ocfs2: Support creation of unwritten extents
[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
38 #define MLOG_MASK_PREFIX ML_INODE
39 #include <cluster/masklog.h>
40
41 #include "ocfs2.h"
42
43 #include "alloc.h"
44 #include "aops.h"
45 #include "dir.h"
46 #include "dlmglue.h"
47 #include "extent_map.h"
48 #include "file.h"
49 #include "sysfile.h"
50 #include "inode.h"
51 #include "ioctl.h"
52 #include "journal.h"
53 #include "mmap.h"
54 #include "suballoc.h"
55 #include "super.h"
56
57 #include "buffer_head_io.h"
58
59 static int ocfs2_sync_inode(struct inode *inode)
60 {
61         filemap_fdatawrite(inode->i_mapping);
62         return sync_mapping_buffers(inode->i_mapping);
63 }
64
65 static int ocfs2_file_open(struct inode *inode, struct file *file)
66 {
67         int status;
68         int mode = file->f_flags;
69         struct ocfs2_inode_info *oi = OCFS2_I(inode);
70
71         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
72                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
73
74         spin_lock(&oi->ip_lock);
75
76         /* Check that the inode hasn't been wiped from disk by another
77          * node. If it hasn't then we're safe as long as we hold the
78          * spin lock until our increment of open count. */
79         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
80                 spin_unlock(&oi->ip_lock);
81
82                 status = -ENOENT;
83                 goto leave;
84         }
85
86         if (mode & O_DIRECT)
87                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
88
89         oi->ip_open_count++;
90         spin_unlock(&oi->ip_lock);
91         status = 0;
92 leave:
93         mlog_exit(status);
94         return status;
95 }
96
97 static int ocfs2_file_release(struct inode *inode, struct file *file)
98 {
99         struct ocfs2_inode_info *oi = OCFS2_I(inode);
100
101         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
102                        file->f_path.dentry->d_name.len,
103                        file->f_path.dentry->d_name.name);
104
105         spin_lock(&oi->ip_lock);
106         if (!--oi->ip_open_count)
107                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
108         spin_unlock(&oi->ip_lock);
109
110         mlog_exit(0);
111
112         return 0;
113 }
114
115 static int ocfs2_sync_file(struct file *file,
116                            struct dentry *dentry,
117                            int datasync)
118 {
119         int err = 0;
120         journal_t *journal;
121         struct inode *inode = dentry->d_inode;
122         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
123
124         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
125                    dentry->d_name.len, dentry->d_name.name);
126
127         err = ocfs2_sync_inode(dentry->d_inode);
128         if (err)
129                 goto bail;
130
131         journal = osb->journal->j_journal;
132         err = journal_force_commit(journal);
133
134 bail:
135         mlog_exit(err);
136
137         return (err < 0) ? -EIO : 0;
138 }
139
140 int ocfs2_should_update_atime(struct inode *inode,
141                               struct vfsmount *vfsmnt)
142 {
143         struct timespec now;
144         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
145
146         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
147                 return 0;
148
149         if ((inode->i_flags & S_NOATIME) ||
150             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
151                 return 0;
152
153         /*
154          * We can be called with no vfsmnt structure - NFSD will
155          * sometimes do this.
156          *
157          * Note that our action here is different than touch_atime() -
158          * if we can't tell whether this is a noatime mount, then we
159          * don't know whether to trust the value of s_atime_quantum.
160          */
161         if (vfsmnt == NULL)
162                 return 0;
163
164         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
165             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
166                 return 0;
167
168         if (vfsmnt->mnt_flags & MNT_RELATIME) {
169                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
170                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
171                         return 1;
172
173                 return 0;
174         }
175
176         now = CURRENT_TIME;
177         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
178                 return 0;
179         else
180                 return 1;
181 }
182
183 int ocfs2_update_inode_atime(struct inode *inode,
184                              struct buffer_head *bh)
185 {
186         int ret;
187         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
188         handle_t *handle;
189
190         mlog_entry_void();
191
192         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
193         if (handle == NULL) {
194                 ret = -ENOMEM;
195                 mlog_errno(ret);
196                 goto out;
197         }
198
199         inode->i_atime = CURRENT_TIME;
200         ret = ocfs2_mark_inode_dirty(handle, inode, bh);
201         if (ret < 0)
202                 mlog_errno(ret);
203
204         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
205 out:
206         mlog_exit(ret);
207         return ret;
208 }
209
210 static int ocfs2_set_inode_size(handle_t *handle,
211                                 struct inode *inode,
212                                 struct buffer_head *fe_bh,
213                                 u64 new_i_size)
214 {
215         int status;
216
217         mlog_entry_void();
218         i_size_write(inode, new_i_size);
219         inode->i_blocks = ocfs2_inode_sector_count(inode);
220         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
221
222         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
223         if (status < 0) {
224                 mlog_errno(status);
225                 goto bail;
226         }
227
228 bail:
229         mlog_exit(status);
230         return status;
231 }
232
233 static int ocfs2_simple_size_update(struct inode *inode,
234                                     struct buffer_head *di_bh,
235                                     u64 new_i_size)
236 {
237         int ret;
238         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
239         handle_t *handle = NULL;
240
241         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
242         if (handle == NULL) {
243                 ret = -ENOMEM;
244                 mlog_errno(ret);
245                 goto out;
246         }
247
248         ret = ocfs2_set_inode_size(handle, inode, di_bh,
249                                    new_i_size);
250         if (ret < 0)
251                 mlog_errno(ret);
252
253         ocfs2_commit_trans(osb, handle);
254 out:
255         return ret;
256 }
257
258 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
259                                      struct inode *inode,
260                                      struct buffer_head *fe_bh,
261                                      u64 new_i_size)
262 {
263         int status;
264         handle_t *handle;
265         struct ocfs2_dinode *di;
266
267         mlog_entry_void();
268
269         /* TODO: This needs to actually orphan the inode in this
270          * transaction. */
271
272         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
273         if (IS_ERR(handle)) {
274                 status = PTR_ERR(handle);
275                 mlog_errno(status);
276                 goto out;
277         }
278
279         status = ocfs2_journal_access(handle, inode, fe_bh,
280                                       OCFS2_JOURNAL_ACCESS_WRITE);
281         if (status < 0) {
282                 mlog_errno(status);
283                 goto out_commit;
284         }
285
286         /*
287          * Do this before setting i_size.
