ocfs2: CoW a reflinked cluster when it is truncated.
[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         spin_lock(&oi->ip_lock);
111
112         /* Check that the inode hasn't been wiped from disk by another
113          * node. If it hasn't then we're safe as long as we hold the
114          * spin lock until our increment of open count. */
115         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
116                 spin_unlock(&oi->ip_lock);
117
118                 status = -ENOENT;
119                 goto leave;
120         }
121
122         if (mode & O_DIRECT)
123                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
124
125         oi->ip_open_count++;
126         spin_unlock(&oi->ip_lock);
127
128         status = ocfs2_init_file_private(inode, file);
129         if (status) {
130                 /*
131                  * We want to set open count back if we're failing the
132                  * open.
133                  */
134                 spin_lock(&oi->ip_lock);
135                 oi->ip_open_count--;
136                 spin_unlock(&oi->ip_lock);
137         }
138
139 leave:
140         mlog_exit(status);
141         return status;
142 }
143
144 static int ocfs2_file_release(struct inode *inode, struct file *file)
145 {
146         struct ocfs2_inode_info *oi = OCFS2_I(inode);
147
148         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
149                        file->f_path.dentry->d_name.len,
150                        file->f_path.dentry->d_name.name);
151
152         spin_lock(&oi->ip_lock);
153         if (!--oi->ip_open_count)
154                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
155         spin_unlock(&oi->ip_lock);
156
157         ocfs2_free_file_private(inode, file);
158
159         mlog_exit(0);
160
161         return 0;
162 }
163
164 static int ocfs2_dir_open(struct inode *inode, struct file *file)
165 {
166         return ocfs2_init_file_private(inode, file);
167 }
168
169 static int ocfs2_dir_release(struct inode *inode, struct file *file)
170 {
171         ocfs2_free_file_private(inode, file);
172         return 0;
173 }
174
175 static int ocfs2_sync_file(struct file *file,
176                            struct dentry *dentry,
177                            int datasync)
178 {
179         int err = 0;
180         journal_t *journal;
181         struct inode *inode = dentry->d_inode;
182         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
183
184         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
185                    dentry->d_name.len, dentry->d_name.name);
186
187         err = ocfs2_sync_inode(dentry->d_inode);
188         if (err)
189                 goto bail;
190
191         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
192                 goto bail;
193
194         journal = osb->journal->j_journal;
195         err = jbd2_journal_force_commit(journal);
196
197 bail:
198         mlog_exit(err);
199
200         return (err < 0) ? -EIO : 0;
201 }
202
203 int ocfs2_should_update_atime(struct inode *inode,
204                               struct vfsmount *vfsmnt)
205 {
206         struct timespec now;
207         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
208
209         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
210                 return 0;
211
212         if ((inode->i_flags & S_NOATIME) ||
213             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
214                 return 0;
215
216         /*
217          * We can be called with no vfsmnt structure - NFSD will
218          * sometimes do this.
219          *
220          * Note that our action here is different than touch_atime() -
221          * if we can't tell whether this is a noatime mount, then we
222          * don't know whether to trust the value of s_atime_quantum.
223          */
224         if (vfsmnt == NULL)
225                 return 0;
226
227         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
228             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
229                 return 0;
230
231         if (vfsmnt->mnt_flags & MNT_RELATIME) {
232                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
233                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
234                         return 1;
235
236                 return 0;
237         }
238
239         now = CURRENT_TIME;
240         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
241                 return 0;
242         else
243                 return 1;
244 }
245
246 int ocfs2_update_inode_atime(struct inode *inode,
247                              struct buffer_head *bh)
248 {
249         int ret;
250         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
251         handle_t *handle;
252         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
253
254         mlog_entry_void();
255
256         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
257         if (IS_ERR(handle)) {
258                 ret = PTR_ERR(handle);
259                 mlog_errno(ret);
260                 goto out;
261         }
262
263         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
264                                       OCFS2_JOURNAL_ACCESS_WRITE);
265         if (ret) {
266                 mlog_errno(ret);
267                 goto out_commit;
268         }
269
270         /*
271          * Don't use ocfs2_mark_inode_dirty() here as we don't always
272          * have i_mutex to guard against concurrent changes to other
273          * inode fields.
274          */
275         inode->i_atime = CURRENT_TIME;
276         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
277         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
278
279         ret = ocfs2_journal_dirty(handle, bh);
280         if (ret < 0)
281                 mlog_errno(ret);
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         status = ocfs2_journal_dirty(handle, fe_bh);
431         if (status < 0)
432                 mlog_errno(status);
433
434 out_commit:
435         ocfs2_commit_trans(osb, handle);
436 out:
437
438         mlog_exit(status);
439         return status;
440 }
441
442 static int ocfs2_truncate_file(struct inode *inode,
443                                struct buffer_head *di_bh,
444                                u64 new_i_size)
445 {
446         int status = 0;
447         struct ocfs2_dinode *fe = NULL;
448         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
449         struct ocfs2_truncate_context *tc = NULL;
450
451         mlog_entry("(inode = %llu, new_i_size = %llu\n",
452                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
453                    (unsigned long long)new_i_size);
454
455         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
456          * already validated it */
457         fe = (struct ocfs2_dinode *) di_bh->b_data;
458
459         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
460                         "Inode %llu, inode i_size = %lld != di "
461                         "i_size = %llu, i_flags = 0x%x\n",
462                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
463                         i_size_read(inode),
464                         (unsigned long long)le64_to_cpu(fe->i_size),
465                         le32_to_cpu(fe->i_flags));
466
467         if (new_i_size > le64_to_cpu(fe->i_size)) {
468                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
469                      (unsigned long long)le64_to_cpu(fe->i_size),
470                      (unsigned long long)new_i_size);
471                 status = -EINVAL;
472                 mlog_errno(status);
473                 goto bail;
474         }
475
476         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
477              (unsigned long long)le64_to_cpu(fe->i_blkno),
478              (unsigned long long)le64_to_cpu(fe->i_size),
479              (unsigned long long)new_i_size);
480
481         /* lets handle the simple truncate cases before doing any more
482          * cluster locking. */
483         if (new_i_size == le64_to_cpu(fe->i_size))
484                 goto bail;
485
486         down_write(&OCFS2_I(inode)->ip_alloc_sem);
487
488         /*
489          * The inode lock forced other nodes to sync and drop their
490          * pages, which (correctly) happens even if we have a truncate
491          * without allocation change - ocfs2 cluster sizes can be much
492          * greater than page size, so we have to truncate them
493          * anyway.
