inode)) {
1157                                 status = ocfs2_begin_ordered_truncate(inode,
1158                                                                       attr->ia_size);
1159                                 if (status)
1160                                         goto bail_unlock;
1161                         }
1162                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1163                 } else
1164                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1165                 if (status < 0) {
1166                         if (status != -ENOSPC)
1167                                 mlog_errno(status);
1168                         status = -ENOSPC;
1169                         goto bail_unlock;
1170                 }
1171         }
1172
1173         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1174             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1175                 /*
1176                  * Gather pointers to quota structures so that allocation /
1177                  * freeing of quota structures happens here and not inside
1178                  * dquot_transfer() where we have problems with lock ordering
1179                  */
1180                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1181                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1182                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1183                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1184                                                       USRQUOTA);
1185                         if (!transfer_to[USRQUOTA]) {
1186                                 status = -ESRCH;
1187                                 goto bail_unlock;
1188                         }
1189                 }
1190                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1191                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1192                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1193                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1194                                                       GRPQUOTA);
1195                         if (!transfer_to[GRPQUOTA]) {
1196                                 status = -ESRCH;
1197                                 goto bail_unlock;
1198                         }
1199                 }
1200                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1201                                            2 * ocfs2_quota_trans_credits(sb));
1202                 if (IS_ERR(handle)) {
1203                         status = PTR_ERR(handle);
1204                         mlog_errno(status);
1205                         goto bail_unlock;
1206                 }
1207                 status = __dquot_transfer(inode, transfer_to);
1208                 if (status < 0)
1209                         goto bail_commit;
1210         } else {
1211                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1212                 if (IS_ERR(handle)) {
1213                         status = PTR_ERR(handle);
1214                         mlog_errno(status);
1215                         goto bail_unlock;
1216                 }
1217         }
1218
1219         /*
1220          * This will intentionally not wind up calling truncate_setsize(),
1221          * since all the work for a size change has been done above.
1222          * Otherwise, we could get into problems with truncate as
1223          * ip_alloc_sem is used there to protect against i_size
1224          * changes.
1225          *
1226          * XXX: this means the conditional below can probably be removed.
1227          */
1228         if ((attr->ia_valid & ATTR_SIZE) &&
1229             attr->ia_size != i_size_read(inode)) {
1230                 status = vmtruncate(inode, attr->ia_size);
1231                 if (status) {
1232                         mlog_errno(status);
1233                         goto bail_commit;
1234                 }
1235         }
1236
1237         setattr_copy(inode, attr);
1238         mark_inode_dirty(inode);
1239
1240         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1241         if (status < 0)
1242                 mlog_errno(status);
1243
1244 bail_commit:
1245         ocfs2_commit_trans(osb, handle);
1246 bail_unlock:
1247         ocfs2_inode_unlock(inode, 1);
1248 bail_unlock_rw:
1249         if (size_change)
1250                 ocfs2_rw_unlock(inode, 1);
1251 bail:
1252         brelse(bh);
1253
1254         /* Release quota pointers in case we acquired them */
1255         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1256                 dqput(transfer_to[qtype]);
1257
1258         if (!status && attr->ia_valid & ATTR_MODE) {
1259                 status = ocfs2_acl_chmod(inode);
1260                 if (status < 0)
1261                         mlog_errno(status);
1262         }
1263
1264         return status;
1265 }
1266
1267 int ocfs2_getattr(struct vfsmount *mnt,
1268                   struct dentry *dentry,
1269                   struct kstat *stat)
1270 {
1271         struct inode *inode = dentry->d_inode;
1272         struct super_block *sb = dentry->d_inode->i_sb;
1273         struct ocfs2_super *osb = sb->s_fs_info;
1274         int err;
1275
1276         err = ocfs2_inode_revalidate(dentry);
1277         if (err) {
1278                 if (err != -ENOENT)
1279                         mlog_errno(err);
1280                 goto bail;
1281         }
1282
1283         generic_fillattr(inode, stat);
1284
1285         /* We set the blksize from the cluster size for performance */
1286         stat->blksize = osb->s_clustersize;
1287
1288 bail:
1289         return err;
1290 }
1291
1292 int ocfs2_permission(struct inode *inode, int mask, unsigned int flags)
1293 {
1294         int ret;
1295
1296         if (flags & IPERM_FLAG_RCU)
1297                 return -ECHILD;
1298
1299         ret = ocfs2_inode_lock(inode, NULL, 0);
1300         if (ret) {
1301                 if (ret != -ENOENT)
1302                         mlog_errno(ret);
1303                 goto out;
1304         }
1305
1306         ret = generic_permission(inode, mask, flags, ocfs2_check_acl);
1307
1308         ocfs2_inode_unlock(inode, 0);
1309 out:
1310         return ret;
1311 }
1312
1313 static int __ocfs2_write_remove_suid(struct inode *inode,
1314                                      struct buffer_head *bh)
1315 {
1316         int ret;
1317         handle_t *handle;
1318         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1319         struct ocfs2_dinode *di;
1320
1321         mlog(0, "(Inode %llu, mode 0%o)\n",
1322              (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1323
1324         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1325         if (IS_ERR(handle)) {
1326                 ret = PTR_ERR(handle);
1327                 mlog_errno(ret);
1328                 goto out;
1329         }
1330
1331         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1332                                       OCFS2_JOURNAL_ACCESS_WRITE);
1333         if (ret < 0) {
1334                 mlog_errno(ret);
1335                 goto out_trans;
1336         }
1337
1338         inode->i_mode &= ~S_ISUID;
1339         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1340                 inode->i_mode &= ~S_ISGID;
1341
1342         di = (struct ocfs2_dinode *) bh->b_data;
1343         di->i_mode = cpu_to_le16(inode->i_mode);
1344
1345         ocfs2_journal_dirty(handle, bh);
1346
1347 out_trans:
1348         ocfs2_commit_trans(osb, handle);
1349 out:
1350         return ret;
1351 }
1352
1353 /*
1354  * Will look for holes and unwritten extents in the range starting at
1355  * pos for count bytes (inclusive).
