19d16f2ef81ee5bc22fb54bb06a1c69472061fb6
[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         ocfs2_resv_discard(&osb->osb_la_resmap,
487                            &OCFS2_I(inode)->ip_la_data_resv);
488
489         /*
490          * The inode lock forced other nodes to sync and drop their
491          * pages, which (correctly) happens even if we have a truncate
492          * without allocation change - ocfs2 cluster sizes can be much
493          * greater than page size, so we have to truncate them
494          * anyway.
495          */
496         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
497         truncate_inode_pages(inode->i_mapping, new_i_size);
498
499         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
500                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
501                                                i_size_read(inode), 1);
502                 if (status)
503                         mlog_errno(status);
504
505                 goto bail_unlock_sem;
506         }
507
508         /* alright, we're going to need to do a full blown alloc size
509          * change. Orphan the inode so that recovery can complete the
510          * truncate if necessary. This does the task of marking
511          * i_size. */
512         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
513         if (status < 0) {
514                 mlog_errno(status);
515                 goto bail_unlock_sem;
516         }
517
518         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
519         if (status < 0) {
520                 mlog_errno(status);
521                 goto bail_unlock_sem;
522         }
523
524         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
525         if (status < 0) {
526                 mlog_errno(status);
527                 goto bail_unlock_sem;
528         }
529
530         /* TODO: orphan dir cleanup here. */
531 bail_unlock_sem:
532         up_write(&OCFS2_I(inode)->ip_alloc_sem);
533
534 bail:
535         if (!status && OCFS2_I(inode)->ip_clusters == 0)
536                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
537
538         mlog_exit(status);
539         return status;
540 }
541
542 /*
543  * extend file allocation only here.
544  * we'll update all the disk stuff, and oip->alloc_size
545  *
546  * expect stuff to be locked, a transaction started and enough data /
547  * metadata reservations in the contexts.
548  *
549  * Will return -EAGAIN, and a reason if a restart is needed.
550  * If passed in, *reason will always be set, even in error.
551  */
552 int ocfs2_add_inode_data(struct ocfs2_super *osb,
553                          struct inode *inode,
554                          u32 *logical_offset,
555                          u32 clusters_to_add,
556                          int mark_unwritten,
557                          struct buffer_head *fe_bh,
558                          handle_t *handle,
559                          struct ocfs2_alloc_context *data_ac,
560                          struct ocfs2_alloc_context *meta_ac,
561                          enum ocfs2_alloc_restarted *reason_ret)
562 {
563         int ret;
564         struct ocfs2_extent_tree et;
565
566         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
567         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
568                                           clusters_to_add, mark_unwritten,
569                                           data_ac, meta_ac, reason_ret);
570
571         return ret;
572 }
573
574 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
575                                      u32 clusters_to_add, int mark_unwritten)
576 {
577         int status = 0;
578         int restart_func = 0;
579         int credits;
580         u32 prev_clusters;
581         struct buffer_head *bh = NULL;
582         struct ocfs2_dinode *fe = NULL;
583         handle_t *handle = NULL;
584         struct ocfs2_alloc_context *data_ac = NULL;
585         struct ocfs2_alloc_context *meta_ac = NULL;
586         enum ocfs2_alloc_restarted why;
587         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
588         struct ocfs2_extent_tree et;
589         int did_quota = 0;
590
591         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
592
593         /*
594          * This function only exists for file systems which don't
595          * support holes.
596          */
597         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
598
599         status = ocfs2_read_inode_block(inode, &bh);
600         if (status < 0) {
601                 mlog_errno(status);
602                 goto leave;
603         }
604         fe = (struct ocfs2_dinode *) bh->b_data;
605
606 restart_all:
607         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
608
609         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
610              "clusters_to_add = %u\n",
611              (unsigned long long)OCFS2_I(inode)->ip_blkno,
612              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
613              clusters_to_add);
614         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
615         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
616                                        &data_ac, &meta_ac);
617         if (status) {
618                 mlog_errno(status);
619                 goto leave;
620         }
621
622         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
623                                             clusters_to_add);
624         handle = ocfs2_start_trans(osb, credits);
625         if (IS_ERR(handle)) {
626                 status = PTR_ERR(handle);
627                 handle = NULL;
628                 mlog_errno(status);
629                 goto leave;
630         }
631
632 restarted_transaction:
633         status = dquot_alloc_space_nodirty(inode,
634                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
635         if (status)
636                 goto leave;
637         did_quota = 1;
638
639         /* reserve a write to the file entry early on - that we if we
640          * run out of credits in the allocation path, we can still
641          * update i_size. */
642         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
643                                          OCFS2_JOURNAL_ACCESS_WRITE);
644         if (status < 0) {
645                 mlog_errno(status);
646                 goto leave;
647         }
648
649         prev_clusters = OCFS2_I(inode)->ip_clusters;
650
651         status = ocfs2_add_inode_data(osb,
652                                       inode,
653                                       &logical_start,
654                                       clusters_to_add,
655                                       mark_unwritten,
656                                       bh,
657                                       handle,
658                                       data_ac,
659                                       meta_ac,
660                                       &why);
661         if ((status < 0) && (status != -EAGAIN)) {
662                 if (status != -ENOSPC)
663                         mlog_errno(status);
664                 goto leave;
665         }
666
667         ocfs2_journal_dirty(handle, bh);
668
669         spin_lock(&OCFS2_I(inode)->ip_lock);
670         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
671         spin_unlock(&OCFS2_I(inode)->ip_lock);
672         /* Release unused quota reservation */
673         dquot_free_space(inode,
674                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
675         did_quota = 0;
676
677         if (why != RESTART_NONE && clusters_to_add) {
678                 if (why == RESTART_META) {
679                         mlog(0, "restarting function.\n");
680                         restart_func = 1;
681                 } else {
682                         BUG_ON(why != RESTART_TRANS);