494          */
495         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
496         truncate_inode_pages(inode->i_mapping, new_i_size);
497
498         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
499                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
500                                                i_size_read(inode), 1);
501                 if (status)
502                         mlog_errno(status);
503
504                 goto bail_unlock_sem;
505         }
506
507         /* alright, we're going to need to do a full blown alloc size
508          * change. Orphan the inode so that recovery can complete the
509          * truncate if necessary. This does the task of marking
510          * i_size. */
511         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
512         if (status < 0) {
513                 mlog_errno(status);
514                 goto bail_unlock_sem;
515         }
516
517         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
518         if (status < 0) {
519                 mlog_errno(status);
520                 goto bail_unlock_sem;
521         }
522
523         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
524         if (status < 0) {
525                 mlog_errno(status);
526                 goto bail_unlock_sem;
527         }
528
529         /* TODO: orphan dir cleanup here. */
530 bail_unlock_sem:
531         up_write(&OCFS2_I(inode)->ip_alloc_sem);
532
533 bail:
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         if (vfs_dq_alloc_space_nodirty(inode, ocfs2_clusters_to_bytes(osb->sb,
631             clusters_to_add))) {
632                 status = -EDQUOT;
633                 goto leave;
634         }
635         did_quota = 1;
636
637         /* reserve a write to the file entry early on - that we if we
638          * run out of credits in the allocation path, we can still
639          * update i_size. */
640         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
641                                          OCFS2_JOURNAL_ACCESS_WRITE);
642         if (status < 0) {
643                 mlog_errno(status);
644                 goto leave;
645         }
646
647         prev_clusters = OCFS2_I(inode)->ip_clusters;
648
649         status = ocfs2_add_inode_data(osb,
650                                       inode,
651                                       &logical_start,
652                                       clusters_to_add,
653                                       mark_unwritten,
654                                       bh,
655                                       handle,
656                                       data_ac,
657                                       meta_ac,
658                                       &why);
659         if ((status < 0) && (status != -EAGAIN)) {
660                 if (status != -ENOSPC)
661                         mlog_errno(status);
662                 goto leave;
663         }
664
665         status = ocfs2_journal_dirty(handle, bh);
666         if (status < 0) {
667                 mlog_errno(status);
668                 goto leave;
669         }
670
671         spin_lock(&OCFS2_I(inode)->ip_lock);
672         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
673         spin_unlock(&OCFS2_I(inode)->ip_lock);
674         /* Release unused quota reservation */
675         vfs_dq_free_space(inode,
676                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
677         did_quota = 0;
678
679         if (why != RESTART_NONE && clusters_to_add) {
680                 if (why == RESTART_META) {
681                         mlog(0, "restarting function.\n");
682                         restart_func = 1;
683                 } else {
684                         BUG_ON(why != RESTART_TRANS);
685
686                         mlog(0, "restarting transaction.\n");
687                         /* TODO: This can be more intelligent. */
688                         credits = ocfs2_calc_extend_credits(osb->sb,
689                                                             &fe->id2.i_list,
690                                                             clusters_to_add);
691                         status = ocfs2_extend_trans(handle, credits);
692                         if (status < 0) {
693                                 /* handle still has to be committed at
694                                  * this point. */
695                                 status = -ENOMEM;
696                                 mlog_errno(status);
697                                 goto leave;
698                         }
699                         goto restarted_transaction;
700                 }
701         }
702
703         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
704              le32_to_cpu(fe->i_clusters),
705              (unsigned long long)le64_to_cpu(fe->i_size));
706         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
707              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
708
709 leave:
710         if (status < 0 && did_quota)
711                 vfs_dq_free_space(inode,
712                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
713         if (handle) {
714                 ocfs2_commit_trans(osb, handle);
715                 handle = NULL;
716         }
717         if (data_ac) {
718                 ocfs2_free_alloc_context(data_ac);
719                 data_ac = NULL;
720         }
721         if (meta_ac) {
722                 ocfs2_free_alloc_context(meta_ac);
723                 meta_ac = NULL;
724         }
725         if ((!status) && restart_func) {
726                 restart_func = 0;
727                 goto restart_all;
728         }
729         brelse(bh);
730         bh = NULL;
731
732         mlog_exit(status);
733         return status;
734 }
735
736 /* Some parts of this taken from generic_cont_expand, which turned out
737  * to be too fragile to do exactly what we need without us having to
738  * worry about recursive locking in ->write_begin() and ->write_end(). */
739 static int ocfs2_write_zero_page(struct inode *inode,
740                                  u64 size)
741 {
742         struct address_space *mapping = inode->i_mapping;
743         struct page *page;
744         unsigned long index;
745         unsigned int offset;
746         handle_t *handle = NULL;
747         int ret;
748
749         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
750         /* ugh.  in prepare/commit_write, if from==to==start of block, we 
751         ** skip the prepare.  make sure we never send an offset for the start
752         ** of a block
753         */
754         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
755                 offset++;
756         }
757         index = size >> PAGE_CACHE_SHIFT;
758
759         page = grab_cache_page(mapping, index);
760         if (!page) {
761                 ret = -ENOMEM;
762                 mlog_errno(ret);
763                 goto out;
764         }
765
766         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
767         if (ret < 0) {
768                 mlog_errno(ret);
769                 goto out_unlock;
770         }
771
772         if (ocfs2_should_order_data(inode)) {
773                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
774                                                      offset);
775                 if (IS_ERR(handle)) {
776                         ret = PTR_ERR(handle);
777                         handle = NULL;
778                         goto out_unlock;
779                 }
780         }
781
782         /* must not update i_size! */
783         ret = block_commit_write(page, offset, offset);
784         if (ret < 0)
785                 mlog_errno(ret);
786         else
787                 ret = 0;
788
789         if (handle)
790                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
791 out_unlock:
792         unlock_page(page);
793         page_cache_release(page);
794 out:
795         return ret;
796 }
797
798 static int ocfs2_zero_extend(struct inode *inode,
799                              u64 zero_to_size)
800 {
801         int ret = 0;
802         u64 start_off;
803         struct super_block *sb = inode->i_sb;
804
805         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
806         while (start_off < zero_to_size) {
807                 ret = ocfs2_write_zero_page(inode, start_off);
808                 if (ret < 0) {
809                         mlog_errno(ret);
810                         goto out;
811                 }
812
813                 start_off += sb->s_blocksize;
814
815                 /*
816                  * Very large extends have the potential to lock up
817                  * the cpu for extended periods of time.
