ocfs2: Make ocfs2_journal_dirty() void.
[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
40 #define MLOG_MASK_PREFIX ML_INODE
41 #include <cluster/masklog.h>
42
43 #include "ocfs2.h"
44
45 #include "alloc.h"
46 #include "aops.h"
47 #include "dir.h"
48 #include "dlmglue.h"
49 #include "extent_map.h"
50 #include "file.h"
51 #include "sysfile.h"
52 #include "inode.h"
53 #include "ioctl.h"
54 #include "journal.h"
55 #include "locks.h"
56 #include "mmap.h"
57 #include "suballoc.h"
58 #include "super.h"
59 #include "xattr.h"
60 #include "acl.h"
61 #include "quota.h"
62 #include "refcounttree.h"
63
64 #include "buffer_head_io.h"
65
66 static int ocfs2_sync_inode(struct inode *inode)
67 {
68         filemap_fdatawrite(inode->i_mapping);
69         return sync_mapping_buffers(inode->i_mapping);
70 }
71
72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
73 {
74         struct ocfs2_file_private *fp;
75
76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
77         if (!fp)
78                 return -ENOMEM;
79
80         fp->fp_file = file;
81         mutex_init(&fp->fp_mutex);
82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
83         file->private_data = fp;
84
85         return 0;
86 }
87
88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
89 {
90         struct ocfs2_file_private *fp = file->private_data;
91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
92
93         if (fp) {
94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
95                 ocfs2_lock_res_free(&fp->fp_flock);
96                 kfree(fp);
97                 file->private_data = NULL;
98         }
99 }
100
101 static int ocfs2_file_open(struct inode *inode, struct file *file)
102 {
103         int status;
104         int mode = file->f_flags;
105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
106
107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
109
110         if (file->f_mode & FMODE_WRITE)
111                 dquot_initialize(inode);
112
113         spin_lock(&oi->ip_lock);
114
115         /* Check that the inode hasn't been wiped from disk by another
116          * node. If it hasn't then we're safe as long as we hold the
117          * spin lock until our increment of open count. */
118         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
119                 spin_unlock(&oi->ip_lock);
120
121                 status = -ENOENT;
122                 goto leave;
123         }
124
125         if (mode & O_DIRECT)
126                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
127
128         oi->ip_open_count++;
129         spin_unlock(&oi->ip_lock);
130
131         status = ocfs2_init_file_private(inode, file);
132         if (status) {
133                 /*
134                  * We want to set open count back if we're failing the
135                  * open.
136                  */
137                 spin_lock(&oi->ip_lock);
138                 oi->ip_open_count--;
139                 spin_unlock(&oi->ip_lock);
140         }
141
142 leave:
143         mlog_exit(status);
144         return status;
145 }
146
147 static int ocfs2_file_release(struct inode *inode, struct file *file)
148 {
149         struct ocfs2_inode_info *oi = OCFS2_I(inode);
150
151         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
152                        file->f_path.dentry->d_name.len,
153                        file->f_path.dentry->d_name.name);
154
155         spin_lock(&oi->ip_lock);
156         if (!--oi->ip_open_count)
157                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
158         spin_unlock(&oi->ip_lock);
159
160         ocfs2_free_file_private(inode, file);
161
162         mlog_exit(0);
163
164         return 0;
165 }
166
167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
168 {
169         return ocfs2_init_file_private(inode, file);
170 }
171
172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
173 {
174         ocfs2_free_file_private(inode, file);
175         return 0;
176 }
177
178 static int ocfs2_sync_file(struct file *file,
179                            struct dentry *dentry,
180                            int datasync)
181 {
182         int err = 0;
183         journal_t *journal;
184         struct inode *inode = dentry->d_inode;
185         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
186
187         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
188                    dentry->d_name.len, dentry->d_name.name);
189
190         err = ocfs2_sync_inode(dentry->d_inode);
191         if (err)
192                 goto bail;
193
194         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
195                 goto bail;
196
197         journal = osb->journal->j_journal;
198         err = jbd2_journal_force_commit(journal);
199
200 bail:
201         mlog_exit(err);
202
203         return (err < 0) ? -EIO : 0;
204 }
205
206 int ocfs2_should_update_atime(struct inode *inode,
207                               struct vfsmount *vfsmnt)
208 {
209         struct timespec now;
210         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
211
212         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
213                 return 0;
214
215         if ((inode->i_flags & S_NOATIME) ||
216             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
217                 return 0;
218
219         /*
220          * We can be called with no vfsmnt structure - NFSD will
221          * sometimes do this.
222          *
223          * Note that our action here is different than touch_atime() -
224          * if we can't tell whether this is a noatime mount, then we
225          * don't know whether to trust the value of s_atime_quantum.
226          */
227         if (vfsmnt == NULL)
228                 return 0;
229
230         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
231             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
232                 return 0;
233
234         if (vfsmnt->mnt_flags & MNT_RELATIME) {
235                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
236                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
237                         return 1;
238
239                 return 0;
240         }
241
242         now = CURRENT_TIME;
243         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
244                 return 0;
245         else
246                 return 1;
247 }
248
249 int ocfs2_update_inode_atime(struct inode *inode,
250                              struct buffer_head *bh)
251 {
252         int ret;
253         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
254         handle_t *handle;
255         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
256
257         mlog_entry_void();
258
259         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
260         if (IS_ERR(handle)) {
261                 ret = PTR_ERR(handle);
262                 mlog_errno(ret);
263                 goto out;
264         }
265
266         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
267                                       OCFS2_JOURNAL_ACCESS_WRITE);
268         if (ret) {
269                 mlog_errno(ret);
270                 goto out_commit;
271         }
272
273         /*
274          * Don't use ocfs2_mark_inode_dirty() here as we don't always
275          * have i_mutex to guard against concurrent changes to other
276          * inode fields.
277          */
278         inode->i_atime = CURRENT_TIME;
279         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
280         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
281         ocfs2_journal_dirty(handle, bh);
282
283 out_commit:
284         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
285 out:
286         mlog_exit(ret);
287         return ret;
288 }
289
290 static int ocfs2_set_inode_size(handle_t *handle,
291                                 struct inode *inode,
292                                 struct buffer_head *fe_bh,
293                                 u64 new_i_size)
294 {
295         int status;
296
297         mlog_entry_void();
298         i_size_write(inode, new_i_size);
299         inode->i_blocks = ocfs2_inode_sector_count(inode);
300         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
301
302         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
303         if (status < 0) {
304                 mlog_errno(status);
305                 goto bail;
306         }
307
308 bail:
309         mlog_exit(status);
310         return status;
311 }
312
313 int ocfs2_simple_size_update(struct inode *inode,
314                              struct buffer_head *di_bh,
315                              u64 new_i_size)
316 {
317         int ret;
318         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
319         handle_t *handle = NULL;
320
321         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
322         if (IS_ERR(handle)) {
323                 ret = PTR_ERR(handle);
324                 mlog_errno(ret);
325                 goto out;
326         }
327
328         ret = ocfs2_set_inode_size(handle, inode, di_bh,
329                                    new_i_size);
330         if (ret < 0)
331                 mlog_errno(ret);
332
333         ocfs2_commit_trans(osb, handle);
334 out:
335         return ret;
336 }
337
338 static int ocfs2_cow_file_pos(struct inode *inode,
339                               struct buffer_head *fe_bh,
340                               u64 offset)
341 {
342         int status;
343         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
344         unsigned int num_clusters = 0;
345         unsigned int ext_flags = 0;
346
347         /*
348          * If the new offset is aligned to the range of the cluster, there is
349          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
350          * CoW either.
351          */
352         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
353                 return 0;
354
355         status = ocfs2_get_clusters(inode, cpos, &phys,
356                                     &num_clusters, &ext_flags);
357         if (status) {
358                 mlog_errno(status);
359                 goto out;
360         }
361
362         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
363                 goto out;
364
365         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
366
367 out:
368         return status;
369 }
370
371 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
372                                      struct inode *inode,
373                                      struct buffer_head *fe_bh,
374                                      u64 new_i_size)
375 {
376         int status;
377         handle_t *handle;
378         struct ocfs2_dinode *di;
379         u64 cluster_bytes;
380
381         mlog_entry_void();
382
383         /*
384          * We need to CoW the cluster contains the offset if it is reflinked
385          * since we will call ocfs2_zero_range_for_truncate later which will
386          * write "0" from offset to the end of the cluster.
