Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[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(inode, OCFS2_I(inode)->ip_blkno, &bh);
72         if (status < 0) {
73                 mlog_errno(status);
74                 goto bail;
75         }
76         fe = (struct ocfs2_dinode *) bh->b_data;
77
78         if (!OCFS2_IS_VALID_DINODE(fe)) {
79                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
80                      (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
81                      fe->i_signature);
82                 goto bail;
83         }
84
85         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
86                                                     le32_to_cpu(fe->i_clusters))) {
87                 mlog(ML_ERROR, "block offset is outside the allocated size: "
88                      "%llu\n", (unsigned long long)iblock);
89                 goto bail;
90         }
91
92         /* We don't use the page cache to create symlink data, so if
93          * need be, copy it over from the buffer cache. */
94         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
95                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
96                             iblock;
97                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
98                 if (!buffer_cache_bh) {
99                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
100                         goto bail;
101                 }
102
103                 /* we haven't locked out transactions, so a commit
104                  * could've happened. Since we've got a reference on
105                  * the bh, even if it commits while we're doing the
106                  * copy, the data is still good. */
107                 if (buffer_jbd(buffer_cache_bh)
108                     && ocfs2_inode_is_new(inode)) {
109                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
110                         if (!kaddr) {
111                                 mlog(ML_ERROR, "couldn't kmap!\n");
112                                 goto bail;
113                         }
114                         memcpy(kaddr + (bh_result->b_size * iblock),
115                                buffer_cache_bh->b_data,
116                                bh_result->b_size);
117                         kunmap_atomic(kaddr, KM_USER0);
118                         set_buffer_uptodate(bh_result);
119                 }
120                 brelse(buffer_cache_bh);
121         }
122
123         map_bh(bh_result, inode->i_sb,
124                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
125
126         err = 0;
127
128 bail:
129         brelse(bh);
130
131         mlog_exit(err);
132         return err;
133 }
134
135 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
136                            struct buffer_head *bh_result, int create)
137 {
138         int err = 0;
139         unsigned int ext_flags;
140         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
141         u64 p_blkno, count, past_eof;
142         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
143
144         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
145                    (unsigned long long)iblock, bh_result, create);
146
147         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
148                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
149                      inode, inode->i_ino);
150
151         if (S_ISLNK(inode->i_mode)) {
152                 /* this always does I/O for some reason. */
153                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
154                 goto bail;
155         }
156
157         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
158                                           &ext_flags);
159         if (err) {
160                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
161                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
162                      (unsigned long long)p_blkno);
163                 goto bail;
164         }
165
166         if (max_blocks < count)
167                 count = max_blocks;
168
169         /*
170          * ocfs2 never allocates in this function - the only time we
171          * need to use BH_New is when we're extending i_size on a file
172          * system which doesn't support holes, in which case BH_New
173          * allows block_prepare_write() to zero.
174          *
175          * If we see this on a sparse file system, then a truncate has
176          * raced us and removed the cluster. In this case, we clear
177          * the buffers dirty and uptodate bits and let the buffer code
178          * ignore it as a hole.
179          */
180         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
181                 clear_buffer_dirty(bh_result);
182                 clear_buffer_uptodate(bh_result);
183                 goto bail;
184         }
185
186         /* Treat the unwritten extent as a hole for zeroing purposes. */
187         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
188                 map_bh(bh_result, inode->i_sb, p_blkno);
189
190         bh_result->b_size = count << inode->i_blkbits;
191
192         if (!ocfs2_sparse_alloc(osb)) {
193                 if (p_blkno == 0) {
194                         err = -EIO;
195                         mlog(ML_ERROR,
196                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
197                              (unsigned long long)iblock,
198                              (unsigned long long)p_blkno,
199                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
200                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
201                         dump_stack();
202                 }
203
204                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
205                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
206                      (unsigned long long)past_eof);
207
208                 if (create && (iblock >= past_eof))
209                         set_buffer_new(bh_result);
210         }
211
212 bail:
213         if (err < 0)
214                 err = -EIO;
215
216         mlog_exit(err);
217         return err;
218 }
219
220 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
221                            struct buffer_head *di_bh)
222 {
223         void *kaddr;
224         loff_t size;
225         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
226
227         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
228                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
229                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
230                 return -EROFS;
231         }
232
233         size = i_size_read(inode);
234
235         if (size > PAGE_CACHE_SIZE ||
236             size > ocfs2_max_inline_data(inode->i_sb)) {
237                 ocfs2_error(inode->i_sb,
238                             "Inode %llu has with inline data has bad size: %Lu",
239                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
240                             (unsigned long long)size);
241                 return -EROFS;
242         }
243
244         kaddr = kmap_atomic(page, KM_USER0);
245         if (size)
246                 memcpy(kaddr, di->id2.i_data.id_data, size);
247         /* Clear the remaining part of the page */
248         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
249         flush_dcache_page(page);
250         kunmap_atomic(kaddr, KM_USER0);
251
252         SetPageUptodate(page);
253
254         return 0;
255 }
256
257 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
258 {
259         int ret;
260         struct buffer_head *di_bh = NULL;
261
262         BUG_ON(!PageLocked(page));
263         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
264
265         ret = ocfs2_read_block(inode, OCFS2_I(inode)->ip_blkno, &di_bh);
266         if (ret) {
267                 mlog_errno(ret);
268                 goto out;
269         }
270
271         ret = ocfs2_read_inline_data(inode, page, di_bh);
272 out:
273         unlock_page(page);
274
275         brelse(di_bh);
276         return ret;
277 }
278
279 static int ocfs2_readpage(struct file *file, struct page *page)
280 {
281         struct inode *inode = page->mapping->host;
282         struct ocfs2_inode_info *oi = OCFS2_I(inode);
283         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
284         int ret, unlock = 1;
285
286         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
287
288         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
289         if (ret != 0) {
290                 if (ret == AOP_TRUNCATED_PAGE)
291                         unlock = 0;
292                 mlog_errno(ret);
293                 goto out;
294         }
295
296         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
297                 ret = AOP_TRUNCATED_PAGE;
298                 goto out_inode_unlock;
299         }
300
301         /*
302          * i_size might have just been updated as we grabed the meta lock.  We
303          * might now be discovering a truncate that hit on another node.
304          * block_read_full_page->get_block freaks out if it is asked to read
305          * beyond the end of a file, so we check here.  Callers
306          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
307          * and notice that the page they just read isn't needed.
308          *
309          * XXX sys_readahead() seems to get that wrong?