288          */
289         status = ocfs2_zero_tail_for_truncate(inode, handle, new_i_size);
290         if (status) {
291                 mlog_errno(status);
292                 goto out_commit;
293         }
294
295         i_size_write(inode, new_i_size);
296         inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
297         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
298
299         di = (struct ocfs2_dinode *) fe_bh->b_data;
300         di->i_size = cpu_to_le64(new_i_size);
301         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
302         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
303
304         status = ocfs2_journal_dirty(handle, fe_bh);
305         if (status < 0)
306                 mlog_errno(status);
307
308 out_commit:
309         ocfs2_commit_trans(osb, handle);
310 out:
311
312         mlog_exit(status);
313         return status;
314 }
315
316 static int ocfs2_truncate_file(struct inode *inode,
317                                struct buffer_head *di_bh,
318                                u64 new_i_size)
319 {
320         int status = 0;
321         struct ocfs2_dinode *fe = NULL;
322         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
323         struct ocfs2_truncate_context *tc = NULL;
324
325         mlog_entry("(inode = %llu, new_i_size = %llu\n",
326                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
327                    (unsigned long long)new_i_size);
328
329         fe = (struct ocfs2_dinode *) di_bh->b_data;
330         if (!OCFS2_IS_VALID_DINODE(fe)) {
331                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
332                 status = -EIO;
333                 goto bail;
334         }
335
336         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
337                         "Inode %llu, inode i_size = %lld != di "
338                         "i_size = %llu, i_flags = 0x%x\n",
339                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
340                         i_size_read(inode),
341                         (unsigned long long)le64_to_cpu(fe->i_size),
342                         le32_to_cpu(fe->i_flags));
343
344         if (new_i_size > le64_to_cpu(fe->i_size)) {
345                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
346                      (unsigned long long)le64_to_cpu(fe->i_size),
347                      (unsigned long long)new_i_size);
348                 status = -EINVAL;
349                 mlog_errno(status);
350                 goto bail;
351         }
352
353         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
354              (unsigned long long)le64_to_cpu(fe->i_blkno),
355              (unsigned long long)le64_to_cpu(fe->i_size),
356              (unsigned long long)new_i_size);
357
358         /* lets handle the simple truncate cases before doing any more
359          * cluster locking. */
360         if (new_i_size == le64_to_cpu(fe->i_size))
361                 goto bail;
362
363         down_write(&OCFS2_I(inode)->ip_alloc_sem);
364
365         /* This forces other nodes to sync and drop their pages. Do
366          * this even if we have a truncate without allocation change -
367          * ocfs2 cluster sizes can be much greater than page size, so
368          * we have to truncate them anyway.  */
369         status = ocfs2_data_lock(inode, 1);
370         if (status < 0) {
371                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
372
373                 mlog_errno(status);
374                 goto bail;
375         }
376
377         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
378         truncate_inode_pages(inode->i_mapping, new_i_size);
379
380         /* alright, we're going to need to do a full blown alloc size
381          * change. Orphan the inode so that recovery can complete the
382          * truncate if necessary. This does the task of marking
383          * i_size. */
384         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
385         if (status < 0) {
386                 mlog_errno(status);
387                 goto bail_unlock_data;
388         }
389
390         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
391         if (status < 0) {
392                 mlog_errno(status);
393                 goto bail_unlock_data;
394         }
395
396         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
397         if (status < 0) {
398                 mlog_errno(status);
399                 goto bail_unlock_data;
400         }
401
402         /* TODO: orphan dir cleanup here. */
403 bail_unlock_data:
404         ocfs2_data_unlock(inode, 1);
405
406         up_write(&OCFS2_I(inode)->ip_alloc_sem);
407
408 bail:
409
410         mlog_exit(status);
411         return status;
412 }
413
414 /*
415  * extend allocation only here.
416  * we'll update all the disk stuff, and oip->alloc_size
417  *
418  * expect stuff to be locked, a transaction started and enough data /
419  * metadata reservations in the contexts.
420  *
421  * Will return -EAGAIN, and a reason if a restart is needed.
422  * If passed in, *reason will always be set, even in error.
423  */
424 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
425                                struct inode *inode,
426                                u32 *logical_offset,
427                                u32 clusters_to_add,
428                                int mark_unwritten,
429                                struct buffer_head *fe_bh,
430                                handle_t *handle,
431                                struct ocfs2_alloc_context *data_ac,
432                                struct ocfs2_alloc_context *meta_ac,
433                                enum ocfs2_alloc_restarted *reason_ret)
434 {
435         int status = 0;
436         int free_extents;
437         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
438         enum ocfs2_alloc_restarted reason = RESTART_NONE;
439         u32 bit_off, num_bits;
440         u64 block;
441         u8 flags = 0;
442
443         BUG_ON(!clusters_to_add);
444
445         if (mark_unwritten)
446                 flags = OCFS2_EXT_UNWRITTEN;
447
448         free_extents = ocfs2_num_free_extents(osb, inode, fe);
449         if (free_extents < 0) {
450                 status = free_extents;
451                 mlog_errno(status);
452                 goto leave;
453         }
454
455         /* there are two cases which could cause us to EAGAIN in the
456          * we-need-more-metadata case:
457          * 1) we haven't reserved *any*
458          * 2) we are so fragmented, we've needed to add metadata too
459          *    many times. */
460         if (!free_extents && !meta_ac) {
461                 mlog(0, "we haven't reserved any metadata!\n");
462                 status = -EAGAIN;
463                 reason = RESTART_META;
464                 goto leave;
465         } else if ((!free_extents)
466                    && (ocfs2_alloc_context_bits_left(meta_ac)
467                        < ocfs2_extend_meta_needed(fe))) {
468                 mlog(0, "filesystem is really fragmented...\n");
469                 status = -EAGAIN;
470                 reason = RESTART_META;
471                 goto leave;
472         }
473
474         status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
475                                       &bit_off, &num_bits);
476         if (status < 0) {
477                 if (status != -ENOSPC)
478                         mlog_errno(status);
479                 goto leave;
480         }
481
482         BUG_ON(num_bits > clusters_to_add);
483
484         /* reserve our write early -- insert_extent may update the inode */
485         status = ocfs2_journal_access(handle, inode, fe_bh,
486                                       OCFS2_JOURNAL_ACCESS_WRITE);
487         if (status < 0) {
488                 mlog_errno(status);
489                 goto leave;
490         }
491
492         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
493         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
494              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
495         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
496                                      *logical_offset, block, num_bits,
497                                      flags, meta_ac);
498         if (status < 0) {
499                 mlog_errno(status);
500                 goto leave;
501         }
502
503         status = ocfs2_journal_dirty(handle, fe_bh);
504         if (status < 0) {
505                 mlog_errno(status);
506                 goto leave;
507         }
508
509         clusters_to_add -= num_bits;
510         *logical_offset += num_bits;
511
512         if (clusters_to_add) {
513                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
514                      "%u\n", fe->i_clusters, clusters_to_add);
515                 status = -EAGAIN;
516                 reason = RESTART_TRANS;
517         }
518
519 leave:
520         mlog_exit(status);
521         if (reason_ret)
522                 *reason_ret = reason;
523         return status;
524 }
525
526 /*
527  * For a given allocation, determine which allocators will need to be
528  * accessed, and lock them, reserving the appropriate number of bits.
529  *
530  * Sparse file systems call this from ocfs2_write_begin_nolock()
531  * and ocfs2_allocate_unwritten_extents().
532  *
533  * File systems which don't support holes call this from
534  * ocfs2_extend_allocation().
