ocfs2: ocfs2_write_begin_nolock() should handle len=0
[linux-2.6.git] / 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 #include <linux/quotaops.h>
31
32 #define MLOG_MASK_PREFIX ML_FILE_IO
33 #include <cluster/masklog.h>
34
35 #include "ocfs2.h"
36
37 #include "alloc.h"
38 #include "aops.h"
39 #include "dlmglue.h"
40 #include "extent_map.h"
41 #include "file.h"
42 #include "inode.h"
43 #include "journal.h"
44 #include "suballoc.h"
45 #include "super.h"
46 #include "symlink.h"
47
48 #include "buffer_head_io.h"
49
50 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
51                                    struct buffer_head *bh_result, int create)
52 {
53         int err = -EIO;
54         int status;
55         struct ocfs2_dinode *fe = NULL;
56         struct buffer_head *bh = NULL;
57         struct buffer_head *buffer_cache_bh = NULL;
58         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
59         void *kaddr;
60
61         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
62                    (unsigned long long)iblock, bh_result, create);
63
64         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
65
66         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
67                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
68                      (unsigned long long)iblock);
69                 goto bail;
70         }
71
72         status = ocfs2_read_inode_block(inode, &bh);
73         if (status < 0) {
74                 mlog_errno(status);
75                 goto bail;
76         }
77         fe = (struct ocfs2_dinode *) bh->b_data;
78
79         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
80                                                     le32_to_cpu(fe->i_clusters))) {
81                 mlog(ML_ERROR, "block offset is outside the allocated size: "
82                      "%llu\n", (unsigned long long)iblock);
83                 goto bail;
84         }
85
86         /* We don't use the page cache to create symlink data, so if
87          * need be, copy it over from the buffer cache. */
88         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
89                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
90                             iblock;
91                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
92                 if (!buffer_cache_bh) {
93                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
94                         goto bail;
95                 }
96
97                 /* we haven't locked out transactions, so a commit
98                  * could've happened. Since we've got a reference on
99                  * the bh, even if it commits while we're doing the
100                  * copy, the data is still good. */
101                 if (buffer_jbd(buffer_cache_bh)
102                     && ocfs2_inode_is_new(inode)) {
103                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
104                         if (!kaddr) {
105                                 mlog(ML_ERROR, "couldn't kmap!\n");
106                                 goto bail;
107                         }
108                         memcpy(kaddr + (bh_result->b_size * iblock),
109                                buffer_cache_bh->b_data,
110                                bh_result->b_size);
111                         kunmap_atomic(kaddr, KM_USER0);
112                         set_buffer_uptodate(bh_result);
113                 }
114                 brelse(buffer_cache_bh);
115         }
116
117         map_bh(bh_result, inode->i_sb,
118                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
119
120         err = 0;
121
122 bail:
123         brelse(bh);
124
125         mlog_exit(err);
126         return err;
127 }
128
129 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
130                            struct buffer_head *bh_result, int create)
131 {
132         int err = 0;
133         unsigned int ext_flags;
134         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
135         u64 p_blkno, count, past_eof;
136         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
137
138         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
139                    (unsigned long long)iblock, bh_result, create);
140
141         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
142                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
143                      inode, inode->i_ino);
144
145         if (S_ISLNK(inode->i_mode)) {
146                 /* this always does I/O for some reason. */
147                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
148                 goto bail;
149         }
150
151         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
152                                           &ext_flags);
153         if (err) {
154                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
155                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
156                      (unsigned long long)p_blkno);
157                 goto bail;
158         }
159
160         if (max_blocks < count)
161                 count = max_blocks;
162
163         /*
164          * ocfs2 never allocates in this function - the only time we
165          * need to use BH_New is when we're extending i_size on a file
166          * system which doesn't support holes, in which case BH_New
167          * allows block_prepare_write() to zero.
168          *
169          * If we see this on a sparse file system, then a truncate has
170          * raced us and removed the cluster. In this case, we clear
171          * the buffers dirty and uptodate bits and let the buffer code
172          * ignore it as a hole.
173          */
174         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
175                 clear_buffer_dirty(bh_result);
176                 clear_buffer_uptodate(bh_result);
177                 goto bail;
178         }
179
180         /* Treat the unwritten extent as a hole for zeroing purposes. */
181         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
182                 map_bh(bh_result, inode->i_sb, p_blkno);
183
184         bh_result->b_size = count << inode->i_blkbits;
185
186         if (!ocfs2_sparse_alloc(osb)) {
187                 if (p_blkno == 0) {
188                         err = -EIO;
189                         mlog(ML_ERROR,
190                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
191                              (unsigned long long)iblock,
192                              (unsigned long long)p_blkno,
193                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
194                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
195                         dump_stack();
196                         goto bail;
197                 }
198
199                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
200                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
201                      (unsigned long long)past_eof);
202
203                 if (create && (iblock >= past_eof))
204                         set_buffer_new(bh_result);
205         }
206
207 bail:
208         if (err < 0)
209                 err = -EIO;
210
211         mlog_exit(err);
212         return err;
213 }
214
215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
216                            struct buffer_head *di_bh)
217 {
218         void *kaddr;
219         loff_t size;
220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
221
222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
225                 return -EROFS;
226         }
227
228         size = i_size_read(inode);
229
230         if (size > PAGE_CACHE_SIZE ||
231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
232                 ocfs2_error(inode->i_sb,
233                             "Inode %llu has with inline data has bad size: %Lu",
234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
235                             (unsigned long long)size);
236                 return -EROFS;
237         }
238
239         kaddr = kmap_atomic(page, KM_USER0);
240         if (size)
241                 memcpy(kaddr, di->id2.i_data.id_data, size);
242         /* Clear the remaining part of the page */
243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
244         flush_dcache_page(page);
245         kunmap_atomic(kaddr, KM_USER0);
246
247         SetPageUptodate(page);
248
249         return 0;
250 }
251
252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
253 {
254         int ret;
255         struct buffer_head *di_bh = NULL;
256
257         BUG_ON(!PageLocked(page));
258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
259
260         ret = ocfs2_read_inode_block(inode, &di_bh);
261         if (ret) {
262                 mlog_errno(ret);
263                 goto out;
264         }
265
266         ret = ocfs2_read_inline_data(inode, page, di_bh);
267 out:
268         unlock_page(page);
269
270         brelse(di_bh);
271         return ret;
272 }
273
274 static int ocfs2_readpage(struct file *file, struct page *page)
275 {
276         struct inode *inode = page->mapping->host;
277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
279         int ret, unlock = 1;
280
281         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
282
283         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
284         if (ret != 0) {
285                 if (ret == AOP_TRUNCATED_PAGE)
286                         unlock = 0;
287                 mlog_errno(ret);
288                 goto out;
289         }
290
291         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
292                 ret = AOP_TRUNCATED_PAGE;
293                 goto out_inode_unlock;
294         }
295
296         /*
297          * i_size might have just been updated as we grabed the meta lock.  We
298          * might now be discovering a truncate that hit on another node.