310          */
311         if (start >= i_size_read(inode)) {
312                 zero_user(page, 0, PAGE_SIZE);
313                 SetPageUptodate(page);
314                 ret = 0;
315                 goto out_alloc;
316         }
317
318         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
319                 ret = ocfs2_readpage_inline(inode, page);
320         else
321                 ret = block_read_full_page(page, ocfs2_get_block);
322         unlock = 0;
323
324 out_alloc:
325         up_read(&OCFS2_I(inode)->ip_alloc_sem);
326 out_inode_unlock:
327         ocfs2_inode_unlock(inode, 0);
328 out:
329         if (unlock)
330                 unlock_page(page);
331         mlog_exit(ret);
332         return ret;
333 }
334
335 /*
336  * This is used only for read-ahead. Failures or difficult to handle
337  * situations are safe to ignore.
338  *
339  * Right now, we don't bother with BH_Boundary - in-inode extent lists
340  * are quite large (243 extents on 4k blocks), so most inodes don't
341  * grow out to a tree. If need be, detecting boundary extents could
342  * trivially be added in a future version of ocfs2_get_block().
343  */
344 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
345                            struct list_head *pages, unsigned nr_pages)
346 {
347         int ret, err = -EIO;
348         struct inode *inode = mapping->host;
349         struct ocfs2_inode_info *oi = OCFS2_I(inode);
350         loff_t start;
351         struct page *last;
352
353         /*
354          * Use the nonblocking flag for the dlm code to avoid page
355          * lock inversion, but don't bother with retrying.
356          */
357         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
358         if (ret)
359                 return err;
360
361         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
362                 ocfs2_inode_unlock(inode, 0);
363                 return err;
364         }
365
366         /*
367          * Don't bother with inline-data. There isn't anything
368          * to read-ahead in that case anyway...
369          */
370         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
371                 goto out_unlock;
372
373         /*
374          * Check whether a remote node truncated this file - we just
375          * drop out in that case as it's not worth handling here.
376          */
377         last = list_entry(pages->prev, struct page, lru);
378         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
379         if (start >= i_size_read(inode))
380                 goto out_unlock;
381
382         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
383
384 out_unlock:
385         up_read(&oi->ip_alloc_sem);
386         ocfs2_inode_unlock(inode, 0);
387
388         return err;
389 }
390
391 /* Note: Because we don't support holes, our allocation has
392  * already happened (allocation writes zeros to the file data)
393  * so we don't have to worry about ordered writes in
394  * ocfs2_writepage.
395  *
396  * ->writepage is called during the process of invalidating the page cache
397  * during blocked lock processing.  It can't block on any cluster locks
398  * to during block mapping.  It's relying on the fact that the block
399  * mapping can't have disappeared under the dirty pages that it is
400  * being asked to write back.
401  */
402 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
403 {
404         int ret;
405
406         mlog_entry("(0x%p)\n", page);
407
408         ret = block_write_full_page(page, ocfs2_get_block, wbc);
409
410         mlog_exit(ret);
411
412         return ret;
413 }
414
415 /*
416  * This is called from ocfs2_write_zero_page() which has handled it's
417  * own cluster locking and has ensured allocation exists for those
418  * blocks to be written.
419  */
420 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
421                                unsigned from, unsigned to)
422 {
423         int ret;
424
425         ret = block_prepare_write(page, from, to, ocfs2_get_block);
426
427         return ret;
428 }
429
430 /* Taken from ext3. We don't necessarily need the full blown
431  * functionality yet, but IMHO it's better to cut and paste the whole
432  * thing so we can avoid introducing our own bugs (and easily pick up
433  * their fixes when they happen) --Mark */
434 int walk_page_buffers(  handle_t *handle,
435                         struct buffer_head *head,
436                         unsigned from,
437                         unsigned to,
438                         int *partial,
439                         int (*fn)(      handle_t *handle,
440                                         struct buffer_head *bh))
441 {
442         struct buffer_head *bh;
443         unsigned block_start, block_end;
444         unsigned blocksize = head->b_size;
445         int err, ret = 0;
446         struct buffer_head *next;
447
448         for (   bh = head, block_start = 0;
449                 ret == 0 && (bh != head || !block_start);
450                 block_start = block_end, bh = next)
451         {
452                 next = bh->b_this_page;
453                 block_end = block_start + blocksize;
454                 if (block_end <= from || block_start >= to) {
455                         if (partial && !buffer_uptodate(bh))
456                                 *partial = 1;
457                         continue;
458                 }
459                 err = (*fn)(handle, bh);
460                 if (!ret)
461                         ret = err;
462         }
463         return ret;
464 }
465
466 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
467                                                          struct page *page,
468                                                          unsigned from,
469                                                          unsigned to)
470 {
471         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
472         handle_t *handle;
473         int ret = 0;
474
475         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
476         if (IS_ERR(handle)) {
477                 ret = -ENOMEM;
478                 mlog_errno(ret);
479                 goto out;
480         }
481
482         if (ocfs2_should_order_data(inode)) {
483                 ret = ocfs2_jbd2_file_inode(handle, inode);
484 #ifdef CONFIG_OCFS2_COMPAT_JBD
485                 ret = walk_page_buffers(handle,
486                                         page_buffers(page),
487                                         from, to, NULL,
488                                         ocfs2_journal_dirty_data);
489 #endif
490                 if (ret < 0)
491                         mlog_errno(ret);
492         }
493 out:
494         if (ret) {
495                 if (!IS_ERR(handle))
496                         ocfs2_commit_trans(osb, handle);
497                 handle = ERR_PTR(ret);
498         }
499         return handle;
500 }
501
502 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
503 {
504         sector_t status;
505         u64 p_blkno = 0;
506         int err = 0;
507         struct inode *inode = mapping->host;
508
509         mlog_entry("(block = %llu)\n", (unsigned long long)block);
510
511         /* We don't need to lock journal system files, since they aren't
512          * accessed concurrently from multiple nodes.
513          */
514         if (!INODE_JOURNAL(inode)) {
515                 err = ocfs2_inode_lock(inode, NULL, 0);
516                 if (err) {
517                         if (err != -ENOENT)
518                                 mlog_errno(err);
519                         goto bail;
520                 }
521                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
522         }
523
524         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
525                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
526                                                   NULL);
527
528         if (!INODE_JOURNAL(inode)) {
529                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
530                 ocfs2_inode_unlock(inode, 0);
531         }
532
533         if (err) {
534                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
535                      (unsigned long long)block);
536                 mlog_errno(err);
537                 goto bail;
538         }
539
540 bail:
541         status = err ? 0 : p_blkno;
542
543         mlog_exit((int)status);
544
545         return status;
546 }
547
548 /*
549  * TODO: Make this into a generic get_blocks function.
550  *
551  * From do_direct_io in direct-io.c:
552  *  "So what we do is to permit the ->get_blocks function to populate
553  *   bh.b_size with the size of IO which is permitted at this offset and
554  *   this i_blkbits."
555  *
556  * This function is called directly from get_more_blocks in direct-io.c.