683
684                         mlog(0, "restarting transaction.\n");
685                         /* TODO: This can be more intelligent. */
686                         credits = ocfs2_calc_extend_credits(osb->sb,
687                                                             &fe->id2.i_list,
688                                                             clusters_to_add);
689                         status = ocfs2_extend_trans(handle, credits);
690                         if (status < 0) {
691                                 /* handle still has to be committed at
692                                  * this point. */
693                                 status = -ENOMEM;
694                                 mlog_errno(status);
695                                 goto leave;
696                         }
697                         goto restarted_transaction;
698                 }
699         }
700
701         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
702              le32_to_cpu(fe->i_clusters),
703              (unsigned long long)le64_to_cpu(fe->i_size));
704         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
705              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
706
707 leave:
708         if (status < 0 && did_quota)
709                 dquot_free_space(inode,
710                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
711         if (handle) {
712                 ocfs2_commit_trans(osb, handle);
713                 handle = NULL;
714         }
715         if (data_ac) {
716                 ocfs2_free_alloc_context(data_ac);
717                 data_ac = NULL;
718         }
719         if (meta_ac) {
720                 ocfs2_free_alloc_context(meta_ac);
721                 meta_ac = NULL;
722         }
723         if ((!status) && restart_func) {
724                 restart_func = 0;
725                 goto restart_all;
726         }
727         brelse(bh);
728         bh = NULL;
729
730         mlog_exit(status);
731         return status;
732 }
733
734 /* Some parts of this taken from generic_cont_expand, which turned out
735  * to be too fragile to do exactly what we need without us having to
736  * worry about recursive locking in ->write_begin() and ->write_end(). */
737 static int ocfs2_write_zero_page(struct inode *inode,
738                                  u64 size)
739 {
740         struct address_space *mapping = inode->i_mapping;
741         struct page *page;
742         unsigned long index;
743         unsigned int offset;
744         handle_t *handle = NULL;
745         int ret;
746
747         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
748         /* ugh.  in prepare/commit_write, if from==to==start of block, we
749         ** skip the prepare.  make sure we never send an offset for the start
750         ** of a block
751         */
752         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
753                 offset++;
754         }
755         index = size >> PAGE_CACHE_SHIFT;
756
757         page = grab_cache_page(mapping, index);
758         if (!page) {
759                 ret = -ENOMEM;
760                 mlog_errno(ret);
761                 goto out;
762         }
763
764         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
765         if (ret < 0) {
766                 mlog_errno(ret);
767                 goto out_unlock;
768         }
769
770         if (ocfs2_should_order_data(inode)) {
771                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
772                                                      offset);
773                 if (IS_ERR(handle)) {
774                         ret = PTR_ERR(handle);
775                         handle = NULL;
776                         goto out_unlock;
777                 }
778         }
779
780         /* must not update i_size! */
781         ret = block_commit_write(page, offset, offset);
782         if (ret < 0)
783                 mlog_errno(ret);
784         else
785                 ret = 0;
786
787         if (handle)
788                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
789 out_unlock:
790         unlock_page(page);
791         page_cache_release(page);
792 out:
793         return ret;
794 }
795
796 static int ocfs2_zero_extend(struct inode *inode,
797                              u64 zero_to_size)
798 {
799         int ret = 0;
800         u64 start_off;
801         struct super_block *sb = inode->i_sb;
802
803         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
804         while (start_off < zero_to_size) {
805                 ret = ocfs2_write_zero_page(inode, start_off);
806                 if (ret < 0) {
807                         mlog_errno(ret);
808                         goto out;
809                 }
810
811                 start_off += sb->s_blocksize;
812
813                 /*
814                  * Very large extends have the potential to lock up
815                  * the cpu for extended periods of time.
816                  */
817                 cond_resched();
818         }
819
820 out:
821         return ret;
822 }
823
824 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
825 {
826         int ret;
827         u32 clusters_to_add;
828         struct ocfs2_inode_info *oi = OCFS2_I(inode);
829
830         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
831         if (clusters_to_add < oi->ip_clusters)
832                 clusters_to_add = 0;
833         else
834                 clusters_to_add -= oi->ip_clusters;
835
836         if (clusters_to_add) {
837                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
838                                                 clusters_to_add, 0);
839                 if (ret) {
840                         mlog_errno(ret);
841                         goto out;
842                 }
843         }
844
845         /*
846          * Call this even if we don't add any clusters to the tree. We
847          * still need to zero the area between the old i_size and the
848          * new i_size.
849          */
850         ret = ocfs2_zero_extend(inode, zero_to);
851         if (ret < 0)
852                 mlog_errno(ret);
853
854 out:
855         return ret;
856 }
857
858 static int ocfs2_extend_file(struct inode *inode,
859                              struct buffer_head *di_bh,
860                              u64 new_i_size)
861 {
862         int ret = 0;
863         struct ocfs2_inode_info *oi = OCFS2_I(inode);
864
865         BUG_ON(!di_bh);
866
867         /* setattr sometimes calls us like this. */
868         if (new_i_size == 0)
869                 goto out;
870
871         if (i_size_read(inode) == new_i_size)
872                 goto out;
873         BUG_ON(new_i_size < i_size_read(inode));
874
875         /*
876          * Fall through for converting inline data, even if the fs
877          * supports sparse files.
878          *
879          * The check for inline data here is legal - nobody can add
880          * the feature since we have i_mutex. We must check it again
881          * after acquiring ip_alloc_sem though, as paths like mmap
882          * might have raced us to converting the inode to extents.
883          */
884         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
885             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
886                 goto out_update_size;
887
888         /*
889          * The alloc sem blocks people in read/write from reading our
890          * allocation until we're done changing it. We depend on
891          * i_mutex to block other extend/truncate calls while we're
892          * here.
893          */
894         down_write(&oi->ip_alloc_sem);
895
896         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
897                 /*
898                  * We can optimize small extends by keeping the inodes
899                  * inline data.
