splice: divorce the splice structure/function definitions from the pipe header
[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         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
330         truncate_inode_pages(inode->i_mapping, new_i_size);
331
332         fe = (struct ocfs2_dinode *) di_bh->b_data;
333         if (!OCFS2_IS_VALID_DINODE(fe)) {
334                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
335                 status = -EIO;
336                 goto bail;
337         }
338
339         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
340                         "Inode %llu, inode i_size = %lld != di "
341                         "i_size = %llu, i_flags = 0x%x\n",
342                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
343                         i_size_read(inode),
344                         (unsigned long long)le64_to_cpu(fe->i_size),
345                         le32_to_cpu(fe->i_flags));
346
347         if (new_i_size > le64_to_cpu(fe->i_size)) {
348                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
349                      (unsigned long long)le64_to_cpu(fe->i_size),
350                      (unsigned long long)new_i_size);
351                 status = -EINVAL;
352                 mlog_errno(status);
353                 goto bail;
354         }
355
356         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
357              (unsigned long long)le64_to_cpu(fe->i_blkno),
358              (unsigned long long)le64_to_cpu(fe->i_size),
359              (unsigned long long)new_i_size);
360
361         /* lets handle the simple truncate cases before doing any more
362          * cluster locking. */
363         if (new_i_size == le64_to_cpu(fe->i_size))
364                 goto bail;
365
366         /* This forces other nodes to sync and drop their pages. Do
367          * this even if we have a truncate without allocation change -
368          * ocfs2 cluster sizes can be much greater than page size, so
369          * we have to truncate them anyway.  */
370         status = ocfs2_data_lock(inode, 1);
371         if (status < 0) {
372                 mlog_errno(status);
373                 goto bail;
374         }
375
376         /* alright, we're going to need to do a full blown alloc size
377          * change. Orphan the inode so that recovery can complete the
378          * truncate if necessary. This does the task of marking
379          * i_size. */
380         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
381         if (status < 0) {
382                 mlog_errno(status);
383                 goto bail_unlock_data;
384         }
385
386         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
387         if (status < 0) {
388                 mlog_errno(status);
389                 goto bail_unlock_data;
390         }
391
392         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
393         if (status < 0) {
394                 mlog_errno(status);
395                 goto bail_unlock_data;
396         }
397
398         /* TODO: orphan dir cleanup here. */
399 bail_unlock_data:
400         ocfs2_data_unlock(inode, 1);
401
402 bail:
403
404         mlog_exit(status);
405         return status;
406 }
407
408 /*
409  * extend allocation only here.
410  * we'll update all the disk stuff, and oip->alloc_size
411  *
412  * expect stuff to be locked, a transaction started and enough data /
413  * metadata reservations in the contexts.
414  *
415  * Will return -EAGAIN, and a reason if a restart is needed.
416  * If passed in, *reason will always be set, even in error.
417  */
418 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
419                                struct inode *inode,
420                                u32 *logical_offset,
421                                u32 clusters_to_add,
422                                struct buffer_head *fe_bh,
423                                handle_t *handle,
424                                struct ocfs2_alloc_context *data_ac,
425                                struct ocfs2_alloc_context *meta_ac,
426                                enum ocfs2_alloc_restarted *reason_ret)
427 {
428         int status = 0;
429         int free_extents;
430         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
431         enum ocfs2_alloc_restarted reason = RESTART_NONE;
432         u32 bit_off, num_bits;
433         u64 block;
434
435         BUG_ON(!clusters_to_add);
436
437         free_extents = ocfs2_num_free_extents(osb, inode, fe);
438         if (free_extents < 0) {
439                 status = free_extents;
440                 mlog_errno(status);
441                 goto leave;
442         }
443
444         /* there are two cases which could cause us to EAGAIN in the
445          * we-need-more-metadata case:
446          * 1) we haven't reserved *any*
447          * 2) we are so fragmented, we've needed to add metadata too
448          *    many times. */
449         if (!free_extents && !meta_ac) {
450                 mlog(0, "we haven't reserved any metadata!\n");
451                 status = -EAGAIN;
452                 reason = RESTART_META;
453                 goto leave;
454         } else if ((!free_extents)
455                    && (ocfs2_alloc_context_bits_left(meta_ac)
456                        < ocfs2_extend_meta_needed(fe))) {
457                 mlog(0, "filesystem is really fragmented...\n");
458                 status = -EAGAIN;
459                 reason = RESTART_META;
460                 goto leave;
461         }
462
463         status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
464                                       &bit_off, &num_bits);
465         if (status < 0) {
466                 if (status != -ENOSPC)
467                         mlog_errno(status);
468                 goto leave;
469         }
470
471         BUG_ON(num_bits > clusters_to_add);
472
473         /* reserve our write early -- insert_extent may update the inode */
474         status = ocfs2_journal_access(handle, inode, fe_bh,
475                                       OCFS2_JOURNAL_ACCESS_WRITE);
476         if (status < 0) {
477                 mlog_errno(status);
478                 goto leave;
479         }
480
481         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
482         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
483              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
484         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
485                                      *logical_offset, block, num_bits,
486                                      meta_ac);
487         if (status < 0) {
488                 mlog_errno(status);
489                 goto leave;
490         }
491
492         status = ocfs2_journal_dirty(handle, fe_bh);
493         if (status < 0) {
494                 mlog_errno(status);
495                 goto leave;
496         }
497
498         clusters_to_add -= num_bits;
499         *logical_offset += num_bits;
500
501         if (clusters_to_add) {
502                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
503                      "%u\n", fe->i_clusters, clusters_to_add);
504                 status = -EAGAIN;
505                 reason = RESTART_TRANS;
506         }
507
508 leave:
509         mlog_exit(status);
510         if (reason_ret)
511                 *reason_ret = reason;
512         return status;
513 }
514
515 /*
516  * For a given allocation, determine which allocators will need to be
517  * accessed, and lock them, reserving the appropriate number of bits.