557  *
558  * called like this: dio->get_blocks(dio->inode, fs_startblk,
559  *                                      fs_count, map_bh, dio->rw == WRITE);
560  */
561 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
562                                      struct buffer_head *bh_result, int create)
563 {
564         int ret;
565         u64 p_blkno, inode_blocks, contig_blocks;
566         unsigned int ext_flags;
567         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
568         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
569
570         /* This function won't even be called if the request isn't all
571          * nicely aligned and of the right size, so there's no need
572          * for us to check any of that. */
573
574         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
575
576         /*
577          * Any write past EOF is not allowed because we'd be extending.
578          */
579         if (create && (iblock + max_blocks) > inode_blocks) {
580                 ret = -EIO;
581                 goto bail;
582         }
583
584         /* This figures out the size of the next contiguous block, and
585          * our logical offset */
586         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
587                                           &contig_blocks, &ext_flags);
588         if (ret) {
589                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
590                      (unsigned long long)iblock);
591                 ret = -EIO;
592                 goto bail;
593         }
594
595         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
596                 ocfs2_error(inode->i_sb,
597                             "Inode %llu has a hole at block %llu\n",
598                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
599                             (unsigned long long)iblock);
600                 ret = -EROFS;
601                 goto bail;
602         }
603
604         /*
605          * get_more_blocks() expects us to describe a hole by clearing
606          * the mapped bit on bh_result().
607          *
608          * Consider an unwritten extent as a hole.
609          */
610         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
611                 map_bh(bh_result, inode->i_sb, p_blkno);
612         else {
613                 /*
614                  * ocfs2_prepare_inode_for_write() should have caught
615                  * the case where we'd be filling a hole and triggered
616                  * a buffered write instead.
617                  */
618                 if (create) {
619                         ret = -EIO;
620                         mlog_errno(ret);
621                         goto bail;
622                 }
623
624                 clear_buffer_mapped(bh_result);
625         }
626
627         /* make sure we don't map more than max_blocks blocks here as
628            that's all the kernel will handle at this point. */
629         if (max_blocks < contig_blocks)
630                 contig_blocks = max_blocks;
631         bh_result->b_size = contig_blocks << blocksize_bits;
632 bail:
633         return ret;
634 }
635
636 /* 
637  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
638  * particularly interested in the aio/dio case.  Like the core uses
639  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
640  * truncation on another.
641  */
642 static void ocfs2_dio_end_io(struct kiocb *iocb,
643                              loff_t offset,
644                              ssize_t bytes,
645                              void *private)
646 {
647         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
648         int level;
649
650         /* this io's submitter should not have unlocked this before we could */
651         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
652
653         ocfs2_iocb_clear_rw_locked(iocb);
654
655         level = ocfs2_iocb_rw_locked_level(iocb);
656         if (!level)
657                 up_read(&inode->i_alloc_sem);
658         ocfs2_rw_unlock(inode, level);
659 }
660
661 /*
662  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
663  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
664  * do journalled data.
665  */
666 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
667 {
668         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
669
670         jbd2_journal_invalidatepage(journal, page, offset);
671 }
672
673 static int ocfs2_releasepage(struct page *page, gfp_t wait)
674 {
675         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
676
677         if (!page_has_buffers(page))
678                 return 0;
679         return jbd2_journal_try_to_free_buffers(journal, page, wait);
680 }
681
682 static ssize_t ocfs2_direct_IO(int rw,
683                                struct kiocb *iocb,
684                                const struct iovec *iov,
685                                loff_t offset,
686                                unsigned long nr_segs)
687 {
688         struct file *file = iocb->ki_filp;
689         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
690         int ret;
691
692         mlog_entry_void();
693
694         /*
695          * Fallback to buffered I/O if we see an inode without
696          * extents.
697          */
698         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
699                 return 0;
700
701         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
702                                             inode->i_sb->s_bdev, iov, offset,
703                                             nr_segs, 
704                                             ocfs2_direct_IO_get_blocks,
705                                             ocfs2_dio_end_io);
706
707         mlog_exit(ret);
708         return ret;
709 }
710
711 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
712                                             u32 cpos,
713                                             unsigned int *start,
714                                             unsigned int *end)
715 {
716         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
717
718         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
719                 unsigned int cpp;
720
721                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
722
723                 cluster_start = cpos % cpp;
724                 cluster_start = cluster_start << osb->s_clustersize_bits;
725
726                 cluster_end = cluster_start + osb->s_clustersize;
727         }
728
729         BUG_ON(cluster_start > PAGE_SIZE);
730         BUG_ON(cluster_end > PAGE_SIZE);
731
732         if (start)
733                 *start = cluster_start;
734         if (end)
735                 *end = cluster_end;
736 }
737
738 /*
739  * 'from' and 'to' are the region in the page to avoid zeroing.
740  *
741  * If pagesize > clustersize, this function will avoid zeroing outside
742  * of the cluster boundary.
743  *
744  * from == to == 0 is code for "zero the entire cluster region"
745  */
746 static void ocfs2_clear_page_regions(struct page *page,
747                                      struct ocfs2_super *osb, u32 cpos,
748                                      unsigned from, unsigned to)
749 {
750         void *kaddr;
751         unsigned int cluster_start, cluster_end;
752
753         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
754
755         kaddr = kmap_atomic(page, KM_USER0);
756
757         if (from || to) {
758                 if (from > cluster_start)
759                         memset(kaddr + cluster_start, 0, from - cluster_start);
760                 if (to < cluster_end)
761                         memset(kaddr + to, 0, cluster_end - to);
762         } else {
763                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
764         }
765
766         kunmap_atomic(kaddr, KM_USER0);
767 }
768
769 /*
770  * Nonsparse file systems fully allocate before we get to the write
771  * code. This prevents ocfs2_write() from tagging the write as an
772  * allocating one, which means ocfs2_map_page_blocks() might try to
773  * read-in the blocks at the tail of our file. Avoid reading them by
774  * testing i_size against each block offset.
775  */
776 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
777                                  unsigned int block_start)
778 {
779         u64 offset = page_offset(page) + block_start;
780
781         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
782                 return 1;
783
784         if (i_size_read(inode) > offset)
785                 return 1;
786
787         return 0;
788 }
789
790 /*
791  * Some of this taken from block_prepare_write(). We already have our
792  * mapping by now though, and the entire write will be allocating or
793  * it won't, so not much need to use BH_New.
794  *
795  * This will also skip zeroing, which is handled externally.