387          */
388         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
389         if (status) {
390                 mlog_errno(status);
391                 return status;
392         }
393
394         /* TODO: This needs to actually orphan the inode in this
395          * transaction. */
396
397         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
398         if (IS_ERR(handle)) {
399                 status = PTR_ERR(handle);
400                 mlog_errno(status);
401                 goto out;
402         }
403
404         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
405                                          OCFS2_JOURNAL_ACCESS_WRITE);
406         if (status < 0) {
407                 mlog_errno(status);
408                 goto out_commit;
409         }
410
411         /*
412          * Do this before setting i_size.
413          */
414         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
415         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
416                                                cluster_bytes);
417         if (status) {
418                 mlog_errno(status);
419                 goto out_commit;
420         }
421
422         i_size_write(inode, new_i_size);
423         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
424
425         di = (struct ocfs2_dinode *) fe_bh->b_data;
426         di->i_size = cpu_to_le64(new_i_size);
427         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
428         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
429
430         ocfs2_journal_dirty(handle, fe_bh);
431
432 out_commit:
433         ocfs2_commit_trans(osb, handle);
434 out:
435
436         mlog_exit(status);
437         return status;
438 }
439
440 static int ocfs2_truncate_file(struct inode *inode,
441                                struct buffer_head *di_bh,
442                                u64 new_i_size)
443 {
444         int status = 0;
445         struct ocfs2_dinode *fe = NULL;
446         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
447         struct ocfs2_truncate_context *tc = NULL;
448
449         mlog_entry("(inode = %llu, new_i_size = %llu\n",
450                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
451                    (unsigned long long)new_i_size);
452
453         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
454          * already validated it */
455         fe = (struct ocfs2_dinode *) di_bh->b_data;
456
457         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
458                         "Inode %llu, inode i_size = %lld != di "
459                         "i_size = %llu, i_flags = 0x%x\n",
460                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
461                         i_size_read(inode),
462                         (unsigned long long)le64_to_cpu(fe->i_size),
463                         le32_to_cpu(fe->i_flags));
464
465         if (new_i_size > le64_to_cpu(fe->i_size)) {
466                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
467                      (unsigned long long)le64_to_cpu(fe->i_size),
468                      (unsigned long long)new_i_size);
469                 status = -EINVAL;
470                 mlog_errno(status);
471                 goto bail;
472         }
473
474         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
475              (unsigned long long)le64_to_cpu(fe->i_blkno),
476              (unsigned long long)le64_to_cpu(fe->i_size),
477              (unsigned long long)new_i_size);
478
479         /* lets handle the simple truncate cases before doing any more
480          * cluster locking. */
481         if (new_i_size == le64_to_cpu(fe->i_size))
482                 goto bail;
483
484         down_write(&OCFS2_I(inode)->ip_alloc_sem);
485
486         /*
487          * The inode lock forced other nodes to sync and drop their
488          * pages, which (correctly) happens even if we have a truncate
489          * without allocation change - ocfs2 cluster sizes can be much
490          * greater than page size, so we have to truncate them
491          * anyway.
492          */
493         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
494         truncate_inode_pages(inode->i_mapping, new_i_size);
495
496         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
497                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
498                                                i_size_read(inode), 1);
499                 if (status)
500                         mlog_errno(status);
501
502                 goto bail_unlock_sem;
503         }
504
505         /* alright, we're going to need to do a full blown alloc size
506          * change. Orphan the inode so that recovery can complete the
507          * truncate if necessary. This does the task of marking
508          * i_size. */
509         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
510         if (status < 0) {
511                 mlog_errno(status);
512                 goto bail_unlock_sem;
513         }
514
515         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
516         if (status < 0) {
517                 mlog_errno(status);
518                 goto bail_unlock_sem;
519         }
520
521         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
522         if (status < 0) {
523                 mlog_errno(status);
524                 goto bail_unlock_sem;
525         }
526
527         /* TODO: orphan dir cleanup here. */
528 bail_unlock_sem:
529         up_write(&OCFS2_I(inode)->ip_alloc_sem);
530
531 bail:
532         if (!status && OCFS2_I(inode)->ip_clusters == 0)
533                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
534
535         mlog_exit(status);
536         return status;
537 }
538
539 /*
540  * extend file allocation only here.
541  * we'll update all the disk stuff, and oip->alloc_size
542  *
543  * expect stuff to be locked, a transaction started and enough data /
544  * metadata reservations in the contexts.
545  *
546  * Will return -EAGAIN, and a reason if a restart is needed.
547  * If passed in, *reason will always be set, even in error.
548  */
549 int ocfs2_add_inode_data(struct ocfs2_super *osb,
550                          struct inode *inode,
551                          u32 *logical_offset,
552                          u32 clusters_to_add,
553                          int mark_unwritten,
554                          struct buffer_head *fe_bh,
555                          handle_t *handle,
556                          struct ocfs2_alloc_context *data_ac,
557                          struct ocfs2_alloc_context *meta_ac,
558                          enum ocfs2_alloc_restarted *reason_ret)
559 {
560         int ret;
561         struct ocfs2_extent_tree et;
562
563         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
564         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
565                                           clusters_to_add, mark_unwritten,
566                                           data_ac, meta_ac, reason_ret);
567
568         return ret;
569 }
570
571 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
572                                      u32 clusters_to_add, int mark_unwritten)
573 {
574         int status = 0;
575         int restart_func = 0;
576         int credits;
577         u32 prev_clusters;
578         struct buffer_head *bh = NULL;
579         struct ocfs2_dinode *fe = NULL;
580         handle_t *handle = NULL;
581         struct ocfs2_alloc_context *data_ac = NULL;
582         struct ocfs2_alloc_context *meta_ac = NULL;
583         enum ocfs2_alloc_restarted why;
584         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
585         struct ocfs2_extent_tree et;
586         int did_quota = 0;
587
588         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
589
590         /*
591          * This function only exists for file systems which don't
592          * support holes.