900                  */
901                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
902                         up_write(&oi->ip_alloc_sem);
903                         goto out_update_size;
904                 }
905
906                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
907                 if (ret) {
908                         up_write(&oi->ip_alloc_sem);
909
910                         mlog_errno(ret);
911                         goto out;
912                 }
913         }
914
915         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
916                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
917
918         up_write(&oi->ip_alloc_sem);
919
920         if (ret < 0) {
921                 mlog_errno(ret);
922                 goto out;
923         }
924
925 out_update_size:
926         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
927         if (ret < 0)
928                 mlog_errno(ret);
929
930 out:
931         return ret;
932 }
933
934 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
935 {
936         int status = 0, size_change;
937         struct inode *inode = dentry->d_inode;
938         struct super_block *sb = inode->i_sb;
939         struct ocfs2_super *osb = OCFS2_SB(sb);
940         struct buffer_head *bh = NULL;
941         handle_t *handle = NULL;
942         int qtype;
943         struct dquot *transfer_from[MAXQUOTAS] = { };
944         struct dquot *transfer_to[MAXQUOTAS] = { };
945
946         mlog_entry("(0x%p, '%.*s')\n", dentry,
947                    dentry->d_name.len, dentry->d_name.name);
948
949         /* ensuring we don't even attempt to truncate a symlink */
950         if (S_ISLNK(inode->i_mode))
951                 attr->ia_valid &= ~ATTR_SIZE;
952
953         if (attr->ia_valid & ATTR_MODE)
954                 mlog(0, "mode change: %d\n", attr->ia_mode);
955         if (attr->ia_valid & ATTR_UID)
956                 mlog(0, "uid change: %d\n", attr->ia_uid);
957         if (attr->ia_valid & ATTR_GID)
958                 mlog(0, "gid change: %d\n", attr->ia_gid);
959         if (attr->ia_valid & ATTR_SIZE)
960                 mlog(0, "size change...\n");
961         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
962                 mlog(0, "time change...\n");
963
964 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
965                            | ATTR_GID | ATTR_UID | ATTR_MODE)
966         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
967                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
968                 return 0;
969         }
970
971         status = inode_change_ok(inode, attr);
972         if (status)
973                 return status;
974
975         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
976         if (size_change) {
977                 dquot_initialize(inode);
978
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                 status = inode_newsize_ok(inode, attr->ia_size);
995                 if (status)
996                         goto bail_unlock;
997
998                 if (i_size_read(inode) > attr->ia_size) {
999                         if (ocfs2_should_order_data(inode)) {
1000                                 status = ocfs2_begin_ordered_truncate(inode,
1001                                                                       attr->ia_size);
1002                                 if (status)
1003                                         goto bail_unlock;
1004                         }
1005                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1006                 } else
1007                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1008                 if (status < 0) {
1009                         if (status != -ENOSPC)
1010                                 mlog_errno(status);
1011                         status = -ENOSPC;
1012                         goto bail_unlock;
1013                 }
1014         }
1015
1016         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1017             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1018                 /*
1019                  * Gather pointers to quota structures so that allocation /
1020                  * freeing of quota structures happens here and not inside
1021                  * dquot_transfer() where we have problems with lock ordering
1022                  */
1023                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1024                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1025                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1026                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1027                                                       USRQUOTA);
1028                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1029                                                         USRQUOTA);
1030                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1031                                 status = -ESRCH;
1032                                 goto bail_unlock;
1033                         }
1034                 }
1035                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1036                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1037                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1038                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1039                                                       GRPQUOTA);
1040                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1041                                                         GRPQUOTA);
1042                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1043                                 status = -ESRCH;
1044                                 goto bail_unlock;
1045                         }
1046                 }
1047                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1048                                            2 * ocfs2_quota_trans_credits(sb));
1049                 if (IS_ERR(handle)) {
1050                         status = PTR_ERR(handle);
1051                         mlog_errno(status);
1052                         goto bail_unlock;
1053                 }
1054                 status = dquot_transfer(inode, attr);
1055                 if (status < 0)
1056                         goto bail_commit;
1057         } else {
1058                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1059                 if (IS_ERR(handle)) {
1060                         status = PTR_ERR(handle);
1061                         mlog_errno(status);
1062                         goto bail_unlock;
1063                 }
1064         }
1065
1066         /*
1067          * This will intentionally not wind up calling vmtruncate(),
1068          * since all the work for a size change has been done above.
1069          * Otherwise, we could get into problems with truncate as
1070          * ip_alloc_sem is used there to protect against i_size
1071          * changes.
1072          */
1073         status = inode_setattr(inode, attr);
1074         if (status < 0) {
1075                 mlog_errno(status);
1076                 goto bail_commit;
1077         }
1078
1079         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1080         if (status < 0)
1081                 mlog_errno(status);
1082
1083 bail_commit:
1084         ocfs2_commit_trans(osb, handle);
1085 bail_unlock:
1086         ocfs2_inode_unlock(inode, 1);
1087 bail_unlock_rw:
1088         if (size_change)
1089                 ocfs2_rw_unlock(inode, 1);
1090 bail:
1091         brelse(bh);
1092
1093         /* Release quota pointers in case we acquired them */
1094         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1095                 dqput(transfer_to[qtype]);
1096                 dqput(transfer_from[qtype]);
1097         }
1098
1099         if (!status && attr->ia_valid & ATTR_MODE) {
1100                 status = ocfs2_acl_chmod(inode);
1101                 if (status < 0)
1102                         mlog_errno(status);
1103         }
1104
1105         mlog_exit(status);
1106         return status;
1107 }
1108
1109 int ocfs2_getattr(struct vfsmount *mnt,
1110                   struct dentry *dentry,
1111                   struct kstat *stat)
1112 {
1113         struct inode *inode = dentry->d_inode;
1114         struct super_block *sb = dentry->d_inode->i_sb;
1115         struct ocfs2_super *osb = sb->s_fs_info;
1116         int err;
1117
1118         mlog_entry_void();
1119
1120         err = ocfs2_inode_revalidate(dentry);
1121         if (err) {
1122                 if (err != -ENOENT)
1123                         mlog_errno(err);
1124                 goto bail;
1125         }
1126
1127         generic_fillattr(inode, stat);
1128
1129         /* We set the blksize from the cluster size for performance */
1130         stat->blksize = osb->s_clustersize;
1131
1132 bail:
1133         mlog_exit(err);
1134
1135         return err;
1136 }
1137
1138 int ocfs2_permission(struct inode *inode, int mask)
1139 {
1140         int ret;
1141
1142         mlog_entry_void();
1143
1144         ret = ocfs2_inode_lock(inode, NULL, 0);
1145         if (ret) {
1146                 if (ret != -ENOENT)
1147                         mlog_errno(ret);
1148                 goto out;
1149         }
1150
1151         ret = generic_permission(inode, mask, ocfs2_check_acl);
1152
1153         ocfs2_inode_unlock(inode, 0);
1154 out:
1155         mlog_exit(ret);
1156         return ret;
1157 }
1158
1159 static int __ocfs2_write_remove_suid(struct inode *inode,
1160                                      struct buffer_head *bh)
1161 {
1162         int ret;
1163         handle_t *handle;
1164         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1165         struct ocfs2_dinode *di;
1166
1167         mlog_entry("(Inode %llu, mode 0%o)\n",
1168                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1169
1170         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1171         if (IS_ERR(handle)) {
1172                 ret = PTR_ERR(handle);
1173                 mlog_errno(ret);
1174                 goto out;
1175         }
1176
1177         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1178                                       OCFS2_JOURNAL_ACCESS_WRITE);
1179         if (ret < 0) {
1180                 mlog_errno(ret);
1181                 goto out_trans;
1182         }
1183
1184         inode->i_mode &= ~S_ISUID;
1185         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1186                 inode->i_mode &= ~S_ISGID;
1187
1188         di = (struct ocfs2_dinode *) bh->b_data;
1189         di->i_mode = cpu_to_le16(inode->i_mode);
1190
1191         ocfs2_journal_dirty(handle, bh);
1192
1193 out_trans:
1194         ocfs2_commit_trans(osb, handle);
1195 out:
1196         mlog_exit(ret);
1197         return ret;
1198 }
1199
1200 /*
1201  * Will look for holes and unwritten extents in the range starting at
1202  * pos for count bytes (inclusive).