796  */
797 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
798                           struct inode *inode, unsigned int from,
799                           unsigned int to, int new)
800 {
801         int ret = 0;
802         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
803         unsigned int block_end, block_start;
804         unsigned int bsize = 1 << inode->i_blkbits;
805
806         if (!page_has_buffers(page))
807                 create_empty_buffers(page, bsize, 0);
808
809         head = page_buffers(page);
810         for (bh = head, block_start = 0; bh != head || !block_start;
811              bh = bh->b_this_page, block_start += bsize) {
812                 block_end = block_start + bsize;
813
814                 clear_buffer_new(bh);
815
816                 /*
817                  * Ignore blocks outside of our i/o range -
818                  * they may belong to unallocated clusters.
819                  */
820                 if (block_start >= to || block_end <= from) {
821                         if (PageUptodate(page))
822                                 set_buffer_uptodate(bh);
823                         continue;
824                 }
825
826                 /*
827                  * For an allocating write with cluster size >= page
828                  * size, we always write the entire page.
829                  */
830                 if (new)
831                         set_buffer_new(bh);
832
833                 if (!buffer_mapped(bh)) {
834                         map_bh(bh, inode->i_sb, *p_blkno);
835                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
836                 }
837
838                 if (PageUptodate(page)) {
839                         if (!buffer_uptodate(bh))
840                                 set_buffer_uptodate(bh);
841                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
842                            !buffer_new(bh) &&
843                            ocfs2_should_read_blk(inode, page, block_start) &&
844                            (block_start < from || block_end > to)) {
845                         ll_rw_block(READ, 1, &bh);
846                         *wait_bh++=bh;
847                 }
848
849                 *p_blkno = *p_blkno + 1;
850         }
851
852         /*
853          * If we issued read requests - let them complete.
854          */
855         while(wait_bh > wait) {
856                 wait_on_buffer(*--wait_bh);
857                 if (!buffer_uptodate(*wait_bh))
858                         ret = -EIO;
859         }
860
861         if (ret == 0 || !new)
862                 return ret;
863
864         /*
865          * If we get -EIO above, zero out any newly allocated blocks
866          * to avoid exposing stale data.
867          */
868         bh = head;
869         block_start = 0;
870         do {
871                 block_end = block_start + bsize;
872                 if (block_end <= from)
873                         goto next_bh;
874                 if (block_start >= to)
875                         break;
876
877                 zero_user(page, block_start, bh->b_size);
878                 set_buffer_uptodate(bh);
879                 mark_buffer_dirty(bh);
880
881 next_bh:
882                 block_start = block_end;
883                 bh = bh->b_this_page;
884         } while (bh != head);
885
886         return ret;
887 }
888
889 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
890 #define OCFS2_MAX_CTXT_PAGES    1
891 #else
892 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
893 #endif
894
895 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
896
897 /*
898  * Describe the state of a single cluster to be written to.
899  */
900 struct ocfs2_write_cluster_desc {
901         u32             c_cpos;
902         u32             c_phys;
903         /*
904          * Give this a unique field because c_phys eventually gets
905          * filled.
906          */
907         unsigned        c_new;
908         unsigned        c_unwritten;
909 };
910
911 static inline int ocfs2_should_zero_cluster(struct ocfs2_write_cluster_desc *d)
912 {
913         return d->c_new || d->c_unwritten;
914 }
915
916 struct ocfs2_write_ctxt {
917         /* Logical cluster position / len of write */
918         u32                             w_cpos;
919         u32                             w_clen;
920
921         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
922
923         /*
924          * This is true if page_size > cluster_size.
925          *
926          * It triggers a set of special cases during write which might
927          * have to deal with allocating writes to partial pages.
928          */
929         unsigned int                    w_large_pages;
930
931         /*
932          * Pages involved in this write.
933          *
934          * w_target_page is the page being written to by the user.
935          *
936          * w_pages is an array of pages which always contains
937          * w_target_page, and in the case of an allocating write with
938          * page_size < cluster size, it will contain zero'd and mapped
939          * pages adjacent to w_target_page which need to be written
940          * out in so that future reads from that region will get
941          * zero's.
942          */
943         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
944         unsigned int                    w_num_pages;
945         struct page                     *w_target_page;
946
947         /*
948          * ocfs2_write_end() uses this to know what the real range to
949          * write in the target should be.
950          */
951         unsigned int                    w_target_from;
952         unsigned int                    w_target_to;
953
954         /*
955          * We could use journal_current_handle() but this is cleaner,
956          * IMHO -Mark
957          */
958         handle_t                        *w_handle;
959
960         struct buffer_head              *w_di_bh;
961
962         struct ocfs2_cached_dealloc_ctxt w_dealloc;
963 };
964
965 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
966 {
967         int i;
968
969         for(i = 0; i < num_pages; i++) {
970                 if (pages[i]) {
971                         unlock_page(pages[i]);
972                         mark_page_accessed(pages[i]);
973                         page_cache_release(pages[i]);
974                 }
975         }
976 }
977
978 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
979 {
980         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
981
982         brelse(wc->w_di_bh);
983         kfree(wc);
984 }
985
986 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
987                                   struct ocfs2_super *osb, loff_t pos,
988                                   unsigned len, struct buffer_head *di_bh)
989 {
990         u32 cend;
991         struct ocfs2_write_ctxt *wc;
992
993         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
994         if (!wc)
995                 return -ENOMEM;
996
997         wc->w_cpos = pos >> osb->s_clustersize_bits;
998         cend = (pos + len - 1) >> osb->s_clustersize_bits;
999         wc->w_clen = cend - wc->w_cpos + 1;
1000         get_bh(di_bh);
1001         wc->w_di_bh = di_bh;
1002
1003         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1004                 wc->w_large_pages = 1;
1005         else
1006                 wc->w_large_pages = 0;
1007
1008         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1009
1010         *wcp = wc;
1011
1012         return 0;
1013 }
1014
1015 /*
1016  * If a page has any new buffers, zero them out here, and mark them uptodate
1017  * and dirty so they'll be written out (in order to prevent uninitialised
1018  * block data from leaking). And clear the new bit.
1019  */
1020 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1021 {
1022         unsigned int block_start, block_end;
1023         struct buffer_head *head, *bh;
1024
1025         BUG_ON(!PageLocked(page));
1026         if (!page_has_buffers(page))
1027                 return;
1028
1029         bh = head = page_buffers(page);
1030         block_start = 0;
1031         do {
1032                 block_end = block_start + bh->b_size;
1033
1034                 if (buffer_new(bh)) {
1035                         if (block_end > from && block_start < to) {
1036                                 if (!PageUptodate(page)) {
1037                                         unsigned start, end;
1038
1039                                         start = max(from, block_start);
1040                                         end = min(to, block_end);
1041
1042                                         zero_user_segment(page, start, end);
1043                                         set_buffer_uptodate(bh);
1044                                 }
1045
1046                                 clear_buffer_new(bh);
1047                                 mark_buffer_dirty(bh);
1048                         }
1049                 }
1050
1051                 block_start = block_end;
1052                 bh = bh->b_this_page;
1053         } while (bh != head);
1054 }
1055
1056 /*
1057  * Only called when we have a failure during allocating write to write
1058  * zero's to the newly allocated region.