1356  */
1357 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1358                                        size_t count)
1359 {
1360         int ret = 0;
1361         unsigned int extent_flags;
1362         u32 cpos, clusters, extent_len, phys_cpos;
1363         struct super_block *sb = inode->i_sb;
1364
1365         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1366         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1367
1368         while (clusters) {
1369                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1370                                          &extent_flags);
1371                 if (ret < 0) {
1372                         mlog_errno(ret);
1373                         goto out;
1374                 }
1375
1376                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1377                         ret = 1;
1378                         break;
1379                 }
1380
1381                 if (extent_len > clusters)
1382                         extent_len = clusters;
1383
1384                 clusters -= extent_len;
1385                 cpos += extent_len;
1386         }
1387 out:
1388         return ret;
1389 }
1390
1391 static int ocfs2_write_remove_suid(struct inode *inode)
1392 {
1393         int ret;
1394         struct buffer_head *bh = NULL;
1395
1396         ret = ocfs2_read_inode_block(inode, &bh);
1397         if (ret < 0) {
1398                 mlog_errno(ret);
1399                 goto out;
1400         }
1401
1402         ret =  __ocfs2_write_remove_suid(inode, bh);
1403 out:
1404         brelse(bh);
1405         return ret;
1406 }
1407
1408 /*
1409  * Allocate enough extents to cover the region starting at byte offset
1410  * start for len bytes. Existing extents are skipped, any extents
1411  * added are marked as "unwritten".
1412  */
1413 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1414                                             u64 start, u64 len)
1415 {
1416         int ret;
1417         u32 cpos, phys_cpos, clusters, alloc_size;
1418         u64 end = start + len;
1419         struct buffer_head *di_bh = NULL;
1420
1421         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1422                 ret = ocfs2_read_inode_block(inode, &di_bh);
1423                 if (ret) {
1424                         mlog_errno(ret);
1425                         goto out;
1426                 }
1427
1428                 /*
1429                  * Nothing to do if the requested reservation range
1430                  * fits within the inode.
1431                  */
1432                 if (ocfs2_size_fits_inline_data(di_bh, end))
1433                         goto out;
1434
1435                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1436                 if (ret) {
1437                         mlog_errno(ret);
1438                         goto out;
1439                 }
1440         }
1441
1442         /*
1443          * We consider both start and len to be inclusive.
1444          */
1445         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1446         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1447         clusters -= cpos;
1448
1449         while (clusters) {
1450                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1451                                          &alloc_size, NULL);
1452                 if (ret) {
1453                         mlog_errno(ret);
1454                         goto out;
1455                 }
1456
1457                 /*
1458                  * Hole or existing extent len can be arbitrary, so
1459                  * cap it to our own allocation request.
1460                  */
1461                 if (alloc_size > clusters)
1462                         alloc_size = clusters;
1463
1464                 if (phys_cpos) {
1465                         /*
1466                          * We already have an allocation at this
1467                          * region so we can safely skip it.
1468                          */
1469                         goto next;
1470                 }
1471
1472                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1473                 if (ret) {
1474                         if (ret != -ENOSPC)
1475                                 mlog_errno(ret);
1476                         goto out;
1477                 }
1478
1479 next:
1480                 cpos += alloc_size;
1481                 clusters -= alloc_size;
1482         }
1483
1484         ret = 0;
1485 out:
1486
1487         brelse(di_bh);
1488         return ret;
1489 }
1490
1491 /*
1492  * Truncate a byte range, avoiding pages within partial clusters. This
1493  * preserves those pages for the zeroing code to write to.
1494  */
1495 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1496                                          u64 byte_len)
1497 {
1498         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1499         loff_t start, end;
1500         struct address_space *mapping = inode->i_mapping;
1501
1502         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1503         end = byte_start + byte_len;
1504         end = end & ~(osb->s_clustersize - 1);
1505
1506         if (start < end) {
1507                 unmap_mapping_range(mapping, start, end - start, 0);
1508                 truncate_inode_pages_range(mapping, start, end - 1);
1509         }
1510 }
1511
1512 static int ocfs2_zero_partial_clusters(struct inode *inode,
1513                                        u64 start, u64 len)
1514 {
1515         int ret = 0;
1516         u64 tmpend, end = start + len;
1517         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1518         unsigned int csize = osb->s_clustersize;
1519         handle_t *handle;
1520
1521         /*
1522          * The "start" and "end" values are NOT necessarily part of
1523          * the range whose allocation is being deleted. Rather, this
1524          * is what the user passed in with the request. We must zero
1525          * partial clusters here. There's no need to worry about
1526          * physical allocation - the zeroing code knows to skip holes.
1527          */
1528         mlog(0, "byte start: %llu, end: %llu\n",
1529              (unsigned long long)start, (unsigned long long)end);
1530
1531         /*
1532          * If both edges are on a cluster boundary then there's no
1533          * zeroing required as the region is part of the allocation to
1534          * be truncated.
1535          */
1536         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1537                 goto out;
1538
1539         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1540         if (IS_ERR(handle)) {
1541                 ret = PTR_ERR(handle);
1542                 mlog_errno(ret);
1543                 goto out;
1544         }
1545
1546         /*
1547          * We want to get the byte offset of the end of the 1st cluster.