1203  */
1204 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1205                                        size_t count)
1206 {
1207         int ret = 0;
1208         unsigned int extent_flags;
1209         u32 cpos, clusters, extent_len, phys_cpos;
1210         struct super_block *sb = inode->i_sb;
1211
1212         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1213         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1214
1215         while (clusters) {
1216                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1217                                          &extent_flags);
1218                 if (ret < 0) {
1219                         mlog_errno(ret);
1220                         goto out;
1221                 }
1222
1223                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1224                         ret = 1;
1225                         break;
1226                 }
1227
1228                 if (extent_len > clusters)
1229                         extent_len = clusters;
1230
1231                 clusters -= extent_len;
1232                 cpos += extent_len;
1233         }
1234 out:
1235         return ret;
1236 }
1237
1238 static int ocfs2_write_remove_suid(struct inode *inode)
1239 {
1240         int ret;
1241         struct buffer_head *bh = NULL;
1242
1243         ret = ocfs2_read_inode_block(inode, &bh);
1244         if (ret < 0) {
1245                 mlog_errno(ret);
1246                 goto out;
1247         }
1248
1249         ret =  __ocfs2_write_remove_suid(inode, bh);
1250 out:
1251         brelse(bh);
1252         return ret;
1253 }
1254
1255 /*
1256  * Allocate enough extents to cover the region starting at byte offset
1257  * start for len bytes. Existing extents are skipped, any extents
1258  * added are marked as "unwritten".
1259  */
1260 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1261                                             u64 start, u64 len)
1262 {
1263         int ret;
1264         u32 cpos, phys_cpos, clusters, alloc_size;
1265         u64 end = start + len;
1266         struct buffer_head *di_bh = NULL;
1267
1268         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1269                 ret = ocfs2_read_inode_block(inode, &di_bh);
1270                 if (ret) {
1271                         mlog_errno(ret);
1272                         goto out;
1273                 }
1274
1275                 /*
1276                  * Nothing to do if the requested reservation range
1277                  * fits within the inode.
1278                  */
1279                 if (ocfs2_size_fits_inline_data(di_bh, end))
1280                         goto out;
1281
1282                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1283                 if (ret) {
1284                         mlog_errno(ret);
1285                         goto out;
1286                 }
1287         }
1288
1289         /*
1290          * We consider both start and len to be inclusive.
1291          */
1292         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1293         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1294         clusters -= cpos;
1295
1296         while (clusters) {
1297                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1298                                          &alloc_size, NULL);
1299                 if (ret) {
1300                         mlog_errno(ret);
1301                         goto out;
1302                 }
1303
1304                 /*
1305                  * Hole or existing extent len can be arbitrary, so
1306                  * cap it to our own allocation request.
1307                  */
1308                 if (alloc_size > clusters)
1309                         alloc_size = clusters;
1310
1311                 if (phys_cpos) {
1312                         /*
1313                          * We already have an allocation at this
1314                          * region so we can safely skip it.
1315                          */
1316                         goto next;
1317                 }
1318
1319                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1320                 if (ret) {
1321                         if (ret != -ENOSPC)
1322                                 mlog_errno(ret);
1323                         goto out;
1324                 }
1325
1326 next:
1327                 cpos += alloc_size;
1328                 clusters -= alloc_size;
1329         }
1330
1331         ret = 0;
1332 out:
1333
1334         brelse(di_bh);
1335         return ret;
1336 }
1337
1338 /*
1339  * Truncate a byte range, avoiding pages within partial clusters. This
1340  * preserves those pages for the zeroing code to write to.
1341  */
1342 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1343                                          u64 byte_len)
1344 {
1345         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1346         loff_t start, end;
1347         struct address_space *mapping = inode->i_mapping;
1348
1349         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1350         end = byte_start + byte_len;
1351         end = end & ~(osb->s_clustersize - 1);
1352
1353         if (start < end) {
1354                 unmap_mapping_range(mapping, start, end - start, 0);
1355                 truncate_inode_pages_range(mapping, start, end - 1);
1356         }
1357 }
1358
1359 static int ocfs2_zero_partial_clusters(struct inode *inode,
1360                                        u64 start, u64 len)
1361 {
1362         int ret = 0;
1363         u64 tmpend, end = start + len;
1364         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1365         unsigned int csize = osb->s_clustersize;
1366         handle_t *handle;
1367
1368         /*
1369          * The "start" and "end" values are NOT necessarily part of
1370          * the range whose allocation is being deleted. Rather, this
1371          * is what the user passed in with the request. We must zero
1372          * partial clusters here. There's no need to worry about
1373          * physical allocation - the zeroing code knows to skip holes.
1374          */
1375         mlog(0, "byte start: %llu, end: %llu\n",
1376              (unsigned long long)start, (unsigned long long)end);
1377
1378         /*
1379          * If both edges are on a cluster boundary then there's no
1380          * zeroing required as the region is part of the allocation to
1381          * be truncated.
1382          */
1383         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1384                 goto out;
1385
1386         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1387         if (IS_ERR(handle)) {
1388                 ret = PTR_ERR(handle);
1389                 mlog_errno(ret);
1390                 goto out;
1391         }
1392
1393         /*
1394          * We want to get the byte offset of the end of the 1st cluster.
1395          */
1396         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1397         if (tmpend > end)
1398                 tmpend = end;
1399
1400         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1401              (unsigned long long)start, (unsigned long long)tmpend);
1402
1403         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1404         if (ret)
1405                 mlog_errno(ret);
1406
1407         if (tmpend < end) {
1408                 /*
1409                  * This may make start and end equal, but the zeroing
1410                  * code will skip any work in that case so there's no
1411                  * need to catch it up here.
1412                  */
1413                 start = end & ~(osb->s_clustersize - 1);
1414
1415                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1416                      (unsigned long long)start, (unsigned long long)end);
1417
1418                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1419                 if (ret)
1420                         mlog_errno(ret);
1421         }
1422
1423         ocfs2_commit_trans(osb, handle);
1424 out:
1425         return ret;
1426 }
1427
1428 static int ocfs2_remove_inode_range(struct inode *inode,
1429                                     struct buffer_head *di_bh, u64 byte_start,
1430                                     u64 byte_len)
1431 {
1432         int ret = 0;
1433         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1434         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1435         struct ocfs2_cached_dealloc_ctxt dealloc;
1436         struct address_space *mapping = inode->i_mapping;
1437         struct ocfs2_extent_tree et;
1438
1439         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1440         ocfs2_init_dealloc_ctxt(&dealloc);
1441
1442         if (byte_len == 0)
1443                 return 0;
1444
1445         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1446                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1447                                             byte_start + byte_len, 0);
1448                 if (ret) {
1449                         mlog_errno(ret);
1450                         goto out;
1451                 }
1452                 /*
1453                  * There's no need to get fancy with the page cache
1454                  * truncate of an inline-data inode. We're talking
1455                  * about less than a page here, which will be cached
1456                  * in the dinode buffer anyway.