1059  */
1060 static void ocfs2_write_failure(struct inode *inode,
1061                                 struct ocfs2_write_ctxt *wc,
1062                                 loff_t user_pos, unsigned user_len)
1063 {
1064         int i;
1065         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1066                 to = user_pos + user_len;
1067         struct page *tmppage;
1068
1069         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1070
1071         for(i = 0; i < wc->w_num_pages; i++) {
1072                 tmppage = wc->w_pages[i];
1073
1074                 if (page_has_buffers(tmppage)) {
1075                         if (ocfs2_should_order_data(inode)) {
1076                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1077 #ifdef CONFIG_OCFS2_COMPAT_JBD
1078                                 walk_page_buffers(wc->w_handle,
1079                                                   page_buffers(tmppage),
1080                                                   from, to, NULL,
1081                                                   ocfs2_journal_dirty_data);
1082 #endif
1083                         }
1084
1085                         block_commit_write(tmppage, from, to);
1086                 }
1087         }
1088 }
1089
1090 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1091                                         struct ocfs2_write_ctxt *wc,
1092                                         struct page *page, u32 cpos,
1093                                         loff_t user_pos, unsigned user_len,
1094                                         int new)
1095 {
1096         int ret;
1097         unsigned int map_from = 0, map_to = 0;
1098         unsigned int cluster_start, cluster_end;
1099         unsigned int user_data_from = 0, user_data_to = 0;
1100
1101         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1102                                         &cluster_start, &cluster_end);
1103
1104         if (page == wc->w_target_page) {
1105                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1106                 map_to = map_from + user_len;
1107
1108                 if (new)
1109                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1110                                                     cluster_start, cluster_end,
1111                                                     new);
1112                 else
1113                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1114                                                     map_from, map_to, new);
1115                 if (ret) {
1116                         mlog_errno(ret);
1117                         goto out;
1118                 }
1119
1120                 user_data_from = map_from;
1121                 user_data_to = map_to;
1122                 if (new) {
1123                         map_from = cluster_start;
1124                         map_to = cluster_end;
1125                 }
1126         } else {
1127                 /*
1128                  * If we haven't allocated the new page yet, we
1129                  * shouldn't be writing it out without copying user
1130                  * data. This is likely a math error from the caller.
1131                  */
1132                 BUG_ON(!new);
1133
1134                 map_from = cluster_start;
1135                 map_to = cluster_end;
1136
1137                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1138                                             cluster_start, cluster_end, new);
1139                 if (ret) {
1140                         mlog_errno(ret);
1141                         goto out;
1142                 }
1143         }
1144
1145         /*
1146          * Parts of newly allocated pages need to be zero'd.
1147          *
1148          * Above, we have also rewritten 'to' and 'from' - as far as
1149          * the rest of the function is concerned, the entire cluster
1150          * range inside of a page needs to be written.
1151          *
1152          * We can skip this if the page is up to date - it's already
1153          * been zero'd from being read in as a hole.
1154          */
1155         if (new && !PageUptodate(page))
1156                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1157                                          cpos, user_data_from, user_data_to);
1158
1159         flush_dcache_page(page);
1160
1161 out:
1162         return ret;
1163 }
1164
1165 /*
1166  * This function will only grab one clusters worth of pages.
1167  */
1168 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1169                                       struct ocfs2_write_ctxt *wc,
1170                                       u32 cpos, loff_t user_pos, int new,
1171                                       struct page *mmap_page)
1172 {
1173         int ret = 0, i;
1174         unsigned long start, target_index, index;
1175         struct inode *inode = mapping->host;
1176
1177         target_index = user_pos >> PAGE_CACHE_SHIFT;
1178
1179         /*
1180          * Figure out how many pages we'll be manipulating here. For
1181          * non allocating write, we just change the one
1182          * page. Otherwise, we'll need a whole clusters worth.
1183          */
1184         if (new) {
1185                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1186                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1187         } else {
1188                 wc->w_num_pages = 1;
1189                 start = target_index;
1190         }
1191
1192         for(i = 0; i < wc->w_num_pages; i++) {
1193                 index = start + i;
1194
1195                 if (index == target_index && mmap_page) {
1196                         /*
1197                          * ocfs2_pagemkwrite() is a little different
1198                          * and wants us to directly use the page
1199                          * passed in.
1200                          */
1201                         lock_page(mmap_page);
1202
1203                         if (mmap_page->mapping != mapping) {
1204                                 unlock_page(mmap_page);
1205                                 /*
1206                                  * Sanity check - the locking in
1207                                  * ocfs2_pagemkwrite() should ensure
1208                                  * that this code doesn't trigger.
1209                                  */
1210                                 ret = -EINVAL;
1211                                 mlog_errno(ret);
1212                                 goto out;
1213                         }
1214
1215                         page_cache_get(mmap_page);
1216                         wc->w_pages[i] = mmap_page;
1217                 } else {
1218                         wc->w_pages[i] = find_or_create_page(mapping, index,
1219                                                              GFP_NOFS);
1220                         if (!wc->w_pages[i]) {
1221                                 ret = -ENOMEM;
1222                                 mlog_errno(ret);
1223                                 goto out;
1224                         }
1225                 }
1226
1227                 if (index == target_index)
1228                         wc->w_target_page = wc->w_pages[i];
1229         }
1230 out:
1231         return ret;
1232 }
1233
1234 /*
1235  * Prepare a single cluster for write one cluster into the file.
1236  */
1237 static int ocfs2_write_cluster(struct address_space *mapping,
1238                                u32 phys, unsigned int unwritten,
1239                                struct ocfs2_alloc_context *data_ac,
1240                                struct ocfs2_alloc_context *meta_ac,
1241                                struct ocfs2_write_ctxt *wc, u32 cpos,
1242                                loff_t user_pos, unsigned user_len)
1243 {
1244         int ret, i, new, should_zero = 0;
1245         u64 v_blkno, p_blkno;
1246         struct inode *inode = mapping->host;
1247         struct ocfs2_extent_tree et;
1248
1249         new = phys == 0 ? 1 : 0;
1250         if (new || unwritten)
1251                 should_zero = 1;
1252
1253         if (new) {
1254                 u32 tmp_pos;
1255
1256                 /*
1257                  * This is safe to call with the page locks - it won't take
1258                  * any additional semaphores or cluster locks.
1259                  */
1260                 tmp_pos = cpos;
1261                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1262                                            &tmp_pos, 1, 0, wc->w_di_bh,
1263                                            wc->w_handle, data_ac,
1264                                            meta_ac, NULL);
1265                 /*
1266                  * This shouldn't happen because we must have already
1267                  * calculated the correct meta data allocation required. The
1268                  * internal tree allocation code should know how to increase
1269                  * transaction credits itself.