299          * block_read_full_page->get_block freaks out if it is asked to read
300          * beyond the end of a file, so we check here.  Callers
301          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
302          * and notice that the page they just read isn't needed.
303          *
304          * XXX sys_readahead() seems to get that wrong?
305          */
306         if (start >= i_size_read(inode)) {
307                 zero_user(page, 0, PAGE_SIZE);
308                 SetPageUptodate(page);
309                 ret = 0;
310                 goto out_alloc;
311         }
312
313         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
314                 ret = ocfs2_readpage_inline(inode, page);
315         else
316                 ret = block_read_full_page(page, ocfs2_get_block);
317         unlock = 0;
318
319 out_alloc:
320         up_read(&OCFS2_I(inode)->ip_alloc_sem);
321 out_inode_unlock:
322         ocfs2_inode_unlock(inode, 0);
323 out:
324         if (unlock)
325                 unlock_page(page);
326         mlog_exit(ret);
327         return ret;
328 }
329
330 /*
331  * This is used only for read-ahead. Failures or difficult to handle
332  * situations are safe to ignore.
333  *
334  * Right now, we don't bother with BH_Boundary - in-inode extent lists
335  * are quite large (243 extents on 4k blocks), so most inodes don't
336  * grow out to a tree. If need be, detecting boundary extents could
337  * trivially be added in a future version of ocfs2_get_block().
338  */
339 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
340                            struct list_head *pages, unsigned nr_pages)
341 {
342         int ret, err = -EIO;
343         struct inode *inode = mapping->host;
344         struct ocfs2_inode_info *oi = OCFS2_I(inode);
345         loff_t start;
346         struct page *last;
347
348         /*
349          * Use the nonblocking flag for the dlm code to avoid page
350          * lock inversion, but don't bother with retrying.
351          */
352         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
353         if (ret)
354                 return err;
355
356         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
357                 ocfs2_inode_unlock(inode, 0);
358                 return err;
359         }
360
361         /*
362          * Don't bother with inline-data. There isn't anything
363          * to read-ahead in that case anyway...
364          */
365         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
366                 goto out_unlock;
367
368         /*
369          * Check whether a remote node truncated this file - we just
370          * drop out in that case as it's not worth handling here.
371          */
372         last = list_entry(pages->prev, struct page, lru);
373         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
374         if (start >= i_size_read(inode))
375                 goto out_unlock;
376
377         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
378
379 out_unlock:
380         up_read(&oi->ip_alloc_sem);
381         ocfs2_inode_unlock(inode, 0);
382
383         return err;
384 }
385
386 /* Note: Because we don't support holes, our allocation has
387  * already happened (allocation writes zeros to the file data)
388  * so we don't have to worry about ordered writes in
389  * ocfs2_writepage.
390  *
391  * ->writepage is called during the process of invalidating the page cache
392  * during blocked lock processing.  It can't block on any cluster locks
393  * to during block mapping.  It's relying on the fact that the block
394  * mapping can't have disappeared under the dirty pages that it is
395  * being asked to write back.
396  */
397 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
398 {
399         int ret;
400
401         mlog_entry("(0x%p)\n", page);
402
403         ret = block_write_full_page(page, ocfs2_get_block, wbc);
404
405         mlog_exit(ret);
406
407         return ret;
408 }
409
410 /*
411  * This is called from ocfs2_write_zero_page() which has handled it's
412  * own cluster locking and has ensured allocation exists for those
413  * blocks to be written.
414  */
415 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
416                                unsigned from, unsigned to)
417 {
418         int ret;
419
420         ret = block_prepare_write(page, from, to, ocfs2_get_block);
421
422         return ret;
423 }
424
425 /* Taken from ext3. We don't necessarily need the full blown
426  * functionality yet, but IMHO it's better to cut and paste the whole
427  * thing so we can avoid introducing our own bugs (and easily pick up
428  * their fixes when they happen) --Mark */
429 int walk_page_buffers(  handle_t *handle,
430                         struct buffer_head *head,
431                         unsigned from,
432                         unsigned to,
433                         int *partial,
434                         int (*fn)(      handle_t *handle,
435                                         struct buffer_head *bh))
436 {
437         struct buffer_head *bh;
438         unsigned block_start, block_end;
439         unsigned blocksize = head->b_size;
440         int err, ret = 0;
441         struct buffer_head *next;
442
443         for (   bh = head, block_start = 0;
444                 ret == 0 && (bh != head || !block_start);
445                 block_start = block_end, bh = next)
446         {
447                 next = bh->b_this_page;
448                 block_end = block_start + blocksize;
449                 if (block_end <= from || block_start >= to) {
450                         if (partial && !buffer_uptodate(bh))
451                                 *partial = 1;
452                         continue;
453                 }
454                 err = (*fn)(handle, bh);
455                 if (!ret)
456                         ret = err;
457         }
458         return ret;
459 }
460
461 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
462                                                          struct page *page,
463                                                          unsigned from,
464                                                          unsigned to)
465 {
466         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
467         handle_t *handle;
468         int ret = 0;
469
470         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
471         if (IS_ERR(handle)) {
472                 ret = -ENOMEM;
473                 mlog_errno(ret);
474                 goto out;
475         }
476
477         if (ocfs2_should_order_data(inode)) {
478                 ret = ocfs2_jbd2_file_inode(handle, inode);
479                 if (ret < 0)
480                         mlog_errno(ret);
481         }
482 out:
483         if (ret) {
484                 if (!IS_ERR(handle))
485                         ocfs2_commit_trans(osb, handle);
486                 handle = ERR_PTR(ret);
487         }
488         return handle;
489 }
490
491 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
492 {
493         sector_t status;
494         u64 p_blkno = 0;
495         int err = 0;
496         struct inode *inode = mapping->host;
497
498         mlog_entry("(block = %llu)\n", (unsigned long long)block);
499
500         /* We don't need to lock journal system files, since they aren't
501          * accessed concurrently from multiple nodes.
502          */
503         if (!INODE_JOURNAL(inode)) {
504                 err = ocfs2_inode_lock(inode, NULL, 0);
505                 if (err) {
506                         if (err != -ENOENT)
507                                 mlog_errno(err);
508                         goto bail;
509                 }
510                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
511         }
512
513         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
514                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
515                                                   NULL);
516
517         if (!INODE_JOURNAL(inode)) {
518                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
519                 ocfs2_inode_unlock(inode, 0);
520         }
521
522         if (err) {
523                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
524                      (unsigned long long)block);
525                 mlog_errno(err);
526                 goto bail;
527         }
528
529 bail:
530         status = err ? 0 : p_blkno;
531
532         mlog_exit((int)status);
533
534         return status;
535 }
536
537 /*
538  * TODO: Make this into a generic get_blocks function.
539  *
540  * From do_direct_io in direct-io.c:
541  *  "So what we do is to permit the ->get_blocks function to populate
542  *   bh.b_size with the size of IO which is permitted at this offset and
543  *   this i_blkbits."
544  *
545  * This function is called directly from get_more_blocks in direct-io.c.