818                  */
819                 cond_resched();
820         }
821
822 out:
823         return ret;
824 }
825
826 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
827 {
828         int ret;
829         u32 clusters_to_add;
830         struct ocfs2_inode_info *oi = OCFS2_I(inode);
831
832         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
833         if (clusters_to_add < oi->ip_clusters)
834                 clusters_to_add = 0;
835         else
836                 clusters_to_add -= oi->ip_clusters;
837
838         if (clusters_to_add) {
839                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
840                                                 clusters_to_add, 0);
841                 if (ret) {
842                         mlog_errno(ret);
843                         goto out;
844                 }
845         }
846
847         /*
848          * Call this even if we don't add any clusters to the tree. We
849          * still need to zero the area between the old i_size and the
850          * new i_size.
851          */
852         ret = ocfs2_zero_extend(inode, zero_to);
853         if (ret < 0)
854                 mlog_errno(ret);
855
856 out:
857         return ret;
858 }
859
860 static int ocfs2_extend_file(struct inode *inode,
861                              struct buffer_head *di_bh,
862                              u64 new_i_size)
863 {
864         int ret = 0;
865         struct ocfs2_inode_info *oi = OCFS2_I(inode);
866
867         BUG_ON(!di_bh);
868
869         /* setattr sometimes calls us like this. */
870         if (new_i_size == 0)
871                 goto out;
872
873         if (i_size_read(inode) == new_i_size)
874                 goto out;
875         BUG_ON(new_i_size < i_size_read(inode));
876
877         /*
878          * Fall through for converting inline data, even if the fs
879          * supports sparse files.
880          *
881          * The check for inline data here is legal - nobody can add
882          * the feature since we have i_mutex. We must check it again
883          * after acquiring ip_alloc_sem though, as paths like mmap
884          * might have raced us to converting the inode to extents.
885          */
886         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
887             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
888                 goto out_update_size;
889
890         /*
891          * The alloc sem blocks people in read/write from reading our
892          * allocation until we're done changing it. We depend on
893          * i_mutex to block other extend/truncate calls while we're
894          * here.
895          */
896         down_write(&oi->ip_alloc_sem);
897
898         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
899                 /*
900                  * We can optimize small extends by keeping the inodes
901                  * inline data.
902                  */
903                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
904                         up_write(&oi->ip_alloc_sem);
905                         goto out_update_size;
906                 }
907
908                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
909                 if (ret) {
910                         up_write(&oi->ip_alloc_sem);
911
912                         mlog_errno(ret);
913                         goto out;
914                 }
915         }
916
917         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
918                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
919
920         up_write(&oi->ip_alloc_sem);
921
922         if (ret < 0) {
923                 mlog_errno(ret);
924                 goto out;
925         }
926
927 out_update_size:
928         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
929         if (ret < 0)
930                 mlog_errno(ret);
931
932 out:
933         return ret;
934 }
935
936 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
937 {
938         int status = 0, size_change;
939         struct inode *inode = dentry->d_inode;
940         struct super_block *sb = inode->i_sb;
941         struct ocfs2_super *osb = OCFS2_SB(sb);
942         struct buffer_head *bh = NULL;
943         handle_t *handle = NULL;
944         int qtype;
945         struct dquot *transfer_from[MAXQUOTAS] = { };
946         struct dquot *transfer_to[MAXQUOTAS] = { };
947
948         mlog_entry("(0x%p, '%.*s')\n", dentry,
949                    dentry->d_name.len, dentry->d_name.name);
950
951         /* ensuring we don't even attempt to truncate a symlink */
952         if (S_ISLNK(inode->i_mode))
953                 attr->ia_valid &= ~ATTR_SIZE;
954
955         if (attr->ia_valid & ATTR_MODE)
956                 mlog(0, "mode change: %d\n", attr->ia_mode);
957         if (attr->ia_valid & ATTR_UID)
958                 mlog(0, "uid change: %d\n", attr->ia_uid);
959         if (attr->ia_valid & ATTR_GID)
960                 mlog(0, "gid change: %d\n", attr->ia_gid);
961         if (attr->ia_valid & ATTR_SIZE)
962                 mlog(0, "size change...\n");
963         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
964                 mlog(0, "time change...\n");
965
966 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
967                            | ATTR_GID | ATTR_UID | ATTR_MODE)
968         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
969                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
970                 return 0;
971         }
972
973         status = inode_change_ok(inode, attr);
974         if (status)
975                 return status;
976
977         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
978         if (size_change) {
979                 status = ocfs2_rw_lock(inode, 1);
980                 if (status < 0) {
981                         mlog_errno(status);
982                         goto bail;
983                 }
984         }
985
986         status = ocfs2_inode_lock(inode, &bh, 1);
987         if (status < 0) {
988                 if (status != -ENOENT)
989                         mlog_errno(status);
990                 goto bail_unlock_rw;
991         }
992
993         if (size_change && attr->ia_size != i_size_read(inode)) {
994                 if (attr->ia_size > sb->s_maxbytes) {
995                         status = -EFBIG;
996                         goto bail_unlock;
997                 }
998
999                 if (i_size_read(inode) > attr->ia_size) {
1000                         if (ocfs2_should_order_data(inode)) {
1001                                 status = ocfs2_begin_ordered_truncate(inode,
1002                                                                       attr->ia_size);
1003                                 if (status)
1004                                         goto bail_unlock;
1005                         }
1006                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1007                 } else
1008                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1009                 if (status < 0) {
1010                         if (status != -ENOSPC)
1011                                 mlog_errno(status);
1012                         status = -ENOSPC;
1013                         goto bail_unlock;
1014                 }
1015         }
1016
1017         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1018             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1019                 /*
1020                  * Gather pointers to quota structures so that allocation /
1021                  * freeing of quota structures happens here and not inside
1022                  * vfs_dq_transfer() where we have problems with lock ordering
1023                  */
1024                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1025                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1026                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1027                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1028                                                       USRQUOTA);
1029                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1030                                                         USRQUOTA);
1031                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1032                                 status = -ESRCH;
1033                                 goto bail_unlock;
1034                         }
1035                 }
1036                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1037                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1038                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1039                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1040                                                       GRPQUOTA);
1041                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1042                                                         GRPQUOTA);
1043                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1044                                 status = -ESRCH;
1045                                 goto bail_unlock;
1046                         }
1047                 }
1048                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1049                                            2 * ocfs2_quota_trans_credits(sb));
1050                 if (IS_ERR(handle)) {
1051                         status = PTR_ERR(handle);
1052                         mlog_errno(status);
1053                         goto bail_unlock;
1054                 }
1055                 status = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
1056                 if (status < 0)
1057                         goto bail_commit;
1058         } else {
1059                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1060                 if (IS_ERR(handle)) {
1061                         status = PTR_ERR(handle);
1062                         mlog_errno(status);
1063                         goto bail_unlock;
1064                 }
1065         }
1066
1067         /*
1068          * This will intentionally not wind up calling vmtruncate(),
1069          * since all the work for a size change has been done above.