593          */
594         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
595
596         status = ocfs2_read_inode_block(inode, &bh);
597         if (status < 0) {
598                 mlog_errno(status);
599                 goto leave;
600         }
601         fe = (struct ocfs2_dinode *) bh->b_data;
602
603 restart_all:
604         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
605
606         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
607              "clusters_to_add = %u\n",
608              (unsigned long long)OCFS2_I(inode)->ip_blkno,
609              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
610              clusters_to_add);
611         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
612         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
613                                        &data_ac, &meta_ac);
614         if (status) {
615                 mlog_errno(status);
616                 goto leave;
617         }
618
619         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
620                                             clusters_to_add);
621         handle = ocfs2_start_trans(osb, credits);
622         if (IS_ERR(handle)) {
623                 status = PTR_ERR(handle);
624                 handle = NULL;
625                 mlog_errno(status);
626                 goto leave;
627         }
628
629 restarted_transaction:
630         status = dquot_alloc_space_nodirty(inode,
631                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
632         if (status)
633                 goto leave;
634         did_quota = 1;
635
636         /* reserve a write to the file entry early on - that we if we
637          * run out of credits in the allocation path, we can still
638          * update i_size. */
639         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
640                                          OCFS2_JOURNAL_ACCESS_WRITE);
641         if (status < 0) {
642                 mlog_errno(status);
643                 goto leave;
644         }
645
646         prev_clusters = OCFS2_I(inode)->ip_clusters;
647
648         status = ocfs2_add_inode_data(osb,
649                                       inode,
650                                       &logical_start,
651                                       clusters_to_add,
652                                       mark_unwritten,
653                                       bh,
654                                       handle,
655                                       data_ac,
656                                       meta_ac,
657                                       &why);
658         if ((status < 0) && (status != -EAGAIN)) {
659                 if (status != -ENOSPC)
660                         mlog_errno(status);
661                 goto leave;
662         }
663
664         ocfs2_journal_dirty(handle, bh);
665
666         spin_lock(&OCFS2_I(inode)->ip_lock);
667         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
668         spin_unlock(&OCFS2_I(inode)->ip_lock);
669         /* Release unused quota reservation */
670         dquot_free_space(inode,
671                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
672         did_quota = 0;
673
674         if (why != RESTART_NONE && clusters_to_add) {
675                 if (why == RESTART_META) {
676                         mlog(0, "restarting function.\n");
677                         restart_func = 1;
678                 } else {
679                         BUG_ON(why != RESTART_TRANS);
680
681                         mlog(0, "restarting transaction.\n");
682                         /* TODO: This can be more intelligent. */
683                         credits = ocfs2_calc_extend_credits(osb->sb,
684                                                             &fe->id2.i_list,
685                                                             clusters_to_add);
686                         status = ocfs2_extend_trans(handle, credits);
687                         if (status < 0) {
688                                 /* handle still has to be committed at
689                                  * this point. */
690                                 status = -ENOMEM;
691                                 mlog_errno(status);
692                                 goto leave;
693                         }
694                         goto restarted_transaction;
695                 }
696         }
697
698         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
699              le32_to_cpu(fe->i_clusters),
700              (unsigned long long)le64_to_cpu(fe->i_size));
701         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
702              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
703
704 leave:
705         if (status < 0 && did_quota)
706                 dquot_free_space(inode,
707                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
708         if (handle) {
709                 ocfs2_commit_trans(osb, handle);
710                 handle = NULL;
711         }
712         if (data_ac) {
713                 ocfs2_free_alloc_context(data_ac);
714                 data_ac = NULL;
715         }
716         if (meta_ac) {
717                 ocfs2_free_alloc_context(meta_ac);
718                 meta_ac = NULL;
719         }
720         if ((!status) && restart_func) {
721                 restart_func = 0;
722                 goto restart_all;
723         }
724         brelse(bh);
725         bh = NULL;
726
727         mlog_exit(status);
728         return status;
729 }
730
731 /* Some parts of this taken from generic_cont_expand, which turned out
732  * to be too fragile to do exactly what we need without us having to
733  * worry about recursive locking in ->write_begin() and ->write_end(). */
734 static int ocfs2_write_zero_page(struct inode *inode,
735                                  u64 size)
736 {
737         struct address_space *mapping = inode->i_mapping;
738         struct page *page;
739         unsigned long index;
740         unsigned int offset;
741         handle_t *handle = NULL;
742         int ret;
743
744         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
745         /* ugh.  in prepare/commit_write, if from==to==start of block, we
746         ** skip the prepare.  make sure we never send an offset for the start
747         ** of a block
748         */
749         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
750                 offset++;
751         }
752         index = size >> PAGE_CACHE_SHIFT;
753
754         page = grab_cache_page(mapping, index);
755         if (!page) {
756                 ret = -ENOMEM;
757                 mlog_errno(ret);
758                 goto out;
759         }
760
761         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
762         if (ret < 0) {
763                 mlog_errno(ret);
764                 goto out_unlock;
765         }
766
767         if (ocfs2_should_order_data(inode)) {
768                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
769                                                      offset);
770                 if (IS_ERR(handle)) {
771                         ret = PTR_ERR(handle);
772                         handle = NULL;
773                         goto out_unlock;
774                 }
775         }
776
777         /* must not update i_size! */
778         ret = block_commit_write(page, offset, offset);
779         if (ret < 0)
780                 mlog_errno(ret);
781         else
782                 ret = 0;
783
784         if (handle)
785                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
786 out_unlock:
787         unlock_page(page);
788         page_cache_release(page);
789 out:
790         return ret;
791 }
792
793 static int ocfs2_zero_extend(struct inode *inode,
794                              u64 zero_to_size)
795 {
796         int ret = 0;
797         u64 start_off;
798         struct super_block *sb = inode->i_sb;
799
800         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
801         while (start_off < zero_to_size) {
802                 ret = ocfs2_write_zero_page(inode, start_off);
803                 if (ret < 0) {
804                         mlog_errno(ret);
805                         goto out;
806                 }
807
808                 start_off += sb->s_blocksize;
809
810                 /*
811                  * Very large extends have the potential to lock up
812                  * the cpu for extended periods of time.
813                  */
814                 cond_resched();
815         }
816
817 out:
818         return ret;
819 }
820
821 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
822 {
823         int ret;
824         u32 clusters_to_add;
825         struct ocfs2_inode_info *oi = OCFS2_I(inode);
826
827         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
828         if (clusters_to_add < oi->ip_clusters)
829                 clusters_to_add = 0;
830         else
831                 clusters_to_add -= oi->ip_clusters;
832
833         if (clusters_to_add) {
834                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
835                                                 clusters_to_add, 0);
836                 if (ret) {
837                         mlog_errno(ret);
838                         goto out;
839                 }
840         }
841
842         /*
843          * Call this even if we don't add any clusters to the tree. We
844          * still need to zero the area between the old i_size and the
845          * new i_size.
846          */
847         ret = ocfs2_zero_extend(inode, zero_to);
848         if (ret < 0)
849                 mlog_errno(ret);
850
851 out:
852         return ret;
853 }
854
855 static int ocfs2_extend_file(struct inode *inode,
856                              struct buffer_head *di_bh,
857                              u64 new_i_size)
858 {
859         int ret = 0;
860         struct ocfs2_inode_info *oi = OCFS2_I(inode);
861
862         BUG_ON(!di_bh);
863
864         /* setattr sometimes calls us like this. */
865         if (new_i_size == 0)
866                 goto out;
867
868         if (i_size_read(inode) == new_i_size)
869                 goto out;
870         BUG_ON(new_i_size < i_size_read(inode));
871
872         /*
873          * Fall through for converting inline data, even if the fs
874          * supports sparse files.
875          *
876          * The check for inline data here is legal - nobody can add
877          * the feature since we have i_mutex. We must check it again
878          * after acquiring ip_alloc_sem though, as paths like mmap
879          * might have raced us to converting the inode to extents.
880          */
881         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
882             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
883                 goto out_update_size;
884
885         /*
886          * The alloc sem blocks people in read/write from reading our
887          * allocation until we're done changing it. We depend on
888          * i_mutex to block other extend/truncate calls while we're
889          * here.
890          */
891         down_write(&oi->ip_alloc_sem);
892
893         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
894                 /*
895                  * We can optimize small extends by keeping the inodes
896                  * inline data.