546  *
547  * called like this: dio->get_blocks(dio->inode, fs_startblk,
548  *                                      fs_count, map_bh, dio->rw == WRITE);
549  */
550 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
551                                      struct buffer_head *bh_result, int create)
552 {
553         int ret;
554         u64 p_blkno, inode_blocks, contig_blocks;
555         unsigned int ext_flags;
556         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
557         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
558
559         /* This function won't even be called if the request isn't all
560          * nicely aligned and of the right size, so there's no need
561          * for us to check any of that. */
562
563         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
564
565         /*
566          * Any write past EOF is not allowed because we'd be extending.
567          */
568         if (create && (iblock + max_blocks) > inode_blocks) {
569                 ret = -EIO;
570                 goto bail;
571         }
572
573         /* This figures out the size of the next contiguous block, and
574          * our logical offset */
575         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
576                                           &contig_blocks, &ext_flags);
577         if (ret) {
578                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
579                      (unsigned long long)iblock);
580                 ret = -EIO;
581                 goto bail;
582         }
583
584         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
585                 ocfs2_error(inode->i_sb,
586                             "Inode %llu has a hole at block %llu\n",
587                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
588                             (unsigned long long)iblock);
589                 ret = -EROFS;
590                 goto bail;
591         }
592
593         /*
594          * get_more_blocks() expects us to describe a hole by clearing
595          * the mapped bit on bh_result().
596          *
597          * Consider an unwritten extent as a hole.
598          */
599         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
600                 map_bh(bh_result, inode->i_sb, p_blkno);
601         else {
602                 /*
603                  * ocfs2_prepare_inode_for_write() should have caught
604                  * the case where we'd be filling a hole and triggered
605                  * a buffered write instead.
606                  */
607                 if (create) {
608                         ret = -EIO;
609                         mlog_errno(ret);
610                         goto bail;
611                 }
612
613                 clear_buffer_mapped(bh_result);
614         }
615
616         /* make sure we don't map more than max_blocks blocks here as
617            that's all the kernel will handle at this point. */
618         if (max_blocks < contig_blocks)
619                 contig_blocks = max_blocks;
620         bh_result->b_size = contig_blocks << blocksize_bits;
621 bail:
622         return ret;
623 }
624
625 /* 
626  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
627  * particularly interested in the aio/dio case.  Like the core uses
628  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
629  * truncation on another.
630  */
631 static void ocfs2_dio_end_io(struct kiocb *iocb,
632                              loff_t offset,
633                              ssize_t bytes,
634                              void *private)
635 {
636         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
637         int level;
638
639         /* this io's submitter should not have unlocked this before we could */
640         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
641
642         ocfs2_iocb_clear_rw_locked(iocb);
643
644         level = ocfs2_iocb_rw_locked_level(iocb);
645         if (!level)
646                 up_read(&inode->i_alloc_sem);
647         ocfs2_rw_unlock(inode, level);
648 }
649
650 /*
651  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
652  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
653  * do journalled data.
654  */
655 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
656 {
657         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
658
659         jbd2_journal_invalidatepage(journal, page, offset);
660 }
661
662 static int ocfs2_releasepage(struct page *page, gfp_t wait)
663 {
664         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
665
666         if (!page_has_buffers(page))
667                 return 0;
668         return jbd2_journal_try_to_free_buffers(journal, page, wait);
669 }
670
671 static ssize_t ocfs2_direct_IO(int rw,
672                                struct kiocb *iocb,
673                                const struct iovec *iov,
674                                loff_t offset,
675                                unsigned long nr_segs)
676 {
677         struct file *file = iocb->ki_filp;
678         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
679         int ret;
680
681         mlog_entry_void();
682
683         /*
684          * Fallback to buffered I/O if we see an inode without
685          * extents.
686          */
687         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
688                 return 0;
689
690         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
691                                             inode->i_sb->s_bdev, iov, offset,
692                                             nr_segs, 
693                                             ocfs2_direct_IO_get_blocks,
694                                             ocfs2_dio_end_io);
695
696         mlog_exit(ret);
697         return ret;
698 }
699
700 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
701                                             u32 cpos,
702                                             unsigned int *start,
703                                             unsigned int *end)
704 {
705         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
706
707         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
708                 unsigned int cpp;
709
710                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
711
712                 cluster_start = cpos % cpp;
713                 cluster_start = cluster_start << osb->s_clustersize_bits;
714
715                 cluster_end = cluster_start + osb->s_clustersize;
716         }
717
718         BUG_ON(cluster_start > PAGE_SIZE);
719         BUG_ON(cluster_end > PAGE_SIZE);
720
721         if (start)
722                 *start = cluster_start;
723         if (end)
724                 *end = cluster_end;
725 }
726
727 /*
728  * 'from' and 'to' are the region in the page to avoid zeroing.
729  *
730  * If pagesize > clustersize, this function will avoid zeroing outside
731  * of the cluster boundary.
732  *
733  * from == to == 0 is code for "zero the entire cluster region"
734  */
735 static void ocfs2_clear_page_regions(struct page *page,
736                                      struct ocfs2_super *osb, u32 cpos,
737                                      unsigned from, unsigned to)
738 {
739         void *kaddr;
740         unsigned int cluster_start, cluster_end;
741
742         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
743
744         kaddr = kmap_atomic(page, KM_USER0);
745
746         if (from || to) {
747                 if (from > cluster_start)
748                         memset(kaddr + cluster_start, 0, from - cluster_start);
749                 if (to < cluster_end)
750                         memset(kaddr + to, 0, cluster_end - to);
751         } else {
752                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
753         }
754
755         kunmap_atomic(kaddr, KM_USER0);
756 }
757
758 /*
759  * Nonsparse file systems fully allocate before we get to the write
760  * code. This prevents ocfs2_write() from tagging the write as an
761  * allocating one, which means ocfs2_map_page_blocks() might try to
762  * read-in the blocks at the tail of our file. Avoid reading them by
763  * testing i_size against each block offset.
764  */
765 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
766                                  unsigned int block_start)
767 {
768         u64 offset = page_offset(page) + block_start;
769
770         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
771                 return 1;
772
773         if (i_size_read(inode) > offset)
774                 return 1;
775
776         return 0;
777 }
778
779 /*
780  * Some of this taken from block_prepare_write(). We already have our
781  * mapping by now though, and the entire write will be allocating or
782  * it won't, so not much need to use BH_New.
783  *
784  * This will also skip zeroing, which is handled externally.
785  */
786 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
787                           struct inode *inode, unsigned int from,
788                           unsigned int to, int new)
789 {
790         int ret = 0;
791         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
792         unsigned int block_end, block_start;
793         unsigned int bsize = 1 << inode->i_blkbits;
794
795         if (!page_has_buffers(page))
796                 create_empty_buffers(page, bsize, 0);
797
798         head = page_buffers(page);
799         for (bh = head, block_start = 0; bh != head || !block_start;
800              bh = bh->b_this_page, block_start += bsize) {
801                 block_end = block_start + bsize;
802
803                 clear_buffer_new(bh);
804
805                 /*
806                  * Ignore blocks outside of our i/o range -
807                  * they may belong to unallocated clusters.