535  */
536 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
537                           u32 clusters_to_add, u32 extents_to_split,
538                           struct ocfs2_alloc_context **data_ac,
539                           struct ocfs2_alloc_context **meta_ac)
540 {
541         int ret, num_free_extents;
542         unsigned int max_recs_needed = clusters_to_add + 2 * extents_to_split;
543         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
544
545         *meta_ac = NULL;
546         *data_ac = NULL;
547
548         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
549              "clusters_to_add = %u, extents_to_split = %u\n",
550              (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
551              le32_to_cpu(di->i_clusters), clusters_to_add, extents_to_split);
552
553         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
554         if (num_free_extents < 0) {
555                 ret = num_free_extents;
556                 mlog_errno(ret);
557                 goto out;
558         }
559
560         /*
561          * Sparse allocation file systems need to be more conservative
562          * with reserving room for expansion - the actual allocation
563          * happens while we've got a journal handle open so re-taking
564          * a cluster lock (because we ran out of room for another
565          * extent) will violate ordering rules.
566          *
567          * Most of the time we'll only be seeing this 1 cluster at a time
568          * anyway.
569          *
570          * Always lock for any unwritten extents - we might want to
571          * add blocks during a split.
572          */
573         if (!num_free_extents ||
574             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed)) {
575                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
576                 if (ret < 0) {
577                         if (ret != -ENOSPC)
578                                 mlog_errno(ret);
579                         goto out;
580                 }
581         }
582
583         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
584         if (ret < 0) {
585                 if (ret != -ENOSPC)
586                         mlog_errno(ret);
587                 goto out;
588         }
589
590 out:
591         if (ret) {
592                 if (*meta_ac) {
593                         ocfs2_free_alloc_context(*meta_ac);
594                         *meta_ac = NULL;
595                 }
596
597                 /*
598                  * We cannot have an error and a non null *data_ac.
599                  */
600         }
601
602         return ret;
603 }
604
605 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
606                                      u32 clusters_to_add, int mark_unwritten)
607 {
608         int status = 0;
609         int restart_func = 0;
610         int credits;
611         u32 prev_clusters;
612         struct buffer_head *bh = NULL;
613         struct ocfs2_dinode *fe = NULL;
614         handle_t *handle = NULL;
615         struct ocfs2_alloc_context *data_ac = NULL;
616         struct ocfs2_alloc_context *meta_ac = NULL;
617         enum ocfs2_alloc_restarted why;
618         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
619
620         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
621
622         /*
623          * This function only exists for file systems which don't
624          * support holes.
625          */
626         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
627
628         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
629                                   OCFS2_BH_CACHED, inode);
630         if (status < 0) {
631                 mlog_errno(status);
632                 goto leave;
633         }
634
635         fe = (struct ocfs2_dinode *) bh->b_data;
636         if (!OCFS2_IS_VALID_DINODE(fe)) {
637                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
638                 status = -EIO;
639                 goto leave;
640         }
641
642 restart_all:
643         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
644
645         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, 0, &data_ac,
646                                        &meta_ac);
647         if (status) {
648                 mlog_errno(status);
649                 goto leave;
650         }
651
652         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
653         handle = ocfs2_start_trans(osb, credits);
654         if (IS_ERR(handle)) {
655                 status = PTR_ERR(handle);
656                 handle = NULL;
657                 mlog_errno(status);
658                 goto leave;
659         }
660
661 restarted_transaction:
662         /* reserve a write to the file entry early on - that we if we
663          * run out of credits in the allocation path, we can still
664          * update i_size. */
665         status = ocfs2_journal_access(handle, inode, bh,
666                                       OCFS2_JOURNAL_ACCESS_WRITE);
667         if (status < 0) {
668                 mlog_errno(status);
669                 goto leave;
670         }
671
672         prev_clusters = OCFS2_I(inode)->ip_clusters;
673
674         status = ocfs2_do_extend_allocation(osb,
675                                             inode,
676                                             &logical_start,
677                                             clusters_to_add,
678                                             mark_unwritten,
679                                             bh,
680                                             handle,
681                                             data_ac,
682                                             meta_ac,
683                                             &why);
684         if ((status < 0) && (status != -EAGAIN)) {
685                 if (status != -ENOSPC)
686                         mlog_errno(status);
687                 goto leave;
688         }
689
690         status = ocfs2_journal_dirty(handle, bh);
691         if (status < 0) {
692                 mlog_errno(status);
693                 goto leave;
694         }
695
696         spin_lock(&OCFS2_I(inode)->ip_lock);
697         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
698         spin_unlock(&OCFS2_I(inode)->ip_lock);
699
700         if (why != RESTART_NONE && clusters_to_add) {
701                 if (why == RESTART_META) {
702                         mlog(0, "restarting function.\n");
703                         restart_func = 1;
704                 } else {
705                         BUG_ON(why != RESTART_TRANS);
706
707                         mlog(0, "restarting transaction.\n");
708                         /* TODO: This can be more intelligent. */
709                         credits = ocfs2_calc_extend_credits(osb->sb,
710                                                             fe,
711                                                             clusters_to_add);
712                         status = ocfs2_extend_trans(handle, credits);
713                         if (status < 0) {
714                                 /* handle still has to be committed at
715                                  * this point. */
716                                 status = -ENOMEM;
717                                 mlog_errno(status);
718                                 goto leave;
719                         }
720                         goto restarted_transaction;
721                 }
722         }
723
724         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
725              le32_to_cpu(fe->i_clusters),
726              (unsigned long long)le64_to_cpu(fe->i_size));
727         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
728              OCFS2_I(inode)->ip_clusters, i_size_read(inode));
729
730 leave:
731         if (handle) {
732                 ocfs2_commit_trans(osb, handle);
733                 handle = NULL;
734         }
735         if (data_ac) {
736                 ocfs2_free_alloc_context(data_ac);
737                 data_ac = NULL;
738         }
739         if (meta_ac) {
740                 ocfs2_free_alloc_context(meta_ac);
741                 meta_ac = NULL;
742         }
743         if ((!status) && restart_func) {
744                 restart_func = 0;
745                 goto restart_all;
746         }
747         if (bh) {
748                 brelse(bh);
749                 bh = NULL;
750         }
751
752         mlog_exit(status);
753         return status;
754 }
755
756 static int ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
757                                    u32 clusters_to_add, int mark_unwritten)
758 {
759         int ret;
760
761         /*
762          * The alloc sem blocks peope in read/write from reading our
763          * allocation until we're done changing it. We depend on
764          * i_mutex to block other extend/truncate calls while we're
765          * here.