1070          * Otherwise, we could get into problems with truncate as
1071          * ip_alloc_sem is used there to protect against i_size
1072          * changes.
1073          */
1074         status = inode_setattr(inode, attr);
1075         if (status < 0) {
1076                 mlog_errno(status);
1077                 goto bail_commit;
1078         }
1079
1080         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1081         if (status < 0)
1082                 mlog_errno(status);
1083
1084 bail_commit:
1085         ocfs2_commit_trans(osb, handle);
1086 bail_unlock:
1087         ocfs2_inode_unlock(inode, 1);
1088 bail_unlock_rw:
1089         if (size_change)
1090                 ocfs2_rw_unlock(inode, 1);
1091 bail:
1092         brelse(bh);
1093
1094         /* Release quota pointers in case we acquired them */
1095         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1096                 dqput(transfer_to[qtype]);
1097                 dqput(transfer_from[qtype]);
1098         }
1099
1100         if (!status && attr->ia_valid & ATTR_MODE) {
1101                 status = ocfs2_acl_chmod(inode);
1102                 if (status < 0)
1103                         mlog_errno(status);
1104         }
1105
1106         mlog_exit(status);
1107         return status;
1108 }
1109
1110 int ocfs2_getattr(struct vfsmount *mnt,
1111                   struct dentry *dentry,
1112                   struct kstat *stat)
1113 {
1114         struct inode *inode = dentry->d_inode;
1115         struct super_block *sb = dentry->d_inode->i_sb;
1116         struct ocfs2_super *osb = sb->s_fs_info;
1117         int err;
1118
1119         mlog_entry_void();
1120
1121         err = ocfs2_inode_revalidate(dentry);
1122         if (err) {
1123                 if (err != -ENOENT)
1124                         mlog_errno(err);
1125                 goto bail;
1126         }
1127
1128         generic_fillattr(inode, stat);
1129
1130         /* We set the blksize from the cluster size for performance */
1131         stat->blksize = osb->s_clustersize;
1132
1133 bail:
1134         mlog_exit(err);
1135
1136         return err;
1137 }
1138
1139 int ocfs2_permission(struct inode *inode, int mask)
1140 {
1141         int ret;
1142
1143         mlog_entry_void();
1144
1145         ret = ocfs2_inode_lock(inode, NULL, 0);
1146         if (ret) {
1147                 if (ret != -ENOENT)
1148                         mlog_errno(ret);
1149                 goto out;
1150         }
1151
1152         ret = generic_permission(inode, mask, ocfs2_check_acl);
1153
1154         ocfs2_inode_unlock(inode, 0);
1155 out:
1156         mlog_exit(ret);
1157         return ret;
1158 }
1159
1160 static int __ocfs2_write_remove_suid(struct inode *inode,
1161                                      struct buffer_head *bh)
1162 {
1163         int ret;
1164         handle_t *handle;
1165         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1166         struct ocfs2_dinode *di;
1167
1168         mlog_entry("(Inode %llu, mode 0%o)\n",
1169                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1170
1171         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1172         if (IS_ERR(handle)) {
1173                 ret = PTR_ERR(handle);
1174                 mlog_errno(ret);
1175                 goto out;
1176         }
1177
1178         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1179                                       OCFS2_JOURNAL_ACCESS_WRITE);
1180         if (ret < 0) {
1181                 mlog_errno(ret);
1182                 goto out_trans;
1183         }
1184
1185         inode->i_mode &= ~S_ISUID;
1186         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1187                 inode->i_mode &= ~S_ISGID;
1188
1189         di = (struct ocfs2_dinode *) bh->b_data;
1190         di->i_mode = cpu_to_le16(inode->i_mode);
1191
1192         ret = ocfs2_journal_dirty(handle, bh);
1193         if (ret < 0)
1194                 mlog_errno(ret);
1195
1196 out_trans:
1197         ocfs2_commit_trans(osb, handle);
1198 out:
1199         mlog_exit(ret);
1200         return ret;
1201 }
1202
1203 /*
1204  * Will look for holes and unwritten extents in the range starting at
1205  * pos for count bytes (inclusive).
1206  */
1207 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1208                                        size_t count)
1209 {
1210         int ret = 0;
1211         unsigned int extent_flags;
1212         u32 cpos, clusters, extent_len, phys_cpos;
1213         struct super_block *sb = inode->i_sb;
1214
1215         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1216         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1217
1218         while (clusters) {
1219                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1220                                          &extent_flags);
1221                 if (ret < 0) {
1222                         mlog_errno(ret);
1223                         goto out;
1224                 }
1225
1226                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1227                         ret = 1;
1228                         break;
1229                 }
1230
1231                 if (extent_len > clusters)
1232                         extent_len = clusters;
1233
1234                 clusters -= extent_len;
1235                 cpos += extent_len;
1236         }
1237 out:
1238         return ret;
1239 }
1240
1241 static int ocfs2_write_remove_suid(struct inode *inode)
1242 {
1243         int ret;
1244         struct buffer_head *bh = NULL;
1245
1246         ret = ocfs2_read_inode_block(inode, &bh);
1247         if (ret < 0) {
1248                 mlog_errno(ret);
1249                 goto out;
1250         }
1251
1252         ret =  __ocfs2_write_remove_suid(inode, bh);
1253 out:
1254         brelse(bh);
1255         return ret;
1256 }
1257
1258 /*
1259  * Allocate enough extents to cover the region starting at byte offset
1260  * start for len bytes. Existing extents are skipped, any extents
1261  * added are marked as "unwritten".
1262  */
1263 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1264                                             u64 start, u64 len)
1265 {
1266         int ret;
1267         u32 cpos, phys_cpos, clusters, alloc_size;
1268         u64 end = start + len;
1269         struct buffer_head *di_bh = NULL;
1270
1271         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1272                 ret = ocfs2_read_inode_block(inode, &di_bh);
1273                 if (ret) {
1274                         mlog_errno(ret);
1275                         goto out;
1276                 }
1277
1278                 /*
1279                  * Nothing to do if the requested reservation range
1280                  * fits within the inode.
1281                  */
1282                 if (ocfs2_size_fits_inline_data(di_bh, end))
1283                         goto out;
1284
1285                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1286                 if (ret) {
1287                         mlog_errno(ret);
1288                         goto out;
1289                 }
1290         }
1291
1292         /*
1293          * We consider both start and len to be inclusive.
1294          */
1295         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1296         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1297         clusters -= cpos;
1298
1299         while (clusters) {
1300                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1301                                          &alloc_size, NULL);
1302                 if (ret) {
1303                         mlog_errno(ret);
1304                         goto out;
1305                 }
1306
1307                 /*
1308                  * Hole or existing extent len can be arbitrary, so
1309                  * cap it to our own allocation request.