808                  */
809                 if (block_start >= to || block_end <= from) {
810                         if (PageUptodate(page))
811                                 set_buffer_uptodate(bh);
812                         continue;
813                 }
814
815                 /*
816                  * For an allocating write with cluster size >= page
817                  * size, we always write the entire page.
818                  */
819                 if (new)
820                         set_buffer_new(bh);
821
822                 if (!buffer_mapped(bh)) {
823                         map_bh(bh, inode->i_sb, *p_blkno);
824                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
825                 }
826
827                 if (PageUptodate(page)) {
828                         if (!buffer_uptodate(bh))
829                                 set_buffer_uptodate(bh);
830                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
831                            !buffer_new(bh) &&
832                            ocfs2_should_read_blk(inode, page, block_start) &&
833                            (block_start < from || block_end > to)) {
834                         ll_rw_block(READ, 1, &bh);
835                         *wait_bh++=bh;
836                 }
837
838                 *p_blkno = *p_blkno + 1;
839         }
840
841         /*
842          * If we issued read requests - let them complete.
843          */
844         while(wait_bh > wait) {
845                 wait_on_buffer(*--wait_bh);
846                 if (!buffer_uptodate(*wait_bh))
847                         ret = -EIO;
848         }
849
850         if (ret == 0 || !new)
851                 return ret;
852
853         /*
854          * If we get -EIO above, zero out any newly allocated blocks
855          * to avoid exposing stale data.
856          */
857         bh = head;
858         block_start = 0;
859         do {
860                 block_end = block_start + bsize;
861                 if (block_end <= from)
862                         goto next_bh;
863                 if (block_start >= to)
864                         break;
865
866                 zero_user(page, block_start, bh->b_size);
867                 set_buffer_uptodate(bh);
868                 mark_buffer_dirty(bh);
869
870 next_bh:
871                 block_start = block_end;
872                 bh = bh->b_this_page;
873         } while (bh != head);
874
875         return ret;
876 }
877
878 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
879 #define OCFS2_MAX_CTXT_PAGES    1
880 #else
881 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
882 #endif
883
884 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
885
886 /*
887  * Describe the state of a single cluster to be written to.
888  */
889 struct ocfs2_write_cluster_desc {
890         u32             c_cpos;
891         u32             c_phys;
892         /*
893          * Give this a unique field because c_phys eventually gets
894          * filled.
895          */
896         unsigned        c_new;
897         unsigned        c_unwritten;
898         unsigned        c_needs_zero;
899 };
900
901 struct ocfs2_write_ctxt {
902         /* Logical cluster position / len of write */
903         u32                             w_cpos;
904         u32                             w_clen;
905
906         /* First cluster allocated in a nonsparse extend */
907         u32                             w_first_new_cpos;
908
909         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
910
911         /*
912          * This is true if page_size > cluster_size.
913          *
914          * It triggers a set of special cases during write which might
915          * have to deal with allocating writes to partial pages.
916          */
917         unsigned int                    w_large_pages;
918
919         /*
920          * Pages involved in this write.
921          *
922          * w_target_page is the page being written to by the user.
923          *
924          * w_pages is an array of pages which always contains
925          * w_target_page, and in the case of an allocating write with
926          * page_size < cluster size, it will contain zero'd and mapped
927          * pages adjacent to w_target_page which need to be written
928          * out in so that future reads from that region will get
929          * zero's.
930          */
931         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
932         unsigned int                    w_num_pages;
933         struct page                     *w_target_page;
934
935         /*
936          * ocfs2_write_end() uses this to know what the real range to
937          * write in the target should be.
938          */
939         unsigned int                    w_target_from;
940         unsigned int                    w_target_to;
941
942         /*
943          * We could use journal_current_handle() but this is cleaner,
944          * IMHO -Mark
945          */
946         handle_t                        *w_handle;
947
948         struct buffer_head              *w_di_bh;
949
950         struct ocfs2_cached_dealloc_ctxt w_dealloc;
951 };
952
953 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
954 {
955         int i;
956
957         for(i = 0; i < num_pages; i++) {
958                 if (pages[i]) {
959                         unlock_page(pages[i]);
960                         mark_page_accessed(pages[i]);
961                         page_cache_release(pages[i]);
962                 }
963         }
964 }
965
966 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
967 {
968         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
969
970         brelse(wc->w_di_bh);
971         kfree(wc);
972 }
973
974 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
975                                   struct ocfs2_super *osb, loff_t pos,
976                                   unsigned len, struct buffer_head *di_bh)
977 {
978         u32 cend;
979         struct ocfs2_write_ctxt *wc;
980
981         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
982         if (!wc)
983                 return -ENOMEM;
984
985         wc->w_cpos = pos >> osb->s_clustersize_bits;
986         wc->w_first_new_cpos = UINT_MAX;
987         cend = (pos + len - 1) >> osb->s_clustersize_bits;
988         wc->w_clen = cend - wc->w_cpos + 1;
989         get_bh(di_bh);
990         wc->w_di_bh = di_bh;
991
992         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
993                 wc->w_large_pages = 1;
994         else
995                 wc->w_large_pages = 0;
996
997         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
998
999         *wcp = wc;
1000
1001         return 0;
1002 }
1003
1004 /*
1005  * If a page has any new buffers, zero them out here, and mark them uptodate
1006  * and dirty so they'll be written out (in order to prevent uninitialised
1007  * block data from leaking). And clear the new bit.
1008  */
1009 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1010 {
1011         unsigned int block_start, block_end;
1012         struct buffer_head *head, *bh;
1013
1014         BUG_ON(!PageLocked(page));
1015         if (!page_has_buffers(page))
1016                 return;
1017
1018         bh = head = page_buffers(page);
1019         block_start = 0;
1020         do {
1021                 block_end = block_start + bh->b_size;
1022
1023                 if (buffer_new(bh)) {
1024                         if (block_end > from && block_start < to) {
1025                                 if (!PageUptodate(page)) {
1026                                         unsigned start, end;
1027
1028                                         start = max(from, block_start);
1029                                         end = min(to, block_end);
1030
1031                                         zero_user_segment(page, start, end);
1032                                         set_buffer_uptodate(bh);
1033                                 }
1034
1035                                 clear_buffer_new(bh);
1036                                 mark_buffer_dirty(bh);
1037                         }
1038                 }
1039
1040                 block_start = block_end;
1041                 bh = bh->b_this_page;
1042         } while (bh != head);
1043 }
1044
1045 /*
1046  * Only called when we have a failure during allocating write to write
1047  * zero's to the newly allocated region.