1270                  *
1271                  * If need be, we could handle -EAGAIN for a
1272                  * RESTART_TRANS here.
1273                  */
1274                 mlog_bug_on_msg(ret == -EAGAIN,
1275                                 "Inode %llu: EAGAIN return during allocation.\n",
1276                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1277                 if (ret < 0) {
1278                         mlog_errno(ret);
1279                         goto out;
1280                 }
1281         } else if (unwritten) {
1282                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1283                 ret = ocfs2_mark_extent_written(inode, &et,
1284                                                 wc->w_handle, cpos, 1, phys,
1285                                                 meta_ac, &wc->w_dealloc);
1286                 if (ret < 0) {
1287                         mlog_errno(ret);
1288                         goto out;
1289                 }
1290         }
1291
1292         if (should_zero)
1293                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1294         else
1295                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1296
1297         /*
1298          * The only reason this should fail is due to an inability to
1299          * find the extent added.
1300          */
1301         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1302                                           NULL);
1303         if (ret < 0) {
1304                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1305                             "at logical block %llu",
1306                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1307                             (unsigned long long)v_blkno);
1308                 goto out;
1309         }
1310
1311         BUG_ON(p_blkno == 0);
1312
1313         for(i = 0; i < wc->w_num_pages; i++) {
1314                 int tmpret;
1315
1316                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1317                                                       wc->w_pages[i], cpos,
1318                                                       user_pos, user_len,
1319                                                       should_zero);
1320                 if (tmpret) {
1321                         mlog_errno(tmpret);
1322                         if (ret == 0)
1323                                 tmpret = ret;
1324                 }
1325         }
1326
1327         /*
1328          * We only have cleanup to do in case of allocating write.
1329          */
1330         if (ret && new)
1331                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1332
1333 out:
1334
1335         return ret;
1336 }
1337
1338 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1339                                        struct ocfs2_alloc_context *data_ac,
1340                                        struct ocfs2_alloc_context *meta_ac,
1341                                        struct ocfs2_write_ctxt *wc,
1342                                        loff_t pos, unsigned len)
1343 {
1344         int ret, i;
1345         loff_t cluster_off;
1346         unsigned int local_len = len;
1347         struct ocfs2_write_cluster_desc *desc;
1348         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1349
1350         for (i = 0; i < wc->w_clen; i++) {
1351                 desc = &wc->w_desc[i];
1352
1353                 /*
1354                  * We have to make sure that the total write passed in
1355                  * doesn't extend past a single cluster.
1356                  */
1357                 local_len = len;
1358                 cluster_off = pos & (osb->s_clustersize - 1);
1359                 if ((cluster_off + local_len) > osb->s_clustersize)
1360                         local_len = osb->s_clustersize - cluster_off;
1361
1362                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1363                                           desc->c_unwritten, data_ac, meta_ac,
1364                                           wc, desc->c_cpos, pos, local_len);
1365                 if (ret) {
1366                         mlog_errno(ret);
1367                         goto out;
1368                 }
1369
1370                 len -= local_len;
1371                 pos += local_len;
1372         }
1373
1374         ret = 0;
1375 out:
1376         return ret;
1377 }
1378
1379 /*
1380  * ocfs2_write_end() wants to know which parts of the target page it
1381  * should complete the write on. It's easiest to compute them ahead of
1382  * time when a more complete view of the write is available.
1383  */
1384 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1385                                         struct ocfs2_write_ctxt *wc,
1386                                         loff_t pos, unsigned len, int alloc)
1387 {
1388         struct ocfs2_write_cluster_desc *desc;
1389
1390         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1391         wc->w_target_to = wc->w_target_from + len;
1392
1393         if (alloc == 0)
1394                 return;
1395
1396         /*
1397          * Allocating write - we may have different boundaries based
1398          * on page size and cluster size.
1399          *
1400          * NOTE: We can no longer compute one value from the other as
1401          * the actual write length and user provided length may be
1402          * different.
1403          */
1404
1405         if (wc->w_large_pages) {
1406                 /*
1407                  * We only care about the 1st and last cluster within
1408                  * our range and whether they should be zero'd or not. Either
1409                  * value may be extended out to the start/end of a
1410                  * newly allocated cluster.
1411                  */
1412                 desc = &wc->w_desc[0];
1413                 if (ocfs2_should_zero_cluster(desc))
1414                         ocfs2_figure_cluster_boundaries(osb,
1415                                                         desc->c_cpos,
1416                                                         &wc->w_target_from,
1417                                                         NULL);
1418
1419                 desc = &wc->w_desc[wc->w_clen - 1];
1420                 if (ocfs2_should_zero_cluster(desc))
1421                         ocfs2_figure_cluster_boundaries(osb,
1422                                                         desc->c_cpos,
1423                                                         NULL,
1424                                                         &wc->w_target_to);
1425         } else {
1426                 wc->w_target_from = 0;
1427                 wc->w_target_to = PAGE_CACHE_SIZE;
1428         }
1429 }
1430
1431 /*
1432  * Populate each single-cluster write descriptor in the write context
1433  * with information about the i/o to be done.
1434  *
1435  * Returns the number of clusters that will have to be allocated, as
1436  * well as a worst case estimate of the number of extent records that
1437  * would have to be created during a write to an unwritten region.
1438  */
1439 static int ocfs2_populate_write_desc(struct inode *inode,
1440                                      struct ocfs2_write_ctxt *wc,
1441                                      unsigned int *clusters_to_alloc,
1442                                      unsigned int *extents_to_split)
1443 {
1444         int ret;
1445         struct ocfs2_write_cluster_desc *desc;
1446         unsigned int num_clusters = 0;
1447         unsigned int ext_flags = 0;
1448         u32 phys = 0;
1449         int i;
1450
1451         *clusters_to_alloc = 0;
1452         *extents_to_split = 0;
1453
1454         for (i = 0; i < wc->w_clen; i++) {
1455                 desc = &wc->w_desc[i];
1456                 desc->c_cpos = wc->w_cpos + i;
1457
1458                 if (num_clusters == 0) {
1459                         /*
1460                          * Need to look up the next extent record.
1461                          */
1462                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1463                                                  &num_clusters, &ext_flags);
1464                         if (ret) {
1465                                 mlog_errno(ret);
1466                                 goto out;
1467                         }
1468
1469                         /*
1470                          * Assume worst case - that we're writing in
1471                          * the middle of the extent.
1472                          *
1473                          * We can assume that the write proceeds from
1474                          * left to right, in which case the extent
1475                          * insert code is smart enough to coalesce the
1476                          * next splits into the previous records created.