1310                  */
1311                 if (alloc_size > clusters)
1312                         alloc_size = clusters;
1313
1314                 if (phys_cpos) {
1315                         /*
1316                          * We already have an allocation at this
1317                          * region so we can safely skip it.
1318                          */
1319                         goto next;
1320                 }
1321
1322                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1323                 if (ret) {
1324                         if (ret != -ENOSPC)
1325                                 mlog_errno(ret);
1326                         goto out;
1327                 }
1328
1329 next:
1330                 cpos += alloc_size;
1331                 clusters -= alloc_size;
1332         }
1333
1334         ret = 0;
1335 out:
1336
1337         brelse(di_bh);
1338         return ret;
1339 }
1340
1341 /*
1342  * Truncate a byte range, avoiding pages within partial clusters. This
1343  * preserves those pages for the zeroing code to write to.
1344  */
1345 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1346                                          u64 byte_len)
1347 {
1348         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1349         loff_t start, end;
1350         struct address_space *mapping = inode->i_mapping;
1351
1352         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1353         end = byte_start + byte_len;
1354         end = end & ~(osb->s_clustersize - 1);
1355
1356         if (start < end) {
1357                 unmap_mapping_range(mapping, start, end - start, 0);
1358                 truncate_inode_pages_range(mapping, start, end - 1);
1359         }
1360 }
1361
1362 static int ocfs2_zero_partial_clusters(struct inode *inode,
1363                                        u64 start, u64 len)
1364 {
1365         int ret = 0;
1366         u64 tmpend, end = start + len;
1367         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1368         unsigned int csize = osb->s_clustersize;
1369         handle_t *handle;
1370
1371         /*
1372          * The "start" and "end" values are NOT necessarily part of
1373          * the range whose allocation is being deleted. Rather, this
1374          * is what the user passed in with the request. We must zero
1375          * partial clusters here. There's no need to worry about
1376          * physical allocation - the zeroing code knows to skip holes.
1377          */
1378         mlog(0, "byte start: %llu, end: %llu\n",
1379              (unsigned long long)start, (unsigned long long)end);
1380
1381         /*
1382          * If both edges are on a cluster boundary then there's no
1383          * zeroing required as the region is part of the allocation to
1384          * be truncated.
1385          */
1386         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1387                 goto out;
1388
1389         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1390         if (IS_ERR(handle)) {
1391                 ret = PTR_ERR(handle);
1392                 mlog_errno(ret);
1393                 goto out;
1394         }
1395
1396         /*
1397          * We want to get the byte offset of the end of the 1st cluster.
1398          */
1399         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1400         if (tmpend > end)
1401                 tmpend = end;
1402
1403         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1404              (unsigned long long)start, (unsigned long long)tmpend);
1405
1406         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1407         if (ret)
1408                 mlog_errno(ret);
1409
1410         if (tmpend < end) {
1411                 /*
1412                  * This may make start and end equal, but the zeroing
1413                  * code will skip any work in that case so there's no
1414                  * need to catch it up here.
1415                  */
1416                 start = end & ~(osb->s_clustersize - 1);
1417
1418                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1419                      (unsigned long long)start, (unsigned long long)end);
1420
1421                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1422                 if (ret)
1423                         mlog_errno(ret);
1424         }
1425
1426         ocfs2_commit_trans(osb, handle);
1427 out:
1428         return ret;
1429 }
1430
1431 static int ocfs2_remove_inode_range(struct inode *inode,
1432                                     struct buffer_head *di_bh, u64 byte_start,
1433                                     u64 byte_len)
1434 {
1435         int ret = 0;
1436         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1437         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1438         struct ocfs2_cached_dealloc_ctxt dealloc;
1439         struct address_space *mapping = inode->i_mapping;
1440         struct ocfs2_extent_tree et;
1441
1442         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1443         ocfs2_init_dealloc_ctxt(&dealloc);
1444
1445         if (byte_len == 0)
1446                 return 0;
1447
1448         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1449                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1450                                             byte_start + byte_len, 0);
1451                 if (ret) {
1452                         mlog_errno(ret);
1453                         goto out;
1454                 }
1455                 /*
1456                  * There's no need to get fancy with the page cache
1457                  * truncate of an inline-data inode. We're talking
1458                  * about less than a page here, which will be cached
1459                  * in the dinode buffer anyway.
1460                  */
1461                 unmap_mapping_range(mapping, 0, 0, 0);
1462                 truncate_inode_pages(mapping, 0);
1463                 goto out;
1464         }
1465
1466         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1467         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1468         if (trunc_len >= trunc_start)
1469                 trunc_len -= trunc_start;
1470         else
1471                 trunc_len = 0;
1472
1473         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1474              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1475              (unsigned long long)byte_start,
1476              (unsigned long long)byte_len, trunc_start, trunc_len);
1477
1478         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1479         if (ret) {
1480                 mlog_errno(ret);
1481                 goto out;
1482         }
1483
1484         cpos = trunc_start;
1485         while (trunc_len) {
1486                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1487                                          &alloc_size, NULL);
1488                 if (ret) {
1489                         mlog_errno(ret);
1490                         goto out;
1491                 }
1492
1493                 if (alloc_size > trunc_len)
1494                         alloc_size = trunc_len;
1495
1496                 /* Only do work for non-holes */
1497                 if (phys_cpos != 0) {
1498                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1499                                                        phys_cpos, alloc_size,
1500                                                        &dealloc);
1501                         if (ret) {
1502                                 mlog_errno(ret);
1503                                 goto out;
1504                         }
1505                 }
1506
1507                 cpos += alloc_size;
1508                 trunc_len -= alloc_size;
1509         }
1510
1511         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1512
1513 out:
1514         ocfs2_schedule_truncate_log_flush(osb, 1);
1515         ocfs2_run_deallocs(osb, &dealloc);
1516
1517         return ret;
1518 }
1519
1520 /*
1521  * Parts of this function taken from xfs_change_file_space()
1522  */
1523 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1524                                      loff_t f_pos, unsigned int cmd,
1525                                      struct ocfs2_space_resv *sr,
1526                                      int change_size)
1527 {
1528         int ret;
1529         s64 llen;
1530         loff_t size;
1531         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1532         struct buffer_head *di_bh = NULL;
1533         handle_t *handle;
1534         unsigned long long max_off = inode->i_sb->s_maxbytes;
1535
1536         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1537                 return -EROFS;
1538
1539         mutex_lock(&inode->i_mutex);
1540
1541         /*
1542          * This prevents concurrent writes on other nodes
1543          */
1544         ret = ocfs2_rw_lock(inode, 1);
1545         if (ret) {
1546                 mlog_errno(ret);
1547                 goto out;
1548         }
1549
1550         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1551         if (ret) {
1552                 mlog_errno(ret);
1553                 goto out_rw_unlock;
1554         }
1555
1556         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1557                 ret = -EPERM;
1558                 goto out_inode_unlock;
1559         }
1560
1561         switch (sr->l_whence) {
1562         case 0: /*SEEK_SET*/
1563                 break;
1564         case 1: /*SEEK_CUR*/
1565                 sr->l_start += f_pos;
1566                 break;
1567         case 2: /*SEEK_END*/
1568                 sr->l_start += i_size_read(inode);
1569                 break;
1570         default:
1571                 ret = -EINVAL;
1572                 goto out_inode_unlock;
1573         }
1574         sr->l_whence = 0;
1575
1576         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1577
1578         if (sr->l_start < 0
1579             || sr->l_start > max_off
1580             || (sr->l_start + llen) < 0
1581             || (sr->l_start + llen) > max_off) {
1582                 ret = -EINVAL;
1583                 goto out_inode_unlock;
1584         }
1585         size = sr->l_start + sr->l_len;
1586
1587         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1588                 if (sr->l_len <= 0) {
1589                         ret = -EINVAL;
1590                         goto out_inode_unlock;
1591                 }
1592         }
1593
1594         if (file && should_remove_suid(file->f_path.dentry)) {
1595                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1596                 if (ret) {
1597                         mlog_errno(ret);
1598                         goto out_inode_unlock;
1599                 }
1600         }
1601
1602         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1603         switch (cmd) {
1604         case OCFS2_IOC_RESVSP:
1605         case OCFS2_IOC_RESVSP64:
1606                 /*
1607                  * This takes unsigned offsets, but the signed ones we
1608                  * pass have been checked against overflow above.