1048  */
1049 static void ocfs2_write_failure(struct inode *inode,
1050                                 struct ocfs2_write_ctxt *wc,
1051                                 loff_t user_pos, unsigned user_len)
1052 {
1053         int i;
1054         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1055                 to = user_pos + user_len;
1056         struct page *tmppage;
1057
1058         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1059
1060         for(i = 0; i < wc->w_num_pages; i++) {
1061                 tmppage = wc->w_pages[i];
1062
1063                 if (page_has_buffers(tmppage)) {
1064                         if (ocfs2_should_order_data(inode))
1065                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1066
1067                         block_commit_write(tmppage, from, to);
1068                 }
1069         }
1070 }
1071
1072 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1073                                         struct ocfs2_write_ctxt *wc,
1074                                         struct page *page, u32 cpos,
1075                                         loff_t user_pos, unsigned user_len,
1076                                         int new)
1077 {
1078         int ret;
1079         unsigned int map_from = 0, map_to = 0;
1080         unsigned int cluster_start, cluster_end;
1081         unsigned int user_data_from = 0, user_data_to = 0;
1082
1083         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1084                                         &cluster_start, &cluster_end);
1085
1086         if (page == wc->w_target_page) {
1087                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1088                 map_to = map_from + user_len;
1089
1090                 if (new)
1091                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1092                                                     cluster_start, cluster_end,
1093                                                     new);
1094                 else
1095                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1096                                                     map_from, map_to, new);
1097                 if (ret) {
1098                         mlog_errno(ret);
1099                         goto out;
1100                 }
1101
1102                 user_data_from = map_from;
1103                 user_data_to = map_to;
1104                 if (new) {
1105                         map_from = cluster_start;
1106                         map_to = cluster_end;
1107                 }
1108         } else {
1109                 /*
1110                  * If we haven't allocated the new page yet, we
1111                  * shouldn't be writing it out without copying user
1112                  * data. This is likely a math error from the caller.
1113                  */
1114                 BUG_ON(!new);
1115
1116                 map_from = cluster_start;
1117                 map_to = cluster_end;
1118
1119                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1120                                             cluster_start, cluster_end, new);
1121                 if (ret) {
1122                         mlog_errno(ret);
1123                         goto out;
1124                 }
1125         }
1126
1127         /*
1128          * Parts of newly allocated pages need to be zero'd.
1129          *
1130          * Above, we have also rewritten 'to' and 'from' - as far as
1131          * the rest of the function is concerned, the entire cluster
1132          * range inside of a page needs to be written.
1133          *
1134          * We can skip this if the page is up to date - it's already
1135          * been zero'd from being read in as a hole.
1136          */
1137         if (new && !PageUptodate(page))
1138                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1139                                          cpos, user_data_from, user_data_to);
1140
1141         flush_dcache_page(page);
1142
1143 out:
1144         return ret;
1145 }
1146
1147 /*
1148  * This function will only grab one clusters worth of pages.
1149  */
1150 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1151                                       struct ocfs2_write_ctxt *wc,
1152                                       u32 cpos, loff_t user_pos, int new,
1153                                       struct page *mmap_page)
1154 {
1155         int ret = 0, i;
1156         unsigned long start, target_index, index;
1157         struct inode *inode = mapping->host;
1158
1159         target_index = user_pos >> PAGE_CACHE_SHIFT;
1160
1161         /*
1162          * Figure out how many pages we'll be manipulating here. For
1163          * non allocating write, we just change the one
1164          * page. Otherwise, we'll need a whole clusters worth.
1165          */
1166         if (new) {
1167                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1168                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1169         } else {
1170                 wc->w_num_pages = 1;
1171                 start = target_index;
1172         }
1173
1174         for(i = 0; i < wc->w_num_pages; i++) {
1175                 index = start + i;
1176
1177                 if (index == target_index && mmap_page) {
1178                         /*
1179                          * ocfs2_pagemkwrite() is a little different
1180                          * and wants us to directly use the page
1181                          * passed in.
1182                          */
1183                         lock_page(mmap_page);
1184
1185                         if (mmap_page->mapping != mapping) {
1186                                 unlock_page(mmap_page);
1187                                 /*
1188                                  * Sanity check - the locking in
1189                                  * ocfs2_pagemkwrite() should ensure
1190                                  * that this code doesn't trigger.
1191                                  */
1192                                 ret = -EINVAL;
1193                                 mlog_errno(ret);
1194                                 goto out;
1195                         }
1196
1197                         page_cache_get(mmap_page);
1198                         wc->w_pages[i] = mmap_page;
1199                 } else {
1200                         wc->w_pages[i] = find_or_create_page(mapping, index,
1201                                                              GFP_NOFS);
1202                         if (!wc->w_pages[i]) {
1203                                 ret = -ENOMEM;
1204                                 mlog_errno(ret);
1205                                 goto out;
1206                         }
1207                 }
1208
1209                 if (index == target_index)
1210                         wc->w_target_page = wc->w_pages[i];
1211         }
1212 out:
1213         return ret;
1214 }
1215
1216 /*
1217  * Prepare a single cluster for write one cluster into the file.
1218  */
1219 static int ocfs2_write_cluster(struct address_space *mapping,
1220                                u32 phys, unsigned int unwritten,
1221                                unsigned int should_zero,
1222                                struct ocfs2_alloc_context *data_ac,
1223                                struct ocfs2_alloc_context *meta_ac,
1224                                struct ocfs2_write_ctxt *wc, u32 cpos,
1225                                loff_t user_pos, unsigned user_len)
1226 {
1227         int ret, i, new;
1228         u64 v_blkno, p_blkno;
1229         struct inode *inode = mapping->host;
1230         struct ocfs2_extent_tree et;
1231
1232         new = phys == 0 ? 1 : 0;
1233         if (new) {
1234                 u32 tmp_pos;
1235
1236                 /*
1237                  * This is safe to call with the page locks - it won't take
1238                  * any additional semaphores or cluster locks.
1239                  */
1240                 tmp_pos = cpos;
1241                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1242                                            &tmp_pos, 1, 0, wc->w_di_bh,
1243                                            wc->w_handle, data_ac,
1244                                            meta_ac, NULL);
1245                 /*
1246                  * This shouldn't happen because we must have already
1247                  * calculated the correct meta data allocation required. The
1248                  * internal tree allocation code should know how to increase
1249                  * transaction credits itself.
1250                  *
1251                  * If need be, we could handle -EAGAIN for a
1252                  * RESTART_TRANS here.
1253                  */
1254                 mlog_bug_on_msg(ret == -EAGAIN,
1255                                 "Inode %llu: EAGAIN return during allocation.\n",
1256                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1257                 if (ret < 0) {
1258                         mlog_errno(ret);
1259                         goto out;
1260                 }
1261         } else if (unwritten) {
1262                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1263                 ret = ocfs2_mark_extent_written(inode, &et,
1264                                                 wc->w_handle, cpos, 1, phys,
1265                                                 meta_ac, &wc->w_dealloc);
1266                 if (ret < 0) {
1267                         mlog_errno(ret);
1268                         goto out;
1269                 }
1270         }
1271
1272         if (should_zero)
1273                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1274         else
1275                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1276
1277         /*
1278          * The only reason this should fail is due to an inability to
1279          * find the extent added.