518  *
519  * Called from ocfs2_extend_allocation() for file systems which don't
520  * support holes, and from ocfs2_write() for file systems which
521  * understand sparse inodes.
522  */
523 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
524                           u32 clusters_to_add,
525                           struct ocfs2_alloc_context **data_ac,
526                           struct ocfs2_alloc_context **meta_ac)
527 {
528         int ret, num_free_extents;
529         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
530
531         *meta_ac = NULL;
532         *data_ac = NULL;
533
534         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
535              "clusters_to_add = %u\n",
536              (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
537              le32_to_cpu(di->i_clusters), clusters_to_add);
538
539         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
540         if (num_free_extents < 0) {
541                 ret = num_free_extents;
542                 mlog_errno(ret);
543                 goto out;
544         }
545
546         /*
547          * Sparse allocation file systems need to be more conservative
548          * with reserving room for expansion - the actual allocation
549          * happens while we've got a journal handle open so re-taking
550          * a cluster lock (because we ran out of room for another
551          * extent) will violate ordering rules.
552          *
553          * Most of the time we'll only be seeing this 1 cluster at a time
554          * anyway.
555          */
556         if (!num_free_extents ||
557             (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
558                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
559                 if (ret < 0) {
560                         if (ret != -ENOSPC)
561                                 mlog_errno(ret);
562                         goto out;
563                 }
564         }
565
566         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
567         if (ret < 0) {
568                 if (ret != -ENOSPC)
569                         mlog_errno(ret);
570                 goto out;
571         }
572
573 out:
574         if (ret) {
575                 if (*meta_ac) {
576                         ocfs2_free_alloc_context(*meta_ac);
577                         *meta_ac = NULL;
578                 }
579
580                 /*
581                  * We cannot have an error and a non null *data_ac.
582                  */
583         }
584
585         return ret;
586 }
587
588 static int ocfs2_extend_allocation(struct inode *inode,
589                                    u32 clusters_to_add)
590 {
591         int status = 0;
592         int restart_func = 0;
593         int drop_alloc_sem = 0;
594         int credits;
595         u32 prev_clusters, logical_start;
596         struct buffer_head *bh = NULL;
597         struct ocfs2_dinode *fe = NULL;
598         handle_t *handle = NULL;
599         struct ocfs2_alloc_context *data_ac = NULL;
600         struct ocfs2_alloc_context *meta_ac = NULL;
601         enum ocfs2_alloc_restarted why;
602         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
603
604         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
605
606         /*
607          * This function only exists for file systems which don't
608          * support holes.
609          */
610         BUG_ON(ocfs2_sparse_alloc(osb));
611
612         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
613                                   OCFS2_BH_CACHED, inode);
614         if (status < 0) {
615                 mlog_errno(status);
616                 goto leave;
617         }
618
619         fe = (struct ocfs2_dinode *) bh->b_data;
620         if (!OCFS2_IS_VALID_DINODE(fe)) {
621                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
622                 status = -EIO;
623                 goto leave;
624         }
625
626         logical_start = OCFS2_I(inode)->ip_clusters;
627
628 restart_all:
629         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
630
631         /* blocks peope in read/write from reading our allocation
632          * until we're done changing it. We depend on i_mutex to block
633          * other extend/truncate calls while we're here. Ordering wrt
634          * start_trans is important here -- always do it before! */
635         down_write(&OCFS2_I(inode)->ip_alloc_sem);
636         drop_alloc_sem = 1;
637
638         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
639                                        &meta_ac);
640         if (status) {
641                 mlog_errno(status);
642                 goto leave;
643         }
644
645         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
646         handle = ocfs2_start_trans(osb, credits);
647         if (IS_ERR(handle)) {
648                 status = PTR_ERR(handle);
649                 handle = NULL;
650                 mlog_errno(status);
651                 goto leave;
652         }
653
654 restarted_transaction:
655         /* reserve a write to the file entry early on - that we if we
656          * run out of credits in the allocation path, we can still
657          * update i_size. */
658         status = ocfs2_journal_access(handle, inode, bh,
659                                       OCFS2_JOURNAL_ACCESS_WRITE);
660         if (status < 0) {
661                 mlog_errno(status);
662                 goto leave;
663         }
664
665         prev_clusters = OCFS2_I(inode)->ip_clusters;
666
667         status = ocfs2_do_extend_allocation(osb,
668                                             inode,
669                                             &logical_start,
670                                             clusters_to_add,
671                                             bh,
672                                             handle,
673                                             data_ac,
674                                             meta_ac,
675                                             &why);
676         if ((status < 0) && (status != -EAGAIN)) {
677                 if (status != -ENOSPC)
678                         mlog_errno(status);
679                 goto leave;
680         }
681
682         status = ocfs2_journal_dirty(handle, bh);
683         if (status < 0) {
684                 mlog_errno(status);
685                 goto leave;
686         }
687
688         spin_lock(&OCFS2_I(inode)->ip_lock);
689         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
690         spin_unlock(&OCFS2_I(inode)->ip_lock);
691
692         if (why != RESTART_NONE && clusters_to_add) {
693                 if (why == RESTART_META) {
694                         mlog(0, "restarting function.\n");
695                         restart_func = 1;
696                 } else {
697                         BUG_ON(why != RESTART_TRANS);
698
699                         mlog(0, "restarting transaction.\n");
700                         /* TODO: This can be more intelligent. */
701                         credits = ocfs2_calc_extend_credits(osb->sb,
702                                                             fe,