766          */
767         down_write(&OCFS2_I(inode)->ip_alloc_sem);
768         ret = __ocfs2_extend_allocation(inode, logical_start, clusters_to_add,
769                                         mark_unwritten);
770         up_write(&OCFS2_I(inode)->ip_alloc_sem);
771
772         return ret;
773 }
774
775 /* Some parts of this taken from generic_cont_expand, which turned out
776  * to be too fragile to do exactly what we need without us having to
777  * worry about recursive locking in ->prepare_write() and
778  * ->commit_write(). */
779 static int ocfs2_write_zero_page(struct inode *inode,
780                                  u64 size)
781 {
782         struct address_space *mapping = inode->i_mapping;
783         struct page *page;
784         unsigned long index;
785         unsigned int offset;
786         handle_t *handle = NULL;
787         int ret;
788
789         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
790         /* ugh.  in prepare/commit_write, if from==to==start of block, we 
791         ** skip the prepare.  make sure we never send an offset for the start
792         ** of a block
793         */
794         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
795                 offset++;
796         }
797         index = size >> PAGE_CACHE_SHIFT;
798
799         page = grab_cache_page(mapping, index);
800         if (!page) {
801                 ret = -ENOMEM;
802                 mlog_errno(ret);
803                 goto out;
804         }
805
806         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
807         if (ret < 0) {
808                 mlog_errno(ret);
809                 goto out_unlock;
810         }
811
812         if (ocfs2_should_order_data(inode)) {
813                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
814                                                      offset);
815                 if (IS_ERR(handle)) {
816                         ret = PTR_ERR(handle);
817                         handle = NULL;
818                         goto out_unlock;
819                 }
820         }
821
822         /* must not update i_size! */
823         ret = block_commit_write(page, offset, offset);
824         if (ret < 0)
825                 mlog_errno(ret);
826         else
827                 ret = 0;
828
829         if (handle)
830                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
831 out_unlock:
832         unlock_page(page);
833         page_cache_release(page);
834 out:
835         return ret;
836 }
837
838 static int ocfs2_zero_extend(struct inode *inode,
839                              u64 zero_to_size)
840 {
841         int ret = 0;
842         u64 start_off;
843         struct super_block *sb = inode->i_sb;
844
845         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
846         while (start_off < zero_to_size) {
847                 ret = ocfs2_write_zero_page(inode, start_off);
848                 if (ret < 0) {
849                         mlog_errno(ret);
850                         goto out;
851                 }
852
853                 start_off += sb->s_blocksize;
854
855                 /*
856                  * Very large extends have the potential to lock up
857                  * the cpu for extended periods of time.
858                  */
859                 cond_resched();
860         }
861
862 out:
863         return ret;
864 }
865
866 /* 
867  * A tail_to_skip value > 0 indicates that we're being called from
868  * ocfs2_file_aio_write(). This has the following implications:
869  *
870  * - we don't want to update i_size
871  * - di_bh will be NULL, which is fine because it's only used in the
872  *   case where we want to update i_size.
873  * - ocfs2_zero_extend() will then only be filling the hole created
874  *   between i_size and the start of the write.
875  */
876 static int ocfs2_extend_file(struct inode *inode,
877                              struct buffer_head *di_bh,
878                              u64 new_i_size,
879                              size_t tail_to_skip)
880 {
881         int ret = 0;
882         u32 clusters_to_add = 0;
883
884         BUG_ON(!tail_to_skip && !di_bh);
885
886         /* setattr sometimes calls us like this. */
887         if (new_i_size == 0)
888                 goto out;
889
890         if (i_size_read(inode) == new_i_size)
891                 goto out;
892         BUG_ON(new_i_size < i_size_read(inode));
893
894         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
895                 BUG_ON(tail_to_skip != 0);
896                 goto out_update_size;
897         }
898
899         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) - 
900                 OCFS2_I(inode)->ip_clusters;
901
902         /* 
903          * protect the pages that ocfs2_zero_extend is going to be
904          * pulling into the page cache.. we do this before the
905          * metadata extend so that we don't get into the situation
906          * where we've extended the metadata but can't get the data
907          * lock to zero.
908          */
909         ret = ocfs2_data_lock(inode, 1);
910         if (ret < 0) {
911                 mlog_errno(ret);
912                 goto out;
913         }
914
915         if (clusters_to_add) {
916                 ret = ocfs2_extend_allocation(inode,
917                                               OCFS2_I(inode)->ip_clusters,
918                                               clusters_to_add, 0);
919                 if (ret < 0) {
920                         mlog_errno(ret);
921                         goto out_unlock;
922                 }
923         }
924
925         /*
926          * Call this even if we don't add any clusters to the tree. We
927          * still need to zero the area between the old i_size and the
928          * new i_size.
929          */
930         ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
931         if (ret < 0) {
932                 mlog_errno(ret);
933                 goto out_unlock;
934         }
935
936 out_update_size:
937         if (!tail_to_skip) {
938                 /* We're being called from ocfs2_setattr() which wants
939                  * us to update i_size */
940                 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
941                 if (ret < 0)
942                         mlog_errno(ret);
943         }
944
945 out_unlock:
946         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
947                 ocfs2_data_unlock(inode, 1);
948
949 out:
950         return ret;
951 }
952
953 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
954 {
955         int status = 0, size_change;
956         struct inode *inode = dentry->d_inode;
957         struct super_block *sb = inode->i_sb;
958         struct ocfs2_super *osb = OCFS2_SB(sb);
959         struct buffer_head *bh = NULL;
960         handle_t *handle = NULL;
961
962         mlog_entry("(0x%p, '%.*s')\n", dentry,
963                    dentry->d_name.len, dentry->d_name.name);
964
965         if (attr->ia_valid & ATTR_MODE)
966                 mlog(0, "mode change: %d\n", attr->ia_mode);
967         if (attr->ia_valid & ATTR_UID)
968                 mlog(0, "uid change: %d\n", attr->ia_uid);
969         if (attr->ia_valid & ATTR_GID)
970                 mlog(0, "gid change: %d\n", attr->ia_gid);
971         if (attr->ia_valid & ATTR_SIZE)
972                 mlog(0, "size change...\n");
973         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
974                 mlog(0, "time change...\n");
975
976 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
977                            | ATTR_GID | ATTR_UID | ATTR_MODE)
978         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
979                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
980                 return 0;
981         }
982
983         status = inode_change_ok(inode, attr);
984         if (status)
985                 return status;
986
987         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
988         if (size_change) {
989                 status = ocfs2_rw_lock(inode, 1);
990                 if (status < 0) {
991                         mlog_errno(status);
992                         goto bail;
993                 }
994         }
995
996         status = ocfs2_meta_lock(inode, &bh, 1);
997         if (status < 0) {
998                 if (status != -ENOENT)
999                         mlog_errno(status);
1000                 goto bail_unlock_rw;
1001         }
1002
1003         if (size_change && attr->ia_size != i_size_read(inode)) {
1004                 if (i_size_read(inode) > attr->ia_size)
1005                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1006                 else
1007                         status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
1008                 if (status < 0) {
1009                         if (status != -ENOSPC)
1010                                 mlog_errno(status);
1011                         status = -ENOSPC;
1012                         goto bail_unlock;
1013                 }
1014         }
1015
1016         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1017         if (IS_ERR(handle)) {
1018                 status = PTR_ERR(handle);
1019                 mlog_errno(status);
1020                 goto bail_unlock;
1021         }
1022
1023         /*
1024          * This will intentionally not wind up calling vmtruncate(),
1025          * since all the work for a size change has been done above.