1280          */
1281         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1282                                           NULL);
1283         if (ret < 0) {
1284                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1285                             "at logical block %llu",
1286                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1287                             (unsigned long long)v_blkno);
1288                 goto out;
1289         }
1290
1291         BUG_ON(p_blkno == 0);
1292
1293         for(i = 0; i < wc->w_num_pages; i++) {
1294                 int tmpret;
1295
1296                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1297                                                       wc->w_pages[i], cpos,
1298                                                       user_pos, user_len,
1299                                                       should_zero);
1300                 if (tmpret) {
1301                         mlog_errno(tmpret);
1302                         if (ret == 0)
1303                                 ret = tmpret;
1304                 }
1305         }
1306
1307         /*
1308          * We only have cleanup to do in case of allocating write.
1309          */
1310         if (ret && new)
1311                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1312
1313 out:
1314
1315         return ret;
1316 }
1317
1318 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1319                                        struct ocfs2_alloc_context *data_ac,
1320                                        struct ocfs2_alloc_context *meta_ac,
1321                                        struct ocfs2_write_ctxt *wc,
1322                                        loff_t pos, unsigned len)
1323 {
1324         int ret, i;
1325         loff_t cluster_off;
1326         unsigned int local_len = len;
1327         struct ocfs2_write_cluster_desc *desc;
1328         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1329
1330         for (i = 0; i < wc->w_clen; i++) {
1331                 desc = &wc->w_desc[i];
1332
1333                 /*
1334                  * We have to make sure that the total write passed in
1335                  * doesn't extend past a single cluster.
1336                  */
1337                 local_len = len;
1338                 cluster_off = pos & (osb->s_clustersize - 1);
1339                 if ((cluster_off + local_len) > osb->s_clustersize)
1340                         local_len = osb->s_clustersize - cluster_off;
1341
1342                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1343                                           desc->c_unwritten,
1344                                           desc->c_needs_zero,
1345                                           data_ac, meta_ac,
1346                                           wc, desc->c_cpos, pos, local_len);
1347                 if (ret) {
1348                         mlog_errno(ret);
1349                         goto out;
1350                 }
1351
1352                 len -= local_len;
1353                 pos += local_len;
1354         }
1355
1356         ret = 0;
1357 out:
1358         return ret;
1359 }
1360
1361 /*
1362  * ocfs2_write_end() wants to know which parts of the target page it
1363  * should complete the write on. It's easiest to compute them ahead of
1364  * time when a more complete view of the write is available.
1365  */
1366 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1367                                         struct ocfs2_write_ctxt *wc,
1368                                         loff_t pos, unsigned len, int alloc)
1369 {
1370         struct ocfs2_write_cluster_desc *desc;
1371
1372         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1373         wc->w_target_to = wc->w_target_from + len;
1374
1375         if (alloc == 0)
1376                 return;
1377
1378         /*
1379          * Allocating write - we may have different boundaries based
1380          * on page size and cluster size.
1381          *
1382          * NOTE: We can no longer compute one value from the other as
1383          * the actual write length and user provided length may be
1384          * different.
1385          */
1386
1387         if (wc->w_large_pages) {
1388                 /*
1389                  * We only care about the 1st and last cluster within
1390                  * our range and whether they should be zero'd or not. Either
1391                  * value may be extended out to the start/end of a
1392                  * newly allocated cluster.
1393                  */
1394                 desc = &wc->w_desc[0];
1395                 if (desc->c_needs_zero)
1396                         ocfs2_figure_cluster_boundaries(osb,
1397                                                         desc->c_cpos,
1398                                                         &wc->w_target_from,
1399                                                         NULL);
1400
1401                 desc = &wc->w_desc[wc->w_clen - 1];
1402                 if (desc->c_needs_zero)
1403                         ocfs2_figure_cluster_boundaries(osb,
1404                                                         desc->c_cpos,
1405                                                         NULL,
1406                                                         &wc->w_target_to);
1407         } else {
1408                 wc->w_target_from = 0;
1409                 wc->w_target_to = PAGE_CACHE_SIZE;
1410         }
1411 }
1412
1413 /*
1414  * Populate each single-cluster write descriptor in the write context
1415  * with information about the i/o to be done.
1416  *
1417  * Returns the number of clusters that will have to be allocated, as
1418  * well as a worst case estimate of the number of extent records that
1419  * would have to be created during a write to an unwritten region.
1420  */
1421 static int ocfs2_populate_write_desc(struct inode *inode,
1422                                      struct ocfs2_write_ctxt *wc,
1423                                      unsigned int *clusters_to_alloc,
1424                                      unsigned int *extents_to_split)
1425 {
1426         int ret;
1427         struct ocfs2_write_cluster_desc *desc;
1428         unsigned int num_clusters = 0;
1429         unsigned int ext_flags = 0;
1430         u32 phys = 0;
1431         int i;
1432
1433         *clusters_to_alloc = 0;
1434         *extents_to_split = 0;
1435
1436         for (i = 0; i < wc->w_clen; i++) {
1437                 desc = &wc->w_desc[i];
1438                 desc->c_cpos = wc->w_cpos + i;
1439
1440                 if (num_clusters == 0) {
1441                         /*
1442                          * Need to look up the next extent record.
1443                          */
1444                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1445                                                  &num_clusters, &ext_flags);
1446                         if (ret) {
1447                                 mlog_errno(ret);
1448                                 goto out;
1449                         }
1450
1451                         /*
1452                          * Assume worst case - that we're writing in
1453                          * the middle of the extent.
1454                          *
1455                          * We can assume that the write proceeds from
1456                          * left to right, in which case the extent
1457                          * insert code is smart enough to coalesce the
1458                          * next splits into the previous records created.
1459                          */
1460                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1461                                 *extents_to_split = *extents_to_split + 2;
1462                 } else if (phys) {
1463                         /*
1464                          * Only increment phys if it doesn't describe
1465                          * a hole.
1466                          */
1467                         phys++;
1468                 }
1469
1470                 /*
1471                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1472                  * file that got extended.  w_first_new_cpos tells us
1473                  * where the newly allocated clusters are so we can
1474                  * zero them.
1475                  */
1476                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1477                         BUG_ON(phys == 0);
1478                         desc->c_needs_zero = 1;
1479                 }
1480
1481                 desc->c_phys = phys;
1482                 if (phys == 0) {
1483                         desc->c_new = 1;
1484                         desc->c_needs_zero = 1;
1485                         *clusters_to_alloc = *clusters_to_alloc + 1;
1486                 }
1487
1488                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1489                         desc->c_unwritten = 1;
1490                         desc->c_needs_zero = 1;
1491                 }
1492
1493                 num_clusters--;
1494         }
1495
1496         ret = 0;
1497 out:
1498         return ret;
1499 }
1500
1501 static int ocfs2_write_begin_inline(struct address_space *mapping,
1502                                     struct inode *inode,
1503                                     struct ocfs2_write_ctxt *wc)
1504 {
1505         int ret;
1506         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1507         struct page *page;
1508         handle_t *handle;
1509         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1510
1511         page = find_or_create_page(mapping, 0, GFP_NOFS);
1512         if (!page) {
1513                 ret = -ENOMEM;
1514                 mlog_errno(ret);
1515                 goto out;
1516         }
1517         /*
1518          * If we don't set w_num_pages then this page won't get unlocked
1519          * and freed on cleanup of the write context.
