Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / fs / ocfs2 / aops.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public
8  * License as published by the Free Software Foundation; either
9  * version 2 of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public
17  * License along with this program; if not, write to the
18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19  * Boston, MA 021110-1307, USA.
20  */
21
22 #include <linux/fs.h>
23 #include <linux/slab.h>
24 #include <linux/highmem.h>
25 #include <linux/pagemap.h>
26 #include <asm/byteorder.h>
27 #include <linux/swap.h>
28 #include <linux/pipe_fs_i.h>
29 #include <linux/mpage.h>
30
31 #define MLOG_MASK_PREFIX ML_FILE_IO
32 #include <cluster/masklog.h>
33
34 #include "ocfs2.h"
35
36 #include "alloc.h"
37 #include "aops.h"
38 #include "dlmglue.h"
39 #include "extent_map.h"
40 #include "file.h"
41 #include "inode.h"
42 #include "journal.h"
43 #include "suballoc.h"
44 #include "super.h"
45 #include "symlink.h"
46
47 #include "buffer_head_io.h"
48
49 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
50                                    struct buffer_head *bh_result, int create)
51 {
52         int err = -EIO;
53         int status;
54         struct ocfs2_dinode *fe = NULL;
55         struct buffer_head *bh = NULL;
56         struct buffer_head *buffer_cache_bh = NULL;
57         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
58         void *kaddr;
59
60         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
61                    (unsigned long long)iblock, bh_result, create);
62
63         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
64
65         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
66                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
67                      (unsigned long long)iblock);
68                 goto bail;
69         }
70
71         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
72                                   OCFS2_I(inode)->ip_blkno,
73                                   &bh, OCFS2_BH_CACHED, inode);
74         if (status < 0) {
75                 mlog_errno(status);
76                 goto bail;
77         }
78         fe = (struct ocfs2_dinode *) bh->b_data;
79
80         if (!OCFS2_IS_VALID_DINODE(fe)) {
81                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
82                      (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
83                      fe->i_signature);
84                 goto bail;
85         }
86
87         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
88                                                     le32_to_cpu(fe->i_clusters))) {
89                 mlog(ML_ERROR, "block offset is outside the allocated size: "
90                      "%llu\n", (unsigned long long)iblock);
91                 goto bail;
92         }
93
94         /* We don't use the page cache to create symlink data, so if
95          * need be, copy it over from the buffer cache. */
96         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
97                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
98                             iblock;
99                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
100                 if (!buffer_cache_bh) {
101                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
102                         goto bail;
103                 }
104
105                 /* we haven't locked out transactions, so a commit
106                  * could've happened. Since we've got a reference on
107                  * the bh, even if it commits while we're doing the
108                  * copy, the data is still good. */
109                 if (buffer_jbd(buffer_cache_bh)
110                     && ocfs2_inode_is_new(inode)) {
111                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
112                         if (!kaddr) {
113                                 mlog(ML_ERROR, "couldn't kmap!\n");
114                                 goto bail;
115                         }
116                         memcpy(kaddr + (bh_result->b_size * iblock),
117                                buffer_cache_bh->b_data,
118                                bh_result->b_size);
119                         kunmap_atomic(kaddr, KM_USER0);
120                         set_buffer_uptodate(bh_result);
121                 }
122                 brelse(buffer_cache_bh);
123         }
124
125         map_bh(bh_result, inode->i_sb,
126                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
127
128         err = 0;
129
130 bail:
131         if (bh)
132                 brelse(bh);
133
134         mlog_exit(err);
135         return err;
136 }
137
138 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
139                            struct buffer_head *bh_result, int create)
140 {
141         int err = 0;
142         unsigned int ext_flags;
143         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
144         u64 p_blkno, count, past_eof;
145         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
146
147         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
148                    (unsigned long long)iblock, bh_result, create);
149
150         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
151                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
152                      inode, inode->i_ino);
153
154         if (S_ISLNK(inode->i_mode)) {
155                 /* this always does I/O for some reason. */
156                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
157                 goto bail;
158         }
159
160         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
161                                           &ext_flags);
162         if (err) {
163                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
164                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
165                      (unsigned long long)p_blkno);
166                 goto bail;
167         }
168
169         if (max_blocks < count)
170                 count = max_blocks;
171
172         /*
173          * ocfs2 never allocates in this function - the only time we
174          * need to use BH_New is when we're extending i_size on a file
175          * system which doesn't support holes, in which case BH_New
176          * allows block_prepare_write() to zero.
177          */
178         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
179                         "ino %lu, iblock %llu\n", inode->i_ino,
180                         (unsigned long long)iblock);
181
182         /* Treat the unwritten extent as a hole for zeroing purposes. */
183         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
184                 map_bh(bh_result, inode->i_sb, p_blkno);
185
186         bh_result->b_size = count << inode->i_blkbits;
187
188         if (!ocfs2_sparse_alloc(osb)) {
189                 if (p_blkno == 0) {
190                         err = -EIO;
191                         mlog(ML_ERROR,
192                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
193                              (unsigned long long)iblock,
194                              (unsigned long long)p_blkno,
195                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
196                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
197                         dump_stack();
198                 }
199
200                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
201                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
202                      (unsigned long long)past_eof);
203
204                 if (create && (iblock >= past_eof))
205                         set_buffer_new(bh_result);
206         }
207
208 bail:
209         if (err < 0)
210                 err = -EIO;
211
212         mlog_exit(err);
213         return err;
214 }
215
216 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
217                            struct buffer_head *di_bh)
218 {
219         void *kaddr;
220         loff_t size;
221         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
222
223         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
224                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
225                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
226                 return -EROFS;
227         }
228
229         size = i_size_read(inode);
230
231         if (size > PAGE_CACHE_SIZE ||
232             size > ocfs2_max_inline_data(inode->i_sb)) {
233                 ocfs2_error(inode->i_sb,
234                             "Inode %llu has with inline data has bad size: %Lu",
235                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
236                             (unsigned long long)size);
237                 return -EROFS;
238         }
239
240         kaddr = kmap_atomic(page, KM_USER0);
241         if (size)
242                 memcpy(kaddr, di->id2.i_data.id_data, size);
243         /* Clear the remaining part of the page */
244         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
245         flush_dcache_page(page);
246         kunmap_atomic(kaddr, KM_USER0);
247
248         SetPageUptodate(page);
249
250         return 0;
251 }
252
253 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
254 {
255         int ret;
256         struct buffer_head *di_bh = NULL;
257         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
258
259         BUG_ON(!PageLocked(page));
260         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
261
262         ret = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &di_bh,
263                                OCFS2_BH_CACHED, inode);
264         if (ret) {
265                 mlog_errno(ret);
266                 goto out;
267         }
268
269         ret = ocfs2_read_inline_data(inode, page, di_bh);
270 out:
271         unlock_page(page);
272
273         brelse(di_bh);
274         return ret;
275 }
276
277 static int ocfs2_readpage(struct file *file, struct page *page)
278 {
279         struct inode *inode = page->mapping->host;
280         struct ocfs2_inode_info *oi = OCFS2_I(inode);
281         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
282         int ret, unlock = 1;
283
284         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
285
286         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
287         if (ret != 0) {
288                 if (ret == AOP_TRUNCATED_PAGE)
289                         unlock = 0;
290                 mlog_errno(ret);
291                 goto out;
292         }
293
294         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
295                 ret = AOP_TRUNCATED_PAGE;
296                 goto out_inode_unlock;
297         }
298
299         /*
300          * i_size might have just been updated as we grabed the meta lock.  We
301          * might now be discovering a truncate that hit on another node.