1026          * Otherwise, we could get into problems with truncate as
1027          * ip_alloc_sem is used there to protect against i_size
1028          * changes.
1029          */
1030         status = inode_setattr(inode, attr);
1031         if (status < 0) {
1032                 mlog_errno(status);
1033                 goto bail_commit;
1034         }
1035
1036         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1037         if (status < 0)
1038                 mlog_errno(status);
1039
1040 bail_commit:
1041         ocfs2_commit_trans(osb, handle);
1042 bail_unlock:
1043         ocfs2_meta_unlock(inode, 1);
1044 bail_unlock_rw:
1045         if (size_change)
1046                 ocfs2_rw_unlock(inode, 1);
1047 bail:
1048         if (bh)
1049                 brelse(bh);
1050
1051         mlog_exit(status);
1052         return status;
1053 }
1054
1055 int ocfs2_getattr(struct vfsmount *mnt,
1056                   struct dentry *dentry,
1057                   struct kstat *stat)
1058 {
1059         struct inode *inode = dentry->d_inode;
1060         struct super_block *sb = dentry->d_inode->i_sb;
1061         struct ocfs2_super *osb = sb->s_fs_info;
1062         int err;
1063
1064         mlog_entry_void();
1065
1066         err = ocfs2_inode_revalidate(dentry);
1067         if (err) {
1068                 if (err != -ENOENT)
1069                         mlog_errno(err);
1070                 goto bail;
1071         }
1072
1073         generic_fillattr(inode, stat);
1074
1075         /* We set the blksize from the cluster size for performance */
1076         stat->blksize = osb->s_clustersize;
1077
1078 bail:
1079         mlog_exit(err);
1080
1081         return err;
1082 }
1083
1084 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1085 {
1086         int ret;
1087
1088         mlog_entry_void();
1089
1090         ret = ocfs2_meta_lock(inode, NULL, 0);
1091         if (ret) {
1092                 if (ret != -ENOENT)
1093                         mlog_errno(ret);
1094                 goto out;
1095         }
1096
1097         ret = generic_permission(inode, mask, NULL);
1098
1099         ocfs2_meta_unlock(inode, 0);
1100 out:
1101         mlog_exit(ret);
1102         return ret;
1103 }
1104
1105 static int ocfs2_write_remove_suid(struct inode *inode)
1106 {
1107         int ret;
1108         struct buffer_head *bh = NULL;
1109         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1110         handle_t *handle;
1111         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1112         struct ocfs2_dinode *di;
1113
1114         mlog_entry("(Inode %llu, mode 0%o)\n",
1115                    (unsigned long long)oi->ip_blkno, inode->i_mode);
1116
1117         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1118         if (handle == NULL) {
1119                 ret = -ENOMEM;
1120                 mlog_errno(ret);
1121                 goto out;
1122         }
1123
1124         ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1125         if (ret < 0) {
1126                 mlog_errno(ret);
1127                 goto out_trans;
1128         }
1129
1130         ret = ocfs2_journal_access(handle, inode, bh,
1131                                    OCFS2_JOURNAL_ACCESS_WRITE);
1132         if (ret < 0) {
1133                 mlog_errno(ret);
1134                 goto out_bh;
1135         }
1136
1137         inode->i_mode &= ~S_ISUID;
1138         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1139                 inode->i_mode &= ~S_ISGID;
1140
1141         di = (struct ocfs2_dinode *) bh->b_data;
1142         di->i_mode = cpu_to_le16(inode->i_mode);
1143
1144         ret = ocfs2_journal_dirty(handle, bh);
1145         if (ret < 0)
1146                 mlog_errno(ret);
1147 out_bh:
1148         brelse(bh);
1149 out_trans:
1150         ocfs2_commit_trans(osb, handle);
1151 out:
1152         mlog_exit(ret);
1153         return ret;
1154 }
1155
1156 /*
1157  * Will look for holes and unwritten extents in the range starting at
1158  * pos for count bytes (inclusive).
1159  */
1160 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1161                                        size_t count)
1162 {
1163         int ret = 0;
1164         unsigned int extent_flags;
1165         u32 cpos, clusters, extent_len, phys_cpos;
1166         struct super_block *sb = inode->i_sb;
1167
1168         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1169         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1170
1171         while (clusters) {
1172                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1173                                          &extent_flags);
1174                 if (ret < 0) {
1175                         mlog_errno(ret);
1176                         goto out;
1177                 }
1178
1179                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1180                         ret = 1;
1181                         break;
1182                 }
1183
1184                 if (extent_len > clusters)
1185                         extent_len = clusters;
1186
1187                 clusters -= extent_len;
1188                 cpos += extent_len;
1189         }
1190 out:
1191         return ret;
1192 }
1193
1194 /*
1195  * Allocate enough extents to cover the region starting at byte offset
1196  * start for len bytes. Existing extents are skipped, any extents
1197  * added are marked as "unwritten".
1198  */
1199 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1200                                             u64 start, u64 len)
1201 {
1202         int ret;
1203         u32 cpos, phys_cpos, clusters, alloc_size;
1204
1205         /*
1206          * We consider both start and len to be inclusive.
1207          */
1208         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1209         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1210         clusters -= cpos;
1211
1212         while (clusters) {
1213                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1214                                          &alloc_size, NULL);
1215                 if (ret) {
1216                         mlog_errno(ret);
1217                         goto out;
1218                 }
1219
1220                 /*
1221                  * Hole or existing extent len can be arbitrary, so
1222                  * cap it to our own allocation request.
1223                  */
1224                 if (alloc_size > clusters)
1225                         alloc_size = clusters;
1226
1227                 if (phys_cpos) {
1228                         /*
1229                          * We already have an allocation at this
1230                          * region so we can safely skip it.
1231                          */
1232                         goto next;
1233                 }
1234
1235                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1236                 if (ret) {
1237                         if (ret != -ENOSPC)
1238                                 mlog_errno(ret);
1239                         goto out;
1240                 }
1241
1242 next:
1243                 cpos += alloc_size;
1244                 clusters -= alloc_size;
1245         }
1246
1247         ret = 0;
1248 out:
1249         return ret;
1250 }
1251
1252 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1253                                          loff_t *ppos,
1254                                          size_t count,
1255                                          int appending,
1256                                          int *direct_io)
1257 {
1258         int ret = 0, meta_level = appending;
1259         struct inode *inode = dentry->d_inode;
1260         u32 clusters;
1261         loff_t newsize, saved_pos;
1262
1263         /* 
1264          * We sample i_size under a read level meta lock to see if our write
1265          * is extending the file, if it is we back off and get a write level
1266          * meta lock.