703                                                             clusters_to_add);
704                         status = ocfs2_extend_trans(handle, credits);
705                         if (status < 0) {
706                                 /* handle still has to be committed at
707                                  * this point. */
708                                 status = -ENOMEM;
709                                 mlog_errno(status);
710                                 goto leave;
711                         }
712                         goto restarted_transaction;
713                 }
714         }
715
716         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
717              le32_to_cpu(fe->i_clusters),
718              (unsigned long long)le64_to_cpu(fe->i_size));
719         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
720              OCFS2_I(inode)->ip_clusters, i_size_read(inode));
721
722 leave:
723         if (drop_alloc_sem) {
724                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
725                 drop_alloc_sem = 0;
726         }
727         if (handle) {
728                 ocfs2_commit_trans(osb, handle);
729                 handle = NULL;
730         }
731         if (data_ac) {
732                 ocfs2_free_alloc_context(data_ac);
733                 data_ac = NULL;
734         }
735         if (meta_ac) {
736                 ocfs2_free_alloc_context(meta_ac);
737                 meta_ac = NULL;
738         }
739         if ((!status) && restart_func) {
740                 restart_func = 0;
741                 goto restart_all;
742         }
743         if (bh) {
744                 brelse(bh);
745                 bh = NULL;
746         }
747
748         mlog_exit(status);
749         return status;
750 }
751
752 /* Some parts of this taken from generic_cont_expand, which turned out
753  * to be too fragile to do exactly what we need without us having to
754  * worry about recursive locking in ->prepare_write() and
755  * ->commit_write(). */
756 static int ocfs2_write_zero_page(struct inode *inode,
757                                  u64 size)
758 {
759         struct address_space *mapping = inode->i_mapping;
760         struct page *page;
761         unsigned long index;
762         unsigned int offset;
763         handle_t *handle = NULL;
764         int ret;
765
766         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
767         /* ugh.  in prepare/commit_write, if from==to==start of block, we 
768         ** skip the prepare.  make sure we never send an offset for the start
769         ** of a block
770         */
771         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
772                 offset++;
773         }
774         index = size >> PAGE_CACHE_SHIFT;
775
776         page = grab_cache_page(mapping, index);
777         if (!page) {
778                 ret = -ENOMEM;
779                 mlog_errno(ret);
780                 goto out;
781         }
782
783         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
784         if (ret < 0) {
785                 mlog_errno(ret);
786                 goto out_unlock;
787         }
788
789         if (ocfs2_should_order_data(inode)) {
790                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
791                                                      offset);
792                 if (IS_ERR(handle)) {
793                         ret = PTR_ERR(handle);
794                         handle = NULL;
795                         goto out_unlock;
796                 }
797         }
798
799         /* must not update i_size! */
800         ret = block_commit_write(page, offset, offset);
801         if (ret < 0)
802                 mlog_errno(ret);
803         else
804                 ret = 0;
805
806         if (handle)
807                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
808 out_unlock:
809         unlock_page(page);
810         page_cache_release(page);
811 out:
812         return ret;
813 }
814
815 static int ocfs2_zero_extend(struct inode *inode,
816                              u64 zero_to_size)
817 {
818         int ret = 0;
819         u64 start_off;
820         struct super_block *sb = inode->i_sb;
821
822         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
823         while (start_off < zero_to_size) {
824                 ret = ocfs2_write_zero_page(inode, start_off);
825                 if (ret < 0) {
826                         mlog_errno(ret);
827                         goto out;
828                 }
829
830                 start_off += sb->s_blocksize;
831
832                 /*
833                  * Very large extends have the potential to lock up
834                  * the cpu for extended periods of time.
835                  */
836                 cond_resched();
837         }
838
839 out:
840         return ret;
841 }
842
843 /* 
844  * A tail_to_skip value > 0 indicates that we're being called from
845  * ocfs2_file_aio_write(). This has the following implications:
846  *
847  * - we don't want to update i_size
848  * - di_bh will be NULL, which is fine because it's only used in the
849  *   case where we want to update i_size.
850  * - ocfs2_zero_extend() will then only be filling the hole created
851  *   between i_size and the start of the write.
852  */
853 static int ocfs2_extend_file(struct inode *inode,
854                              struct buffer_head *di_bh,
855                              u64 new_i_size,
856                              size_t tail_to_skip)
857 {
858         int ret = 0;
859         u32 clusters_to_add = 0;
860
861         BUG_ON(!tail_to_skip && !di_bh);
862
863         /* setattr sometimes calls us like this. */
864         if (new_i_size == 0)
865                 goto out;
866
867         if (i_size_read(inode) == new_i_size)
868                 goto out;
869         BUG_ON(new_i_size < i_size_read(inode));
870
871         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
872                 BUG_ON(tail_to_skip != 0);
873                 goto out_update_size;
874         }
875
876         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) - 
877                 OCFS2_I(inode)->ip_clusters;
878
879         /* 
880          * protect the pages that ocfs2_zero_extend is going to be
881          * pulling into the page cache.. we do this before the
882          * metadata extend so that we don't get into the situation
883          * where we've extended the metadata but can't get the data
884          * lock to zero.
885          */
886         ret = ocfs2_data_lock(inode, 1);
887         if (ret < 0) {
888                 mlog_errno(ret);
889                 goto out;
890         }
891
892         if (clusters_to_add) {
893                 ret = ocfs2_extend_allocation(inode, clusters_to_add);
894                 if (ret < 0) {
895                         mlog_errno(ret);
896                         goto out_unlock;
897                 }
898         }
899
900         /*
901          * Call this even if we don't add any clusters to the tree. We
902          * still need to zero the area between the old i_size and the
903          * new i_size.