302          * block_read_full_page->get_block freaks out if it is asked to read
303          * beyond the end of a file, so we check here.  Callers
304          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
305          * and notice that the page they just read isn't needed.
306          *
307          * XXX sys_readahead() seems to get that wrong?
308          */
309         if (start >= i_size_read(inode)) {
310                 zero_user(page, 0, PAGE_SIZE);
311                 SetPageUptodate(page);
312                 ret = 0;
313                 goto out_alloc;
314         }
315
316         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
317                 ret = ocfs2_readpage_inline(inode, page);
318         else
319                 ret = block_read_full_page(page, ocfs2_get_block);
320         unlock = 0;
321
322 out_alloc:
323         up_read(&OCFS2_I(inode)->ip_alloc_sem);
324 out_inode_unlock:
325         ocfs2_inode_unlock(inode, 0);
326 out:
327         if (unlock)
328                 unlock_page(page);
329         mlog_exit(ret);
330         return ret;
331 }
332
333 /*
334  * This is used only for read-ahead. Failures or difficult to handle
335  * situations are safe to ignore.
336  *
337  * Right now, we don't bother with BH_Boundary - in-inode extent lists
338  * are quite large (243 extents on 4k blocks), so most inodes don't
339  * grow out to a tree. If need be, detecting boundary extents could
340  * trivially be added in a future version of ocfs2_get_block().
341  */
342 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
343                            struct list_head *pages, unsigned nr_pages)
344 {
345         int ret, err = -EIO;
346         struct inode *inode = mapping->host;
347         struct ocfs2_inode_info *oi = OCFS2_I(inode);
348         loff_t start;
349         struct page *last;
350
351         /*
352          * Use the nonblocking flag for the dlm code to avoid page
353          * lock inversion, but don't bother with retrying.
354          */
355         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
356         if (ret)
357                 return err;
358
359         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
360                 ocfs2_inode_unlock(inode, 0);
361                 return err;
362         }
363
364         /*
365          * Don't bother with inline-data. There isn't anything
366          * to read-ahead in that case anyway...
367          */
368         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
369                 goto out_unlock;
370
371         /*
372          * Check whether a remote node truncated this file - we just
373          * drop out in that case as it's not worth handling here.
374          */
375         last = list_entry(pages->prev, struct page, lru);
376         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
377         if (start >= i_size_read(inode))
378                 goto out_unlock;
379
380         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
381
382 out_unlock:
383         up_read(&oi->ip_alloc_sem);
384         ocfs2_inode_unlock(inode, 0);
385
386         return err;
387 }
388
389 /* Note: Because we don't support holes, our allocation has
390  * already happened (allocation writes zeros to the file data)
391  * so we don't have to worry about ordered writes in
392  * ocfs2_writepage.
393  *
394  * ->writepage is called during the process of invalidating the page cache
395  * during blocked lock processing.  It can't block on any cluster locks
396  * to during block mapping.  It's relying on the fact that the block
397  * mapping can't have disappeared under the dirty pages that it is
398  * being asked to write back.
399  */
400 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
401 {
402         int ret;
403
404         mlog_entry("(0x%p)\n", page);
405
406         ret = block_write_full_page(page, ocfs2_get_block, wbc);
407
408         mlog_exit(ret);
409
410         return ret;
411 }
412
413 /*
414  * This is called from ocfs2_write_zero_page() which has handled it's
415  * own cluster locking and has ensured allocation exists for those
416  * blocks to be written.
417  */
418 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
419                                unsigned from, unsigned to)
420 {
421         int ret;
422
423         ret = block_prepare_write(page, from, to, ocfs2_get_block);
424
425         return ret;
426 }
427
428 /* Taken from ext3. We don't necessarily need the full blown
429  * functionality yet, but IMHO it's better to cut and paste the whole
430  * thing so we can avoid introducing our own bugs (and easily pick up
431  * their fixes when they happen) --Mark */
432 int walk_page_buffers(  handle_t *handle,
433                         struct buffer_head *head,
434                         unsigned from,
435                         unsigned to,
436                         int *partial,
437                         int (*fn)(      handle_t *handle,
438                                         struct buffer_head *bh))
439 {
440         struct buffer_head *bh;
441         unsigned block_start, block_end;
442         unsigned blocksize = head->b_size;
443         int err, ret = 0;
444         struct buffer_head *next;
445
446         for (   bh = head, block_start = 0;
447                 ret == 0 && (bh != head || !block_start);
448                 block_start = block_end, bh = next)
449         {
450                 next = bh->b_this_page;
451                 block_end = block_start + blocksize;
452                 if (block_end <= from || block_start >= to) {
453                         if (partial && !buffer_uptodate(bh))
454                                 *partial = 1;
455                         continue;
456                 }
457                 err = (*fn)(handle, bh);
458                 if (!ret)
459                         ret = err;
460         }
461         return ret;
462 }
463
464 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
465                                                          struct page *page,
466                                                          unsigned from,
467                                                          unsigned to)
468 {
469         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
470         handle_t *handle = NULL;
471         int ret = 0;
472
473         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
474         if (!handle) {
475                 ret = -ENOMEM;
476                 mlog_errno(ret);
477                 goto out;
478         }
479
480         if (ocfs2_should_order_data(inode)) {
481                 ret = walk_page_buffers(handle,
482                                         page_buffers(page),
483                                         from, to, NULL,
484                                         ocfs2_journal_dirty_data);
485                 if (ret < 0) 
486                         mlog_errno(ret);
487         }
488 out:
489         if (ret) {
490                 if (handle)
491                         ocfs2_commit_trans(osb, handle);
492                 handle = ERR_PTR(ret);
493         }
494         return handle;
495 }
496
497 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
498 {
499         sector_t status;
500         u64 p_blkno = 0;
501         int err = 0;
502         struct inode *inode = mapping->host;
503
504         mlog_entry("(block = %llu)\n", (unsigned long long)block);
505
506         /* We don't need to lock journal system files, since they aren't
507          * accessed concurrently from multiple nodes.