1548          */
1549         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1550         if (tmpend > end)
1551                 tmpend = end;
1552
1553         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1554              (unsigned long long)start, (unsigned long long)tmpend);
1555
1556         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1557         if (ret)
1558                 mlog_errno(ret);
1559
1560         if (tmpend < end) {
1561                 /*
1562                  * This may make start and end equal, but the zeroing
1563                  * code will skip any work in that case so there's no
1564                  * need to catch it up here.
1565                  */
1566                 start = end & ~(osb->s_clustersize - 1);
1567
1568                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1569                      (unsigned long long)start, (unsigned long long)end);
1570
1571                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1572                 if (ret)
1573                         mlog_errno(ret);
1574         }
1575
1576         ocfs2_commit_trans(osb, handle);
1577 out:
1578         return ret;
1579 }
1580
1581 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1582 {
1583         int i;
1584         struct ocfs2_extent_rec *rec = NULL;
1585
1586         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1587
1588                 rec = &el->l_recs[i];
1589
1590                 if (le32_to_cpu(rec->e_cpos) < pos)
1591                         break;
1592         }
1593
1594         return i;
1595 }
1596
1597 /*
1598  * Helper to calculate the punching pos and length in one run, we handle the
1599  * following three cases in order:
1600  *
1601  * - remove the entire record
1602  * - remove a partial record
1603  * - no record needs to be removed (hole-punching completed)
1604 */
1605 static void ocfs2_calc_trunc_pos(struct inode *inode,
1606                                  struct ocfs2_extent_list *el,
1607                                  struct ocfs2_extent_rec *rec,
1608                                  u32 trunc_start, u32 *trunc_cpos,
1609                                  u32 *trunc_len, u32 *trunc_end,
1610                                  u64 *blkno, int *done)
1611 {
1612         int ret = 0;
1613         u32 coff, range;
1614
1615         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1616
1617         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1618                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1619                 /*
1620                  * Skip holes if any.
1621                  */
1622                 if (range < *trunc_end)
1623                         *trunc_end = range;
1624                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1625                 *blkno = le64_to_cpu(rec->e_blkno);
1626                 *trunc_end = le32_to_cpu(rec->e_cpos);
1627         } else if (range > trunc_start) {
1628                 *trunc_cpos = trunc_start;
1629                 *trunc_len = *trunc_end - trunc_start;
1630                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1631                 *blkno = le64_to_cpu(rec->e_blkno) +
1632                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1633                 *trunc_end = trunc_start;
1634         } else {
1635                 /*
1636                  * It may have two following possibilities:
1637                  *
1638                  * - last record has been removed
1639                  * - trunc_start was within a hole
1640                  *
1641                  * both two cases mean the completion of hole punching.
1642                  */
1643                 ret = 1;
1644         }
1645
1646         *done = ret;
1647 }
1648
1649 static int ocfs2_remove_inode_range(struct inode *inode,
1650                                     struct buffer_head *di_bh, u64 byte_start,
1651                                     u64 byte_len)
1652 {
1653         int ret = 0, flags = 0, done = 0, i;
1654         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1655         u32 cluster_in_el;
1656         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1657         struct ocfs2_cached_dealloc_ctxt dealloc;
1658         struct address_space *mapping = inode->i_mapping;
1659         struct ocfs2_extent_tree et;
1660         struct ocfs2_path *path = NULL;
1661         struct ocfs2_extent_list *el = NULL;
1662         struct ocfs2_extent_rec *rec = NULL;
1663         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1664         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1665
1666         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1667         ocfs2_init_dealloc_ctxt(&dealloc);
1668
1669         if (byte_len == 0)
1670                 return 0;
1671
1672         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1673                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1674                                             byte_start + byte_len, 0);
1675                 if (ret) {
1676                         mlog_errno(ret);
1677                         goto out;
1678                 }
1679                 /*
1680                  * There's no need to get fancy with the page cache
1681                  * truncate of an inline-data inode. We're talking
1682                  * about less than a page here, which will be cached
1683                  * in the dinode buffer anyway.
1684                  */
1685                 unmap_mapping_range(mapping, 0, 0, 0);
1686                 truncate_inode_pages(mapping, 0);
1687                 goto out;
1688         }
1689
1690         /*
1691          * For reflinks, we may need to CoW 2 clusters which might be
1692          * partially zero'd later, if hole's start and end offset were
1693          * within one cluster(means is not exactly aligned to clustersize).
1694          */
1695
1696         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1697
1698                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1699                 if (ret) {
1700                         mlog_errno(ret);
1701                         goto out;
1702                 }
1703
1704                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1705                 if (ret) {
1706                         mlog_errno(ret);
1707                         goto out;
1708                 }
1709         }
1710
1711         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1712         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1713         cluster_in_el = trunc_end;
1714
1715         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1716              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1717              (unsigned long long)byte_start,
1718              (unsigned long long)byte_len, trunc_start, trunc_end);
1719
1720         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1721         if (ret) {
1722                 mlog_errno(ret);
1723                 goto out;
1724         }
1725
1726         path = ocfs2_new_path_from_et(&et);
1727         if (!path) {
1728                 ret = -ENOMEM;
1729                 mlog_errno(ret);
1730                 goto out;
1731         }
1732
1733         while (trunc_end > trunc_start) {
1734
1735                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1736                                       cluster_in_el);
1737                 if (ret) {
1738                         mlog_errno(ret);
1739                         goto out;
1740                 }
1741
1742                 el = path_leaf_el(path);
1743
1744                 i = ocfs2_find_rec(el, trunc_end);
1745                 /*
1746                  * Need to go to previous extent block.