1267          */
1268         for(;;) {
1269                 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1270                 if (ret < 0) {
1271                         meta_level = -1;
1272                         mlog_errno(ret);
1273                         goto out;
1274                 }
1275
1276                 /* Clear suid / sgid if necessary. We do this here
1277                  * instead of later in the write path because
1278                  * remove_suid() calls ->setattr without any hint that
1279                  * we may have already done our cluster locking. Since
1280                  * ocfs2_setattr() *must* take cluster locks to
1281                  * proceeed, this will lead us to recursively lock the
1282                  * inode. There's also the dinode i_size state which
1283                  * can be lost via setattr during extending writes (we
1284                  * set inode->i_size at the end of a write. */
1285                 if (should_remove_suid(dentry)) {
1286                         if (meta_level == 0) {
1287                                 ocfs2_meta_unlock(inode, meta_level);
1288                                 meta_level = 1;
1289                                 continue;
1290                         }
1291
1292                         ret = ocfs2_write_remove_suid(inode);
1293                         if (ret < 0) {
1294                                 mlog_errno(ret);
1295                                 goto out_unlock;
1296                         }
1297                 }
1298
1299                 /* work on a copy of ppos until we're sure that we won't have
1300                  * to recalculate it due to relocking. */
1301                 if (appending) {
1302                         saved_pos = i_size_read(inode);
1303                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1304                 } else {
1305                         saved_pos = *ppos;
1306                 }
1307
1308                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1309                         loff_t end = saved_pos + count;
1310
1311                         /*
1312                          * Skip the O_DIRECT checks if we don't need
1313                          * them.
1314                          */
1315                         if (!direct_io || !(*direct_io))
1316                                 break;
1317
1318                         /*
1319                          * Allowing concurrent direct writes means
1320                          * i_size changes wouldn't be synchronized, so
1321                          * one node could wind up truncating another
1322                          * nodes writes.
1323                          */
1324                         if (end > i_size_read(inode)) {
1325                                 *direct_io = 0;
1326                                 break;
1327                         }
1328
1329                         /*
1330                          * We don't fill holes during direct io, so
1331                          * check for them here. If any are found, the
1332                          * caller will have to retake some cluster
1333                          * locks and initiate the io as buffered.
1334                          */
1335                         ret = ocfs2_check_range_for_holes(inode, saved_pos,
1336                                                           count);
1337                         if (ret == 1) {
1338                                 *direct_io = 0;
1339                                 ret = 0;
1340                         } else if (ret < 0)
1341                                 mlog_errno(ret);
1342                         break;
1343                 }
1344
1345                 /*
1346                  * The rest of this loop is concerned with legacy file
1347                  * systems which don't support sparse files.
1348                  */
1349
1350                 newsize = count + saved_pos;
1351
1352                 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1353                      (long long) saved_pos, (long long) newsize,
1354                      (long long) i_size_read(inode));
1355
1356                 /* No need for a higher level metadata lock if we're
1357                  * never going past i_size. */
1358                 if (newsize <= i_size_read(inode))
1359                         break;
1360
1361                 if (meta_level == 0) {
1362                         ocfs2_meta_unlock(inode, meta_level);
1363                         meta_level = 1;
1364                         continue;
1365                 }
1366
1367                 spin_lock(&OCFS2_I(inode)->ip_lock);
1368                 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1369                         OCFS2_I(inode)->ip_clusters;
1370                 spin_unlock(&OCFS2_I(inode)->ip_lock);
1371
1372                 mlog(0, "Writing at EOF, may need more allocation: "
1373                      "i_size = %lld, newsize = %lld, need %u clusters\n",
1374                      (long long) i_size_read(inode), (long long) newsize,
1375                      clusters);
1376
1377                 /* We only want to continue the rest of this loop if
1378                  * our extend will actually require more
1379                  * allocation. */
1380                 if (!clusters)
1381                         break;
1382
1383                 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1384                 if (ret < 0) {
1385                         if (ret != -ENOSPC)
1386                                 mlog_errno(ret);
1387                         goto out_unlock;
1388                 }
1389                 break;
1390         }
1391
1392         if (appending)
1393                 *ppos = saved_pos;
1394
1395 out_unlock:
1396         ocfs2_meta_unlock(inode, meta_level);
1397
1398 out:
1399         return ret;
1400 }
1401
1402 static inline void
1403 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1404 {
1405         const struct iovec *iov = *iovp;
1406         size_t base = *basep;
1407
1408         do {
1409                 int copy = min(bytes, iov->iov_len - base);
1410
1411                 bytes -= copy;
1412                 base += copy;
1413                 if (iov->iov_len == base) {
1414                         iov++;
1415                         base = 0;
1416                 }
1417         } while (bytes);
1418         *iovp = iov;
1419         *basep = base;
1420 }
1421
1422 static struct page * ocfs2_get_write_source(char **ret_src_buf,
1423                                             const struct iovec *cur_iov,
1424                                             size_t iov_offset)
1425 {
1426         int ret;
1427         char *buf = cur_iov->iov_base + iov_offset;
1428         struct page *src_page = NULL;
1429         unsigned long off;
1430
1431         off = (unsigned long)(buf) & ~PAGE_CACHE_MASK;
1432
1433         if (!segment_eq(get_fs(), KERNEL_DS)) {
1434                 /*
1435                  * Pull in the user page. We want to do this outside
1436                  * of the meta data locks in order to preserve locking
1437                  * order in case of page fault.
1438                  */
1439                 ret = get_user_pages(current, current->mm,
1440                                      (unsigned long)buf & PAGE_CACHE_MASK, 1,
1441                                      0, 0, &src_page, NULL);
1442                 if (ret == 1)
1443                         *ret_src_buf = kmap(src_page) + off;
1444                 else
1445                         src_page = ERR_PTR(-EFAULT);
1446         } else {
1447                 *ret_src_buf = buf;
1448         }
1449
1450         return src_page;
1451 }
1452
1453 static void ocfs2_put_write_source(struct page *page)
1454 {
1455         if (page) {
1456                 kunmap(page);
1457                 page_cache_release(page);
1458         }
1459 }
1460
1461 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1462                                          const struct iovec *iov,
1463                                          unsigned long nr_segs,
1464                                          size_t count,
1465                                          ssize_t o_direct_written)
1466 {
1467         int ret = 0;
1468         ssize_t copied, total = 0;
1469         size_t iov_offset = 0, bytes;
1470         loff_t pos;
1471         const struct iovec *cur_iov = iov;
1472         struct page *user_page, *page;
1473         char *buf, *dst;
1474         void *fsdata;
1475
1476         /*
1477          * handle partial DIO write.  Adjust cur_iov if needed.