508          */
509         if (!INODE_JOURNAL(inode)) {
510                 err = ocfs2_inode_lock(inode, NULL, 0);
511                 if (err) {
512                         if (err != -ENOENT)
513                                 mlog_errno(err);
514                         goto bail;
515                 }
516                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
517         }
518
519         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
520                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
521                                                   NULL);
522
523         if (!INODE_JOURNAL(inode)) {
524                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
525                 ocfs2_inode_unlock(inode, 0);
526         }
527
528         if (err) {
529                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
530                      (unsigned long long)block);
531                 mlog_errno(err);
532                 goto bail;
533         }
534
535 bail:
536         status = err ? 0 : p_blkno;
537
538         mlog_exit((int)status);
539
540         return status;
541 }
542
543 /*
544  * TODO: Make this into a generic get_blocks function.
545  *
546  * From do_direct_io in direct-io.c:
547  *  "So what we do is to permit the ->get_blocks function to populate
548  *   bh.b_size with the size of IO which is permitted at this offset and
549  *   this i_blkbits."
550  *
551  * This function is called directly from get_more_blocks in direct-io.c.
552  *
553  * called like this: dio->get_blocks(dio->inode, fs_startblk,
554  *                                      fs_count, map_bh, dio->rw == WRITE);
555  */
556 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
557                                      struct buffer_head *bh_result, int create)
558 {
559         int ret;
560         u64 p_blkno, inode_blocks, contig_blocks;
561         unsigned int ext_flags;
562         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
563         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
564
565         /* This function won't even be called if the request isn't all
566          * nicely aligned and of the right size, so there's no need
567          * for us to check any of that. */
568
569         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
570
571         /*
572          * Any write past EOF is not allowed because we'd be extending.
573          */
574         if (create && (iblock + max_blocks) > inode_blocks) {
575                 ret = -EIO;
576                 goto bail;
577         }
578
579         /* This figures out the size of the next contiguous block, and
580          * our logical offset */
581         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
582                                           &contig_blocks, &ext_flags);
583         if (ret) {
584                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
585                      (unsigned long long)iblock);
586                 ret = -EIO;
587                 goto bail;
588         }
589
590         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
591                 ocfs2_error(inode->i_sb,
592                             "Inode %llu has a hole at block %llu\n",
593                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
594                             (unsigned long long)iblock);
595                 ret = -EROFS;
596                 goto bail;
597         }
598
599         /*
600          * get_more_blocks() expects us to describe a hole by clearing
601          * the mapped bit on bh_result().
602          *
603          * Consider an unwritten extent as a hole.
604          */
605         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
606                 map_bh(bh_result, inode->i_sb, p_blkno);
607         else {
608                 /*
609                  * ocfs2_prepare_inode_for_write() should have caught
610                  * the case where we'd be filling a hole and triggered
611                  * a buffered write instead.
612                  */
613                 if (create) {
614                         ret = -EIO;
615                         mlog_errno(ret);
616                         goto bail;
617                 }
618
619                 clear_buffer_mapped(bh_result);
620         }
621
622         /* make sure we don't map more than max_blocks blocks here as
623            that's all the kernel will handle at this point. */
624         if (max_blocks < contig_blocks)
625                 contig_blocks = max_blocks;
626         bh_result->b_size = contig_blocks << blocksize_bits;
627 bail:
628         return ret;
629 }
630
631 /* 
632  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
633  * particularly interested in the aio/dio case.  Like the core uses
634  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
635  * truncation on another.
636  */
637 static void ocfs2_dio_end_io(struct kiocb *iocb,
638                              loff_t offset,
639                              ssize_t bytes,
640                              void *private)
641 {
642         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
643         int level;
644
645         /* this io's submitter should not have unlocked this before we could */
646         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
647
648         ocfs2_iocb_clear_rw_locked(iocb);
649
650         level = ocfs2_iocb_rw_locked_level(iocb);
651         if (!level)
652                 up_read(&inode->i_alloc_sem);
653         ocfs2_rw_unlock(inode, level);
654 }
655
656 /*
657  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
658  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
659  * do journalled data.
660  */
661 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
662 {
663         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
664
665         journal_invalidatepage(journal, page, offset);
666 }
667
668 static int ocfs2_releasepage(struct page *page, gfp_t wait)
669 {
670         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
671
672         if (!page_has_buffers(page))
673                 return 0;
674         return journal_try_to_free_buffers(journal, page, wait);
675 }
676
677 static ssize_t ocfs2_direct_IO(int rw,
678                                struct kiocb *iocb,
679                                const struct iovec *iov,
680                                loff_t offset,
681                                unsigned long nr_segs)
682 {
683         struct file *file = iocb->ki_filp;
684         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
685         int ret;
686
687         mlog_entry_void();
688
689         /*
690          * Fallback to buffered I/O if we see an inode without
691          * extents.
692          */
693         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
694                 return 0;
695
696         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
697                                             inode->i_sb->s_bdev, iov, offset,
698                                             nr_segs, 
699                                             ocfs2_direct_IO_get_blocks,
700                                             ocfs2_dio_end_io);
701
702         mlog_exit(ret);
703         return ret;
704 }
705
706 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
707                                             u32 cpos,
708                                             unsigned int *start,
709                                             unsigned int *end)
710 {
711         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
712
713         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
714                 unsigned int cpp;
715
716                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
717
718                 cluster_start = cpos % cpp;
719                 cluster_start = cluster_start << osb->s_clustersize_bits;
720
721                 cluster_end = cluster_start + osb->s_clustersize;
722         }
723
724         BUG_ON(cluster_start > PAGE_SIZE);
725         BUG_ON(cluster_end > PAGE_SIZE);
726
727         if (start)
728                 *start = cluster_start;
729         if (end)
730                 *end = cluster_end;
731 }
732
733 /*
734  * 'from' and 'to' are the region in the page to avoid zeroing.
735  *
736  * If pagesize > clustersize, this function will avoid zeroing outside
737  * of the cluster boundary.
738  *
739  * from == to == 0 is code for "zero the entire cluster region"
740  */
741 static void ocfs2_clear_page_regions(struct page *page,
742                                      struct ocfs2_super *osb, u32 cpos,
743                                      unsigned from, unsigned to)
744 {
745         void *kaddr;
746         unsigned int cluster_start, cluster_end;
747
748         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
749
750         kaddr = kmap_atomic(page, KM_USER0);
751
752         if (from || to) {
753                 if (from > cluster_start)
754                         memset(kaddr + cluster_start, 0, from - cluster_start);
755                 if (to < cluster_end)
756                         memset(kaddr + to, 0, cluster_end - to);
757         } else {
758                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
759         }
760
761         kunmap_atomic(kaddr, KM_USER0);
762 }
763
764 /*
765  * Nonsparse file systems fully allocate before we get to the write
766  * code. This prevents ocfs2_write() from tagging the write as an
767  * allocating one, which means ocfs2_map_page_blocks() might try to
768  * read-in the blocks at the tail of our file. Avoid reading them by
769  * testing i_size against each block offset.