1747                  */
1748                 if (i < 0) {
1749                         if (path->p_tree_depth == 0)
1750                                 break;
1751
1752                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1753                                                             path,
1754                                                             &cluster_in_el);
1755                         if (ret) {
1756                                 mlog_errno(ret);
1757                                 goto out;
1758                         }
1759
1760                         /*
1761                          * We've reached the leftmost extent block,
1762                          * it's safe to leave.
1763                          */
1764                         if (cluster_in_el == 0)
1765                                 break;
1766
1767                         /*
1768                          * The 'pos' searched for previous extent block is
1769                          * always one cluster less than actual trunc_end.
1770                          */
1771                         trunc_end = cluster_in_el + 1;
1772
1773                         ocfs2_reinit_path(path, 1);
1774
1775                         continue;
1776
1777                 } else
1778                         rec = &el->l_recs[i];
1779
1780                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1781                                      &trunc_len, &trunc_end, &blkno, &done);
1782                 if (done)
1783                         break;
1784
1785                 flags = rec->e_flags;
1786                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1787
1788                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1789                                                phys_cpos, trunc_len, flags,
1790                                                &dealloc, refcount_loc);
1791                 if (ret < 0) {
1792                         mlog_errno(ret);
1793                         goto out;
1794                 }
1795
1796                 cluster_in_el = trunc_end;
1797
1798                 ocfs2_reinit_path(path, 1);
1799         }
1800
1801         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1802
1803 out:
1804         ocfs2_schedule_truncate_log_flush(osb, 1);
1805         ocfs2_run_deallocs(osb, &dealloc);
1806
1807         return ret;
1808 }
1809
1810 /*
1811  * Parts of this function taken from xfs_change_file_space()
1812  */
1813 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1814                                      loff_t f_pos, unsigned int cmd,
1815                                      struct ocfs2_space_resv *sr,
1816                                      int change_size)
1817 {
1818         int ret;
1819         s64 llen;
1820         loff_t size;
1821         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1822         struct buffer_head *di_bh = NULL;
1823         handle_t *handle;
1824         unsigned long long max_off = inode->i_sb->s_maxbytes;
1825
1826         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1827                 return -EROFS;
1828
1829         mutex_lock(&inode->i_mutex);
1830
1831         /*
1832          * This prevents concurrent writes on other nodes
1833          */
1834         ret = ocfs2_rw_lock(inode, 1);
1835         if (ret) {
1836                 mlog_errno(ret);
1837                 goto out;
1838         }
1839
1840         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1841         if (ret) {
1842                 mlog_errno(ret);
1843                 goto out_rw_unlock;
1844         }
1845
1846         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1847                 ret = -EPERM;
1848                 goto out_inode_unlock;
1849         }
1850
1851         switch (sr->l_whence) {
1852         case 0: /*SEEK_SET*/
1853                 break;
1854         case 1: /*SEEK_CUR*/
1855                 sr->l_start += f_pos;
1856                 break;
1857         case 2: /*SEEK_END*/
1858                 sr->l_start += i_size_read(inode);
1859                 break;
1860         default:
1861                 ret = -EINVAL;
1862                 goto out_inode_unlock;
1863         }
1864         sr->l_whence = 0;
1865
1866         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1867
1868         if (sr->l_start < 0
1869             || sr->l_start > max_off
1870             || (sr->l_start + llen) < 0
1871             || (sr->l_start + llen) > max_off) {
1872                 ret = -EINVAL;
1873                 goto out_inode_unlock;
1874         }
1875         size = sr->l_start + sr->l_len;
1876
1877         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1878                 if (sr->l_len <= 0) {
1879                         ret = -EINVAL;
1880                         goto out_inode_unlock;
1881                 }
1882         }
1883
1884         if (file && should_remove_suid(file->f_path.dentry)) {
1885                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1886                 if (ret) {
1887                         mlog_errno(ret);
1888                         goto out_inode_unlock;
1889                 }
1890         }
1891
1892         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1893         switch (cmd) {
1894         case OCFS2_IOC_RESVSP:
1895         case OCFS2_IOC_RESVSP64:
1896                 /*
1897                  * This takes unsigned offsets, but the signed ones we
1898                  * pass have been checked against overflow above.