1457                  */
1458                 unmap_mapping_range(mapping, 0, 0, 0);
1459                 truncate_inode_pages(mapping, 0);
1460                 goto out;
1461         }
1462
1463         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1464         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1465         if (trunc_len >= trunc_start)
1466                 trunc_len -= trunc_start;
1467         else
1468                 trunc_len = 0;
1469
1470         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1471              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1472              (unsigned long long)byte_start,
1473              (unsigned long long)byte_len, trunc_start, trunc_len);
1474
1475         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1476         if (ret) {
1477                 mlog_errno(ret);
1478                 goto out;
1479         }
1480
1481         cpos = trunc_start;
1482         while (trunc_len) {
1483                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1484                                          &alloc_size, NULL);
1485                 if (ret) {
1486                         mlog_errno(ret);
1487                         goto out;
1488                 }
1489
1490                 if (alloc_size > trunc_len)
1491                         alloc_size = trunc_len;
1492
1493                 /* Only do work for non-holes */
1494                 if (phys_cpos != 0) {
1495                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1496                                                        phys_cpos, alloc_size,
1497                                                        &dealloc);
1498                         if (ret) {
1499                                 mlog_errno(ret);
1500                                 goto out;
1501                         }
1502                 }
1503
1504                 cpos += alloc_size;
1505                 trunc_len -= alloc_size;
1506         }
1507
1508         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1509
1510 out:
1511         ocfs2_schedule_truncate_log_flush(osb, 1);
1512         ocfs2_run_deallocs(osb, &dealloc);
1513
1514         return ret;
1515 }
1516
1517 /*
1518  * Parts of this function taken from xfs_change_file_space()
1519  */
1520 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1521                                      loff_t f_pos, unsigned int cmd,
1522                                      struct ocfs2_space_resv *sr,
1523                                      int change_size)
1524 {
1525         int ret;
1526         s64 llen;
1527         loff_t size;
1528         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1529         struct buffer_head *di_bh = NULL;
1530         handle_t *handle;
1531         unsigned long long max_off = inode->i_sb->s_maxbytes;
1532
1533         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1534                 return -EROFS;
1535
1536         mutex_lock(&inode->i_mutex);
1537
1538         /*
1539          * This prevents concurrent writes on other nodes
1540          */
1541         ret = ocfs2_rw_lock(inode, 1);
1542         if (ret) {
1543                 mlog_errno(ret);
1544                 goto out;
1545         }
1546
1547         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1548         if (ret) {
1549                 mlog_errno(ret);
1550                 goto out_rw_unlock;
1551         }
1552
1553         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1554                 ret = -EPERM;
1555                 goto out_inode_unlock;
1556         }
1557
1558         switch (sr->l_whence) {
1559         case 0: /*SEEK_SET*/
1560                 break;
1561         case 1: /*SEEK_CUR*/
1562                 sr->l_start += f_pos;
1563                 break;
1564         case 2: /*SEEK_END*/
1565                 sr->l_start += i_size_read(inode);
1566                 break;
1567         default:
1568                 ret = -EINVAL;
1569                 goto out_inode_unlock;
1570         }
1571         sr->l_whence = 0;
1572
1573         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1574
1575         if (sr->l_start < 0
1576             || sr->l_start > max_off
1577             || (sr->l_start + llen) < 0
1578             || (sr->l_start + llen) > max_off) {
1579                 ret = -EINVAL;
1580                 goto out_inode_unlock;
1581         }
1582         size = sr->l_start + sr->l_len;
1583
1584         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1585                 if (sr->l_len <= 0) {
1586                         ret = -EINVAL;
1587                         goto out_inode_unlock;
1588                 }
1589         }
1590
1591         if (file && should_remove_suid(file->f_path.dentry)) {
1592                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1593                 if (ret) {
1594                         mlog_errno(ret);
1595                         goto out_inode_unlock;
1596                 }
1597         }
1598
1599         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1600         switch (cmd) {
1601         case OCFS2_IOC_RESVSP:
1602         case OCFS2_IOC_RESVSP64:
1603                 /*
1604                  * This takes unsigned offsets, but the signed ones we
1605                  * pass have been checked against overflow above.
1606                  */
1607                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1608                                                        sr->l_len);
1609                 break;
1610         case OCFS2_IOC_UNRESVSP:
1611         case OCFS2_IOC_UNRESVSP64:
1612                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1613                                                sr->l_len);
1614                 break;
1615         default:
1616                 ret = -EINVAL;
1617         }
1618         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1619         if (ret) {
1620                 mlog_errno(ret);
1621                 goto out_inode_unlock;
1622         }
1623
1624         /*
1625          * We update c/mtime for these changes
1626          */
1627         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1628         if (IS_ERR(handle)) {
1629                 ret = PTR_ERR(handle);
1630                 mlog_errno(ret);
1631                 goto out_inode_unlock;
1632         }
1633
1634         if (change_size && i_size_read(inode) < size)
1635                 i_size_write(inode, size);
1636
1637         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1638         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1639         if (ret < 0)
1640                 mlog_errno(ret);
1641
1642         ocfs2_commit_trans(osb, handle);
1643
1644 out_inode_unlock:
1645         brelse(di_bh);
1646         ocfs2_inode_unlock(inode, 1);
1647 out_rw_unlock:
1648         ocfs2_rw_unlock(inode, 1);
1649
1650 out:
1651         mutex_unlock(&inode->i_mutex);
1652         return ret;
1653 }
1654
1655 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1656                             struct ocfs2_space_resv *sr)
1657 {
1658         struct inode *inode = file->f_path.dentry->d_inode;
1659         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1660
1661         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1662             !ocfs2_writes_unwritten_extents(osb))
1663                 return -ENOTTY;
1664         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1665                  !ocfs2_sparse_alloc(osb))
1666                 return -ENOTTY;
1667
1668         if (!S_ISREG(inode->i_mode))
1669                 return -EINVAL;
1670
1671         if (!(file->f_mode & FMODE_WRITE))
1672                 return -EBADF;
1673
1674         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1675 }
1676
1677 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1678                             loff_t len)
1679 {
1680         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1681         struct ocfs2_space_resv sr;
1682         int change_size = 1;
1683
1684         if (!ocfs2_writes_unwritten_extents(osb))
1685                 return -EOPNOTSUPP;
1686
1687         if (S_ISDIR(inode->i_mode))
1688                 return -ENODEV;
1689
1690         if (mode & FALLOC_FL_KEEP_SIZE)
1691                 change_size = 0;
1692
1693         sr.l_whence = 0;
1694         sr.l_start = (s64)offset;
1695         sr.l_len = (s64)len;
1696
1697         return __ocfs2_change_file_space(NULL, inode, offset,
1698                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1699 }
1700
1701 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1702                                    size_t count)
1703 {
1704         int ret = 0;
1705         unsigned int extent_flags;
1706         u32 cpos, clusters, extent_len, phys_cpos;
1707         struct super_block *sb = inode->i_sb;
1708
1709         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1710             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1711             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1712                 return 0;
1713
1714         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1715         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1716
1717         while (clusters) {