770  */
771 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
772                                  unsigned int block_start)
773 {
774         u64 offset = page_offset(page) + block_start;
775
776         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
777                 return 1;
778
779         if (i_size_read(inode) > offset)
780                 return 1;
781
782         return 0;
783 }
784
785 /*
786  * Some of this taken from block_prepare_write(). We already have our
787  * mapping by now though, and the entire write will be allocating or
788  * it won't, so not much need to use BH_New.
789  *
790  * This will also skip zeroing, which is handled externally.
791  */
792 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
793                           struct inode *inode, unsigned int from,
794                           unsigned int to, int new)
795 {
796         int ret = 0;
797         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
798         unsigned int block_end, block_start;
799         unsigned int bsize = 1 << inode->i_blkbits;
800
801         if (!page_has_buffers(page))
802                 create_empty_buffers(page, bsize, 0);
803
804         head = page_buffers(page);
805         for (bh = head, block_start = 0; bh != head || !block_start;
806              bh = bh->b_this_page, block_start += bsize) {
807                 block_end = block_start + bsize;
808
809                 clear_buffer_new(bh);
810
811                 /*
812                  * Ignore blocks outside of our i/o range -
813                  * they may belong to unallocated clusters.
814                  */
815                 if (block_start >= to || block_end <= from) {
816                         if (PageUptodate(page))
817                                 set_buffer_uptodate(bh);
818                         continue;
819                 }
820
821                 /*
822                  * For an allocating write with cluster size >= page
823                  * size, we always write the entire page.
824                  */
825                 if (new)
826                         set_buffer_new(bh);
827
828                 if (!buffer_mapped(bh)) {
829                         map_bh(bh, inode->i_sb, *p_blkno);
830                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
831                 }
832
833                 if (PageUptodate(page)) {
834                         if (!buffer_uptodate(bh))
835                                 set_buffer_uptodate(bh);
836                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
837                            !buffer_new(bh) &&
838                            ocfs2_should_read_blk(inode, page, block_start) &&
839                            (block_start < from || block_end > to)) {
840                         ll_rw_block(READ, 1, &bh);
841                         *wait_bh++=bh;
842                 }
843
844                 *p_blkno = *p_blkno + 1;
845         }
846
847         /*
848          * If we issued read requests - let them complete.
849          */
850         while(wait_bh > wait) {
851                 wait_on_buffer(*--wait_bh);
852                 if (!buffer_uptodate(*wait_bh))
853                         ret = -EIO;
854         }
855
856         if (ret == 0 || !new)
857                 return ret;
858
859         /*
860          * If we get -EIO above, zero out any newly allocated blocks
861          * to avoid exposing stale data.
862          */
863         bh = head;
864         block_start = 0;
865         do {
866                 block_end = block_start + bsize;
867                 if (block_end <= from)
868                         goto next_bh;
869                 if (block_start >= to)
870                         break;
871
872                 zero_user(page, block_start, bh->b_size);
873                 set_buffer_uptodate(bh);
874                 mark_buffer_dirty(bh);
875
876 next_bh:
877                 block_start = block_end;
878                 bh = bh->b_this_page;
879         } while (bh != head);
880
881         return ret;
882 }
883
884 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
885 #define OCFS2_MAX_CTXT_PAGES    1
886 #else
887 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
888 #endif
889
890 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
891
892 /*
893  * Describe the state of a single cluster to be written to.
894  */
895 struct ocfs2_write_cluster_desc {
896         u32             c_cpos;
897         u32             c_phys;
898         /*
899          * Give this a unique field because c_phys eventually gets
900          * filled.
901          */
902         unsigned        c_new;
903         unsigned        c_unwritten;
904 };
905
906 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
907 {
908         return d->c_new || d->c_unwritten;
909 }
910
911 struct ocfs2_write_ctxt {
912         /* Logical cluster position / len of write */
913         u32                             w_cpos;
914         u32                             w_clen;
915
916         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
917
918         /*
919          * This is true if page_size > cluster_size.
920          *
921          * It triggers a set of special cases during write which might
922          * have to deal with allocating writes to partial pages.
923          */
924         unsigned int                    w_large_pages;
925
926         /*
927          * Pages involved in this write.
928          *
929          * w_target_page is the page being written to by the user.
930          *
931          * w_pages is an array of pages which always contains
932          * w_target_page, and in the case of an allocating write with
933          * page_size < cluster size, it will contain zero'd and mapped
934          * pages adjacent to w_target_page which need to be written
935          * out in so that future reads from that region will get
936          * zero's.
937          */
938         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
939         unsigned int                    w_num_pages;
940         struct page                     *w_target_page;
941
942         /*
943          * ocfs2_write_end() uses this to know what the real range to
944          * write in the target should be.
945          */
946         unsigned int                    w_target_from;
947         unsigned int                    w_target_to;
948
949         /*
950          * We could use journal_current_handle() but this is cleaner,
951          * IMHO -Mark
952          */
953         handle_t                        *w_handle;
954
955         struct buffer_head              *w_di_bh;
956
957         struct ocfs2_cached_dealloc_ctxt w_dealloc;
958 };
959
960 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
961 {
962         int i;
963
964         for(i = 0; i < num_pages; i++) {
965                 if (pages[i]) {
966                         unlock_page(pages[i]);
967                         mark_page_accessed(pages[i]);
968                         page_cache_release(pages[i]);
969                 }
970         }
971 }
972
973 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
974 {
975         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
976
977         brelse(wc->w_di_bh);
978         kfree(wc);
979 }
980
981 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
982                                   struct ocfs2_super *osb, loff_t pos,
983                                   unsigned len, struct buffer_head *di_bh)
984 {
985         u32 cend;
986         struct ocfs2_write_ctxt *wc;
987
988         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
989         if (!wc)
990                 return -ENOMEM;
991
992         wc->w_cpos = pos >> osb->s_clustersize_bits;
993         cend = (pos + len - 1) >> osb->s_clustersize_bits;
994         wc->w_clen = cend - wc->w_cpos + 1;
995         get_bh(di_bh);
996         wc->w_di_bh = di_bh;
997
998         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
999                 wc->w_large_pages = 1;
1000         else
1001                 wc->w_large_pages = 0;
1002
1003         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1004
1005         *wcp = wc;
1006
1007         return 0;
1008 }
1009
1010 /*
1011  * If a page has any new buffers, zero them out here, and mark them uptodate
1012  * and dirty so they'll be written out (in order to prevent uninitialised
1013  * block data from leaking). And clear the new bit.