1899                  */
1900                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1901                                                        sr->l_len);
1902                 break;
1903         case OCFS2_IOC_UNRESVSP:
1904         case OCFS2_IOC_UNRESVSP64:
1905                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1906                                                sr->l_len);
1907                 break;
1908         default:
1909                 ret = -EINVAL;
1910         }
1911         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1912         if (ret) {
1913                 mlog_errno(ret);
1914                 goto out_inode_unlock;
1915         }
1916
1917         /*
1918          * We update c/mtime for these changes
1919          */
1920         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1921         if (IS_ERR(handle)) {
1922                 ret = PTR_ERR(handle);
1923                 mlog_errno(ret);
1924                 goto out_inode_unlock;
1925         }
1926
1927         if (change_size && i_size_read(inode) < size)
1928                 i_size_write(inode, size);
1929
1930         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1931         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1932         if (ret < 0)
1933                 mlog_errno(ret);
1934
1935         ocfs2_commit_trans(osb, handle);
1936
1937 out_inode_unlock:
1938         brelse(di_bh);
1939         ocfs2_inode_unlock(inode, 1);
1940 out_rw_unlock:
1941         ocfs2_rw_unlock(inode, 1);
1942
1943 out:
1944         mutex_unlock(&inode->i_mutex);
1945         return ret;
1946 }
1947
1948 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1949                             struct ocfs2_space_resv *sr)
1950 {
1951         struct inode *inode = file->f_path.dentry->d_inode;
1952         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1953
1954         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1955             !ocfs2_writes_unwritten_extents(osb))
1956                 return -ENOTTY;
1957         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1958                  !ocfs2_sparse_alloc(osb))
1959                 return -ENOTTY;
1960
1961         if (!S_ISREG(inode->i_mode))
1962                 return -EINVAL;
1963
1964         if (!(file->f_mode & FMODE_WRITE))
1965                 return -EBADF;
1966
1967         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1968 }
1969
1970 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
1971                             loff_t len)
1972 {
1973         struct inode *inode = file->f_path.dentry->d_inode;
1974         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1975         struct ocfs2_space_resv sr;
1976         int change_size = 1;
1977         int cmd = OCFS2_IOC_RESVSP64;
1978
1979         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1980                 return -EOPNOTSUPP;
1981         if (!ocfs2_writes_unwritten_extents(osb))
1982                 return -EOPNOTSUPP;
1983
1984         if (mode & FALLOC_FL_KEEP_SIZE)
1985                 change_size = 0;
1986
1987         if (mode & FALLOC_FL_PUNCH_HOLE)
1988                 cmd = OCFS2_IOC_UNRESVSP64;
1989
1990         sr.l_whence = 0;
1991         sr.l_start = (s64)offset;
1992         sr.l_len = (s64)len;
1993
1994         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
1995                                          change_size);
1996 }
1997
1998 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1999                                    size_t count)
2000 {
2001         int ret = 0;
2002         unsigned int extent_flags;
2003         u32 cpos, clusters, extent_len, phys_cpos;
2004         struct super_block *sb = inode->i_sb;
2005
2006         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2007             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
2008             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2009                 return 0;
2010
2011         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2012         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2013
2014         while (clusters) {
2015                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2016                                          &extent_flags);
2017                 if (ret < 0) {
2018                         mlog_errno(ret);
2019                         goto out;
2020                 }
2021
2022                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2023                         ret = 1;
2024                         break;
2025                 }
2026
2027                 if (extent_len > clusters)
2028                         extent_len = clusters;
2029
2030                 clusters -= extent_len;
2031                 cpos += extent_len;
2032         }
2033 out:
2034         return ret;
2035 }
2036
2037 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2038                                             struct file *file,
2039                                             loff_t pos, size_t count,
2040                                             int *meta_level)
2041 {
2042         int ret;
2043         struct buffer_head *di_bh = NULL;
2044         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2045         u32 clusters =
2046                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2047
2048         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2049         if (ret) {
2050                 mlog_errno(ret);
2051                 goto out;
2052         }
2053
2054         *meta_level = 1;
2055
2056         ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
2057         if (ret)
2058                 mlog_errno(ret);
2059 out:
2060         brelse(di_bh);
2061         return ret;
2062 }
2063
2064 static int ocfs2_prepare_inode_for_write(struct file *file,
2065                                          loff_t *ppos,
2066                                          size_t count,
2067                                          int appending,
2068                                          int *direct_io,
2069                                          int *has_refcount)
2070 {
2071         int ret = 0, meta_level = 0;
2072         struct dentry *dentry = file->f_path.dentry;
2073         struct inode *inode = dentry->d_inode;
2074         loff_t saved_pos, end;
2075
2076         /*
2077          * We start with a read level meta lock and only jump to an ex
2078          * if we need to make modifications here.
2079          */
2080         for(;;) {
2081                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2082                 if (ret < 0) {
2083                         meta_level = -1;
2084                         mlog_errno(ret);
2085                         goto out;
2086                 }
2087
2088                 /* Clear suid / sgid if necessary. We do this here
2089                  * instead of later in the write path because
2090                  * remove_suid() calls ->setattr without any hint that
2091                  * we may have already done our cluster locking. Since
2092                  * ocfs2_setattr() *must* take cluster locks to
2093                  * proceeed, this will lead us to recursively lock the
2094                  * inode. There's also the dinode i_size state which
2095                  * can be lost via setattr during extending writes (we
2096                  * set inode->i_size at the end of a write. */
2097                 if (should_remove_suid(dentry)) {
2098                         if (meta_level == 0) {
2099                                 ocfs2_inode_unlock(inode, meta_level);
2100                                 meta_level = 1;
2101                                 continue;
2102                         }
2103
2104                         ret = ocfs2_write_remove_suid(inode);
2105                         if (ret < 0) {
2106                                 mlog_errno(ret);
2107                                 goto out_unlock;
2108                         }
2109                 }
2110
2111                 /* work on a copy of ppos until we're sure that we won't have
2112                  * to recalculate it due to relocking. */
2113                 if (appending) {
2114                         saved_pos = i_size_read(inode);
2115                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
2116                 } else {
2117                         saved_pos = *ppos;
2118                 }
2119
2120                 end = saved_pos + count;
2121
2122                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
2123                 if (ret == 1) {
2124                         ocfs2_inode_unlock(inode, meta_level);
2125                         meta_level = -1;
2126
2127                         ret = ocfs2_prepare_inode_for_refcount(inode,
2128                                                                file,
2129                                                                saved_pos,
2130                                                                count,
2131                                                                &meta_level);
2132                         if (has_refcount)
2133                                 *has_refcount = 1;
2134                         if (direct_io)
2135                                 *direct_io = 0;
2136                 }
2137
2138                 if (ret < 0) {
2139                         mlog_errno(ret);
2140                         goto out_unlock;
2141                 }
2142
2143                 /*
2144                  * Skip the O_DIRECT checks if we don't need
2145                  * them.
2146                  */
2147                 if (!direct_io || !(*direct_io))
2148                         break;
2149
2150                 /*
2151                  * There's no sane way to do direct writes to an inode
2152                  * with inline data.