904          */
905         ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
906         if (ret < 0) {
907                 mlog_errno(ret);
908                 goto out_unlock;
909         }
910
911 out_update_size:
912         if (!tail_to_skip) {
913                 /* We're being called from ocfs2_setattr() which wants
914                  * us to update i_size */
915                 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
916                 if (ret < 0)
917                         mlog_errno(ret);
918         }
919
920 out_unlock:
921         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
922                 ocfs2_data_unlock(inode, 1);
923
924 out:
925         return ret;
926 }
927
928 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
929 {
930         int status = 0, size_change;
931         struct inode *inode = dentry->d_inode;
932         struct super_block *sb = inode->i_sb;
933         struct ocfs2_super *osb = OCFS2_SB(sb);
934         struct buffer_head *bh = NULL;
935         handle_t *handle = NULL;
936
937         mlog_entry("(0x%p, '%.*s')\n", dentry,
938                    dentry->d_name.len, dentry->d_name.name);
939
940         if (attr->ia_valid & ATTR_MODE)
941                 mlog(0, "mode change: %d\n", attr->ia_mode);
942         if (attr->ia_valid & ATTR_UID)
943                 mlog(0, "uid change: %d\n", attr->ia_uid);
944         if (attr->ia_valid & ATTR_GID)
945                 mlog(0, "gid change: %d\n", attr->ia_gid);
946         if (attr->ia_valid & ATTR_SIZE)
947                 mlog(0, "size change...\n");
948         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
949                 mlog(0, "time change...\n");
950
951 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
952                            | ATTR_GID | ATTR_UID | ATTR_MODE)
953         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
954                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
955                 return 0;
956         }
957
958         status = inode_change_ok(inode, attr);
959         if (status)
960                 return status;
961
962         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
963         if (size_change) {
964                 status = ocfs2_rw_lock(inode, 1);
965                 if (status < 0) {
966                         mlog_errno(status);
967                         goto bail;
968                 }
969         }
970
971         status = ocfs2_meta_lock(inode, &bh, 1);
972         if (status < 0) {
973                 if (status != -ENOENT)
974                         mlog_errno(status);
975                 goto bail_unlock_rw;
976         }
977
978         if (size_change && attr->ia_size != i_size_read(inode)) {
979                 if (i_size_read(inode) > attr->ia_size)
980                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
981                 else
982                         status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
983                 if (status < 0) {
984                         if (status != -ENOSPC)
985                                 mlog_errno(status);
986                         status = -ENOSPC;
987                         goto bail_unlock;
988                 }
989         }
990
991         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
992         if (IS_ERR(handle)) {
993                 status = PTR_ERR(handle);
994                 mlog_errno(status);
995                 goto bail_unlock;
996         }
997
998         status = inode_setattr(inode, attr);
999         if (status < 0) {
1000                 mlog_errno(status);
1001                 goto bail_commit;
1002         }
1003
1004         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1005         if (status < 0)
1006                 mlog_errno(status);
1007
1008 bail_commit:
1009         ocfs2_commit_trans(osb, handle);
1010 bail_unlock:
1011         ocfs2_meta_unlock(inode, 1);
1012 bail_unlock_rw:
1013         if (size_change)
1014                 ocfs2_rw_unlock(inode, 1);
1015 bail:
1016         if (bh)
1017                 brelse(bh);
1018
1019         mlog_exit(status);
1020         return status;
1021 }
1022
1023 int ocfs2_getattr(struct vfsmount *mnt,
1024                   struct dentry *dentry,
1025                   struct kstat *stat)
1026 {
1027         struct inode *inode = dentry->d_inode;
1028         struct super_block *sb = dentry->d_inode->i_sb;
1029         struct ocfs2_super *osb = sb->s_fs_info;
1030         int err;
1031
1032         mlog_entry_void();
1033
1034         err = ocfs2_inode_revalidate(dentry);
1035         if (err) {
1036                 if (err != -ENOENT)
1037                         mlog_errno(err);
1038                 goto bail;
1039         }
1040
1041         generic_fillattr(inode, stat);
1042
1043         /* We set the blksize from the cluster size for performance */
1044         stat->blksize = osb->s_clustersize;
1045
1046 bail:
1047         mlog_exit(err);
1048
1049         return err;
1050 }
1051
1052 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1053 {
1054         int ret;
1055
1056         mlog_entry_void();
1057
1058         ret = ocfs2_meta_lock(inode, NULL, 0);
1059         if (ret) {
1060                 if (ret != -ENOENT)
1061                         mlog_errno(ret);
1062                 goto out;
1063         }
1064
1065         ret = generic_permission(inode, mask, NULL);
1066
1067         ocfs2_meta_unlock(inode, 0);
1068 out:
1069         mlog_exit(ret);
1070         return ret;
1071 }
1072
1073 static int ocfs2_write_remove_suid(struct inode *inode)
1074 {
1075         int ret;
1076         struct buffer_head *bh = NULL;
1077         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1078         handle_t *handle;
1079         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1080         struct ocfs2_dinode *di;
1081
1082         mlog_entry("(Inode %llu, mode 0%o)\n",
1083                    (unsigned long long)oi->ip_blkno, inode->i_mode);
1084
1085         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1086         if (handle == NULL) {
1087                 ret = -ENOMEM;
1088                 mlog_errno(ret);
1089                 goto out;
1090         }
1091
1092         ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1093         if (ret < 0) {
1094                 mlog_errno(ret);
1095                 goto out_trans;
1096         }
1097
1098         ret = ocfs2_journal_access(handle, inode, bh,
1099                                    OCFS2_JOURNAL_ACCESS_WRITE);
1100         if (ret < 0) {
1101                 mlog_errno(ret);
1102                 goto out_bh;
1103         }
1104
1105         inode->i_mode &= ~S_ISUID;
1106         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1107                 inode->i_mode &= ~S_ISGID;
1108
1109         di = (struct ocfs2_dinode *) bh->b_data;
1110         di->i_mode = cpu_to_le16(inode->i_mode);
1111
1112         ret = ocfs2_journal_dirty(handle, bh);
1113         if (ret < 0)
1114                 mlog_errno(ret);
1115 out_bh:
1116         brelse(bh);
1117 out_trans:
1118         ocfs2_commit_trans(osb, handle);
1119 out:
1120         mlog_exit(ret);
1121         return ret;
1122 }
1123
1124 /*
1125  * Will look for holes and unwritten extents in the range starting at
1126  * pos for count bytes (inclusive).