1478          */
1479         ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1480
1481         do {
1482                 pos = *ppos;
1483
1484                 user_page = ocfs2_get_write_source(&buf, cur_iov, iov_offset);
1485                 if (IS_ERR(user_page)) {
1486                         ret = PTR_ERR(user_page);
1487                         goto out;
1488                 }
1489
1490                 /* Stay within our page boundaries */
1491                 bytes = min((PAGE_CACHE_SIZE - ((unsigned long)pos & ~PAGE_CACHE_MASK)),
1492                             (PAGE_CACHE_SIZE - ((unsigned long)buf & ~PAGE_CACHE_MASK)));
1493                 /* Stay within the vector boundary */
1494                 bytes = min_t(size_t, bytes, cur_iov->iov_len - iov_offset);
1495                 /* Stay within count */
1496                 bytes = min(bytes, count);
1497
1498                 page = NULL;
1499                 ret = ocfs2_write_begin(file, file->f_mapping, pos, bytes, 0,
1500                                         &page, &fsdata);
1501                 if (ret) {
1502                         mlog_errno(ret);
1503                         goto out;
1504                 }
1505
1506                 dst = kmap_atomic(page, KM_USER0);
1507                 memcpy(dst + (pos & (PAGE_CACHE_SIZE - 1)), buf, bytes);
1508                 kunmap_atomic(dst, KM_USER0);
1509                 flush_dcache_page(page);
1510                 ocfs2_put_write_source(user_page);
1511
1512                 copied = ocfs2_write_end(file, file->f_mapping, pos, bytes,
1513                                          bytes, page, fsdata);
1514                 if (copied < 0) {
1515                         mlog_errno(copied);
1516                         ret = copied;
1517                         goto out;
1518                 }
1519
1520                 total += copied;
1521                 *ppos = pos + copied;
1522                 count -= copied;
1523
1524                 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1525         } while(count);
1526
1527 out:
1528         return total ? total : ret;
1529 }
1530
1531 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1532                                     const struct iovec *iov,
1533                                     unsigned long nr_segs,
1534                                     loff_t pos)
1535 {
1536         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1537         int can_do_direct, sync = 0;
1538         ssize_t written = 0;
1539         size_t ocount;          /* original count */
1540         size_t count;           /* after file limit checks */
1541         loff_t *ppos = &iocb->ki_pos;
1542         struct file *file = iocb->ki_filp;
1543         struct inode *inode = file->f_path.dentry->d_inode;
1544
1545         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1546                    (unsigned int)nr_segs,
1547                    file->f_path.dentry->d_name.len,
1548                    file->f_path.dentry->d_name.name);
1549
1550         if (iocb->ki_left == 0)
1551                 return 0;
1552
1553         ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1554         if (ret)
1555                 return ret;
1556
1557         count = ocount;
1558
1559         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1560
1561         appending = file->f_flags & O_APPEND ? 1 : 0;
1562         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1563
1564         mutex_lock(&inode->i_mutex);
1565
1566 relock:
1567         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1568         if (direct_io) {
1569                 down_read(&inode->i_alloc_sem);
1570                 have_alloc_sem = 1;
1571         }
1572
1573         /* concurrent O_DIRECT writes are allowed */
1574         rw_level = !direct_io;
1575         ret = ocfs2_rw_lock(inode, rw_level);
1576         if (ret < 0) {
1577                 mlog_errno(ret);
1578                 goto out_sems;
1579         }
1580
1581         can_do_direct = direct_io;
1582         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1583                                             iocb->ki_left, appending,
1584                                             &can_do_direct);
1585         if (ret < 0) {
1586                 mlog_errno(ret);
1587                 goto out;
1588         }
1589
1590         /*
1591          * We can't complete the direct I/O as requested, fall back to
1592          * buffered I/O.
1593          */
1594         if (direct_io && !can_do_direct) {
1595                 ocfs2_rw_unlock(inode, rw_level);
1596                 up_read(&inode->i_alloc_sem);
1597
1598                 have_alloc_sem = 0;
1599                 rw_level = -1;
1600
1601                 direct_io = 0;
1602                 sync = 1;
1603                 goto relock;
1604         }
1605
1606         if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1607                 sync = 1;
1608
1609         /*
1610          * XXX: Is it ok to execute these checks a second time?
1611          */
1612         ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1613         if (ret)
1614                 goto out;
1615
1616         /*
1617          * Set pos so that sync_page_range_nolock() below understands
1618          * where to start from. We might've moved it around via the
1619          * calls above. The range we want to actually sync starts from
1620          * *ppos here.
1621          *
1622          */
1623         pos = *ppos;
1624
1625         /* communicate with ocfs2_dio_end_io */
1626         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1627
1628         if (direct_io) {
1629                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1630                                                     ppos, count, ocount);
1631                 if (written < 0) {
1632                         ret = written;
1633                         goto out_dio;
1634                 }
1635         } else {
1636                 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1637                                                     count, written);
1638                 if (written < 0) {
1639                         ret = written;
1640                         if (ret != -EFAULT || ret != -ENOSPC)
1641                                 mlog_errno(ret);
1642                         goto out;
1643                 }
1644         }
1645
1646 out_dio:
1647         /* buffered aio wouldn't have proper lock coverage today */
1648         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1649
1650         /* 
1651          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1652          * function pointer which is called when o_direct io completes so that
1653          * it can unlock our rw lock.  (it's the clustered equivalent of
1654          * i_alloc_sem; protects truncate from racing with pending ios).
1655          * Unfortunately there are error cases which call end_io and others
1656          * that don't.  so we don't have to unlock the rw_lock if either an
1657          * async dio is going to do it in the future or an end_io after an
1658          * error has already done it.