2153                  */
2154                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2155                         *direct_io = 0;
2156                         break;
2157                 }
2158
2159                 /*
2160                  * Allowing concurrent direct writes means
2161                  * i_size changes wouldn't be synchronized, so
2162                  * one node could wind up truncating another
2163                  * nodes writes.
2164                  */
2165                 if (end > i_size_read(inode)) {
2166                         *direct_io = 0;
2167                         break;
2168                 }
2169
2170                 /*
2171                  * We don't fill holes during direct io, so
2172                  * check for them here. If any are found, the
2173                  * caller will have to retake some cluster
2174                  * locks and initiate the io as buffered.
2175                  */
2176                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2177                 if (ret == 1) {
2178                         *direct_io = 0;
2179                         ret = 0;
2180                 } else if (ret < 0)
2181                         mlog_errno(ret);
2182                 break;
2183         }
2184
2185         if (appending)
2186                 *ppos = saved_pos;
2187
2188 out_unlock:
2189         if (meta_level >= 0)
2190                 ocfs2_inode_unlock(inode, meta_level);
2191
2192 out:
2193         return ret;
2194 }
2195
2196 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2197                                     const struct iovec *iov,
2198                                     unsigned long nr_segs,
2199                                     loff_t pos)
2200 {
2201         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2202         int can_do_direct, has_refcount = 0;
2203         ssize_t written = 0;
2204         size_t ocount;          /* original count */
2205         size_t count;           /* after file limit checks */
2206         loff_t old_size, *ppos = &iocb->ki_pos;
2207         u32 old_clusters;
2208         struct file *file = iocb->ki_filp;
2209         struct inode *inode = file->f_path.dentry->d_inode;
2210         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2211         int full_coherency = !(osb->s_mount_opt &
2212                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2213
2214         mlog(0, "(0x%p, %u, '%.*s')\n", file,
2215              (unsigned int)nr_segs,
2216              file->f_path.dentry->d_name.len,
2217              file->f_path.dentry->d_name.name);
2218
2219         if (iocb->ki_left == 0)
2220                 return 0;
2221
2222         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2223
2224         appending = file->f_flags & O_APPEND ? 1 : 0;
2225         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2226
2227         mutex_lock(&inode->i_mutex);
2228
2229         ocfs2_iocb_clear_sem_locked(iocb);
2230
2231 relock:
2232         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2233         if (direct_io) {
2234                 down_read(&inode->i_alloc_sem);
2235                 have_alloc_sem = 1;
2236                 /* communicate with ocfs2_dio_end_io */
2237                 ocfs2_iocb_set_sem_locked(iocb);
2238         }
2239
2240         /*
2241          * Concurrent O_DIRECT writes are allowed with
2242          * mount_option "coherency=buffered".
2243          */
2244         rw_level = (!direct_io || full_coherency);
2245
2246         ret = ocfs2_rw_lock(inode, rw_level);
2247         if (ret < 0) {
2248                 mlog_errno(ret);
2249                 goto out_sems;
2250         }
2251
2252         /*
2253          * O_DIRECT writes with "coherency=full" need to take EX cluster
2254          * inode_lock to guarantee coherency.
2255          */
2256         if (direct_io && full_coherency) {
2257                 /*
2258                  * We need to take and drop the inode lock to force
2259                  * other nodes to drop their caches.  Buffered I/O
2260                  * already does this in write_begin().
2261                  */
2262                 ret = ocfs2_inode_lock(inode, NULL, 1);
2263                 if (ret < 0) {
2264                         mlog_errno(ret);
2265                         goto out_sems;
2266                 }
2267
2268                 ocfs2_inode_unlock(inode, 1);
2269         }
2270
2271         can_do_direct = direct_io;
2272         ret = ocfs2_prepare_inode_for_write(file, ppos,
2273                                             iocb->ki_left, appending,
2274                                             &can_do_direct, &has_refcount);
2275         if (ret < 0) {
2276                 mlog_errno(ret);
2277                 goto out;
2278         }
2279
2280         /*
2281          * We can't complete the direct I/O as requested, fall back to
2282          * buffered I/O.
2283          */
2284         if (direct_io && !can_do_direct) {
2285                 ocfs2_rw_unlock(inode, rw_level);
2286                 up_read(&inode->i_alloc_sem);
2287
2288                 have_alloc_sem = 0;
2289                 rw_level = -1;
2290
2291                 direct_io = 0;
2292                 goto relock;
2293         }
2294
2295         /*
2296          * To later detect whether a journal commit for sync writes is
2297          * necessary, we sample i_size, and cluster count here.
2298          */
2299         old_size = i_size_read(inode);
2300         old_clusters = OCFS2_I(inode)->ip_clusters;
2301
2302         /* communicate with ocfs2_dio_end_io */
2303         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2304
2305         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2306                                      VERIFY_READ);
2307         if (ret)
2308                 goto out_dio;
2309
2310         count = ocount;
2311         ret = generic_write_checks(file, ppos, &count,
2312                                    S_ISBLK(inode->i_mode));
2313         if (ret)
2314                 goto out_dio;
2315
2316         if (direct_io) {
2317                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2318                                                     ppos, count, ocount);
2319                 if (written < 0) {
2320                         ret = written;
2321                         goto out_dio;
2322                 }
2323         } else {
2324                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2325                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2326                                                       ppos, count, 0);
2327                 current->backing_dev_info = NULL;
2328         }
2329
2330 out_dio:
2331         /* buffered aio wouldn't have proper lock coverage today */
2332         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2333
2334         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2335             ((file->f_flags & O_DIRECT) && !direct_io)) {
2336                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2337                                                pos + count - 1);
2338                 if (ret < 0)
2339                         written = ret;
2340
2341                 if (!ret && ((old_size != i_size_read(inode)) ||
2342                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2343                              has_refcount)) {
2344                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2345                         if (ret < 0)
2346                                 written = ret;
2347                 }
2348
2349                 if (!ret)
2350                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2351                                                       pos + count - 1);
2352         }
2353
2354         /*
2355          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2356          * function pointer which is called when o_direct io completes so that
2357          * it can unlock our rw lock.  (it's the clustered equivalent of
2358          * i_alloc_sem; protects truncate from racing with pending ios).