1127  */
1128 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1129                                        size_t count)
1130 {
1131         int ret = 0;
1132         unsigned int extent_flags;
1133         u32 cpos, clusters, extent_len, phys_cpos;
1134         struct super_block *sb = inode->i_sb;
1135
1136         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1137         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1138
1139         while (clusters) {
1140                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1141                                          &extent_flags);
1142                 if (ret < 0) {
1143                         mlog_errno(ret);
1144                         goto out;
1145                 }
1146
1147                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1148                         ret = 1;
1149                         break;
1150                 }
1151
1152                 if (extent_len > clusters)
1153                         extent_len = clusters;
1154
1155                 clusters -= extent_len;
1156                 cpos += extent_len;
1157         }
1158 out:
1159         return ret;
1160 }
1161
1162 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1163                                          loff_t *ppos,
1164                                          size_t count,
1165                                          int appending,
1166                                          int *direct_io)
1167 {
1168         int ret = 0, meta_level = appending;
1169         struct inode *inode = dentry->d_inode;
1170         u32 clusters;
1171         loff_t newsize, saved_pos;
1172
1173         /* 
1174          * We sample i_size under a read level meta lock to see if our write
1175          * is extending the file, if it is we back off and get a write level
1176          * meta lock.
1177          */
1178         for(;;) {
1179                 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1180                 if (ret < 0) {
1181                         meta_level = -1;
1182                         mlog_errno(ret);
1183                         goto out;
1184                 }
1185
1186                 /* Clear suid / sgid if necessary. We do this here
1187                  * instead of later in the write path because
1188                  * remove_suid() calls ->setattr without any hint that
1189                  * we may have already done our cluster locking. Since
1190                  * ocfs2_setattr() *must* take cluster locks to
1191                  * proceeed, this will lead us to recursively lock the
1192                  * inode. There's also the dinode i_size state which
1193                  * can be lost via setattr during extending writes (we
1194                  * set inode->i_size at the end of a write. */
1195                 if (should_remove_suid(dentry)) {
1196                         if (meta_level == 0) {
1197                                 ocfs2_meta_unlock(inode, meta_level);
1198                                 meta_level = 1;
1199                                 continue;
1200                         }
1201
1202                         ret = ocfs2_write_remove_suid(inode);
1203                         if (ret < 0) {
1204                                 mlog_errno(ret);
1205                                 goto out_unlock;
1206                         }
1207                 }
1208
1209                 /* work on a copy of ppos until we're sure that we won't have
1210                  * to recalculate it due to relocking. */
1211                 if (appending) {
1212                         saved_pos = i_size_read(inode);
1213                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1214                 } else {
1215                         saved_pos = *ppos;
1216                 }
1217
1218                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1219                         loff_t end = saved_pos + count;
1220
1221                         /*
1222                          * Skip the O_DIRECT checks if we don't need
1223                          * them.
1224                          */
1225                         if (!direct_io || !(*direct_io))
1226                                 break;
1227
1228                         /*
1229                          * Allowing concurrent direct writes means
1230                          * i_size changes wouldn't be synchronized, so
1231                          * one node could wind up truncating another
1232                          * nodes writes.
1233                          */
1234                         if (end > i_size_read(inode)) {
1235                                 *direct_io = 0;
1236                                 break;
1237                         }
1238
1239                         /*
1240                          * We don't fill holes during direct io, so
1241                          * check for them here. If any are found, the
1242                          * caller will have to retake some cluster
1243                          * locks and initiate the io as buffered.
1244                          */
1245                         ret = ocfs2_check_range_for_holes(inode, saved_pos,
1246                                                           count);
1247                         if (ret == 1) {
1248                                 *direct_io = 0;
1249                                 ret = 0;
1250                         } else if (ret < 0)
1251                                 mlog_errno(ret);
1252                         break;
1253                 }
1254
1255                 /*
1256                  * The rest of this loop is concerned with legacy file
1257                  * systems which don't support sparse files.