1477                          */
1478                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1479                                 *extents_to_split = *extents_to_split + 2;
1480                 } else if (phys) {
1481                         /*
1482                          * Only increment phys if it doesn't describe
1483                          * a hole.
1484                          */
1485                         phys++;
1486                 }
1487
1488                 desc->c_phys = phys;
1489                 if (phys == 0) {
1490                         desc->c_new = 1;
1491                         *clusters_to_alloc = *clusters_to_alloc + 1;
1492                 }
1493                 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1494                         desc->c_unwritten = 1;
1495
1496                 num_clusters--;
1497         }
1498
1499         ret = 0;
1500 out:
1501         return ret;
1502 }
1503
1504 static int ocfs2_write_begin_inline(struct address_space *mapping,
1505                                     struct inode *inode,
1506                                     struct ocfs2_write_ctxt *wc)
1507 {
1508         int ret;
1509         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1510         struct page *page;
1511         handle_t *handle;
1512         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1513
1514         page = find_or_create_page(mapping, 0, GFP_NOFS);
1515         if (!page) {
1516                 ret = -ENOMEM;
1517                 mlog_errno(ret);
1518                 goto out;
1519         }
1520         /*
1521          * If we don't set w_num_pages then this page won't get unlocked
1522          * and freed on cleanup of the write context.
1523          */
1524         wc->w_pages[0] = wc->w_target_page = page;
1525         wc->w_num_pages = 1;
1526
1527         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1528         if (IS_ERR(handle)) {
1529                 ret = PTR_ERR(handle);
1530                 mlog_errno(ret);
1531                 goto out;
1532         }
1533
1534         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1535                                    OCFS2_JOURNAL_ACCESS_WRITE);
1536         if (ret) {
1537                 ocfs2_commit_trans(osb, handle);
1538
1539                 mlog_errno(ret);
1540                 goto out;
1541         }
1542
1543         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1544                 ocfs2_set_inode_data_inline(inode, di);
1545
1546         if (!PageUptodate(page)) {
1547                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1548                 if (ret) {
1549                         ocfs2_commit_trans(osb, handle);
1550
1551                         goto out;
1552                 }
1553         }
1554
1555         wc->w_handle = handle;
1556 out:
1557         return ret;
1558 }
1559
1560 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1561 {
1562         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1563
1564         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1565                 return 1;
1566         return 0;
1567 }
1568
1569 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1570                                           struct inode *inode, loff_t pos,
1571                                           unsigned len, struct page *mmap_page,
1572                                           struct ocfs2_write_ctxt *wc)
1573 {
1574         int ret, written = 0;
1575         loff_t end = pos + len;
1576         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1577
1578         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1579              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1580              oi->ip_dyn_features);
1581
1582         /*
1583          * Handle inodes which already have inline data 1st.
1584          */
1585         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1586                 if (mmap_page == NULL &&
1587                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1588                         goto do_inline_write;
1589
1590                 /*
1591                  * The write won't fit - we have to give this inode an
1592                  * inline extent list now.
1593                  */
1594                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1595                 if (ret)
1596                         mlog_errno(ret);
1597                 goto out;
1598         }
1599
1600         /*
1601          * Check whether the inode can accept inline data.
1602          */
1603         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1604                 return 0;
1605
1606         /*
1607          * Check whether the write can fit.
1608          */
1609         if (mmap_page || end > ocfs2_max_inline_data(inode->i_sb))
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, newsize - len);
1652         if (ret)
1653                 mlog_errno(ret);
1654
1655         return ret;
1656 }
1657
1658 int ocfs2_write_begin_nolock(struct address_space *mapping,
1659                              loff_t pos, unsigned len, unsigned flags,
1660                              struct page **pagep, void **fsdata,
1661                              struct buffer_head *di_bh, struct page *mmap_page)
1662 {
1663         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1664         unsigned int clusters_to_alloc, extents_to_split;
1665         struct ocfs2_write_ctxt *wc;
1666         struct inode *inode = mapping->host;
1667         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1668         struct ocfs2_dinode *di;
1669         struct ocfs2_alloc_context *data_ac = NULL;
1670         struct ocfs2_alloc_context *meta_ac = NULL;
1671         handle_t *handle;
1672         struct ocfs2_extent_tree et;
1673
1674         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1675         if (ret) {
1676                 mlog_errno(ret);
1677                 return ret;
1678         }
1679
1680         if (ocfs2_supports_inline_data(osb)) {
1681                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1682                                                      mmap_page, wc);
1683                 if (ret == 1) {
1684                         ret = 0;
1685                         goto success;
1686                 }
1687                 if (ret < 0) {
1688                         mlog_errno(ret);
1689                         goto out;
1690                 }
1691         }
1692
1693         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1694         if (ret) {
1695                 mlog_errno(ret);
1696                 goto out;
1697         }
1698
1699         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1700                                         &extents_to_split);
1701         if (ret) {
1702                 mlog_errno(ret);
1703                 goto out;
1704         }
1705
1706         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1707
1708         /*
1709          * We set w_target_from, w_target_to here so that
1710          * ocfs2_write_end() knows which range in the target page to
1711          * write out. An allocation requires that we write the entire
1712          * cluster range.
1713          */
1714         if (clusters_to_alloc || extents_to_split) {
1715                 /*
1716                  * XXX: We are stretching the limits of
1717                  * ocfs2_lock_allocators(). It greatly over-estimates
1718                  * the work to be done.
1719                  */
1720                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1721                      " clusters_to_add = %u, extents_to_split = %u\n",
1722                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1723                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1724                      clusters_to_alloc, extents_to_split);
1725
1726                 ocfs2_init_dinode_extent_tree(&et, inode, wc->w_di_bh);
1727                 ret = ocfs2_lock_allocators(inode, &et,
1728                                             clusters_to_alloc, extents_to_split,
1729                                             &data_ac, &meta_ac);
1730                 if (ret) {
1731                         mlog_errno(ret);
1732                         goto out;
1733                 }
1734
1735                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1736                                                     &di->id2.i_list,
1737                                                     clusters_to_alloc);
1738
1739         }
1740
1741         ocfs2_set_target_boundaries(osb, wc, pos, len,
1742                                     clusters_to_alloc + extents_to_split);
1743
1744         handle = ocfs2_start_trans(osb, credits);
1745         if (IS_ERR(handle)) {
1746                 ret = PTR_ERR(handle);
1747                 mlog_errno(ret);
1748                 goto out;
1749         }
1750
1751         wc->w_handle = handle;
1752
1753         /*
1754          * We don't want this to fail in ocfs2_write_end(), so do it
1755          * here.