2359          * Unfortunately there are error cases which call end_io and others
2360          * that don't.  so we don't have to unlock the rw_lock if either an
2361          * async dio is going to do it in the future or an end_io after an
2362          * error has already done it.
2363          */
2364         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2365                 rw_level = -1;
2366                 have_alloc_sem = 0;
2367         }
2368
2369 out:
2370         if (rw_level != -1)
2371                 ocfs2_rw_unlock(inode, rw_level);
2372
2373 out_sems:
2374         if (have_alloc_sem) {
2375                 up_read(&inode->i_alloc_sem);
2376                 ocfs2_iocb_clear_sem_locked(iocb);
2377         }
2378
2379         mutex_unlock(&inode->i_mutex);
2380
2381         if (written)
2382                 ret = written;
2383         return ret;
2384 }
2385
2386 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2387                                 struct file *out,
2388                                 struct splice_desc *sd)
2389 {
2390         int ret;
2391
2392         ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
2393                                             sd->total_len, 0, NULL, NULL);
2394         if (ret < 0) {
2395                 mlog_errno(ret);
2396                 return ret;
2397         }
2398
2399         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2400 }
2401
2402 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2403                                        struct file *out,
2404                                        loff_t *ppos,
2405                                        size_t len,
2406                                        unsigned int flags)
2407 {
2408         int ret;
2409         struct address_space *mapping = out->f_mapping;
2410         struct inode *inode = mapping->host;
2411         struct splice_desc sd = {
2412                 .total_len = len,
2413                 .flags = flags,
2414                 .pos = *ppos,
2415                 .u.file = out,
2416         };
2417
2418         mlog(0, "(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2419              (unsigned int)len,
2420              out->f_path.dentry->d_name.len,
2421              out->f_path.dentry->d_name.name);
2422
2423         if (pipe->inode)
2424                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2425
2426         splice_from_pipe_begin(&sd);
2427         do {
2428                 ret = splice_from_pipe_next(pipe, &sd);
2429                 if (ret <= 0)
2430                         break;
2431
2432                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2433                 ret = ocfs2_rw_lock(inode, 1);
2434                 if (ret < 0)
2435                         mlog_errno(ret);
2436                 else {
2437                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2438                         ocfs2_rw_unlock(inode, 1);
2439                 }
2440                 mutex_unlock(&inode->i_mutex);
2441         } while (ret > 0);
2442         splice_from_pipe_end(pipe, &sd);
2443
2444         if (pipe->inode)
2445                 mutex_unlock(&pipe->inode->i_mutex);
2446
2447         if (sd.num_spliced)
2448                 ret = sd.num_spliced;
2449
2450         if (ret > 0) {
2451                 unsigned long nr_pages;
2452                 int err;
2453
2454                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2455
2456                 err = generic_write_sync(out, *ppos, ret);
2457                 if (err)
2458                         ret = err;
2459                 else
2460                         *ppos += ret;
2461
2462                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2463         }
2464
2465         return ret;
2466 }
2467
2468 static ssize_t ocfs2_file_splice_read(struct file *in,
2469                                       loff_t *ppos,
2470                                       struct pipe_inode_info *pipe,
2471                                       size_t len,
2472                                       unsigned int flags)
2473 {
2474         int ret = 0, lock_level = 0;
2475         struct inode *inode = in->f_path.dentry->d_inode;
2476
2477         mlog(0, "(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2478              (unsigned int)len,
2479              in->f_path.dentry->d_name.len,
2480              in->f_path.dentry->d_name.name);
2481
2482         /*
2483          * See the comment in ocfs2_file_aio_read()
2484          */
2485         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2486         if (ret < 0) {
2487                 mlog_errno(ret);
2488                 goto bail;
2489         }
2490         ocfs2_inode_unlock(inode, lock_level);
2491
2492         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2493
2494 bail:
2495         return ret;
2496 }
2497
2498 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2499                                    const struct iovec *iov,
2500                                    unsigned long nr_segs,
2501                                    loff_t pos)
2502 {
2503         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2504         struct file *filp = iocb->ki_filp;
2505         struct inode *inode = filp->f_path.dentry->d_inode;
2506
2507         mlog(0, "(0x%p, %u, '%.*s')\n", filp,
2508              (unsigned int)nr_segs,
2509              filp->f_path.dentry->d_name.len,
2510              filp->f_path.dentry->d_name.name);
2511
2512         if (!inode) {
2513                 ret = -EINVAL;
2514                 mlog_errno(ret);
2515                 goto bail;
2516         }
2517
2518         ocfs2_iocb_clear_sem_locked(iocb);
2519
2520         /*
2521          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2522          * need locks to protect pending reads from racing with truncate.
2523          */
2524         if (filp->f_flags & O_DIRECT) {
2525                 down_read(&inode->i_alloc_sem);
2526                 have_alloc_sem = 1;
2527                 ocfs2_iocb_set_sem_locked(iocb);
2528
2529                 ret = ocfs2_rw_lock(inode, 0);
2530                 if (ret < 0) {
2531                         mlog_errno(ret);
2532                         goto bail;
2533                 }
2534                 rw_level = 0;
2535                 /* communicate with ocfs2_dio_end_io */
2536                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2537         }
2538
2539         /*
2540          * We're fine letting folks race truncates and extending
2541          * writes with read across the cluster, just like they can
2542          * locally. Hence no rw_lock during read.