1014  */
1015 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1016 {
1017         unsigned int block_start, block_end;
1018         struct buffer_head *head, *bh;
1019
1020         BUG_ON(!PageLocked(page));
1021         if (!page_has_buffers(page))
1022                 return;
1023
1024         bh = head = page_buffers(page);
1025         block_start = 0;
1026         do {
1027                 block_end = block_start + bh->b_size;
1028
1029                 if (buffer_new(bh)) {
1030                         if (block_end > from && block_start < to) {
1031                                 if (!PageUptodate(page)) {
1032                                         unsigned start, end;
1033
1034                                         start = max(from, block_start);
1035                                         end = min(to, block_end);
1036
1037                                         zero_user_segment(page, start, end);
1038                                         set_buffer_uptodate(bh);
1039                                 }
1040
1041                                 clear_buffer_new(bh);
1042                                 mark_buffer_dirty(bh);
1043                         }
1044                 }
1045
1046                 block_start = block_end;
1047                 bh = bh->b_this_page;
1048         } while (bh != head);
1049 }
1050
1051 /*
1052  * Only called when we have a failure during allocating write to write
1053  * zero's to the newly allocated region.
1054  */
1055 static void ocfs2_write_failure(struct inode *inode,
1056                                 struct ocfs2_write_ctxt *wc,
1057                                 loff_t user_pos, unsigned user_len)
1058 {
1059         int i;
1060         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1061                 to = user_pos + user_len;
1062         struct page *tmppage;
1063
1064         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1065
1066         for(i = 0; i < wc->w_num_pages; i++) {
1067                 tmppage = wc->w_pages[i];
1068
1069                 if (ocfs2_should_order_data(inode))
1070                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1071                                           from, to, NULL,
1072                                           ocfs2_journal_dirty_data);
1073
1074                 block_commit_write(tmppage, from, to);
1075         }
1076 }
1077
1078 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1079                                         struct ocfs2_write_ctxt *wc,
1080                                         struct page *page, u32 cpos,
1081                                         loff_t user_pos, unsigned user_len,
1082                                         int new)
1083 {
1084         int ret;
1085         unsigned int map_from = 0, map_to = 0;
1086         unsigned int cluster_start, cluster_end;
1087         unsigned int user_data_from = 0, user_data_to = 0;
1088
1089         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1090                                         &cluster_start, &cluster_end);
1091
1092         if (page == wc->w_target_page) {
1093                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1094                 map_to = map_from + user_len;
1095
1096                 if (new)
1097                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1098                                                     cluster_start, cluster_end,
1099                                                     new);
1100                 else
1101                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1102                                                     map_from, map_to, new);
1103                 if (ret) {
1104                         mlog_errno(ret);
1105                         goto out;
1106                 }
1107
1108                 user_data_from = map_from;
1109                 user_data_to = map_to;
1110                 if (new) {
1111                         map_from = cluster_start;
1112                         map_to = cluster_end;
1113                 }
1114         } else {
1115                 /*
1116                  * If we haven't allocated the new page yet, we
1117                  * shouldn't be writing it out without copying user
1118                  * data. This is likely a math error from the caller.
1119                  */
1120                 BUG_ON(!new);
1121
1122                 map_from = cluster_start;
1123                 map_to = cluster_end;
1124
1125                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1126                                             cluster_start, cluster_end, new);
1127                 if (ret) {
1128                         mlog_errno(ret);
1129                         goto out;
1130                 }
1131         }
1132
1133         /*
1134          * Parts of newly allocated pages need to be zero'd.
1135          *
1136          * Above, we have also rewritten 'to' and 'from' - as far as
1137          * the rest of the function is concerned, the entire cluster
1138          * range inside of a page needs to be written.
1139          *
1140          * We can skip this if the page is up to date - it's already
1141          * been zero'd from being read in as a hole.
1142          */
1143         if (new && !PageUptodate(page))
1144                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1145                                          cpos, user_data_from, user_data_to);
1146
1147         flush_dcache_page(page);
1148
1149 out:
1150         return ret;
1151 }
1152
1153 /*
1154  * This function will only grab one clusters worth of pages.
1155  */
1156 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1157                                       struct ocfs2_write_ctxt *wc,
1158                                       u32 cpos, loff_t user_pos, int new,
1159                                       struct page *mmap_page)
1160 {
1161         int ret = 0, i;
1162         unsigned long start, target_index, index;
1163         struct inode *inode = mapping->host;
1164
1165         target_index = user_pos >> PAGE_CACHE_SHIFT;
1166
1167         /*
1168          * Figure out how many pages we'll be manipulating here. For
1169          * non allocating write, we just change the one
1170          * page. Otherwise, we'll need a whole clusters worth.
1171          */
1172         if (new) {
1173                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1174                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1175         } else {
1176                 wc->w_num_pages = 1;
1177                 start = target_index;
1178         }
1179
1180         for(i = 0; i < wc->w_num_pages; i++) {
1181                 index = start + i;
1182
1183                 if (index == target_index && mmap_page) {
1184                         /*
1185                          * ocfs2_pagemkwrite() is a little different
1186                          * and wants us to directly use the page
1187                          * passed in.
1188                          */
1189                         lock_page(mmap_page);
1190
1191                         if (mmap_page->mapping != mapping) {
1192                                 unlock_page(mmap_page);
1193                                 /*
1194                                  * Sanity check - the locking in
1195                                  * ocfs2_pagemkwrite() should ensure
1196                                  * that this code doesn't trigger.
1197                                  */
1198                                 ret = -EINVAL;
1199                                 mlog_errno(ret);
1200                                 goto out;
1201                         }
1202
1203                         page_cache_get(mmap_page);
1204                         wc->w_pages[i] = mmap_page;
1205                 } else {
1206                         wc->w_pages[i] = find_or_create_page(mapping, index,
1207                                                              GFP_NOFS);
1208                         if (!wc->w_pages[i]) {
1209                                 ret = -ENOMEM;
1210                                 mlog_errno(ret);
1211                                 goto out;
1212                         }
1213                 }
1214
1215                 if (index == target_index)
1216                         wc->w_target_page = wc->w_pages[i];
1217         }
1218 out:
1219         return ret;
1220 }
1221
1222 /*
1223  * Prepare a single cluster for write one cluster into the file.
1224  */
1225 static int ocfs2_write_cluster(struct address_space *mapping,
1226                                u32 phys, unsigned int unwritten,
1227                                struct ocfs2_alloc_context *data_ac,
1228                                struct ocfs2_alloc_context *meta_ac,
1229                                struct ocfs2_write_ctxt *wc, u32 cpos,
1230                                loff_t user_pos, unsigned user_len)
1231 {
1232         int ret, i, new, should_zero = 0;
1233         u64 v_blkno, p_blkno;
1234         struct inode *inode = mapping->host;
1235
1236         new = phys == 0 ? 1 : 0;
1237         if (new || unwritten)
1238                 should_zero = 1;
1239
1240         if (new) {
1241                 u32 tmp_pos;
1242
1243                 /*
1244                  * This is safe to call with the page locks - it won't take
1245                  * any additional semaphores or cluster locks.
1246                  */
1247                 tmp_pos = cpos;
1248                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1249                                                  &tmp_pos, 1, 0, wc->w_di_bh,
1250                                                  wc->w_handle, data_ac,
1251                                                  meta_ac, NULL);
1252                 /*
1253                  * This shouldn't happen because we must have already
1254                  * calculated the correct meta data allocation required. The
1255                  * internal tree allocation code should know how to increase
1256                  * transaction credits itself.