897                  */
898                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
899                         up_write(&oi->ip_alloc_sem);
900                         goto out_update_size;
901                 }
902
903                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
904                 if (ret) {
905                         up_write(&oi->ip_alloc_sem);
906
907                         mlog_errno(ret);
908                         goto out;
909                 }
910         }
911
912         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
913                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
914
915         up_write(&oi->ip_alloc_sem);
916
917         if (ret < 0) {
918                 mlog_errno(ret);
919                 goto out;
920         }
921
922 out_update_size:
923         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
924         if (ret < 0)
925                 mlog_errno(ret);
926
927 out:
928         return ret;
929 }
930
931 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
932 {
933         int status = 0, size_change;
934         struct inode *inode = dentry->d_inode;
935         struct super_block *sb = inode->i_sb;
936         struct ocfs2_super *osb = OCFS2_SB(sb);
937         struct buffer_head *bh = NULL;
938         handle_t *handle = NULL;
939         int qtype;
940         struct dquot *transfer_from[MAXQUOTAS] = { };
941         struct dquot *transfer_to[MAXQUOTAS] = { };
942
943         mlog_entry("(0x%p, '%.*s')\n", dentry,
944                    dentry->d_name.len, dentry->d_name.name);
945
946         /* ensuring we don't even attempt to truncate a symlink */
947         if (S_ISLNK(inode->i_mode))
948                 attr->ia_valid &= ~ATTR_SIZE;
949
950         if (attr->ia_valid & ATTR_MODE)
951                 mlog(0, "mode change: %d\n", attr->ia_mode);
952         if (attr->ia_valid & ATTR_UID)
953                 mlog(0, "uid change: %d\n", attr->ia_uid);
954         if (attr->ia_valid & ATTR_GID)
955                 mlog(0, "gid change: %d\n", attr->ia_gid);
956         if (attr->ia_valid & ATTR_SIZE)
957                 mlog(0, "size change...\n");
958         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
959                 mlog(0, "time change...\n");
960
961 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
962                            | ATTR_GID | ATTR_UID | ATTR_MODE)
963         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
964                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
965                 return 0;
966         }
967
968         status = inode_change_ok(inode, attr);
969         if (status)
970                 return status;
971
972         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
973         if (size_change) {
974                 dquot_initialize(inode);
975
976                 status = ocfs2_rw_lock(inode, 1);
977                 if (status < 0) {
978                         mlog_errno(status);
979                         goto bail;
980                 }
981         }
982
983         status = ocfs2_inode_lock(inode, &bh, 1);
984         if (status < 0) {
985                 if (status != -ENOENT)
986                         mlog_errno(status);
987                 goto bail_unlock_rw;
988         }
989
990         if (size_change && attr->ia_size != i_size_read(inode)) {
991                 status = inode_newsize_ok(inode, attr->ia_size);
992                 if (status)
993                         goto bail_unlock;
994
995                 if (i_size_read(inode) > attr->ia_size) {
996                         if (ocfs2_should_order_data(inode)) {
997                                 status = ocfs2_begin_ordered_truncate(inode,
998                                                                       attr->ia_size);
999                                 if (status)
1000                                         goto bail_unlock;
1001                         }
1002                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1003                 } else
1004                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1005                 if (status < 0) {
1006                         if (status != -ENOSPC)
1007                                 mlog_errno(status);
1008                         status = -ENOSPC;
1009                         goto bail_unlock;
1010                 }
1011         }
1012
1013         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1014             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1015                 /*
1016                  * Gather pointers to quota structures so that allocation /
1017                  * freeing of quota structures happens here and not inside
1018                  * dquot_transfer() where we have problems with lock ordering
1019                  */
1020                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1021                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1022                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1023                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1024                                                       USRQUOTA);
1025                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1026                                                         USRQUOTA);
1027                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1028                                 status = -ESRCH;
1029                                 goto bail_unlock;
1030                         }
1031                 }
1032                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1033                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1034                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1035                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1036                                                       GRPQUOTA);
1037                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1038                                                         GRPQUOTA);
1039                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1040                                 status = -ESRCH;
1041                                 goto bail_unlock;
1042                         }
1043                 }
1044                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1045                                            2 * ocfs2_quota_trans_credits(sb));
1046                 if (IS_ERR(handle)) {
1047                         status = PTR_ERR(handle);
1048                         mlog_errno(status);
1049                         goto bail_unlock;
1050                 }
1051                 status = dquot_transfer(inode, attr);
1052                 if (status < 0)
1053                         goto bail_commit;
1054         } else {
1055                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1056                 if (IS_ERR(handle)) {
1057                         status = PTR_ERR(handle);
1058                         mlog_errno(status);
1059                         goto bail_unlock;
1060                 }
1061         }
1062
1063         /*
1064          * This will intentionally not wind up calling vmtruncate(),
1065          * since all the work for a size change has been done above.
1066          * Otherwise, we could get into problems with truncate as
1067          * ip_alloc_sem is used there to protect against i_size
1068          * changes.
1069          */
1070         status = inode_setattr(inode, attr);
1071         if (status < 0) {
1072                 mlog_errno(status);
1073                 goto bail_commit;
1074         }
1075
1076         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1077         if (status < 0)
1078                 mlog_errno(status);
1079
1080 bail_commit:
1081         ocfs2_commit_trans(osb, handle);
1082 bail_unlock:
1083         ocfs2_inode_unlock(inode, 1);
1084 bail_unlock_rw:
1085         if (size_change)
1086                 ocfs2_rw_unlock(inode, 1);
1087 bail:
1088         brelse(bh);
1089
1090         /* Release quota pointers in case we acquired them */
1091         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1092                 dqput(transfer_to[qtype]);
1093                 dqput(transfer_from[qtype]);
1094         }
1095
1096         if (!status && attr->ia_valid & ATTR_MODE) {
1097                 status = ocfs2_acl_chmod(inode);
1098                 if (status < 0)
1099                         mlog_errno(status);
1100         }
1101
1102         mlog_exit(status);
1103         return status;
1104 }
1105
1106 int ocfs2_getattr(struct vfsmount *mnt,
1107                   struct dentry *dentry,
1108                   struct kstat *stat)
1109 {
1110         struct inode *inode = dentry->d_inode;
1111         struct super_block *sb = dentry->d_inode->i_sb;
1112         struct ocfs2_super *osb = sb->s_fs_info;
1113         int err;
1114
1115         mlog_entry_void();
1116
1117         err = ocfs2_inode_revalidate(dentry);
1118         if (err) {
1119                 if (err != -ENOENT)
1120                         mlog_errno(err);
1121                 goto bail;
1122         }
1123
1124         generic_fillattr(inode, stat);
1125
1126         /* We set the blksize from the cluster size for performance */
1127         stat->blksize = osb->s_clustersize;
1128
1129 bail:
1130         mlog_exit(err);
1131
1132         return err;
1133 }
1134
1135 int ocfs2_permission(struct inode *inode, int mask)
1136 {
1137         int ret;
1138
1139         mlog_entry_void();
1140
1141         ret = ocfs2_inode_lock(inode, NULL, 0);
1142         if (ret) {
1143                 if (ret != -ENOENT)
1144                         mlog_errno(ret);
1145                 goto out;
1146         }
1147
1148         ret = generic_permission(inode, mask, ocfs2_check_acl);
1149
1150         ocfs2_inode_unlock(inode, 0);
1151 out:
1152         mlog_exit(ret);
1153         return ret;
1154 }
1155
1156 static int __ocfs2_write_remove_suid(struct inode *inode,
1157                                      struct buffer_head *bh)
1158 {
1159         int ret;
1160         handle_t *handle;
1161         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1162         struct ocfs2_dinode *di;
1163
1164         mlog_entry("(Inode %llu, mode 0%o)\n",
1165                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1166
1167         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1168         if (IS_ERR(handle)) {
1169                 ret = PTR_ERR(handle);
1170                 mlog_errno(ret);
1171                 goto out;
1172         }
1173
1174         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1175                                       OCFS2_JOURNAL_ACCESS_WRITE);
1176         if (ret < 0) {
1177                 mlog_errno(ret);
1178                 goto out_trans;
1179         }
1180
1181         inode->i_mode &= ~S_ISUID;
1182         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1183                 inode->i_mode &= ~S_ISGID;
1184
1185         di = (struct ocfs2_dinode *) bh->b_data;
1186         di->i_mode = cpu_to_le16(inode->i_mode);
1187
1188         ocfs2_journal_dirty(handle, bh);
1189
1190 out_trans:
1191         ocfs2_commit_trans(osb, handle);
1192 out:
1193         mlog_exit(ret);
1194         return ret;
1195 }
1196
1197 /*
1198  * Will look for holes and unwritten extents in the range starting at
1199  * pos for count bytes (inclusive).
1200  */
1201 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1202                                        size_t count)
1203 {
1204         int ret = 0;
1205         unsigned int extent_flags;
1206         u32 cpos, clusters, extent_len, phys_cpos;
1207         struct super_block *sb = inode->i_sb;
1208
1209         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1210         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1211
1212         while (clusters) {
1213                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1214                                          &extent_flags);
1215                 if (ret < 0) {
1216                         mlog_errno(ret);
1217                         goto out;
1218                 }
1219
1220                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1221                         ret = 1;
1222                         break;
1223                 }
1224
1225                 if (extent_len > clusters)
1226                         extent_len = clusters;
1227
1228                 clusters -= extent_len;
1229                 cpos += extent_len;
1230         }
1231 out:
1232         return ret;
1233 }
1234
1235 static int ocfs2_write_remove_suid(struct inode *inode)
1236 {
1237         int ret;
1238         struct buffer_head *bh = NULL;
1239
1240         ret = ocfs2_read_inode_block(inode, &bh);
1241         if (ret < 0) {
1242                 mlog_errno(ret);
1243                 goto out;
1244         }
1245
1246         ret =  __ocfs2_write_remove_suid(inode, bh);
1247 out:
1248         brelse(bh);
1249         return ret;
1250 }
1251
1252 /*
1253  * Allocate enough extents to cover the region starting at byte offset
1254  * start for len bytes. Existing extents are skipped, any extents
1255  * added are marked as "unwritten".