2543          *
2544          * Take and drop the meta data lock to update inode fields
2545          * like i_size. This allows the checks down below
2546          * generic_file_aio_read() a chance of actually working.
2547          */
2548         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2549         if (ret < 0) {
2550                 mlog_errno(ret);
2551                 goto bail;
2552         }
2553         ocfs2_inode_unlock(inode, lock_level);
2554
2555         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2556         if (ret == -EINVAL)
2557                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2558
2559         /* buffered aio wouldn't have proper lock coverage today */
2560         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2561
2562         /* see ocfs2_file_aio_write */
2563         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2564                 rw_level = -1;
2565                 have_alloc_sem = 0;
2566         }
2567
2568 bail:
2569         if (have_alloc_sem) {
2570                 up_read(&inode->i_alloc_sem);
2571                 ocfs2_iocb_clear_sem_locked(iocb);
2572         }
2573         if (rw_level != -1)
2574                 ocfs2_rw_unlock(inode, rw_level);
2575
2576         return ret;
2577 }
2578
2579 const struct inode_operations ocfs2_file_iops = {
2580         .setattr        = ocfs2_setattr,
2581         .getattr        = ocfs2_getattr,
2582         .permission     = ocfs2_permission,
2583         .setxattr       = generic_setxattr,
2584         .getxattr       = generic_getxattr,
2585         .listxattr      = ocfs2_listxattr,
2586         .removexattr    = generic_removexattr,
2587         .fiemap         = ocfs2_fiemap,
2588 };
2589
2590 const struct inode_operations ocfs2_special_file_iops = {
2591         .setattr        = ocfs2_setattr,
2592         .getattr        = ocfs2_getattr,
2593         .permission     = ocfs2_permission,
2594 };
2595
2596 /*
2597  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2598  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2599  */
2600 const struct file_operations ocfs2_fops = {
2601         .llseek         = generic_file_llseek,
2602         .read           = do_sync_read,
2603         .write          = do_sync_write,
2604         .mmap           = ocfs2_mmap,
2605         .fsync          = ocfs2_sync_file,
2606         .release        = ocfs2_file_release,
2607         .open           = ocfs2_file_open,
2608         .aio_read       = ocfs2_file_aio_read,
2609         .aio_write      = ocfs2_file_aio_write,
2610         .unlocked_ioctl = ocfs2_ioctl,
2611 #ifdef CONFIG_COMPAT
2612         .compat_ioctl   = ocfs2_compat_ioctl,
2613 #endif
2614         .lock           = ocfs2_lock,
2615         .flock          = ocfs2_flock,
2616         .splice_read    = ocfs2_file_splice_read,
2617         .splice_write   = ocfs2_file_splice_write,
2618         .fallocate      = ocfs2_fallocate,
2619 };
2620
2621 const struct file_operations ocfs2_dops = {
2622         .llseek         = generic_file_llseek,
2623         .read           = generic_read_dir,
2624         .readdir        = ocfs2_readdir,
2625         .fsync          = ocfs2_sync_file,
2626         .release        = ocfs2_dir_release,
2627         .open           = ocfs2_dir_open,
2628         .unlocked_ioctl = ocfs2_ioctl,
2629 #ifdef CONFIG_COMPAT
2630         .compat_ioctl   = ocfs2_compat_ioctl,
2631 #endif
2632         .lock           = ocfs2_lock,
2633         .flock          = ocfs2_flock,
2634 };
2635
2636 /*
2637  * POSIX-lockless variants of our file_operations.
2638  *
2639  * These will be used if the underlying cluster stack does not support
2640  * posix file locking, if the user passes the "localflocks" mount
2641  * option, or if we have a local-only fs.
2642  *
2643  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2644  * so we still want it in the case of no stack support for
2645  * plocks. Internally, it will do the right thing when asked to ignore
2646  * the cluster.
2647  */
2648 const struct file_operations ocfs2_fops_no_plocks = {
2649         .llseek         = generic_file_llseek,
2650         .read           = do_sync_read,
2651         .write          = do_sync_write,
2652         .mmap           = ocfs2_mmap,
2653         .fsync          = ocfs2_sync_file,
2654         .release        = ocfs2_file_release,
2655         .open           = ocfs2_file_open,
2656         .aio_read       = ocfs2_file_aio_read,
2657         .aio_write      = ocfs2_file_aio_write,
2658         .unlocked_ioctl = ocfs2_ioctl,
2659 #ifdef CONFIG_COMPAT
2660         .compat_ioctl   = ocfs2_compat_ioctl,
2661 #endif
2662         .flock          = ocfs2_flock,
2663         .splice_read    = ocfs2_file_splice_read,
2664         .splice_write   = ocfs2_file_splice_write,
2665 };
2666
2667 const struct file_operations ocfs2_dops_no_plocks = {
2668         .llseek         = generic_file_llseek,
2669         .read           = generic_read_dir,
2670         .readdir        = ocfs2_readdir,
2671         .fsync          = ocfs2_sync_file,
2672         .release        = ocfs2_dir_release,
2673         .open           = ocfs2_dir_open,
2674         .unlocked_ioctl = ocfs2_ioctl,
2675 #ifdef CONFIG_COMPAT
2676         .compat_ioctl   = ocfs2_compat_ioctl,
2677 #endif
2678         .flock          = ocfs2_flock,
2679 };