1520          */
1521         wc->w_pages[0] = wc->w_target_page = page;
1522         wc->w_num_pages = 1;
1523
1524         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1525         if (IS_ERR(handle)) {
1526                 ret = PTR_ERR(handle);
1527                 mlog_errno(ret);
1528                 goto out;
1529         }
1530
1531         ret = ocfs2_journal_access_di(handle, inode, wc->w_di_bh,
1532                                       OCFS2_JOURNAL_ACCESS_WRITE);
1533         if (ret) {
1534                 ocfs2_commit_trans(osb, handle);
1535
1536                 mlog_errno(ret);
1537                 goto out;
1538         }
1539
1540         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1541                 ocfs2_set_inode_data_inline(inode, di);
1542
1543         if (!PageUptodate(page)) {
1544                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1545                 if (ret) {
1546                         ocfs2_commit_trans(osb, handle);
1547
1548                         goto out;
1549                 }
1550         }
1551
1552         wc->w_handle = handle;
1553 out:
1554         return ret;
1555 }
1556
1557 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1558 {
1559         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1560
1561         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1562                 return 1;
1563         return 0;
1564 }
1565
1566 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1567                                           struct inode *inode, loff_t pos,
1568                                           unsigned len, struct page *mmap_page,
1569                                           struct ocfs2_write_ctxt *wc)
1570 {
1571         int ret, written = 0;
1572         loff_t end = pos + len;
1573         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1574         struct ocfs2_dinode *di = NULL;
1575
1576         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1577              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1578              oi->ip_dyn_features);
1579
1580         /*
1581          * Handle inodes which already have inline data 1st.
1582          */
1583         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1584                 if (mmap_page == NULL &&
1585                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1586                         goto do_inline_write;
1587
1588                 /*
1589                  * The write won't fit - we have to give this inode an
1590                  * inline extent list now.
1591                  */
1592                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1593                 if (ret)
1594                         mlog_errno(ret);
1595                 goto out;
1596         }
1597
1598         /*
1599          * Check whether the inode can accept inline data.
1600          */
1601         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1602                 return 0;
1603
1604         /*
1605          * Check whether the write can fit.
1606          */
1607         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1608         if (mmap_page ||
1609             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1610                 return 0;
1611
1612 do_inline_write:
1613         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1614         if (ret) {
1615                 mlog_errno(ret);
1616                 goto out;
1617         }
1618
1619         /*
1620          * This signals to the caller that the data can be written
1621          * inline.
1622          */
1623         written = 1;
1624 out:
1625         return written ? written : ret;
1626 }
1627
1628 /*
1629  * This function only does anything for file systems which can't
1630  * handle sparse files.
1631  *
1632  * What we want to do here is fill in any hole between the current end
1633  * of allocation and the end of our write. That way the rest of the
1634  * write path can treat it as an non-allocating write, which has no
1635  * special case code for sparse/nonsparse files.
1636  */
1637 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1638                                         unsigned len,
1639                                         struct ocfs2_write_ctxt *wc)
1640 {
1641         int ret;
1642         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1643         loff_t newsize = pos + len;
1644
1645         if (ocfs2_sparse_alloc(osb))
1646                 return 0;
1647
1648         if (newsize <= i_size_read(inode))
1649                 return 0;
1650
1651         ret = ocfs2_extend_no_holes(inode, newsize, pos);
1652         if (ret)
1653                 mlog_errno(ret);
1654
1655         wc->w_first_new_cpos =
1656                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1657
1658         return ret;
1659 }
1660
1661 int ocfs2_write_begin_nolock(struct address_space *mapping,
1662                              loff_t pos, unsigned len, unsigned flags,
1663                              struct page **pagep, void **fsdata,
1664                              struct buffer_head *di_bh, struct page *mmap_page)
1665 {
1666         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1667         unsigned int clusters_to_alloc, extents_to_split;
1668         struct ocfs2_write_ctxt *wc;
1669         struct inode *inode = mapping->host;
1670         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1671         struct ocfs2_dinode *di;
1672         struct ocfs2_alloc_context *data_ac = NULL;
1673         struct ocfs2_alloc_context *meta_ac = NULL;
1674         handle_t *handle;
1675         struct ocfs2_extent_tree et;
1676
1677         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1678         if (ret) {
1679                 mlog_errno(ret);
1680                 return ret;
1681         }
1682
1683         if (ocfs2_supports_inline_data(osb)) {
1684                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1685                                                      mmap_page, wc);
1686                 if (ret == 1) {
1687                         ret = 0;
1688                         goto success;
1689                 }
1690                 if (ret < 0) {
1691                         mlog_errno(ret);
1692                         goto out;
1693                 }
1694         }
1695
1696         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1697         if (ret) {
1698                 mlog_errno(ret);
1699                 goto out;
1700         }
1701
1702         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1703                                         &extents_to_split);
1704         if (ret) {
1705                 mlog_errno(ret);
1706                 goto out;
1707         }
1708
1709         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1710
1711         /*
1712          * We set w_target_from, w_target_to here so that
1713          * ocfs2_write_end() knows which range in the target page to
1714          * write out. An allocation requires that we write the entire
1715          * cluster range.
1716          */
1717         if (clusters_to_alloc || extents_to_split) {
1718                 /*
1719                  * XXX: We are stretching the limits of
1720                  * ocfs2_lock_allocators(). It greatly over-estimates
1721                  * the work to be done.
1722                  */
1723                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1724                      " clusters_to_add = %u, extents_to_split = %u\n",
1725                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1726                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1727                      clusters_to_alloc, extents_to_split);
1728
1729                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1730                 ret = ocfs2_lock_allocators(inode, &et,
1731                                             clusters_to_alloc, extents_to_split,
1732                                             &data_ac, &meta_ac);
1733                 if (ret) {
1734                         mlog_errno(ret);
1735                         goto out;
1736                 }
1737
1738                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1739                                                     &di->id2.i_list,
1740                                                     clusters_to_alloc);
1741
1742         }
1743
1744         /*
1745          * We have to zero sparse allocated clusters, unwritten extent clusters,
1746          * and non-sparse clusters we just extended.  For non-sparse writes,
1747          * we know zeros will only be needed in the first and/or last cluster.
1748          */
1749         if (clusters_to_alloc || extents_to_split ||
1750             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1751                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1752                 cluster_of_pages = 1;
1753         else
1754                 cluster_of_pages = 0;
1755
1756         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1757
1758         handle = ocfs2_start_trans(osb, credits);
1759         if (IS_ERR(handle)) {
1760                 ret = PTR_ERR(handle);
1761                 mlog_errno(ret);
1762                 goto out;
1763         }
1764
1765         wc->w_handle = handle;
1766
1767         if (clusters_to_alloc && vfs_dq_alloc_space_nodirty(inode,
1768                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc))) {
1769                 ret = -EDQUOT;
1770                 goto out_commit;
1771         }
1772         /*
1773          * We don't want this to fail in ocfs2_write_end(), so do it
1774          * here.