1659          */
1660         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1661                 rw_level = -1;
1662                 have_alloc_sem = 0;
1663         }
1664
1665 out:
1666         if (rw_level != -1)
1667                 ocfs2_rw_unlock(inode, rw_level);
1668
1669 out_sems:
1670         if (have_alloc_sem)
1671                 up_read(&inode->i_alloc_sem);
1672
1673         if (written > 0 && sync) {
1674                 ssize_t err;
1675
1676                 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1677                 if (err < 0)
1678                         written = err;
1679         }
1680
1681         mutex_unlock(&inode->i_mutex);
1682
1683         mlog_exit(ret);
1684         return written ? written : ret;
1685 }
1686
1687 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1688                                     struct pipe_buffer *buf,
1689                                     struct splice_desc *sd)
1690 {
1691         int ret, count;
1692         ssize_t copied = 0;
1693         struct file *file = sd->u.file;
1694         unsigned int offset;
1695         struct page *page = NULL;
1696         void *fsdata;
1697         char *src, *dst;
1698
1699         ret = buf->ops->confirm(pipe, buf);
1700         if (ret)
1701                 goto out;
1702
1703         offset = sd->pos & ~PAGE_CACHE_MASK;
1704         count = sd->len;
1705         if (count + offset > PAGE_CACHE_SIZE)
1706                 count = PAGE_CACHE_SIZE - offset;
1707
1708         ret = ocfs2_write_begin(file, file->f_mapping, sd->pos, count, 0,
1709                                 &page, &fsdata);
1710         if (ret) {
1711                 mlog_errno(ret);
1712                 goto out;
1713         }
1714
1715         src = buf->ops->map(pipe, buf, 1);
1716         dst = kmap_atomic(page, KM_USER1);
1717         memcpy(dst + offset, src + buf->offset, count);
1718         kunmap_atomic(page, KM_USER1);
1719         buf->ops->unmap(pipe, buf, src);
1720
1721         copied = ocfs2_write_end(file, file->f_mapping, sd->pos, count, count,
1722                                  page, fsdata);
1723         if (copied < 0) {
1724                 mlog_errno(copied);
1725                 ret = copied;
1726                 goto out;
1727         }
1728 out:
1729
1730         return copied ? copied : ret;
1731 }
1732
1733 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1734                                          struct file *out,
1735                                          loff_t *ppos,
1736                                          size_t len,
1737                                          unsigned int flags)
1738 {
1739         int ret, err;
1740         struct address_space *mapping = out->f_mapping;
1741         struct inode *inode = mapping->host;
1742         struct splice_desc sd = {
1743                 .total_len = len,
1744                 .flags = flags,
1745                 .pos = *ppos,
1746                 .u.file = out,
1747         };
1748
1749         ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
1750         if (ret > 0) {
1751                 *ppos += ret;
1752
1753                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1754                         err = generic_osync_inode(inode, mapping,
1755                                                   OSYNC_METADATA|OSYNC_DATA);
1756                         if (err)
1757                                 ret = err;
1758                 }
1759         }
1760
1761         return ret;
1762 }
1763
1764 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1765                                        struct file *out,
1766                                        loff_t *ppos,
1767                                        size_t len,
1768                                        unsigned int flags)
1769 {
1770         int ret;
1771         struct inode *inode = out->f_path.dentry->d_inode;
1772
1773         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1774                    (unsigned int)len,
1775                    out->f_path.dentry->d_name.len,
1776                    out->f_path.dentry->d_name.name);
1777
1778         inode_double_lock(inode, pipe->inode);
1779
1780         ret = ocfs2_rw_lock(inode, 1);
1781         if (ret < 0) {
1782                 mlog_errno(ret);
1783                 goto out;
1784         }
1785
1786         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1787                                             NULL);
1788         if (ret < 0) {
1789                 mlog_errno(ret);
1790                 goto out_unlock;
1791         }
1792
1793         /* ok, we're done with i_size and alloc work */
1794         ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1795
1796 out_unlock:
1797         ocfs2_rw_unlock(inode, 1);
1798 out:
1799         inode_double_unlock(inode, pipe->inode);
1800
1801         mlog_exit(ret);
1802         return ret;
1803 }
1804
1805 static ssize_t ocfs2_file_splice_read(struct file *in,
1806                                       loff_t *ppos,
1807                                       struct pipe_inode_info *pipe,
1808                                       size_t len,
1809                                       unsigned int flags)
1810 {
1811         int ret = 0;
1812         struct inode *inode = in->f_path.dentry->d_inode;
1813
1814         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1815                    (unsigned int)len,
1816                    in->f_path.dentry->d_name.len,
1817                    in->f_path.dentry->d_name.name);
1818
1819         /*
1820          * See the comment in ocfs2_file_aio_read()
1821          */
1822         ret = ocfs2_meta_lock(inode, NULL, 0);
1823         if (ret < 0) {
1824                 mlog_errno(ret);
1825                 goto bail;
1826         }
1827         ocfs2_meta_unlock(inode, 0);
1828
1829         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1830
1831 bail:
1832         mlog_exit(ret);
1833         return ret;
1834 }
1835
1836 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1837                                    const struct iovec *iov,
1838                                    unsigned long nr_segs,
1839                                    loff_t pos)
1840 {
1841         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1842         struct file *filp = iocb->ki_filp;
1843         struct inode *inode = filp->f_path.dentry->d_inode;
1844
1845         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1846                    (unsigned int)nr_segs,
1847                    filp->f_path.dentry->d_name.len,
1848                    filp->f_path.dentry->d_name.name);
1849
1850         if (!inode) {
1851                 ret = -EINVAL;
1852                 mlog_errno(ret);
1853                 goto bail;
1854         }
1855
1856         /* 
1857          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1858          * need locks to protect pending reads from racing with truncate.
1859          */
1860         if (filp->f_flags & O_DIRECT) {
1861                 down_read(&inode->i_alloc_sem);
1862                 have_alloc_sem = 1;
1863
1864                 ret = ocfs2_rw_lock(inode, 0);
1865                 if (ret < 0) {
1866                         mlog_errno(ret);
1867                         goto bail;
1868                 }
1869                 rw_level = 0;
1870                 /* communicate with ocfs2_dio_end_io */
1871                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1872         }
1873
1874         /*
1875          * We're fine letting folks race truncates and extending
1876          * writes with read across the cluster, just like they can
1877          * locally. Hence no rw_lock during read.
1878          * 
1879          * Take and drop the meta data lock to update inode fields
1880          * like i_size. This allows the checks down below
1881          * generic_file_aio_read() a chance of actually working. 
1882          */
1883         ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1884         if (ret < 0) {
1885                 mlog_errno(ret);
1886                 goto bail;
1887         }
1888         ocfs2_meta_unlock(inode, lock_level);
1889
1890         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1891         if (ret == -EINVAL)
1892                 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1893
1894         /* buffered aio wouldn't have proper lock coverage today */
1895         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1896
1897         /* see ocfs2_file_aio_write */
1898         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1899                 rw_level = -1;
1900                 have_alloc_sem = 0;
1901         }
1902
1903 bail:
1904         if (have_alloc_sem)
1905                 up_read(&inode->i_alloc_sem);
1906         if (rw_level != -1) 
1907                 ocfs2_rw_unlock(inode, rw_level);
1908         mlog_exit(ret);
1909
1910         return ret;
1911 }
1912
1913 const struct inode_operations ocfs2_file_iops = {
1914         .setattr        = ocfs2_setattr,
1915         .getattr        = ocfs2_getattr,
1916         .permission     = ocfs2_permission,
1917 };
1918
1919 const struct inode_operations ocfs2_special_file_iops = {
1920         .setattr        = ocfs2_setattr,
1921         .getattr        = ocfs2_getattr,
1922         .permission     = ocfs2_permission,
1923 };
1924
1925 const struct file_operations ocfs2_fops = {
1926         .read           = do_sync_read,
1927         .write          = do_sync_write,
1928         .mmap           = ocfs2_mmap,
1929         .fsync          = ocfs2_sync_file,
1930         .release        = ocfs2_file_release,
1931         .open           = ocfs2_file_open,
1932         .aio_read       = ocfs2_file_aio_read,
1933         .aio_write      = ocfs2_file_aio_write,
1934         .ioctl          = ocfs2_ioctl,
1935 #ifdef CONFIG_COMPAT
1936         .compat_ioctl   = ocfs2_compat_ioctl,
1937 #endif
1938         .splice_read    = ocfs2_file_splice_read,
1939         .splice_write   = ocfs2_file_splice_write,
1940 };
1941
1942 const struct file_operations ocfs2_dops = {
1943         .read           = generic_read_dir,
1944         .readdir        = ocfs2_readdir,
1945         .fsync          = ocfs2_sync_file,
1946         .ioctl          = ocfs2_ioctl,
1947 #ifdef CONFIG_COMPAT
1948         .compat_ioctl   = ocfs2_compat_ioctl,
1949 #endif
1950 };