1256  */
1257 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1258                                             u64 start, u64 len)
1259 {
1260         int ret;
1261         u32 cpos, phys_cpos, clusters, alloc_size;
1262         u64 end = start + len;
1263         struct buffer_head *di_bh = NULL;
1264
1265         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1266                 ret = ocfs2_read_inode_block(inode, &di_bh);
1267                 if (ret) {
1268                         mlog_errno(ret);
1269                         goto out;
1270                 }
1271
1272                 /*
1273                  * Nothing to do if the requested reservation range
1274                  * fits within the inode.
1275                  */
1276                 if (ocfs2_size_fits_inline_data(di_bh, end))
1277                         goto out;
1278
1279                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1280                 if (ret) {
1281                         mlog_errno(ret);
1282                         goto out;
1283                 }
1284         }
1285
1286         /*
1287          * We consider both start and len to be inclusive.
1288          */
1289         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1290         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1291         clusters -= cpos;
1292
1293         while (clusters) {
1294                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1295                                          &alloc_size, NULL);
1296                 if (ret) {
1297                         mlog_errno(ret);
1298                         goto out;
1299                 }
1300
1301                 /*
1302                  * Hole or existing extent len can be arbitrary, so
1303                  * cap it to our own allocation request.
1304                  */
1305                 if (alloc_size > clusters)
1306                         alloc_size = clusters;
1307
1308                 if (phys_cpos) {
1309                         /*
1310                          * We already have an allocation at this
1311                          * region so we can safely skip it.
1312                          */
1313                         goto next;
1314                 }
1315
1316                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1317                 if (ret) {
1318                         if (ret != -ENOSPC)
1319                                 mlog_errno(ret);
1320                         goto out;
1321                 }
1322
1323 next:
1324                 cpos += alloc_size;
1325                 clusters -= alloc_size;
1326         }
1327
1328         ret = 0;
1329 out:
1330
1331         brelse(di_bh);
1332         return ret;
1333 }
1334
1335 /*
1336  * Truncate a byte range, avoiding pages within partial clusters. This
1337  * preserves those pages for the zeroing code to write to.
1338  */
1339 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1340                                          u64 byte_len)
1341 {
1342         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1343         loff_t start, end;
1344         struct address_space *mapping = inode->i_mapping;
1345
1346         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1347         end = byte_start + byte_len;
1348         end = end & ~(osb->s_clustersize - 1);
1349
1350         if (start < end) {
1351                 unmap_mapping_range(mapping, start, end - start, 0);
1352                 truncate_inode_pages_range(mapping, start, end - 1);
1353         }
1354 }
1355
1356 static int ocfs2_zero_partial_clusters(struct inode *inode,
1357                                        u64 start, u64 len)
1358 {
1359         int ret = 0;
1360         u64 tmpend, end = start + len;
1361         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1362         unsigned int csize = osb->s_clustersize;
1363         handle_t *handle;
1364
1365         /*
1366          * The "start" and "end" values are NOT necessarily part of
1367          * the range whose allocation is being deleted. Rather, this
1368          * is what the user passed in with the request. We must zero
1369          * partial clusters here. There's no need to worry about
1370          * physical allocation - the zeroing code knows to skip holes.
1371          */
1372         mlog(0, "byte start: %llu, end: %llu\n",
1373              (unsigned long long)start, (unsigned long long)end);
1374
1375         /*
1376          * If both edges are on a cluster boundary then there's no
1377          * zeroing required as the region is part of the allocation to
1378          * be truncated.
1379          */
1380         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1381                 goto out;
1382
1383         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1384         if (IS_ERR(handle)) {
1385                 ret = PTR_ERR(handle);
1386                 mlog_errno(ret);
1387                 goto out;
1388         }
1389
1390         /*
1391          * We want to get the byte offset of the end of the 1st cluster.
1392          */
1393         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1394         if (tmpend > end)
1395                 tmpend = end;
1396
1397         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1398              (unsigned long long)start, (unsigned long long)tmpend);
1399
1400         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1401         if (ret)
1402                 mlog_errno(ret);
1403
1404         if (tmpend < end) {
1405                 /*
1406                  * This may make start and end equal, but the zeroing
1407                  * code will skip any work in that case so there's no
1408                  * need to catch it up here.
1409                  */
1410                 start = end & ~(osb->s_clustersize - 1);
1411
1412                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1413                      (unsigned long long)start, (unsigned long long)end);
1414
1415                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1416                 if (ret)
1417                         mlog_errno(ret);
1418         }
1419
1420         ocfs2_commit_trans(osb, handle);
1421 out:
1422         return ret;
1423 }
1424
1425 static int ocfs2_remove_inode_range(struct inode *inode,
1426                                     struct buffer_head *di_bh, u64 byte_start,
1427                                     u64 byte_len)
1428 {
1429         int ret = 0;
1430         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1431         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1432         struct ocfs2_cached_dealloc_ctxt dealloc;
1433         struct address_space *mapping = inode->i_mapping;
1434         struct ocfs2_extent_tree et;
1435
1436         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1437         ocfs2_init_dealloc_ctxt(&dealloc);
1438
1439         if (byte_len == 0)
1440                 return 0;
1441
1442         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1443                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1444                                             byte_start + byte_len, 0);
1445                 if (ret) {
1446                         mlog_errno(ret);
1447                         goto out;
1448                 }
1449                 /*
1450                  * There's no need to get fancy with the page cache
1451                  * truncate of an inline-data inode. We're talking
1452                  * about less than a page here, which will be cached
1453                  * in the dinode buffer anyway.
1454                  */
1455                 unmap_mapping_range(mapping, 0, 0, 0);
1456                 truncate_inode_pages(mapping, 0);
1457                 goto out;
1458         }
1459
1460         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1461         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1462         if (trunc_len >= trunc_start)
1463                 trunc_len -= trunc_start;
1464         else
1465                 trunc_len = 0;
1466
1467         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1468              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1469              (unsigned long long)byte_start,
1470              (unsigned long long)byte_len, trunc_start, trunc_len);
1471
1472         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1473         if (ret) {
1474                 mlog_errno(ret);
1475                 goto out;
1476         }
1477
1478         cpos = trunc_start;
1479         while (trunc_len) {
1480                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1481                                          &alloc_size, NULL);
1482                 if (ret) {
1483                         mlog_errno(ret);
1484                         goto out;
1485                 }
1486
1487                 if (alloc_size > trunc_len)
1488                         alloc_size = trunc_len;
1489
1490                 /* Only do work for non-holes */
1491                 if (phys_cpos != 0) {
1492                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1493                                                        phys_cpos, alloc_size,
1494                                                        &dealloc);
1495                         if (ret) {
1496                                 mlog_errno(ret);
1497                                 goto out;
1498                         }
1499                 }
1500
1501                 cpos += alloc_size;
1502                 trunc_len -= alloc_size;
1503         }
1504
1505         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1506
1507 out:
1508         ocfs2_schedule_truncate_log_flush(osb, 1);
1509         ocfs2_run_deallocs(osb, &dealloc);
1510
1511         return ret;
1512 }
1513
1514 /*
1515  * Parts of this function taken from xfs_change_file_space()
1516  */
1517 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1518                                      loff_t f_pos, unsigned int cmd,
1519                                      struct ocfs2_space_resv *sr,
1520                                      int change_size)
1521 {
1522         int ret;
1523         s64 llen;
1524         loff_t size;
1525         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1526         struct buffer_head *di_bh = NULL;
1527         handle_t *handle;
1528         unsigned long long max_off = inode->i_sb->s_maxbytes;
1529
1530         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1531                 return -EROFS;
1532
1533         mutex_lock(&inode->i_mutex);
1534
1535         /*
1536          * This prevents concurrent writes on other nodes
1537          */
1538         ret = ocfs2_rw_lock(inode, 1);
1539         if (ret) {
1540                 mlog_errno(ret);
1541                 goto out;
1542         }
1543
1544         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1545         if (ret) {
1546                 mlog_errno(ret);
1547                 goto out_rw_unlock;
1548         }
1549
1550         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1551                 ret = -EPERM;
1552                 goto out_inode_unlock;
1553         }
1554
1555         switch (sr->l_whence) {
1556         case 0: /*SEEK_SET*/
1557                 break;
1558         case 1: /*SEEK_CUR*/
1559                 sr->l_start += f_pos;
1560                 break;
1561         case 2: /*SEEK_END*/
1562                 sr->l_start += i_size_read(inode);
1563                 break;
1564         default:
1565                 ret = -EINVAL;
1566                 goto out_inode_unlock;
1567         }
1568         sr->l_whence = 0;
1569
1570         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1571
1572         if (sr->l_start < 0
1573             || sr->l_start > max_off
1574             || (sr->l_start + llen) < 0
1575             || (sr->l_start + llen) > max_off) {
1576                 ret = -EINVAL;
1577                 goto out_inode_unlock;
1578         }
1579         size = sr->l_start + sr->l_len;
1580
1581         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1582                 if (sr->l_len <= 0) {
1583                         ret = -EINVAL;
1584                         goto out_inode_unlock;
1585                 }
1586         }
1587
1588         if (file && should_remove_suid(file->f_path.dentry)) {
1589                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1590                 if (ret) {
1591                         mlog_errno(ret);
1592                         goto out_inode_unlock;
1593                 }
1594         }
1595
1596         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1597         switch (cmd) {
1598         case OCFS2_IOC_RESVSP:
1599         case OCFS2_IOC_RESVSP64:
1600                 /*
1601                  * This takes unsigned offsets, but the signed ones we
1602                  * pass have been checked against overflow above.