1718                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1719                                          &extent_flags);
1720                 if (ret < 0) {
1721                         mlog_errno(ret);
1722                         goto out;
1723                 }
1724
1725                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1726                         ret = 1;
1727                         break;
1728                 }
1729
1730                 if (extent_len > clusters)
1731                         extent_len = clusters;
1732
1733                 clusters -= extent_len;
1734                 cpos += extent_len;
1735         }
1736 out:
1737         return ret;
1738 }
1739
1740 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1741                                             loff_t pos, size_t count,
1742                                             int *meta_level)
1743 {
1744         int ret;
1745         struct buffer_head *di_bh = NULL;
1746         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1747         u32 clusters =
1748                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1749
1750         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1751         if (ret) {
1752                 mlog_errno(ret);
1753                 goto out;
1754         }
1755
1756         *meta_level = 1;
1757
1758         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1759         if (ret)
1760                 mlog_errno(ret);
1761 out:
1762         brelse(di_bh);
1763         return ret;
1764 }
1765
1766 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1767                                          loff_t *ppos,
1768                                          size_t count,
1769                                          int appending,
1770                                          int *direct_io,
1771                                          int *has_refcount)
1772 {
1773         int ret = 0, meta_level = 0;
1774         struct inode *inode = dentry->d_inode;
1775         loff_t saved_pos, end;
1776
1777         /*
1778          * We start with a read level meta lock and only jump to an ex
1779          * if we need to make modifications here.
1780          */
1781         for(;;) {
1782                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1783                 if (ret < 0) {
1784                         meta_level = -1;
1785                         mlog_errno(ret);
1786                         goto out;
1787                 }
1788
1789                 /* Clear suid / sgid if necessary. We do this here
1790                  * instead of later in the write path because
1791                  * remove_suid() calls ->setattr without any hint that
1792                  * we may have already done our cluster locking. Since
1793                  * ocfs2_setattr() *must* take cluster locks to
1794                  * proceeed, this will lead us to recursively lock the
1795                  * inode. There's also the dinode i_size state which
1796                  * can be lost via setattr during extending writes (we
1797                  * set inode->i_size at the end of a write. */
1798                 if (should_remove_suid(dentry)) {
1799                         if (meta_level == 0) {
1800                                 ocfs2_inode_unlock(inode, meta_level);
1801                                 meta_level = 1;
1802                                 continue;
1803                         }
1804
1805                         ret = ocfs2_write_remove_suid(inode);
1806                         if (ret < 0) {
1807                                 mlog_errno(ret);
1808                                 goto out_unlock;
1809                         }
1810                 }
1811
1812                 /* work on a copy of ppos until we're sure that we won't have
1813                  * to recalculate it due to relocking. */
1814                 if (appending) {
1815                         saved_pos = i_size_read(inode);
1816                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1817                 } else {
1818                         saved_pos = *ppos;
1819                 }
1820
1821                 end = saved_pos + count;
1822
1823                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1824                 if (ret == 1) {
1825                         ocfs2_inode_unlock(inode, meta_level);
1826                         meta_level = -1;
1827
1828                         ret = ocfs2_prepare_inode_for_refcount(inode,
1829                                                                saved_pos,
1830                                                                count,
1831                                                                &meta_level);
1832                         if (has_refcount)
1833                                 *has_refcount = 1;
1834                         if (direct_io)
1835                                 *direct_io = 0;
1836                 }
1837
1838                 if (ret < 0) {
1839                         mlog_errno(ret);
1840                         goto out_unlock;
1841                 }
1842
1843                 /*
1844                  * Skip the O_DIRECT checks if we don't need
1845                  * them.
1846                  */
1847                 if (!direct_io || !(*direct_io))
1848                         break;
1849
1850                 /*
1851                  * There's no sane way to do direct writes to an inode
1852                  * with inline data.
1853                  */
1854                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1855                         *direct_io = 0;
1856                         break;
1857                 }
1858
1859                 /*
1860                  * Allowing concurrent direct writes means
1861                  * i_size changes wouldn't be synchronized, so
1862                  * one node could wind up truncating another
1863                  * nodes writes.
1864                  */
1865                 if (end > i_size_read(inode)) {
1866                         *direct_io = 0;
1867                         break;
1868                 }
1869
1870                 /*
1871                  * We don't fill holes during direct io, so
1872                  * check for them here. If any are found, the
1873                  * caller will have to retake some cluster
1874                  * locks and initiate the io as buffered.
1875                  */
1876                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1877                 if (ret == 1) {
1878                         *direct_io = 0;
1879                         ret = 0;
1880                 } else if (ret < 0)
1881                         mlog_errno(ret);
1882                 break;
1883         }
1884
1885         if (appending)
1886                 *ppos = saved_pos;
1887
1888 out_unlock:
1889         if (meta_level >= 0)
1890                 ocfs2_inode_unlock(inode, meta_level);
1891
1892 out:
1893         return ret;
1894 }
1895
1896 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1897                                     const struct iovec *iov,
1898                                     unsigned long nr_segs,
1899                                     loff_t pos)
1900 {
1901         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1902         int can_do_direct, has_refcount = 0;
1903         ssize_t written = 0;
1904         size_t ocount;          /* original count */
1905         size_t count;           /* after file limit checks */
1906         loff_t old_size, *ppos = &iocb->ki_pos;
1907         u32 old_clusters;
1908         struct file *file = iocb->ki_filp;
1909         struct inode *inode = file->f_path.dentry->d_inode;
1910         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1911
1912         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1913                    (unsigned int)nr_segs,
1914                    file->f_path.dentry->d_name.len,
1915                    file->f_path.dentry->d_name.name);
1916
1917         if (iocb->ki_left == 0)
1918                 return 0;
1919
1920         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1921
1922         appending = file->f_flags & O_APPEND ? 1 : 0;
1923         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1924
1925         mutex_lock(&inode->i_mutex);
1926
1927 relock:
1928         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1929         if (direct_io) {
1930                 down_read(&inode->i_alloc_sem);
1931                 have_alloc_sem = 1;
1932         }
1933
1934         /* concurrent O_DIRECT writes are allowed */
1935         rw_level = !direct_io;
1936         ret = ocfs2_rw_lock(inode, rw_level);
1937         if (ret < 0) {
1938                 mlog_errno(ret);
1939                 goto out_sems;
1940         }
1941
1942         can_do_direct = direct_io;
1943         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1944                                             iocb->ki_left, appending,
1945                                             &can_do_direct, &has_refcount);
1946         if (ret < 0) {
1947                 mlog_errno(ret);
1948                 goto out;
1949         }
1950
1951         /*
1952          * We can't complete the direct I/O as requested, fall back to
1953          * buffered I/O.