1257                  *
1258                  * If need be, we could handle -EAGAIN for a
1259                  * RESTART_TRANS here.
1260                  */
1261                 mlog_bug_on_msg(ret == -EAGAIN,
1262                                 "Inode %llu: EAGAIN return during allocation.\n",
1263                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1264                 if (ret < 0) {
1265                         mlog_errno(ret);
1266                         goto out;
1267                 }
1268         } else if (unwritten) {
1269                 ret = ocfs2_mark_extent_written(inode, wc->w_di_bh,
1270                                                 wc->w_handle, cpos, 1, phys,
1271                                                 meta_ac, &wc->w_dealloc);
1272                 if (ret < 0) {
1273                         mlog_errno(ret);
1274                         goto out;
1275                 }
1276         }
1277
1278         if (should_zero)
1279                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1280         else
1281                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1282
1283         /*
1284          * The only reason this should fail is due to an inability to
1285          * find the extent added.
1286          */
1287         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1288                                           NULL);
1289         if (ret < 0) {
1290                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1291                             "at logical block %llu",
1292                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1293                             (unsigned long long)v_blkno);
1294                 goto out;
1295         }
1296
1297         BUG_ON(p_blkno == 0);
1298
1299         for(i = 0; i < wc->w_num_pages; i++) {
1300                 int tmpret;
1301
1302                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1303                                                       wc->w_pages[i], cpos,
1304                                                       user_pos, user_len,
1305                                                       should_zero);
1306                 if (tmpret) {
1307                         mlog_errno(tmpret);
1308                         if (ret == 0)
1309                                 tmpret = ret;
1310                 }
1311         }
1312
1313         /*
1314          * We only have cleanup to do in case of allocating write.
1315          */
1316         if (ret && new)
1317                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1318
1319 out:
1320
1321         return ret;
1322 }
1323
1324 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1325                                        struct ocfs2_alloc_context *data_ac,
1326                                        struct ocfs2_alloc_context *meta_ac,
1327                                        struct ocfs2_write_ctxt *wc,
1328                                        loff_t pos, unsigned len)
1329 {
1330         int ret, i;
1331         loff_t cluster_off;
1332         unsigned int local_len = len;
1333         struct ocfs2_write_cluster_desc *desc;
1334         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1335
1336         for (i = 0; i < wc->w_clen; i++) {
1337                 desc = &wc->w_desc[i];
1338
1339                 /*
1340                  * We have to make sure that the total write passed in
1341                  * doesn't extend past a single cluster.
1342                  */
1343                 local_len = len;
1344                 cluster_off = pos & (osb->s_clustersize - 1);
1345                 if ((cluster_off + local_len) > osb->s_clustersize)
1346                         local_len = osb->s_clustersize - cluster_off;
1347
1348                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1349                                           desc->c_unwritten, data_ac, meta_ac,
1350                                           wc, desc->c_cpos, pos, local_len);
1351                 if (ret) {
1352                         mlog_errno(ret);
1353                         goto out;
1354                 }
1355
1356                 len -= local_len;
1357                 pos += local_len;
1358         }
1359
1360         ret = 0;
1361 out:
1362         return ret;
1363 }
1364
1365 /*
1366  * ocfs2_write_end() wants to know which parts of the target page it
1367  * should complete the write on. It's easiest to compute them ahead of
1368  * time when a more complete view of the write is available.
1369  */
1370 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1371                                         struct ocfs2_write_ctxt *wc,
1372                                         loff_t pos, unsigned len, int alloc)
1373 {
1374         struct ocfs2_write_cluster_desc *desc;
1375
1376         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1377         wc->w_target_to = wc->w_target_from + len;
1378
1379         if (alloc == 0)
1380                 return;
1381
1382         /*
1383          * Allocating write - we may have different boundaries based
1384          * on page size and cluster size.
1385          *
1386          * NOTE: We can no longer compute one value from the other as
1387          * the actual write length and user provided length may be
1388          * different.
1389          */
1390
1391         if (wc->w_large_pages) {
1392                 /*
1393                  * We only care about the 1st and last cluster within
1394                  * our range and whether they should be zero'd or not. Either
1395                  * value may be extended out to the start/end of a
1396                  * newly allocated cluster.
1397                  */
1398                 desc = &wc->w_desc[0];
1399                 if (ocfs2_should_zero_cluster(desc))
1400                         ocfs2_figure_cluster_boundaries(osb,
1401                                                         desc->c_cpos,
1402                                                         &wc->w_target_from,
1403                                                         NULL);
1404
1405                 desc = &wc->w_desc[wc->w_clen - 1];
1406                 if (ocfs2_should_zero_cluster(desc))
1407                         ocfs2_figure_cluster_boundaries(osb,
1408                                                         desc->c_cpos,
1409                                                         NULL,
1410                                                         &wc->w_target_to);
1411         } else {
1412                 wc->w_target_from = 0;
1413                 wc->w_target_to = PAGE_CACHE_SIZE;
1414         }
1415 }
1416
1417 /*
1418  * Populate each single-cluster write descriptor in the write context
1419  * with information about the i/o to be done.
1420  *
1421  * Returns the number of clusters that will have to be allocated, as
1422  * well as a worst case estimate of the number of extent records that
1423  * would have to be created during a write to an unwritten region.
1424  */
1425 static int ocfs2_populate_write_desc(struct inode *inode,
1426                                      struct ocfs2_write_ctxt *wc,
1427                                      unsigned int *clusters_to_alloc,
1428                                      unsigned int *extents_to_split)
1429 {
1430         int ret;
1431         struct ocfs2_write_cluster_desc *desc;
1432         unsigned int num_clusters = 0;
1433         unsigned int ext_flags = 0;
1434         u32 phys = 0;
1435         int i;
1436
1437         *clusters_to_alloc = 0;
1438         *extents_to_split = 0;
1439
1440         for (i = 0; i < wc->w_clen; i++) {
1441                 desc = &wc->w_desc[i];
1442                 desc->c_cpos = wc->w_cpos + i;
1443
1444                 if (num_clusters == 0) {
1445                         /*
1446                          * Need to look up the next extent record.
1447                          */
1448                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1449                                                  &num_clusters, &ext_flags);
1450                         if (ret) {
1451                                 mlog_errno(ret);
1452                                 goto out;
1453                         }
1454
1455                         /*
1456                          * Assume worst case - that we're writing in
1457                          * the middle of the extent.
1458                          *
1459                          * We can assume that the write proceeds from
1460                          * left to right, in which case the extent
1461                          * insert code is smart enough to coalesce the
1462                          * next splits into the previous records created.
1463                          */
1464                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1465                                 *extents_to_split = *extents_to_split + 2;
1466                 } else if (phys) {
1467                         /*
1468                          * Only increment phys if it doesn't describe
1469                          * a hole.