1609                  */
1610                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1611                                                        sr->l_len);
1612                 break;
1613         case OCFS2_IOC_UNRESVSP:
1614         case OCFS2_IOC_UNRESVSP64:
1615                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1616                                                sr->l_len);
1617                 break;
1618         default:
1619                 ret = -EINVAL;
1620         }
1621         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1622         if (ret) {
1623                 mlog_errno(ret);
1624                 goto out_inode_unlock;
1625         }
1626
1627         /*
1628          * We update c/mtime for these changes
1629          */
1630         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1631         if (IS_ERR(handle)) {
1632                 ret = PTR_ERR(handle);
1633                 mlog_errno(ret);
1634                 goto out_inode_unlock;
1635         }
1636
1637         if (change_size && i_size_read(inode) < size)
1638                 i_size_write(inode, size);
1639
1640         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1641         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1642         if (ret < 0)
1643                 mlog_errno(ret);
1644
1645         ocfs2_commit_trans(osb, handle);
1646
1647 out_inode_unlock:
1648         brelse(di_bh);
1649         ocfs2_inode_unlock(inode, 1);
1650 out_rw_unlock:
1651         ocfs2_rw_unlock(inode, 1);
1652
1653 out:
1654         mutex_unlock(&inode->i_mutex);
1655         return ret;
1656 }
1657
1658 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1659                             struct ocfs2_space_resv *sr)
1660 {
1661         struct inode *inode = file->f_path.dentry->d_inode;
1662         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1663
1664         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1665             !ocfs2_writes_unwritten_extents(osb))
1666                 return -ENOTTY;
1667         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1668                  !ocfs2_sparse_alloc(osb))
1669                 return -ENOTTY;
1670
1671         if (!S_ISREG(inode->i_mode))
1672                 return -EINVAL;
1673
1674         if (!(file->f_mode & FMODE_WRITE))
1675                 return -EBADF;
1676
1677         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1678 }
1679
1680 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1681                             loff_t len)
1682 {
1683         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1684         struct ocfs2_space_resv sr;
1685         int change_size = 1;
1686
1687         if (!ocfs2_writes_unwritten_extents(osb))
1688                 return -EOPNOTSUPP;
1689
1690         if (S_ISDIR(inode->i_mode))
1691                 return -ENODEV;
1692
1693         if (mode & FALLOC_FL_KEEP_SIZE)
1694                 change_size = 0;
1695
1696         sr.l_whence = 0;
1697         sr.l_start = (s64)offset;
1698         sr.l_len = (s64)len;
1699
1700         return __ocfs2_change_file_space(NULL, inode, offset,
1701                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1702 }
1703
1704 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1705                                    size_t count)
1706 {
1707         int ret = 0;
1708         unsigned int extent_flags;
1709         u32 cpos, clusters, extent_len, phys_cpos;
1710         struct super_block *sb = inode->i_sb;
1711
1712         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1713             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
1714                 return 0;
1715
1716         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1717         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1718
1719         while (clusters) {
1720                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1721                                          &extent_flags);
1722                 if (ret < 0) {
1723                         mlog_errno(ret);
1724                         goto out;
1725                 }
1726
1727                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1728                         ret = 1;
1729                         break;
1730                 }
1731
1732                 if (extent_len > clusters)
1733                         extent_len = clusters;
1734
1735                 clusters -= extent_len;
1736                 cpos += extent_len;
1737         }
1738 out:
1739         return ret;
1740 }
1741
1742 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1743                                             loff_t pos, size_t count,
1744                                             int *meta_level)
1745 {
1746         int ret;
1747         struct buffer_head *di_bh = NULL;
1748         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1749         u32 clusters =
1750                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1751
1752         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1753         if (ret) {
1754                 mlog_errno(ret);
1755                 goto out;
1756         }
1757
1758         *meta_level = 1;
1759
1760         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1761         if (ret)
1762                 mlog_errno(ret);
1763 out:
1764         brelse(di_bh);
1765         return ret;
1766 }
1767
1768 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1769                                          loff_t *ppos,
1770                                          size_t count,
1771                                          int appending,
1772                                          int *direct_io)
1773 {
1774         int ret = 0, meta_level = 0;
1775         struct inode *inode = dentry->d_inode;
1776         loff_t saved_pos, end;
1777
1778         /* 
1779          * We start with a read level meta lock and only jump to an ex
1780          * if we need to make modifications here.