1603                  */
1604                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1605                                                        sr->l_len);
1606                 break;
1607         case OCFS2_IOC_UNRESVSP:
1608         case OCFS2_IOC_UNRESVSP64:
1609                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1610                                                sr->l_len);
1611                 break;
1612         default:
1613                 ret = -EINVAL;
1614         }
1615         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1616         if (ret) {
1617                 mlog_errno(ret);
1618                 goto out_inode_unlock;
1619         }
1620
1621         /*
1622          * We update c/mtime for these changes
1623          */
1624         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1625         if (IS_ERR(handle)) {
1626                 ret = PTR_ERR(handle);
1627                 mlog_errno(ret);
1628                 goto out_inode_unlock;
1629         }
1630
1631         if (change_size && i_size_read(inode) < size)
1632                 i_size_write(inode, size);
1633
1634         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1635         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1636         if (ret < 0)
1637                 mlog_errno(ret);
1638
1639         ocfs2_commit_trans(osb, handle);
1640
1641 out_inode_unlock:
1642         brelse(di_bh);
1643         ocfs2_inode_unlock(inode, 1);
1644 out_rw_unlock:
1645         ocfs2_rw_unlock(inode, 1);
1646
1647 out:
1648         mutex_unlock(&inode->i_mutex);
1649         return ret;
1650 }
1651
1652 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1653                             struct ocfs2_space_resv *sr)
1654 {
1655         struct inode *inode = file->f_path.dentry->d_inode;
1656         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1657
1658         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1659             !ocfs2_writes_unwritten_extents(osb))
1660                 return -ENOTTY;
1661         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1662                  !ocfs2_sparse_alloc(osb))
1663                 return -ENOTTY;
1664
1665         if (!S_ISREG(inode->i_mode))
1666                 return -EINVAL;
1667
1668         if (!(file->f_mode & FMODE_WRITE))
1669                 return -EBADF;
1670
1671         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1672 }
1673
1674 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1675                             loff_t len)
1676 {
1677         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1678         struct ocfs2_space_resv sr;
1679         int change_size = 1;
1680
1681         if (!ocfs2_writes_unwritten_extents(osb))
1682                 return -EOPNOTSUPP;
1683
1684         if (S_ISDIR(inode->i_mode))
1685                 return -ENODEV;
1686
1687         if (mode & FALLOC_FL_KEEP_SIZE)
1688                 change_size = 0;
1689
1690         sr.l_whence = 0;
1691         sr.l_start = (s64)offset;
1692         sr.l_len = (s64)len;
1693
1694         return __ocfs2_change_file_space(NULL, inode, offset,
1695                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1696 }
1697
1698 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1699                                    size_t count)
1700 {
1701         int ret = 0;
1702         unsigned int extent_flags;
1703         u32 cpos, clusters, extent_len, phys_cpos;
1704         struct super_block *sb = inode->i_sb;
1705
1706         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1707             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1708             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1709                 return 0;
1710
1711         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1712         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1713
1714         while (clusters) {
1715                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1716                                          &extent_flags);
1717                 if (ret < 0) {
1718                         mlog_errno(ret);
1719                         goto out;
1720                 }
1721
1722                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1723                         ret = 1;
1724                         break;
1725                 }
1726
1727                 if (extent_len > clusters)
1728                         extent_len = clusters;
1729
1730                 clusters -= extent_len;
1731                 cpos += extent_len;
1732         }
1733 out:
1734         return ret;
1735 }
1736
1737 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1738                                             loff_t pos, size_t count,
1739                                             int *meta_level)
1740 {
1741         int ret;
1742         struct buffer_head *di_bh = NULL;
1743         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1744         u32 clusters =
1745                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1746
1747         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1748         if (ret) {
1749                 mlog_errno(ret);
1750                 goto out;
1751         }
1752
1753         *meta_level = 1;
1754
1755         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1756         if (ret)
1757                 mlog_errno(ret);
1758 out:
1759         brelse(di_bh);
1760         return ret;
1761 }
1762
1763 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1764                                          loff_t *ppos,
1765                                          size_t count,
1766                                          int appending,
1767                                          int *direct_io,
1768                                          int *has_refcount)
1769 {
1770         int ret = 0, meta_level = 0;
1771         struct inode *inode = dentry->d_inode;
1772         loff_t saved_pos, end;
1773
1774         /*
1775          * We start with a read level meta lock and only jump to an ex
1776          * if we need to make modifications here.
1777          */
1778         for(;;) {
1779                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1780                 if (ret < 0) {
1781                         meta_level = -1;
1782                         mlog_errno(ret);
1783                         goto out;
1784                 }
1785
1786                 /* Clear suid / sgid if necessary. We do this here
1787                  * instead of later in the write path because
1788                  * remove_suid() calls ->setattr without any hint that
1789                  * we may have already done our cluster locking. Since
1790                  * ocfs2_setattr() *must* take cluster locks to
1791                  * proceeed, this will lead us to recursively lock the
1792                  * inode. There's also the dinode i_size state which
1793                  * can be lost via setattr during extending writes (we
1794                  * set inode->i_size at the end of a write. */
1795                 if (should_remove_suid(dentry)) {
1796                         if (meta_level == 0) {
1797                                 ocfs2_inode_unlock(inode, meta_level);
1798                                 meta_level = 1;
1799                                 continue;
1800                         }
1801
1802                         ret = ocfs2_write_remove_suid(inode);
1803                         if (ret < 0) {
1804                                 mlog_errno(ret);
1805                                 goto out_unlock;
1806                         }
1807                 }
1808
1809                 /* work on a copy of ppos until we're sure that we won't have
1810                  * to recalculate it due to relocking. */
1811                 if (appending) {
1812                         saved_pos = i_size_read(inode);
1813                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1814                 } else {
1815                         saved_pos = *ppos;
1816                 }
1817
1818                 end = saved_pos + count;
1819
1820                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1821                 if (ret == 1) {
1822                         ocfs2_inode_unlock(inode, meta_level);
1823                         meta_level = -1;
1824
1825                         ret = ocfs2_prepare_inode_for_refcount(inode,
1826                                                                saved_pos,
1827                                                                count,
1828                                                                &meta_level);
1829                         if (has_refcount)
1830                                 *has_refcount = 1;
1831                         if (direct_io)
1832                                 *direct_io = 0;
1833                 }
1834
1835                 if (ret < 0) {
1836                         mlog_errno(ret);
1837                         goto out_unlock;
1838                 }
1839
1840                 /*
1841                  * Skip the O_DIRECT checks if we don't need
1842                  * them.