1470                          */
1471                         phys++;
1472                 }
1473
1474                 desc->c_phys = phys;
1475                 if (phys == 0) {
1476                         desc->c_new = 1;
1477                         *clusters_to_alloc = *clusters_to_alloc + 1;
1478                 }
1479                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1480                         desc->c_unwritten = 1;
1481
1482                 num_clusters--;
1483         }
1484
1485         ret = 0;
1486 out:
1487         return ret;
1488 }
1489
1490 static int ocfs2_write_begin_inline(struct address_space *mapping,
1491                                     struct inode *inode,
1492                                     struct ocfs2_write_ctxt *wc)
1493 {
1494         int ret;
1495         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1496         struct page *page;
1497         handle_t *handle;
1498         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1499
1500         page = find_or_create_page(mapping, 0, GFP_NOFS);
1501         if (!page) {
1502                 ret = -ENOMEM;
1503                 mlog_errno(ret);
1504                 goto out;
1505         }
1506         /*
1507          * If we don't set w_num_pages then this page won't get unlocked
1508          * and freed on cleanup of the write context.
1509          */
1510         wc->w_pages[0] = wc->w_target_page = page;
1511         wc->w_num_pages = 1;
1512
1513         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1514         if (IS_ERR(handle)) {
1515                 ret = PTR_ERR(handle);
1516                 mlog_errno(ret);
1517                 goto out;
1518         }
1519
1520         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1521                                    OCFS2_JOURNAL_ACCESS_WRITE);
1522         if (ret) {
1523                 ocfs2_commit_trans(osb, handle);
1524
1525                 mlog_errno(ret);
1526                 goto out;
1527         }
1528
1529         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1530                 ocfs2_set_inode_data_inline(inode, di);
1531
1532         if (!PageUptodate(page)) {
1533                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1534                 if (ret) {
1535                         ocfs2_commit_trans(osb, handle);
1536
1537                         goto out;
1538                 }
1539         }
1540
1541         wc->w_handle = handle;
1542 out:
1543         return ret;
1544 }
1545
1546 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1547 {
1548         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1549
1550         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1551                 return 1;
1552         return 0;
1553 }
1554
1555 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1556                                           struct inode *inode, loff_t pos,
1557                                           unsigned len, struct page *mmap_page,
1558                                           struct ocfs2_write_ctxt *wc)
1559 {
1560         int ret, written = 0;
1561         loff_t end = pos + len;
1562         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1563
1564         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1565              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1566              oi->ip_dyn_features);
1567
1568         /*
1569          * Handle inodes which already have inline data 1st.
1570          */
1571         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1572                 if (mmap_page == NULL &&
1573                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1574                         goto do_inline_write;
1575
1576                 /*
1577                  * The write won't fit - we have to give this inode an
1578                  * inline extent list now.
1579                  */
1580                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1581                 if (ret)
1582                         mlog_errno(ret);
1583                 goto out;
1584         }
1585
1586         /*
1587          * Check whether the inode can accept inline data.
1588          */
1589         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1590                 return 0;
1591
1592         /*
1593          * Check whether the write can fit.
1594          */
1595         if (mmap_page || end > ocfs2_max_inline_data(inode->i_sb))
1596                 return 0;
1597
1598 do_inline_write:
1599         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1600         if (ret) {
1601                 mlog_errno(ret);
1602                 goto out;
1603         }
1604
1605         /*
1606          * This signals to the caller that the data can be written
1607          * inline.
1608          */
1609         written = 1;
1610 out:
1611         return written ? written : ret;
1612 }
1613
1614 /*
1615  * This function only does anything for file systems which can't
1616  * handle sparse files.
1617  *
1618  * What we want to do here is fill in any hole between the current end
1619  * of allocation and the end of our write. That way the rest of the
1620  * write path can treat it as an non-allocating write, which has no
1621  * special case code for sparse/nonsparse files.
1622  */
1623 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1624                                         unsigned len,
1625                                         struct ocfs2_write_ctxt *wc)
1626 {
1627         int ret;
1628         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1629         loff_t newsize = pos + len;
1630
1631         if (ocfs2_sparse_alloc(osb))
1632                 return 0;
1633
1634         if (newsize <= i_size_read(inode))
1635                 return 0;
1636
1637         ret = ocfs2_extend_no_holes(inode, newsize, newsize - len);
1638         if (ret)
1639                 mlog_errno(ret);
1640
1641         return ret;
1642 }
1643
1644 int ocfs2_write_begin_nolock(struct address_space *mapping,
1645                              loff_t pos, unsigned len, unsigned flags,
1646                              struct page **pagep, void **fsdata,
1647                              struct buffer_head *di_bh, struct page *mmap_page)
1648 {
1649         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1650         unsigned int clusters_to_alloc, extents_to_split;
1651         struct ocfs2_write_ctxt *wc;
1652         struct inode *inode = mapping->host;
1653         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1654         struct ocfs2_dinode *di;
1655         struct ocfs2_alloc_context *data_ac = NULL;
1656         struct ocfs2_alloc_context *meta_ac = NULL;
1657         handle_t *handle;
1658
1659         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1660         if (ret) {
1661                 mlog_errno(ret);
1662                 return ret;
1663         }
1664
1665         if (ocfs2_supports_inline_data(osb)) {
1666                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1667                                                      mmap_page, wc);
1668                 if (ret == 1) {
1669                         ret = 0;
1670                         goto success;
1671                 }
1672                 if (ret < 0) {
1673                         mlog_errno(ret);
1674                         goto out;
1675                 }
1676         }
1677
1678         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1679         if (ret) {
1680                 mlog_errno(ret);
1681                 goto out;
1682         }
1683
1684         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1685                                         &extents_to_split);
1686         if (ret) {
1687                 mlog_errno(ret);
1688                 goto out;
1689         }
1690
1691         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1692
1693         /*
1694          * We set w_target_from, w_target_to here so that
1695          * ocfs2_write_end() knows which range in the target page to
1696          * write out. An allocation requires that we write the entire
1697          * cluster range.
1698          */
1699         if (clusters_to_alloc || extents_to_split) {
1700                 /*
1701                  * XXX: We are stretching the limits of
1702                  * ocfs2_lock_allocators(). It greatly over-estimates
1703                  * the work to be done.
1704                  */
1705                 ret = ocfs2_lock_allocators(inode, di, clusters_to_alloc,
1706                                             extents_to_split, &data_ac, &meta_ac);
1707                 if (ret) {
1708                         mlog_errno(ret);
1709                         goto out;
1710                 }
1711
1712                 credits = ocfs2_calc_extend_credits(inode->i_sb, di,
1713                                                     clusters_to_alloc);
1714
1715         }
1716
1717         ocfs2_set_target_boundaries(osb, wc, pos, len,
1718                                     clusters_to_alloc + extents_to_split);
1719
1720         handle = ocfs2_start_trans(osb, credits);
1721         if (IS_ERR(handle)) {
1722                 ret = PTR_ERR(handle);
1723                 mlog_errno(ret);
1724                 goto out;
1725         }
1726
1727         wc->w_handle = handle;
1728
1729         /*
1730          * We don't want this to fail in ocfs2_write_end(), so do it
1731          * here.