1258                  */
1259
1260                 newsize = count + saved_pos;
1261
1262                 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1263                      (long long) saved_pos, (long long) newsize,
1264                      (long long) i_size_read(inode));
1265
1266                 /* No need for a higher level metadata lock if we're
1267                  * never going past i_size. */
1268                 if (newsize <= i_size_read(inode))
1269                         break;
1270
1271                 if (meta_level == 0) {
1272                         ocfs2_meta_unlock(inode, meta_level);
1273                         meta_level = 1;
1274                         continue;
1275                 }
1276
1277                 spin_lock(&OCFS2_I(inode)->ip_lock);
1278                 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1279                         OCFS2_I(inode)->ip_clusters;
1280                 spin_unlock(&OCFS2_I(inode)->ip_lock);
1281
1282                 mlog(0, "Writing at EOF, may need more allocation: "
1283                      "i_size = %lld, newsize = %lld, need %u clusters\n",
1284                      (long long) i_size_read(inode), (long long) newsize,
1285                      clusters);
1286
1287                 /* We only want to continue the rest of this loop if
1288                  * our extend will actually require more
1289                  * allocation. */
1290                 if (!clusters)
1291                         break;
1292
1293                 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1294                 if (ret < 0) {
1295                         if (ret != -ENOSPC)
1296                                 mlog_errno(ret);
1297                         goto out_unlock;
1298                 }
1299                 break;
1300         }
1301
1302         if (appending)
1303                 *ppos = saved_pos;
1304
1305 out_unlock:
1306         ocfs2_meta_unlock(inode, meta_level);
1307
1308 out:
1309         return ret;
1310 }
1311
1312 static inline void
1313 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1314 {
1315         const struct iovec *iov = *iovp;
1316         size_t base = *basep;
1317
1318         do {
1319                 int copy = min(bytes, iov->iov_len - base);
1320
1321                 bytes -= copy;
1322                 base += copy;
1323                 if (iov->iov_len == base) {
1324                         iov++;
1325                         base = 0;
1326                 }
1327         } while (bytes);
1328         *iovp = iov;
1329         *basep = base;
1330 }
1331
1332 static struct page * ocfs2_get_write_source(struct ocfs2_buffered_write_priv *bp,
1333                                             const struct iovec *cur_iov,
1334                                             size_t iov_offset)
1335 {
1336         int ret;
1337         char *buf;
1338         struct page *src_page = NULL;
1339
1340         buf = cur_iov->iov_base + iov_offset;
1341
1342         if (!segment_eq(get_fs(), KERNEL_DS)) {
1343                 /*
1344                  * Pull in the user page. We want to do this outside
1345                  * of the meta data locks in order to preserve locking
1346                  * order in case of page fault.
1347                  */
1348                 ret = get_user_pages(current, current->mm,
1349                                      (unsigned long)buf & PAGE_CACHE_MASK, 1,
1350                                      0, 0, &src_page, NULL);
1351                 if (ret == 1)
1352                         bp->b_src_buf = kmap(src_page);
1353                 else
1354                         src_page = ERR_PTR(-EFAULT);
1355         } else {
1356                 bp->b_src_buf = buf;
1357         }
1358
1359         return src_page;
1360 }
1361
1362 static void ocfs2_put_write_source(struct ocfs2_buffered_write_priv *bp,
1363                                    struct page *page)
1364 {
1365         if (page) {
1366                 kunmap(page);
1367                 page_cache_release(page);
1368         }
1369 }
1370
1371 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1372                                          const struct iovec *iov,
1373                                          unsigned long nr_segs,
1374                                          size_t count,
1375                                          ssize_t o_direct_written)
1376 {
1377         int ret = 0;
1378         ssize_t copied, total = 0;
1379         size_t iov_offset = 0;
1380         const struct iovec *cur_iov = iov;
1381         struct ocfs2_buffered_write_priv bp;
1382         struct page *page;
1383
1384         /*
1385          * handle partial DIO write.  Adjust cur_iov if needed.
1386          */
1387         ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1388
1389         do {
1390                 bp.b_cur_off = iov_offset;
1391                 bp.b_cur_iov = cur_iov;
1392
1393                 page = ocfs2_get_write_source(&bp, cur_iov, iov_offset);
1394                 if (IS_ERR(page)) {
1395                         ret = PTR_ERR(page);
1396                         goto out;
1397                 }
1398
1399                 copied = ocfs2_buffered_write_cluster(file, *ppos, count,
1400                                                       ocfs2_map_and_write_user_data,
1401                                                       &bp);
1402
1403                 ocfs2_put_write_source(&bp, page);
1404
1405                 if (copied < 0) {
1406                         mlog_errno(copied);
1407                         ret = copied;
1408                         goto out;
1409                 }
1410
1411                 total += copied;
1412                 *ppos = *ppos + copied;
1413                 count -= copied;
1414
1415                 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1416         } while(count);
1417
1418 out:
1419         return total ? total : ret;
1420 }
1421
1422 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1423                                     const struct iovec *iov,
1424                                     unsigned long nr_segs,
1425                                     loff_t pos)
1426 {
1427         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1428         int can_do_direct, sync = 0;
1429         ssize_t written = 0;
1430         size_t ocount;          /* original count */
1431         size_t count;           /* after file limit checks */
1432         loff_t *ppos = &iocb->ki_pos;
1433         struct file *file = iocb->ki_filp;
1434         struct inode *inode = file->f_path.dentry->d_inode;
1435
1436         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1437                    (unsigned int)nr_segs,
1438                    file->f_path.dentry->d_name.len,
1439                    file->f_path.dentry->d_name.name);
1440
1441         if (iocb->ki_left == 0)
1442                 return 0;
1443
1444         ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1445         if (ret)
1446                 return ret;
1447
1448         count = ocount;
1449
1450         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1451
1452         appending = file->f_flags & O_APPEND ? 1 : 0;
1453         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1454
1455         mutex_lock(&inode->i_mutex);
1456
1457 relock:
1458         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1459         if (direct_io) {
1460                 down_read(&inode->i_alloc_sem);
1461                 have_alloc_sem = 1;
1462         }
1463
1464         /* concurrent O_DIRECT writes are allowed */
1465         rw_level = !direct_io;
1466         ret = ocfs2_rw_lock(inode, rw_level);
1467         if (ret < 0) {
1468                 mlog_errno(ret);
1469                 goto out_sems;
1470         }
1471
1472         can_do_direct = direct_io;
1473         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1474                                             iocb->ki_left, appending,
1475                                             &can_do_direct);
1476         if (ret < 0) {
1477                 mlog_errno(ret);
1478                 goto out;
1479         }
1480
1481         /*
1482          * We can't complete the direct I/O as requested, fall back to
1483          * buffered I/O.