1843                  */
1844                 if (!direct_io || !(*direct_io))
1845                         break;
1846
1847                 /*
1848                  * There's no sane way to do direct writes to an inode
1849                  * with inline data.
1850                  */
1851                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1852                         *direct_io = 0;
1853                         break;
1854                 }
1855
1856                 /*
1857                  * Allowing concurrent direct writes means
1858                  * i_size changes wouldn't be synchronized, so
1859                  * one node could wind up truncating another
1860                  * nodes writes.
1861                  */
1862                 if (end > i_size_read(inode)) {
1863                         *direct_io = 0;
1864                         break;
1865                 }
1866
1867                 /*
1868                  * We don't fill holes during direct io, so
1869                  * check for them here. If any are found, the
1870                  * caller will have to retake some cluster
1871                  * locks and initiate the io as buffered.
1872                  */
1873                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1874                 if (ret == 1) {
1875                         *direct_io = 0;
1876                         ret = 0;
1877                 } else if (ret < 0)
1878                         mlog_errno(ret);
1879                 break;
1880         }
1881
1882         if (appending)
1883                 *ppos = saved_pos;
1884
1885 out_unlock:
1886         if (meta_level >= 0)
1887                 ocfs2_inode_unlock(inode, meta_level);
1888
1889 out:
1890         return ret;
1891 }
1892
1893 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1894                                     const struct iovec *iov,
1895                                     unsigned long nr_segs,
1896                                     loff_t pos)
1897 {
1898         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1899         int can_do_direct, has_refcount = 0;
1900         ssize_t written = 0;
1901         size_t ocount;          /* original count */
1902         size_t count;           /* after file limit checks */
1903         loff_t old_size, *ppos = &iocb->ki_pos;
1904         u32 old_clusters;
1905         struct file *file = iocb->ki_filp;
1906         struct inode *inode = file->f_path.dentry->d_inode;
1907         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1908
1909         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1910                    (unsigned int)nr_segs,
1911                    file->f_path.dentry->d_name.len,
1912                    file->f_path.dentry->d_name.name);
1913
1914         if (iocb->ki_left == 0)
1915                 return 0;
1916
1917         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1918
1919         appending = file->f_flags & O_APPEND ? 1 : 0;
1920         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1921
1922         mutex_lock(&inode->i_mutex);
1923
1924 relock:
1925         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1926         if (direct_io) {
1927                 down_read(&inode->i_alloc_sem);
1928                 have_alloc_sem = 1;
1929         }
1930
1931         /* concurrent O_DIRECT writes are allowed */
1932         rw_level = !direct_io;
1933         ret = ocfs2_rw_lock(inode, rw_level);
1934         if (ret < 0) {
1935                 mlog_errno(ret);
1936                 goto out_sems;
1937         }
1938
1939         can_do_direct = direct_io;
1940         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1941                                             iocb->ki_left, appending,
1942                                             &can_do_direct, &has_refcount);
1943         if (ret < 0) {
1944                 mlog_errno(ret);
1945                 goto out;
1946         }
1947
1948         /*
1949          * We can't complete the direct I/O as requested, fall back to
1950          * buffered I/O.
1951          */
1952         if (direct_io && !can_do_direct) {
1953                 ocfs2_rw_unlock(inode, rw_level);
1954                 up_read(&inode->i_alloc_sem);
1955
1956                 have_alloc_sem = 0;
1957                 rw_level = -1;
1958
1959                 direct_io = 0;
1960                 goto relock;
1961         }
1962
1963         /*
1964          * To later detect whether a journal commit for sync writes is
1965          * necessary, we sample i_size, and cluster count here.
1966          */
1967         old_size = i_size_read(inode);
1968         old_clusters = OCFS2_I(inode)->ip_clusters;
1969
1970         /* communicate with ocfs2_dio_end_io */
1971         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1972
1973         if (direct_io) {
1974                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1975                                              VERIFY_READ);
1976                 if (ret)
1977                         goto out_dio;
1978
1979                 count = ocount;
1980                 ret = generic_write_checks(file, ppos, &count,
1981                                            S_ISBLK(inode->i_mode));
1982                 if (ret)
1983                         goto out_dio;
1984
1985                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1986                                                     ppos, count, ocount);
1987                 if (written < 0) {
1988                         /*
1989                          * direct write may have instantiated a few
1990                          * blocks outside i_size. Trim these off again.
1991                          * Don't need i_size_read because we hold i_mutex.
1992                          */
1993                         if (*ppos + count > inode->i_size)
1994                                 vmtruncate(inode, inode->i_size);
1995                         ret = written;
1996                         goto out_dio;
1997                 }
1998         } else {
1999                 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2000         }
2001
2002 out_dio:
2003         /* buffered aio wouldn't have proper lock coverage today */
2004         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2005
2006         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2007             ((file->f_flags & O_DIRECT) && has_refcount)) {
2008                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2009                                                pos + count - 1);
2010                 if (ret < 0)
2011                         written = ret;
2012
2013                 if (!ret && (old_size != i_size_read(inode) ||
2014                     old_clusters != OCFS2_I(inode)->ip_clusters ||
2015                     has_refcount)) {
2016                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2017                         if (ret < 0)
2018                                 written = ret;
2019                 }
2020
2021                 if (!ret)
2022                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2023                                                       pos + count - 1);
2024         }
2025
2026         /*
2027          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2028          * function pointer which is called when o_direct io completes so that
2029          * it can unlock our rw lock.  (it's the clustered equivalent of
2030          * i_alloc_sem; protects truncate from racing with pending ios).
2031          * Unfortunately there are error cases which call end_io and others
2032          * that don't.  so we don't have to unlock the rw_lock if either an
2033          * async dio is going to do it in the future or an end_io after an
2034          * error has already done it.