1732          */
1733         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1734                                    OCFS2_JOURNAL_ACCESS_WRITE);
1735         if (ret) {
1736                 mlog_errno(ret);
1737                 goto out_commit;
1738         }
1739
1740         /*
1741          * Fill our page array first. That way we've grabbed enough so
1742          * that we can zero and flush if we error after adding the
1743          * extent.
1744          */
1745         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1746                                          clusters_to_alloc + extents_to_split,
1747                                          mmap_page);
1748         if (ret) {
1749                 mlog_errno(ret);
1750                 goto out_commit;
1751         }
1752
1753         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1754                                           len);
1755         if (ret) {
1756                 mlog_errno(ret);
1757                 goto out_commit;
1758         }
1759
1760         if (data_ac)
1761                 ocfs2_free_alloc_context(data_ac);
1762         if (meta_ac)
1763                 ocfs2_free_alloc_context(meta_ac);
1764
1765 success:
1766         *pagep = wc->w_target_page;
1767         *fsdata = wc;
1768         return 0;
1769 out_commit:
1770         ocfs2_commit_trans(osb, handle);
1771
1772 out:
1773         ocfs2_free_write_ctxt(wc);
1774
1775         if (data_ac)
1776                 ocfs2_free_alloc_context(data_ac);
1777         if (meta_ac)
1778                 ocfs2_free_alloc_context(meta_ac);
1779         return ret;
1780 }
1781
1782 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1783                              loff_t pos, unsigned len, unsigned flags,
1784                              struct page **pagep, void **fsdata)
1785 {
1786         int ret;
1787         struct buffer_head *di_bh = NULL;
1788         struct inode *inode = mapping->host;
1789
1790         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1791         if (ret) {
1792                 mlog_errno(ret);
1793                 return ret;
1794         }
1795
1796         /*
1797          * Take alloc sem here to prevent concurrent lookups. That way
1798          * the mapping, zeroing and tree manipulation within
1799          * ocfs2_write() will be safe against ->readpage(). This
1800          * should also serve to lock out allocation from a shared
1801          * writeable region.
1802          */
1803         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1804
1805         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1806                                        fsdata, di_bh, NULL);
1807         if (ret) {
1808                 mlog_errno(ret);
1809                 goto out_fail;
1810         }
1811
1812         brelse(di_bh);
1813
1814         return 0;
1815
1816 out_fail:
1817         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1818
1819         brelse(di_bh);
1820         ocfs2_inode_unlock(inode, 1);
1821
1822         return ret;
1823 }
1824
1825 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1826                                    unsigned len, unsigned *copied,
1827                                    struct ocfs2_dinode *di,
1828                                    struct ocfs2_write_ctxt *wc)
1829 {
1830         void *kaddr;
1831
1832         if (unlikely(*copied < len)) {
1833                 if (!PageUptodate(wc->w_target_page)) {
1834                         *copied = 0;
1835                         return;
1836                 }
1837         }
1838
1839         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1840         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1841         kunmap_atomic(kaddr, KM_USER0);
1842
1843         mlog(0, "Data written to inode at offset %llu. "
1844              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1845              (unsigned long long)pos, *copied,
1846              le16_to_cpu(di->id2.i_data.id_count),
1847              le16_to_cpu(di->i_dyn_features));
1848 }
1849
1850 int ocfs2_write_end_nolock(struct address_space *mapping,
1851                            loff_t pos, unsigned len, unsigned copied,
1852                            struct page *page, void *fsdata)
1853 {
1854         int i;
1855         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1856         struct inode *inode = mapping->host;
1857         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1858         struct ocfs2_write_ctxt *wc = fsdata;
1859         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1860         handle_t *handle = wc->w_handle;
1861         struct page *tmppage;
1862
1863         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1864                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1865                 goto out_write_size;
1866         }
1867
1868         if (unlikely(copied < len)) {
1869                 if (!PageUptodate(wc->w_target_page))
1870                         copied = 0;
1871
1872                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1873                                        start+len);
1874         }
1875         flush_dcache_page(wc->w_target_page);
1876
1877         for(i = 0; i < wc->w_num_pages; i++) {
1878                 tmppage = wc->w_pages[i];
1879
1880                 if (tmppage == wc->w_target_page) {
1881                         from = wc->w_target_from;
1882                         to = wc->w_target_to;
1883
1884                         BUG_ON(from > PAGE_CACHE_SIZE ||
1885                                to > PAGE_CACHE_SIZE ||
1886                                to < from);
1887                 } else {
1888                         /*
1889                          * Pages adjacent to the target (if any) imply
1890                          * a hole-filling write in which case we want
1891                          * to flush their entire range.
1892                          */
1893                         from = 0;
1894                         to = PAGE_CACHE_SIZE;
1895                 }
1896
1897                 if (ocfs2_should_order_data(inode))
1898                         walk_page_buffers(wc->w_handle, page_buffers(tmppage),
1899                                           from, to, NULL,
1900                                           ocfs2_journal_dirty_data);
1901
1902                 block_commit_write(tmppage, from, to);
1903         }
1904
1905 out_write_size:
1906         pos += copied;
1907         if (pos > inode->i_size) {
1908                 i_size_write(inode, pos);
1909                 mark_inode_dirty(inode);
1910         }
1911         inode->i_blocks = ocfs2_inode_sector_count(inode);
1912         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1913         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1914         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1915         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1916         ocfs2_journal_dirty(handle, wc->w_di_bh);
1917
1918         ocfs2_commit_trans(osb, handle);
1919
1920         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1921
1922         ocfs2_free_write_ctxt(wc);
1923
1924         return copied;
1925 }
1926
1927 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1928                            loff_t pos, unsigned len, unsigned copied,
1929                            struct page *page, void *fsdata)
1930 {
1931         int ret;
1932         struct inode *inode = mapping->host;
1933
1934         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1935
1936         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1937         ocfs2_inode_unlock(inode, 1);
1938
1939         return ret;
1940 }
1941
1942 const struct address_space_operations ocfs2_aops = {
1943         .readpage       = ocfs2_readpage,
1944         .readpages      = ocfs2_readpages,
1945         .writepage      = ocfs2_writepage,
1946         .write_begin    = ocfs2_write_begin,
1947         .write_end      = ocfs2_write_end,
1948         .bmap           = ocfs2_bmap,
1949         .sync_page      = block_sync_page,
1950         .direct_IO      = ocfs2_direct_IO,
1951         .invalidatepage = ocfs2_invalidatepage,
1952         .releasepage    = ocfs2_releasepage,
1953         .migratepage    = buffer_migrate_page,
1954 };