1781          */
1782         for(;;) {
1783                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1784                 if (ret < 0) {
1785                         meta_level = -1;
1786                         mlog_errno(ret);
1787                         goto out;
1788                 }
1789
1790                 /* Clear suid / sgid if necessary. We do this here
1791                  * instead of later in the write path because
1792                  * remove_suid() calls ->setattr without any hint that
1793                  * we may have already done our cluster locking. Since
1794                  * ocfs2_setattr() *must* take cluster locks to
1795                  * proceeed, this will lead us to recursively lock the
1796                  * inode. There's also the dinode i_size state which
1797                  * can be lost via setattr during extending writes (we
1798                  * set inode->i_size at the end of a write. */
1799                 if (should_remove_suid(dentry)) {
1800                         if (meta_level == 0) {
1801                                 ocfs2_inode_unlock(inode, meta_level);
1802                                 meta_level = 1;
1803                                 continue;
1804                         }
1805
1806                         ret = ocfs2_write_remove_suid(inode);
1807                         if (ret < 0) {
1808                                 mlog_errno(ret);
1809                                 goto out_unlock;
1810                         }
1811                 }
1812
1813                 /* work on a copy of ppos until we're sure that we won't have
1814                  * to recalculate it due to relocking. */
1815                 if (appending) {
1816                         saved_pos = i_size_read(inode);
1817                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1818                 } else {
1819                         saved_pos = *ppos;
1820                 }
1821
1822                 end = saved_pos + count;
1823
1824                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1825                 if (ret == 1) {
1826                         ocfs2_inode_unlock(inode, meta_level);
1827                         meta_level = -1;
1828
1829                         ret = ocfs2_prepare_inode_for_refcount(inode,
1830                                                                saved_pos,
1831                                                                count,
1832                                                                &meta_level);
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;
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);
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_nolock(iocb, iov, nr_segs,
2000                                                         *ppos);
2001         }
2002
2003 out_dio:
2004         /* buffered aio wouldn't have proper lock coverage today */
2005         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2006
2007         if ((file->f_flags & O_SYNC && !direct_io) || IS_SYNC(inode)) {
2008                 /*
2009                  * The generic write paths have handled getting data
2010                  * to disk, but since we don't make use of the dirty
2011                  * inode list, a manual journal commit is necessary
2012                  * here.
2013                  */
2014                 if (old_size != i_size_read(inode) ||
2015                     old_clusters != OCFS2_I(inode)->ip_clusters) {
2016                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2017                         if (ret < 0)
2018                                 written = ret;
2019                 }
2020         }
2021
2022         /* 
2023          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2024          * function pointer which is called when o_direct io completes so that
2025          * it can unlock our rw lock.  (it's the clustered equivalent of
2026          * i_alloc_sem; protects truncate from racing with pending ios).
2027          * Unfortunately there are error cases which call end_io and others
2028          * that don't.  so we don't have to unlock the rw_lock if either an
2029          * async dio is going to do it in the future or an end_io after an
2030          * error has already done it.
2031          */
2032         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2033                 rw_level = -1;
2034                 have_alloc_sem = 0;
2035         }
2036
2037 out:
2038         if (rw_level != -1)
2039                 ocfs2_rw_unlock(inode, rw_level);
2040
2041 out_sems:
2042         if (have_alloc_sem)
2043                 up_read(&inode->i_alloc_sem);
2044
2045         mutex_unlock(&inode->i_mutex);
2046
2047         if (written)
2048                 ret = written;
2049         mlog_exit(ret);
2050         return ret;
2051 }
2052
2053 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2054                                 struct file *out,
2055                                 struct splice_desc *sd)
2056 {
2057         int ret;
2058
2059         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2060                                             sd->total_len, 0, NULL);
2061         if (ret < 0) {
2062                 mlog_errno(ret);
2063                 return ret;
2064         }
2065
2066         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2067 }
2068
2069 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2070                                        struct file *out,
2071                                        loff_t *ppos,
2072                                        size_t len,
2073                                        unsigned int flags)
2074 {
2075         int ret;
2076         struct address_space *mapping = out->f_mapping;
2077         struct inode *inode = mapping->host;
2078         struct splice_desc sd = {
2079                 .total_len = len,
2080                 .flags = flags,
2081                 .pos = *ppos,
2082                 .u.file = out,
2083         };
2084
2085         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2086                    (unsigned int)len,
2087                    out->f_path.dentry->d_name.len,
2088                    out->f_path.dentry->d_name.name);
2089
2090         if (pipe->inode)
2091                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2092
2093         splice_from_pipe_begin(&sd);
2094         do {
2095                 ret = splice_from_pipe_next(pipe, &sd);
2096                 if (ret <= 0)
2097                         break;
2098
2099                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2100                 ret = ocfs2_rw_lock(inode, 1);
2101                 if (ret < 0)
2102                         mlog_errno(ret);
2103                 else {
2104                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2105                         ocfs2_rw_unlock(inode, 1);
2106                 }
2107                 mutex_unlock(&inode->i_mutex);
2108         } while (ret > 0);
2109         splice_from_pipe_end(pipe, &sd);
2110
2111         if (pipe->inode)
2112                 mutex_unlock(&pipe->inode->i_mutex);
2113
2114         if (sd.num_spliced)
2115                 ret = sd.num_spliced;
2116
2117         if (ret > 0) {
2118                 unsigned long nr_pages;
2119
2120                 *ppos += ret;
2121                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2122
2123                 /*
2124                  * If file or inode is SYNC and we actually wrote some data,
2125                  * sync it.
2126                  */
2127                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
2128                         int err;
2129
2130                         mutex_lock(&inode->i_mutex);
2131                         err = ocfs2_rw_lock(inode, 1);
2132                         if (err < 0) {
2133                                 mlog_errno(err);
2134                         } else {
2135                                 err = generic_osync_inode(inode, mapping,
2136                                                   OSYNC_METADATA|OSYNC_DATA);
2137                                 ocfs2_rw_unlock(inode, 1);
2138                         }
2139                         mutex_unlock(&inode->i_mutex);
2140
2141                         if (err)
2142                                 ret = err;
2143                 }
2144                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2145         }
2146
2147         mlog_exit(ret);
2148         return ret;
2149 }
2150
2151 static ssize_t ocfs2_file_splice_read(struct file *in,
2152                                       loff_t *ppos,
2153                                       struct pipe_inode_info *pipe,
2154                                       size_t len,
2155                                       unsigned int flags)
2156 {
2157         int ret = 0, lock_level = 0;
2158         struct inode *inode = in->f_path.dentry->d_inode;
2159
2160         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2161                    (unsigned int)len,
2162                    in->f_path.dentry->d_name.len,
2163                    in->f_path.dentry->d_name.name);
2164
2165         /*
2166          * See the comment in ocfs2_file_aio_read()
2167          */
2168         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2169         if (ret < 0) {
2170                 mlog_errno(ret);
2171                 goto bail;
2172         }
2173         ocfs2_inode_unlock(inode, lock_level);
2174
2175         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2176
2177 bail:
2178         mlog_exit(ret);
2179         return ret;
2180 }
2181
2182 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2183                                    const struct iovec *iov,
2184                                    unsigned long nr_segs,
2185                                    loff_t pos)
2186 {
2187         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2188         struct file *filp = iocb->ki_filp;
2189         struct inode *inode = filp->f_path.dentry->d_inode;
2190
2191         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2192                    (unsigned int)nr_segs,
2193                    filp->f_path.dentry->d_name.len,
2194                    filp->f_path.dentry->d_name.name);
2195
2196         if (!inode) {
2197                 ret = -EINVAL;
2198                 mlog_errno(ret);
2199                 goto bail;
2200         }
2201
2202         /* 
2203          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2204          * need locks to protect pending reads from racing with truncate.