1756          */
1757         ret = ocfs2_journal_access(handle, inode, wc->w_di_bh,
1758                                    OCFS2_JOURNAL_ACCESS_WRITE);
1759         if (ret) {
1760                 mlog_errno(ret);
1761                 goto out_commit;
1762         }
1763
1764         /*
1765          * Fill our page array first. That way we've grabbed enough so
1766          * that we can zero and flush if we error after adding the
1767          * extent.
1768          */
1769         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1770                                          clusters_to_alloc + extents_to_split,
1771                                          mmap_page);
1772         if (ret) {
1773                 mlog_errno(ret);
1774                 goto out_commit;
1775         }
1776
1777         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1778                                           len);
1779         if (ret) {
1780                 mlog_errno(ret);
1781                 goto out_commit;
1782         }
1783
1784         if (data_ac)
1785                 ocfs2_free_alloc_context(data_ac);
1786         if (meta_ac)
1787                 ocfs2_free_alloc_context(meta_ac);
1788
1789 success:
1790         *pagep = wc->w_target_page;
1791         *fsdata = wc;
1792         return 0;
1793 out_commit:
1794         ocfs2_commit_trans(osb, handle);
1795
1796 out:
1797         ocfs2_free_write_ctxt(wc);
1798
1799         if (data_ac)
1800                 ocfs2_free_alloc_context(data_ac);
1801         if (meta_ac)
1802                 ocfs2_free_alloc_context(meta_ac);
1803         return ret;
1804 }
1805
1806 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1807                              loff_t pos, unsigned len, unsigned flags,
1808                              struct page **pagep, void **fsdata)
1809 {
1810         int ret;
1811         struct buffer_head *di_bh = NULL;
1812         struct inode *inode = mapping->host;
1813
1814         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1815         if (ret) {
1816                 mlog_errno(ret);
1817                 return ret;
1818         }
1819
1820         /*
1821          * Take alloc sem here to prevent concurrent lookups. That way
1822          * the mapping, zeroing and tree manipulation within
1823          * ocfs2_write() will be safe against ->readpage(). This
1824          * should also serve to lock out allocation from a shared
1825          * writeable region.
1826          */
1827         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1828
1829         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1830                                        fsdata, di_bh, NULL);
1831         if (ret) {
1832                 mlog_errno(ret);
1833                 goto out_fail;
1834         }
1835
1836         brelse(di_bh);
1837
1838         return 0;
1839
1840 out_fail:
1841         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1842
1843         brelse(di_bh);
1844         ocfs2_inode_unlock(inode, 1);
1845
1846         return ret;
1847 }
1848
1849 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1850                                    unsigned len, unsigned *copied,
1851                                    struct ocfs2_dinode *di,
1852                                    struct ocfs2_write_ctxt *wc)
1853 {
1854         void *kaddr;
1855
1856         if (unlikely(*copied < len)) {
1857                 if (!PageUptodate(wc->w_target_page)) {
1858                         *copied = 0;
1859                         return;
1860                 }
1861         }
1862
1863         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1864         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1865         kunmap_atomic(kaddr, KM_USER0);
1866
1867         mlog(0, "Data written to inode at offset %llu. "
1868              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1869              (unsigned long long)pos, *copied,
1870              le16_to_cpu(di->id2.i_data.id_count),
1871              le16_to_cpu(di->i_dyn_features));
1872 }
1873
1874 int ocfs2_write_end_nolock(struct address_space *mapping,
1875                            loff_t pos, unsigned len, unsigned copied,
1876                            struct page *page, void *fsdata)
1877 {
1878         int i;
1879         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1880         struct inode *inode = mapping->host;
1881         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1882         struct ocfs2_write_ctxt *wc = fsdata;
1883         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1884         handle_t *handle = wc->w_handle;
1885         struct page *tmppage;
1886
1887         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1888                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1889                 goto out_write_size;
1890         }
1891
1892         if (unlikely(copied < len)) {
1893                 if (!PageUptodate(wc->w_target_page))
1894                         copied = 0;
1895
1896                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1897                                        start+len);
1898         }
1899         flush_dcache_page(wc->w_target_page);
1900
1901         for(i = 0; i < wc->w_num_pages; i++) {
1902                 tmppage = wc->w_pages[i];
1903
1904                 if (tmppage == wc->w_target_page) {
1905                         from = wc->w_target_from;
1906                         to = wc->w_target_to;
1907
1908                         BUG_ON(from > PAGE_CACHE_SIZE ||
1909                                to > PAGE_CACHE_SIZE ||
1910                                to < from);
1911                 } else {
1912                         /*
1913                          * Pages adjacent to the target (if any) imply
1914                          * a hole-filling write in which case we want
1915                          * to flush their entire range.
1916                          */
1917                         from = 0;
1918                         to = PAGE_CACHE_SIZE;
1919                 }
1920
1921                 if (page_has_buffers(tmppage)) {
1922                         if (ocfs2_should_order_data(inode)) {
1923                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1924 #ifdef CONFIG_OCFS2_COMPAT_JBD
1925                                 walk_page_buffers(wc->w_handle,
1926                                                   page_buffers(tmppage),
1927                                                   from, to, NULL,
1928                                                   ocfs2_journal_dirty_data);
1929 #endif
1930                         }
1931                         block_commit_write(tmppage, from, to);
1932                 }
1933         }
1934
1935 out_write_size:
1936         pos += copied;
1937         if (pos > inode->i_size) {
1938                 i_size_write(inode, pos);
1939                 mark_inode_dirty(inode);
1940         }
1941         inode->i_blocks = ocfs2_inode_sector_count(inode);
1942         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1943         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1944         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1945         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1946         ocfs2_journal_dirty(handle, wc->w_di_bh);
1947
1948         ocfs2_commit_trans(osb, handle);
1949
1950         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1951
1952         ocfs2_free_write_ctxt(wc);
1953
1954         return copied;
1955 }
1956
1957 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1958                            loff_t pos, unsigned len, unsigned copied,
1959                            struct page *page, void *fsdata)
1960 {
1961         int ret;
1962         struct inode *inode = mapping->host;
1963
1964         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1965
1966         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1967         ocfs2_inode_unlock(inode, 1);
1968
1969         return ret;
1970 }
1971
1972 const struct address_space_operations ocfs2_aops = {
1973         .readpage       = ocfs2_readpage,
1974         .readpages      = ocfs2_readpages,
1975         .writepage      = ocfs2_writepage,
1976         .write_begin    = ocfs2_write_begin,
1977         .write_end      = ocfs2_write_end,
1978         .bmap           = ocfs2_bmap,
1979         .sync_page      = block_sync_page,
1980         .direct_IO      = ocfs2_direct_IO,
1981         .invalidatepage = ocfs2_invalidatepage,
1982         .releasepage    = ocfs2_releasepage,
1983         .migratepage    = buffer_migrate_page,
1984 };