1484          */
1485         if (direct_io && !can_do_direct) {
1486                 ocfs2_rw_unlock(inode, rw_level);
1487                 up_read(&inode->i_alloc_sem);
1488
1489                 have_alloc_sem = 0;
1490                 rw_level = -1;
1491
1492                 direct_io = 0;
1493                 sync = 1;
1494                 goto relock;
1495         }
1496
1497         if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1498                 sync = 1;
1499
1500         /*
1501          * XXX: Is it ok to execute these checks a second time?
1502          */
1503         ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1504         if (ret)
1505                 goto out;
1506
1507         /*
1508          * Set pos so that sync_page_range_nolock() below understands
1509          * where to start from. We might've moved it around via the
1510          * calls above. The range we want to actually sync starts from
1511          * *ppos here.
1512          *
1513          */
1514         pos = *ppos;
1515
1516         /* communicate with ocfs2_dio_end_io */
1517         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1518
1519         if (direct_io) {
1520                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1521                                                     ppos, count, ocount);
1522                 if (written < 0) {
1523                         ret = written;
1524                         goto out_dio;
1525                 }
1526         } else {
1527                 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1528                                                     count, written);
1529                 if (written < 0) {
1530                         ret = written;
1531                         if (ret != -EFAULT || ret != -ENOSPC)
1532                                 mlog_errno(ret);
1533                         goto out;
1534                 }
1535         }
1536
1537 out_dio:
1538         /* buffered aio wouldn't have proper lock coverage today */
1539         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1540
1541         /* 
1542          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1543          * function pointer which is called when o_direct io completes so that
1544          * it can unlock our rw lock.  (it's the clustered equivalent of
1545          * i_alloc_sem; protects truncate from racing with pending ios).
1546          * Unfortunately there are error cases which call end_io and others
1547          * that don't.  so we don't have to unlock the rw_lock if either an
1548          * async dio is going to do it in the future or an end_io after an
1549          * error has already done it.
1550          */
1551         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1552                 rw_level = -1;
1553                 have_alloc_sem = 0;
1554         }
1555
1556 out:
1557         if (rw_level != -1)
1558                 ocfs2_rw_unlock(inode, rw_level);
1559
1560 out_sems:
1561         if (have_alloc_sem)
1562                 up_read(&inode->i_alloc_sem);
1563
1564         if (written > 0 && sync) {
1565                 ssize_t err;
1566
1567                 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1568                 if (err < 0)
1569                         written = err;
1570         }
1571
1572         mutex_unlock(&inode->i_mutex);
1573
1574         mlog_exit(ret);
1575         return written ? written : ret;
1576 }
1577
1578 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1579                                     struct pipe_buffer *buf,
1580                                     struct splice_desc *sd)
1581 {
1582         int ret, count, total = 0;
1583         ssize_t copied = 0;
1584         struct ocfs2_splice_write_priv sp;
1585
1586         ret = buf->ops->pin(pipe, buf);
1587         if (ret)
1588                 goto out;
1589
1590         sp.s_sd = sd;
1591         sp.s_buf = buf;
1592         sp.s_pipe = pipe;
1593         sp.s_offset = sd->pos & ~PAGE_CACHE_MASK;
1594         sp.s_buf_offset = buf->offset;
1595
1596         count = sd->len;
1597         if (count + sp.s_offset > PAGE_CACHE_SIZE)
1598                 count = PAGE_CACHE_SIZE - sp.s_offset;
1599
1600         do {
1601                 /*
1602                  * splice wants us to copy up to one page at a
1603                  * time. For pagesize > cluster size, this means we
1604                  * might enter ocfs2_buffered_write_cluster() more
1605                  * than once, so keep track of our progress here.
1606                  */
1607                 copied = ocfs2_buffered_write_cluster(sd->file,
1608                                                       (loff_t)sd->pos + total,
1609                                                       count,
1610                                                       ocfs2_map_and_write_splice_data,
1611                                                       &sp);
1612                 if (copied < 0) {
1613                         mlog_errno(copied);
1614                         ret = copied;
1615                         goto out;
1616                 }
1617
1618                 count -= copied;
1619                 sp.s_offset += copied;
1620                 sp.s_buf_offset += copied;
1621                 total += copied;
1622         } while (count);
1623
1624         ret = 0;
1625 out:
1626
1627         return total ? total : ret;
1628 }
1629
1630 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1631                                          struct file *out,
1632                                          loff_t *ppos,
1633                                          size_t len,
1634                                          unsigned int flags)
1635 {
1636         int ret, err;
1637         struct address_space *mapping = out->f_mapping;
1638         struct inode *inode = mapping->host;
1639         struct splice_desc sd = {
1640                 .total_len = len,
1641                 .flags = flags,
1642                 .pos = *ppos,
1643                 .u.file = out,
1644         };
1645
1646         ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
1647         if (ret > 0) {
1648                 *ppos += ret;
1649
1650                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1651                         err = generic_osync_inode(inode, mapping,
1652                                                   OSYNC_METADATA|OSYNC_DATA);
1653                         if (err)
1654                                 ret = err;
1655                 }
1656         }
1657
1658         return ret;
1659 }
1660
1661 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1662                                        struct file *out,