1954          */
1955         if (direct_io && !can_do_direct) {
1956                 ocfs2_rw_unlock(inode, rw_level);
1957                 up_read(&inode->i_alloc_sem);
1958
1959                 have_alloc_sem = 0;
1960                 rw_level = -1;
1961
1962                 direct_io = 0;
1963                 goto relock;
1964         }
1965
1966         /*
1967          * To later detect whether a journal commit for sync writes is
1968          * necessary, we sample i_size, and cluster count here.
1969          */
1970         old_size = i_size_read(inode);
1971         old_clusters = OCFS2_I(inode)->ip_clusters;
1972
1973         /* communicate with ocfs2_dio_end_io */
1974         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1975
1976         if (direct_io) {
1977                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1978                                              VERIFY_READ);
1979                 if (ret)
1980                         goto out_dio;
1981
1982                 count = ocount;
1983                 ret = generic_write_checks(file, ppos, &count,
1984                                            S_ISBLK(inode->i_mode));
1985                 if (ret)
1986                         goto out_dio;
1987
1988                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1989                                                     ppos, count, ocount);
1990                 if (written < 0) {
1991                         /*
1992                          * direct write may have instantiated a few
1993                          * blocks outside i_size. Trim these off again.
1994                          * Don't need i_size_read because we hold i_mutex.
1995                          */
1996                         if (*ppos + count > inode->i_size)
1997                                 vmtruncate(inode, inode->i_size);
1998                         ret = written;
1999                         goto out_dio;
2000                 }
2001         } else {
2002                 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2003         }
2004
2005 out_dio:
2006         /* buffered aio wouldn't have proper lock coverage today */
2007         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2008
2009         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2010             ((file->f_flags & O_DIRECT) && has_refcount)) {
2011                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2012                                                pos + count - 1);
2013                 if (ret < 0)
2014                         written = ret;
2015
2016                 if (!ret && (old_size != i_size_read(inode) ||
2017                     old_clusters != OCFS2_I(inode)->ip_clusters ||
2018                     has_refcount)) {
2019                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2020                         if (ret < 0)
2021                                 written = ret;
2022                 }
2023
2024                 if (!ret)
2025                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2026                                                       pos + count - 1);
2027         }
2028
2029         /*
2030          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2031          * function pointer which is called when o_direct io completes so that
2032          * it can unlock our rw lock.  (it's the clustered equivalent of
2033          * i_alloc_sem; protects truncate from racing with pending ios).
2034          * Unfortunately there are error cases which call end_io and others
2035          * that don't.  so we don't have to unlock the rw_lock if either an
2036          * async dio is going to do it in the future or an end_io after an
2037          * error has already done it.
2038          */
2039         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2040                 rw_level = -1;
2041                 have_alloc_sem = 0;
2042         }
2043
2044 out:
2045         if (rw_level != -1)
2046                 ocfs2_rw_unlock(inode, rw_level);
2047
2048 out_sems:
2049         if (have_alloc_sem)
2050                 up_read(&inode->i_alloc_sem);
2051
2052         mutex_unlock(&inode->i_mutex);
2053
2054         if (written)
2055                 ret = written;
2056         mlog_exit(ret);
2057         return ret;
2058 }
2059
2060 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2061                                 struct file *out,
2062                                 struct splice_desc *sd)
2063 {
2064         int ret;
2065
2066         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2067                                             sd->total_len, 0, NULL, NULL);
2068         if (ret < 0) {
2069                 mlog_errno(ret);
2070                 return ret;
2071         }
2072
2073         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2074 }
2075
2076 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2077                                        struct file *out,
2078                                        loff_t *ppos,
2079                                        size_t len,
2080                                        unsigned int flags)
2081 {
2082         int ret;
2083         struct address_space *mapping = out->f_mapping;
2084         struct inode *inode = mapping->host;
2085         struct splice_desc sd = {
2086                 .total_len = len,
2087                 .flags = flags,
2088                 .pos = *ppos,
2089                 .u.file = out,
2090         };
2091
2092         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2093                    (unsigned int)len,
2094                    out->f_path.dentry->d_name.len,
2095                    out->f_path.dentry->d_name.name);
2096
2097         if (pipe->inode)
2098                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2099
2100         splice_from_pipe_begin(&sd);
2101         do {
2102                 ret = splice_from_pipe_next(pipe, &sd);
2103                 if (ret <= 0)
2104                         break;
2105
2106                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2107                 ret = ocfs2_rw_lock(inode, 1);
2108                 if (ret < 0)
2109                         mlog_errno(ret);
2110                 else {
2111                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2112                         ocfs2_rw_unlock(inode, 1);
2113                 }
2114                 mutex_unlock(&inode->i_mutex);
2115         } while (ret > 0);
2116         splice_from_pipe_end(pipe, &sd);
2117
2118         if (pipe->inode)
2119                 mutex_unlock(&pipe->inode->i_mutex);
2120
2121         if (sd.num_spliced)
2122                 ret = sd.num_spliced;
2123
2124         if (ret > 0) {
2125                 unsigned long nr_pages;
2126                 int err;
2127
2128                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2129
2130                 err = generic_write_sync(out, *ppos, ret);
2131                 if (err)
2132                         ret = err;
2133                 else
2134                         *ppos += ret;
2135
2136                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2137         }
2138
2139         mlog_exit(ret);
2140         return ret;
2141 }
2142
2143 static ssize_t ocfs2_file_splice_read(struct file *in,
2144                                       loff_t *ppos,
2145                                       struct pipe_inode_info *pipe,
2146                                       size_t len,
2147                                       unsigned int flags)
2148 {
2149         int ret = 0, lock_level = 0;
2150         struct inode *inode = in->f_path.dentry->d_inode;
2151
2152         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2153                    (unsigned int)len,
2154                    in->f_path.dentry->d_name.len,
2155                    in->f_path.dentry->d_name.name);
2156
2157         /*
2158          * See the comment in ocfs2_file_aio_read()
2159          */
2160         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2161         if (ret < 0) {
2162                 mlog_errno(ret);
2163                 goto bail;
2164         }
2165         ocfs2_inode_unlock(inode, lock_level);
2166
2167         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2168
2169 bail:
2170         mlog_exit(ret);
2171         return ret;
2172 }
2173
2174 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2175                                    const struct iovec *iov,
2176                                    unsigned long nr_segs,
2177                                    loff_t pos)
2178 {
2179         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2180         struct file *filp = iocb->ki_filp;
2181         struct inode *inode = filp->f_path.dentry->d_inode;
2182
2183         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2184                    (unsigned int)nr_segs,
2185                    filp->f_path.dentry->d_name.len,
2186                    filp->f_path.dentry->d_name.name);
2187
2188         if (!inode) {
2189                 ret = -EINVAL;
2190                 mlog_errno(ret);
2191                 goto bail;
2192         }
2193
2194         /*
2195          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2196          * need locks to protect pending reads from racing with truncate.