1775          */
1776         ret = ocfs2_journal_access_di(handle, inode, wc->w_di_bh,
1777                                       OCFS2_JOURNAL_ACCESS_WRITE);
1778         if (ret) {
1779                 mlog_errno(ret);
1780                 goto out_quota;
1781         }
1782
1783         /*
1784          * Fill our page array first. That way we've grabbed enough so
1785          * that we can zero and flush if we error after adding the
1786          * extent.
1787          */
1788         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1789                                          cluster_of_pages, mmap_page);
1790         if (ret) {
1791                 mlog_errno(ret);
1792                 goto out_quota;
1793         }
1794
1795         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1796                                           len);
1797         if (ret) {
1798                 mlog_errno(ret);
1799                 goto out_quota;
1800         }
1801
1802         if (data_ac)
1803                 ocfs2_free_alloc_context(data_ac);
1804         if (meta_ac)
1805                 ocfs2_free_alloc_context(meta_ac);
1806
1807 success:
1808         *pagep = wc->w_target_page;
1809         *fsdata = wc;
1810         return 0;
1811 out_quota:
1812         if (clusters_to_alloc)
1813                 vfs_dq_free_space(inode,
1814                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1815 out_commit:
1816         ocfs2_commit_trans(osb, handle);
1817
1818 out:
1819         ocfs2_free_write_ctxt(wc);
1820
1821         if (data_ac)
1822                 ocfs2_free_alloc_context(data_ac);
1823         if (meta_ac)
1824                 ocfs2_free_alloc_context(meta_ac);
1825         return ret;
1826 }
1827
1828 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1829                              loff_t pos, unsigned len, unsigned flags,
1830                              struct page **pagep, void **fsdata)
1831 {
1832         int ret;
1833         struct buffer_head *di_bh = NULL;
1834         struct inode *inode = mapping->host;
1835
1836         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1837         if (ret) {
1838                 mlog_errno(ret);
1839                 return ret;
1840         }
1841
1842         /*
1843          * Take alloc sem here to prevent concurrent lookups. That way
1844          * the mapping, zeroing and tree manipulation within
1845          * ocfs2_write() will be safe against ->readpage(). This
1846          * should also serve to lock out allocation from a shared
1847          * writeable region.
1848          */
1849         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1850
1851         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1852                                        fsdata, di_bh, NULL);
1853         if (ret) {
1854                 mlog_errno(ret);
1855                 goto out_fail;
1856         }
1857
1858         brelse(di_bh);
1859
1860         return 0;
1861
1862 out_fail:
1863         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1864
1865         brelse(di_bh);
1866         ocfs2_inode_unlock(inode, 1);
1867
1868         return ret;
1869 }
1870
1871 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1872                                    unsigned len, unsigned *copied,
1873                                    struct ocfs2_dinode *di,
1874                                    struct ocfs2_write_ctxt *wc)
1875 {
1876         void *kaddr;
1877
1878         if (unlikely(*copied < len)) {
1879                 if (!PageUptodate(wc->w_target_page)) {
1880                         *copied = 0;
1881                         return;
1882                 }
1883         }
1884
1885         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1886         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1887         kunmap_atomic(kaddr, KM_USER0);
1888
1889         mlog(0, "Data written to inode at offset %llu. "
1890              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1891              (unsigned long long)pos, *copied,
1892              le16_to_cpu(di->id2.i_data.id_count),
1893              le16_to_cpu(di->i_dyn_features));
1894 }
1895
1896 int ocfs2_write_end_nolock(struct address_space *mapping,
1897                            loff_t pos, unsigned len, unsigned copied,
1898                            struct page *page, void *fsdata)
1899 {
1900         int i;
1901         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1902         struct inode *inode = mapping->host;
1903         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1904         struct ocfs2_write_ctxt *wc = fsdata;
1905         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1906         handle_t *handle = wc->w_handle;
1907         struct page *tmppage;
1908
1909         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1910                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1911                 goto out_write_size;
1912         }
1913
1914         if (unlikely(copied < len)) {
1915                 if (!PageUptodate(wc->w_target_page))
1916                         copied = 0;
1917
1918                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1919                                        start+len);
1920         }
1921         flush_dcache_page(wc->w_target_page);
1922
1923         for(i = 0; i < wc->w_num_pages; i++) {
1924                 tmppage = wc->w_pages[i];
1925
1926                 if (tmppage == wc->w_target_page) {
1927                         from = wc->w_target_from;
1928                         to = wc->w_target_to;
1929
1930                         BUG_ON(from > PAGE_CACHE_SIZE ||
1931                                to > PAGE_CACHE_SIZE ||
1932                                to < from);
1933                 } else {
1934                         /*
1935                          * Pages adjacent to the target (if any) imply
1936                          * a hole-filling write in which case we want
1937                          * to flush their entire range.
1938                          */
1939                         from = 0;
1940                         to = PAGE_CACHE_SIZE;
1941                 }
1942
1943                 if (page_has_buffers(tmppage)) {
1944                         if (ocfs2_should_order_data(inode))
1945                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1946                         block_commit_write(tmppage, from, to);
1947                 }
1948         }
1949
1950 out_write_size:
1951         pos += copied;
1952         if (pos > inode->i_size) {
1953                 i_size_write(inode, pos);
1954                 mark_inode_dirty(inode);
1955         }
1956         inode->i_blocks = ocfs2_inode_sector_count(inode);
1957         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1958         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1959         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1960         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1961         ocfs2_journal_dirty(handle, wc->w_di_bh);
1962
1963         ocfs2_commit_trans(osb, handle);
1964
1965         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1966
1967         ocfs2_free_write_ctxt(wc);
1968
1969         return copied;
1970 }
1971
1972 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1973                            loff_t pos, unsigned len, unsigned copied,
1974                            struct page *page, void *fsdata)
1975 {
1976         int ret;
1977         struct inode *inode = mapping->host;
1978
1979         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1980
1981         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1982         ocfs2_inode_unlock(inode, 1);
1983
1984         return ret;
1985 }
1986
1987 const struct address_space_operations ocfs2_aops = {
1988         .readpage               = ocfs2_readpage,
1989         .readpages              = ocfs2_readpages,
1990         .writepage              = ocfs2_writepage,
1991         .write_begin            = ocfs2_write_begin,
1992         .write_end              = ocfs2_write_end,
1993         .bmap                   = ocfs2_bmap,
1994         .sync_page              = block_sync_page,
1995         .direct_IO              = ocfs2_direct_IO,
1996         .invalidatepage         = ocfs2_invalidatepage,
1997         .releasepage            = ocfs2_releasepage,
1998         .migratepage            = buffer_migrate_page,
1999         .is_partially_uptodate  = block_is_partially_uptodate,
2000 };