1663                                        loff_t *ppos,
1664                                        size_t len,
1665                                        unsigned int flags)
1666 {
1667         int ret;
1668         struct inode *inode = out->f_path.dentry->d_inode;
1669
1670         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1671                    (unsigned int)len,
1672                    out->f_path.dentry->d_name.len,
1673                    out->f_path.dentry->d_name.name);
1674
1675         inode_double_lock(inode, pipe->inode);
1676
1677         ret = ocfs2_rw_lock(inode, 1);
1678         if (ret < 0) {
1679                 mlog_errno(ret);
1680                 goto out;
1681         }
1682
1683         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1684                                             NULL);
1685         if (ret < 0) {
1686                 mlog_errno(ret);
1687                 goto out_unlock;
1688         }
1689
1690         /* ok, we're done with i_size and alloc work */
1691         ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1692
1693 out_unlock:
1694         ocfs2_rw_unlock(inode, 1);
1695 out:
1696         inode_double_unlock(inode, pipe->inode);
1697
1698         mlog_exit(ret);
1699         return ret;
1700 }
1701
1702 static ssize_t ocfs2_file_splice_read(struct file *in,
1703                                       loff_t *ppos,
1704                                       struct pipe_inode_info *pipe,
1705                                       size_t len,
1706                                       unsigned int flags)
1707 {
1708         int ret = 0;
1709         struct inode *inode = in->f_path.dentry->d_inode;
1710
1711         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1712                    (unsigned int)len,
1713                    in->f_path.dentry->d_name.len,
1714                    in->f_path.dentry->d_name.name);
1715
1716         /*
1717          * See the comment in ocfs2_file_aio_read()
1718          */
1719         ret = ocfs2_meta_lock(inode, NULL, 0);
1720         if (ret < 0) {
1721                 mlog_errno(ret);
1722                 goto bail;
1723         }
1724         ocfs2_meta_unlock(inode, 0);
1725
1726         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1727
1728 bail:
1729         mlog_exit(ret);
1730         return ret;
1731 }
1732
1733 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1734                                    const struct iovec *iov,
1735                                    unsigned long nr_segs,
1736                                    loff_t pos)
1737 {
1738         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1739         struct file *filp = iocb->ki_filp;
1740         struct inode *inode = filp->f_path.dentry->d_inode;
1741
1742         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1743                    (unsigned int)nr_segs,
1744                    filp->f_path.dentry->d_name.len,
1745                    filp->f_path.dentry->d_name.name);
1746
1747         if (!inode) {
1748                 ret = -EINVAL;
1749                 mlog_errno(ret);
1750                 goto bail;
1751         }
1752
1753         /* 
1754          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1755          * need locks to protect pending reads from racing with truncate.
1756          */
1757         if (filp->f_flags & O_DIRECT) {
1758                 down_read(&inode->i_alloc_sem);
1759                 have_alloc_sem = 1;
1760
1761                 ret = ocfs2_rw_lock(inode, 0);
1762                 if (ret < 0) {
1763                         mlog_errno(ret);
1764                         goto bail;
1765                 }
1766                 rw_level = 0;
1767                 /* communicate with ocfs2_dio_end_io */
1768                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1769         }
1770
1771         /*
1772          * We're fine letting folks race truncates and extending
1773          * writes with read across the cluster, just like they can
1774          * locally. Hence no rw_lock during read.
1775          * 
1776          * Take and drop the meta data lock to update inode fields
1777          * like i_size. This allows the checks down below
1778          * generic_file_aio_read() a chance of actually working. 
1779          */
1780         ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1781         if (ret < 0) {
1782                 mlog_errno(ret);
1783                 goto bail;
1784         }
1785         ocfs2_meta_unlock(inode, lock_level);
1786
1787         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1788         if (ret == -EINVAL)
1789                 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1790
1791         /* buffered aio wouldn't have proper lock coverage today */
1792         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1793
1794         /* see ocfs2_file_aio_write */
1795         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1796                 rw_level = -1;
1797                 have_alloc_sem = 0;
1798         }
1799
1800 bail:
1801         if (have_alloc_sem)
1802                 up_read(&inode->i_alloc_sem);
1803         if (rw_level != -1) 
1804                 ocfs2_rw_unlock(inode, rw_level);
1805         mlog_exit(ret);
1806
1807         return ret;
1808 }
1809
1810 const struct inode_operations ocfs2_file_iops = {
1811         .setattr        = ocfs2_setattr,
1812         .getattr        = ocfs2_getattr,
1813         .permission     = ocfs2_permission,
1814 };
1815
1816 const struct inode_operations ocfs2_special_file_iops = {
1817         .setattr        = ocfs2_setattr,
1818         .getattr        = ocfs2_getattr,
1819         .permission     = ocfs2_permission,
1820 };
1821
1822 const struct file_operations ocfs2_fops = {
1823         .read           = do_sync_read,
1824         .write          = do_sync_write,
1825         .mmap           = ocfs2_mmap,
1826         .fsync          = ocfs2_sync_file,
1827         .release        = ocfs2_file_release,
1828         .open           = ocfs2_file_open,
1829         .aio_read       = ocfs2_file_aio_read,
1830         .aio_write      = ocfs2_file_aio_write,
1831         .ioctl          = ocfs2_ioctl,
1832 #ifdef CONFIG_COMPAT
1833         .compat_ioctl   = ocfs2_compat_ioctl,
1834 #endif
1835         .splice_read    = ocfs2_file_splice_read,
1836         .splice_write   = ocfs2_file_splice_write,
1837 };
1838
1839 const struct file_operations ocfs2_dops = {
1840         .read           = generic_read_dir,
1841         .readdir        = ocfs2_readdir,
1842         .fsync          = ocfs2_sync_file,
1843         .ioctl          = ocfs2_ioctl,
1844 #ifdef CONFIG_COMPAT
1845         .compat_ioctl   = ocfs2_compat_ioctl,
1846 #endif
1847 };