2197          */
2198         if (filp->f_flags & O_DIRECT) {
2199                 down_read(&inode->i_alloc_sem);
2200                 have_alloc_sem = 1;
2201
2202                 ret = ocfs2_rw_lock(inode, 0);
2203                 if (ret < 0) {
2204                         mlog_errno(ret);
2205                         goto bail;
2206                 }
2207                 rw_level = 0;
2208                 /* communicate with ocfs2_dio_end_io */
2209                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2210         }
2211
2212         /*
2213          * We're fine letting folks race truncates and extending
2214          * writes with read across the cluster, just like they can
2215          * locally. Hence no rw_lock during read.
2216          *
2217          * Take and drop the meta data lock to update inode fields
2218          * like i_size. This allows the checks down below
2219          * generic_file_aio_read() a chance of actually working.
2220          */
2221         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2222         if (ret < 0) {
2223                 mlog_errno(ret);
2224                 goto bail;
2225         }
2226         ocfs2_inode_unlock(inode, lock_level);
2227
2228         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2229         if (ret == -EINVAL)
2230                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2231
2232         /* buffered aio wouldn't have proper lock coverage today */
2233         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2234
2235         /* see ocfs2_file_aio_write */
2236         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2237                 rw_level = -1;
2238                 have_alloc_sem = 0;
2239         }
2240
2241 bail:
2242         if (have_alloc_sem)
2243                 up_read(&inode->i_alloc_sem);
2244         if (rw_level != -1)
2245                 ocfs2_rw_unlock(inode, rw_level);
2246         mlog_exit(ret);
2247
2248         return ret;
2249 }
2250
2251 const struct inode_operations ocfs2_file_iops = {
2252         .setattr        = ocfs2_setattr,
2253         .getattr        = ocfs2_getattr,
2254         .permission     = ocfs2_permission,
2255         .setxattr       = generic_setxattr,
2256         .getxattr       = generic_getxattr,
2257         .listxattr      = ocfs2_listxattr,
2258         .removexattr    = generic_removexattr,
2259         .fallocate      = ocfs2_fallocate,
2260         .fiemap         = ocfs2_fiemap,
2261 };
2262
2263 const struct inode_operations ocfs2_special_file_iops = {
2264         .setattr        = ocfs2_setattr,
2265         .getattr        = ocfs2_getattr,
2266         .permission     = ocfs2_permission,
2267 };
2268
2269 /*
2270  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2271  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2272  */
2273 const struct file_operations ocfs2_fops = {
2274         .llseek         = generic_file_llseek,
2275         .read           = do_sync_read,
2276         .write          = do_sync_write,
2277         .mmap           = ocfs2_mmap,
2278         .fsync          = ocfs2_sync_file,
2279         .release        = ocfs2_file_release,
2280         .open           = ocfs2_file_open,
2281         .aio_read       = ocfs2_file_aio_read,
2282         .aio_write      = ocfs2_file_aio_write,
2283         .unlocked_ioctl = ocfs2_ioctl,
2284 #ifdef CONFIG_COMPAT
2285         .compat_ioctl   = ocfs2_compat_ioctl,
2286 #endif
2287         .lock           = ocfs2_lock,
2288         .flock          = ocfs2_flock,
2289         .splice_read    = ocfs2_file_splice_read,
2290         .splice_write   = ocfs2_file_splice_write,
2291 };
2292
2293 const struct file_operations ocfs2_dops = {
2294         .llseek         = generic_file_llseek,
2295         .read           = generic_read_dir,
2296         .readdir        = ocfs2_readdir,
2297         .fsync          = ocfs2_sync_file,
2298         .release        = ocfs2_dir_release,
2299         .open           = ocfs2_dir_open,
2300         .unlocked_ioctl = ocfs2_ioctl,
2301 #ifdef CONFIG_COMPAT
2302         .compat_ioctl   = ocfs2_compat_ioctl,
2303 #endif
2304         .lock           = ocfs2_lock,
2305         .flock          = ocfs2_flock,
2306 };
2307
2308 /*
2309  * POSIX-lockless variants of our file_operations.
2310  *
2311  * These will be used if the underlying cluster stack does not support
2312  * posix file locking, if the user passes the "localflocks" mount
2313  * option, or if we have a local-only fs.
2314  *
2315  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2316  * so we still want it in the case of no stack support for
2317  * plocks. Internally, it will do the right thing when asked to ignore
2318  * the cluster.
2319  */
2320 const struct file_operations ocfs2_fops_no_plocks = {
2321         .llseek         = generic_file_llseek,
2322         .read           = do_sync_read,
2323         .write          = do_sync_write,
2324         .mmap           = ocfs2_mmap,
2325         .fsync          = ocfs2_sync_file,
2326         .release        = ocfs2_file_release,
2327         .open           = ocfs2_file_open,
2328         .aio_read       = ocfs2_file_aio_read,
2329         .aio_write      = ocfs2_file_aio_write,
2330         .unlocked_ioctl = ocfs2_ioctl,
2331 #ifdef CONFIG_COMPAT
2332         .compat_ioctl   = ocfs2_compat_ioctl,
2333 #endif
2334         .flock          = ocfs2_flock,
2335         .splice_read    = ocfs2_file_splice_read,
2336         .splice_write   = ocfs2_file_splice_write,
2337 };
2338
2339 const struct file_operations ocfs2_dops_no_plocks = {
2340         .llseek         = generic_file_llseek,
2341         .read           = generic_read_dir,
2342         .readdir        = ocfs2_readdir,
2343         .fsync          = ocfs2_sync_file,
2344         .release        = ocfs2_dir_release,
2345         .open           = ocfs2_dir_open,
2346         .unlocked_ioctl = ocfs2_ioctl,
2347 #ifdef CONFIG_COMPAT
2348         .compat_ioctl   = ocfs2_compat_ioctl,
2349 #endif
2350         .flock          = ocfs2_flock,
2351 };