2035          */
2036         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2037                 rw_level = -1;
2038                 have_alloc_sem = 0;
2039         }
2040
2041 out:
2042         if (rw_level != -1)
2043                 ocfs2_rw_unlock(inode, rw_level);
2044
2045 out_sems:
2046         if (have_alloc_sem)
2047                 up_read(&inode->i_alloc_sem);
2048
2049         mutex_unlock(&inode->i_mutex);
2050
2051         if (written)
2052                 ret = written;
2053         mlog_exit(ret);
2054         return ret;
2055 }
2056
2057 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2058                                 struct file *out,
2059                                 struct splice_desc *sd)
2060 {
2061         int ret;
2062
2063         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2064                                             sd->total_len, 0, NULL, NULL);
2065         if (ret < 0) {
2066                 mlog_errno(ret);
2067                 return ret;
2068         }
2069
2070         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2071 }
2072
2073 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2074                                        struct file *out,
2075                                        loff_t *ppos,
2076                                        size_t len,
2077                                        unsigned int flags)
2078 {
2079         int ret;
2080         struct address_space *mapping = out->f_mapping;
2081         struct inode *inode = mapping->host;
2082         struct splice_desc sd = {
2083                 .total_len = len,
2084                 .flags = flags,
2085                 .pos = *ppos,
2086                 .u.file = out,
2087         };
2088
2089         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2090                    (unsigned int)len,
2091                    out->f_path.dentry->d_name.len,
2092                    out->f_path.dentry->d_name.name);
2093
2094         if (pipe->inode)
2095                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2096
2097         splice_from_pipe_begin(&sd);
2098         do {
2099                 ret = splice_from_pipe_next(pipe, &sd);
2100                 if (ret <= 0)
2101                         break;
2102
2103                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2104                 ret = ocfs2_rw_lock(inode, 1);
2105                 if (ret < 0)
2106                         mlog_errno(ret);
2107                 else {
2108                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2109                         ocfs2_rw_unlock(inode, 1);
2110                 }
2111                 mutex_unlock(&inode->i_mutex);
2112         } while (ret > 0);
2113         splice_from_pipe_end(pipe, &sd);
2114
2115         if (pipe->inode)
2116                 mutex_unlock(&pipe->inode->i_mutex);
2117
2118         if (sd.num_spliced)
2119                 ret = sd.num_spliced;
2120
2121         if (ret > 0) {
2122                 unsigned long nr_pages;
2123                 int err;
2124
2125                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2126
2127                 err = generic_write_sync(out, *ppos, ret);
2128                 if (err)
2129                         ret = err;
2130                 else
2131                         *ppos += ret;
2132
2133                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2134         }
2135
2136         mlog_exit(ret);
2137         return ret;
2138 }
2139
2140 static ssize_t ocfs2_file_splice_read(struct file *in,
2141                                       loff_t *ppos,
2142                                       struct pipe_inode_info *pipe,
2143                                       size_t len,
2144                                       unsigned int flags)
2145 {
2146         int ret = 0, lock_level = 0;
2147         struct inode *inode = in->f_path.dentry->d_inode;
2148
2149         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2150                    (unsigned int)len,
2151                    in->f_path.dentry->d_name.len,
2152                    in->f_path.dentry->d_name.name);
2153
2154         /*
2155          * See the comment in ocfs2_file_aio_read()
2156          */
2157         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2158         if (ret < 0) {
2159                 mlog_errno(ret);
2160                 goto bail;
2161         }
2162         ocfs2_inode_unlock(inode, lock_level);
2163
2164         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2165
2166 bail:
2167         mlog_exit(ret);
2168         return ret;
2169 }
2170
2171 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2172                                    const struct iovec *iov,
2173                                    unsigned long nr_segs,
2174                                    loff_t pos)
2175 {
2176         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2177         struct file *filp = iocb->ki_filp;
2178         struct inode *inode = filp->f_path.dentry->d_inode;
2179
2180         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2181                    (unsigned int)nr_segs,
2182                    filp->f_path.dentry->d_name.len,
2183                    filp->f_path.dentry->d_name.name);
2184
2185         if (!inode) {
2186                 ret = -EINVAL;
2187                 mlog_errno(ret);
2188                 goto bail;
2189         }
2190
2191         /*
2192          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2193          * need locks to protect pending reads from racing with truncate.
2194          */
2195         if (filp->f_flags & O_DIRECT) {
2196                 down_read(&inode->i_alloc_sem);
2197                 have_alloc_sem = 1;
2198
2199                 ret = ocfs2_rw_lock(inode, 0);
2200                 if (ret < 0) {
2201                         mlog_errno(ret);
2202                         goto bail;
2203                 }
2204                 rw_level = 0;
2205                 /* communicate with ocfs2_dio_end_io */
2206                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2207         }
2208
2209         /*
2210          * We're fine letting folks race truncates and extending
2211          * writes with read across the cluster, just like they can
2212          * locally. Hence no rw_lock during read.
2213          *
2214          * Take and drop the meta data lock to update inode fields
2215          * like i_size. This allows the checks down below
2216          * generic_file_aio_read() a chance of actually working.
2217          */
2218         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2219         if (ret < 0) {
2220                 mlog_errno(ret);
2221                 goto bail;
2222         }
2223         ocfs2_inode_unlock(inode, lock_level);
2224
2225         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2226         if (ret == -EINVAL)
2227                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2228
2229         /* buffered aio wouldn't have proper lock coverage today */
2230         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2231
2232         /* see ocfs2_file_aio_write */
2233         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2234                 rw_level = -1;
2235                 have_alloc_sem = 0;
2236         }
2237
2238 bail:
2239         if (have_alloc_sem)
2240                 up_read(&inode->i_alloc_sem);
2241         if (rw_level != -1)
2242                 ocfs2_rw_unlock(inode, rw_level);
2243         mlog_exit(ret);
2244
2245         return ret;
2246 }
2247
2248 const struct inode_operations ocfs2_file_iops = {
2249         .setattr        = ocfs2_setattr,
2250         .getattr        = ocfs2_getattr,
2251         .permission     = ocfs2_permission,
2252         .setxattr       = generic_setxattr,
2253         .getxattr       = generic_getxattr,
2254         .listxattr      = ocfs2_listxattr,
2255         .removexattr    = generic_removexattr,
2256         .fallocate      = ocfs2_fallocate,
2257         .fiemap         = ocfs2_fiemap,
2258 };
2259
2260 const struct inode_operations ocfs2_special_file_iops = {
2261         .setattr        = ocfs2_setattr,
2262         .getattr        = ocfs2_getattr,
2263         .permission     = ocfs2_permission,
2264 };
2265
2266 /*
2267  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2268  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2269  */
2270 const struct file_operations ocfs2_fops = {
2271         .llseek         = generic_file_llseek,
2272         .read           = do_sync_read,
2273         .write          = do_sync_write,
2274         .mmap           = ocfs2_mmap,
2275         .fsync          = ocfs2_sync_file,
2276         .release        = ocfs2_file_release,
2277         .open           = ocfs2_file_open,
2278         .aio_read       = ocfs2_file_aio_read,
2279         .aio_write      = ocfs2_file_aio_write,
2280         .unlocked_ioctl = ocfs2_ioctl,
2281 #ifdef CONFIG_COMPAT
2282         .compat_ioctl   = ocfs2_compat_ioctl,
2283 #endif
2284         .lock           = ocfs2_lock,
2285         .flock          = ocfs2_flock,
2286         .splice_read    = ocfs2_file_splice_read,
2287         .splice_write   = ocfs2_file_splice_write,
2288 };
2289
2290 const struct file_operations ocfs2_dops = {
2291         .llseek         = generic_file_llseek,
2292         .read           = generic_read_dir,
2293         .readdir        = ocfs2_readdir,
2294         .fsync          = ocfs2_sync_file,
2295         .release        = ocfs2_dir_release,
2296         .open           = ocfs2_dir_open,
2297         .unlocked_ioctl = ocfs2_ioctl,
2298 #ifdef CONFIG_COMPAT
2299         .compat_ioctl   = ocfs2_compat_ioctl,
2300 #endif
2301         .lock           = ocfs2_lock,
2302         .flock          = ocfs2_flock,
2303 };
2304
2305 /*
2306  * POSIX-lockless variants of our file_operations.
2307  *
2308  * These will be used if the underlying cluster stack does not support
2309  * posix file locking, if the user passes the "localflocks" mount
2310  * option, or if we have a local-only fs.
2311  *
2312  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2313  * so we still want it in the case of no stack support for
2314  * plocks. Internally, it will do the right thing when asked to ignore
2315  * the cluster.
2316  */
2317 const struct file_operations ocfs2_fops_no_plocks = {
2318         .llseek         = generic_file_llseek,
2319         .read           = do_sync_read,
2320         .write          = do_sync_write,
2321         .mmap           = ocfs2_mmap,
2322         .fsync          = ocfs2_sync_file,
2323         .release        = ocfs2_file_release,
2324         .open           = ocfs2_file_open,
2325         .aio_read       = ocfs2_file_aio_read,
2326         .aio_write      = ocfs2_file_aio_write,
2327         .unlocked_ioctl = ocfs2_ioctl,
2328 #ifdef CONFIG_COMPAT
2329         .compat_ioctl   = ocfs2_compat_ioctl,
2330 #endif
2331         .flock          = ocfs2_flock,
2332         .splice_read    = ocfs2_file_splice_read,
2333         .splice_write   = ocfs2_file_splice_write,
2334 };
2335
2336 const struct file_operations ocfs2_dops_no_plocks = {
2337         .llseek         = generic_file_llseek,
2338         .read           = generic_read_dir,
2339         .readdir        = ocfs2_readdir,
2340         .fsync          = ocfs2_sync_file,
2341         .release        = ocfs2_dir_release,
2342         .open           = ocfs2_dir_open,
2343         .unlocked_ioctl = ocfs2_ioctl,
2344 #ifdef CONFIG_COMPAT
2345         .compat_ioctl   = ocfs2_compat_ioctl,
2346 #endif
2347         .flock          = ocfs2_flock,
2348 };