2205          */
2206         if (filp->f_flags & O_DIRECT) {
2207                 down_read(&inode->i_alloc_sem);
2208                 have_alloc_sem = 1;
2209
2210                 ret = ocfs2_rw_lock(inode, 0);
2211                 if (ret < 0) {
2212                         mlog_errno(ret);
2213                         goto bail;
2214                 }
2215                 rw_level = 0;
2216                 /* communicate with ocfs2_dio_end_io */
2217                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2218         }
2219
2220         /*
2221          * We're fine letting folks race truncates and extending
2222          * writes with read across the cluster, just like they can
2223          * locally. Hence no rw_lock during read.
2224          * 
2225          * Take and drop the meta data lock to update inode fields
2226          * like i_size. This allows the checks down below
2227          * generic_file_aio_read() a chance of actually working. 
2228          */
2229         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2230         if (ret < 0) {
2231                 mlog_errno(ret);
2232                 goto bail;
2233         }
2234         ocfs2_inode_unlock(inode, lock_level);
2235
2236         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2237         if (ret == -EINVAL)
2238                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2239
2240         /* buffered aio wouldn't have proper lock coverage today */
2241         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2242
2243         /* see ocfs2_file_aio_write */
2244         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2245                 rw_level = -1;
2246                 have_alloc_sem = 0;
2247         }
2248
2249 bail:
2250         if (have_alloc_sem)
2251                 up_read(&inode->i_alloc_sem);
2252         if (rw_level != -1) 
2253                 ocfs2_rw_unlock(inode, rw_level);
2254         mlog_exit(ret);
2255
2256         return ret;
2257 }
2258
2259 const struct inode_operations ocfs2_file_iops = {
2260         .setattr        = ocfs2_setattr,
2261         .getattr        = ocfs2_getattr,
2262         .permission     = ocfs2_permission,
2263         .setxattr       = generic_setxattr,
2264         .getxattr       = generic_getxattr,
2265         .listxattr      = ocfs2_listxattr,
2266         .removexattr    = generic_removexattr,
2267         .fallocate      = ocfs2_fallocate,
2268         .fiemap         = ocfs2_fiemap,
2269 };
2270
2271 const struct inode_operations ocfs2_special_file_iops = {
2272         .setattr        = ocfs2_setattr,
2273         .getattr        = ocfs2_getattr,
2274         .permission     = ocfs2_permission,
2275 };
2276
2277 /*
2278  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2279  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2280  */
2281 const struct file_operations ocfs2_fops = {
2282         .llseek         = generic_file_llseek,
2283         .read           = do_sync_read,
2284         .write          = do_sync_write,
2285         .mmap           = ocfs2_mmap,
2286         .fsync          = ocfs2_sync_file,
2287         .release        = ocfs2_file_release,
2288         .open           = ocfs2_file_open,
2289         .aio_read       = ocfs2_file_aio_read,
2290         .aio_write      = ocfs2_file_aio_write,
2291         .unlocked_ioctl = ocfs2_ioctl,
2292 #ifdef CONFIG_COMPAT
2293         .compat_ioctl   = ocfs2_compat_ioctl,
2294 #endif
2295         .lock           = ocfs2_lock,
2296         .flock          = ocfs2_flock,
2297         .splice_read    = ocfs2_file_splice_read,
2298         .splice_write   = ocfs2_file_splice_write,
2299 };
2300
2301 const struct file_operations ocfs2_dops = {
2302         .llseek         = generic_file_llseek,
2303         .read           = generic_read_dir,
2304         .readdir        = ocfs2_readdir,
2305         .fsync          = ocfs2_sync_file,
2306         .release        = ocfs2_dir_release,
2307         .open           = ocfs2_dir_open,
2308         .unlocked_ioctl = ocfs2_ioctl,
2309 #ifdef CONFIG_COMPAT
2310         .compat_ioctl   = ocfs2_compat_ioctl,
2311 #endif
2312         .lock           = ocfs2_lock,
2313         .flock          = ocfs2_flock,
2314 };
2315
2316 /*
2317  * POSIX-lockless variants of our file_operations.
2318  *
2319  * These will be used if the underlying cluster stack does not support
2320  * posix file locking, if the user passes the "localflocks" mount
2321  * option, or if we have a local-only fs.
2322  *
2323  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2324  * so we still want it in the case of no stack support for
2325  * plocks. Internally, it will do the right thing when asked to ignore
2326  * the cluster.
2327  */
2328 const struct file_operations ocfs2_fops_no_plocks = {
2329         .llseek         = generic_file_llseek,
2330         .read           = do_sync_read,
2331         .write          = do_sync_write,
2332         .mmap           = ocfs2_mmap,
2333         .fsync          = ocfs2_sync_file,
2334         .release        = ocfs2_file_release,
2335         .open           = ocfs2_file_open,
2336         .aio_read       = ocfs2_file_aio_read,
2337         .aio_write      = ocfs2_file_aio_write,
2338         .unlocked_ioctl = ocfs2_ioctl,
2339 #ifdef CONFIG_COMPAT
2340         .compat_ioctl   = ocfs2_compat_ioctl,
2341 #endif
2342         .flock          = ocfs2_flock,
2343         .splice_read    = ocfs2_file_splice_read,
2344         .splice_write   = ocfs2_file_splice_write,
2345 };
2346
2347 const struct file_operations ocfs2_dops_no_plocks = {
2348         .llseek         = generic_file_llseek,
2349         .read           = generic_read_dir,
2350         .readdir        = ocfs2_readdir,
2351         .fsync          = ocfs2_sync_file,
2352         .release        = ocfs2_dir_release,
2353         .open           = ocfs2_dir_open,
2354         .unlocked_ioctl = ocfs2_ioctl,
2355 #ifdef CONFIG_COMPAT
2356         .compat_ioctl   = ocfs2_compat_ioctl,
2357 #endif
2358         .flock          = ocfs2_flock,
2359 };