deb2b132ae5ed42b68fd11f58413f2ffa4779b83
[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 #include "refcounttree.h"
48
49 #include "buffer_head_io.h"
50
51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
52                                    struct buffer_head *bh_result, int create)
53 {
54         int err = -EIO;
55         int status;
56         struct ocfs2_dinode *fe = NULL;
57         struct buffer_head *bh = NULL;
58         struct buffer_head *buffer_cache_bh = NULL;
59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
60         void *kaddr;
61
62         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
63                    (unsigned long long)iblock, bh_result, create);
64
65         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
66
67         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
68                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
69                      (unsigned long long)iblock);
70                 goto bail;
71         }
72
73         status = ocfs2_read_inode_block(inode, &bh);
74         if (status < 0) {
75                 mlog_errno(status);
76                 goto bail;
77         }
78         fe = (struct ocfs2_dinode *) bh->b_data;
79
80         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
81                                                     le32_to_cpu(fe->i_clusters))) {
82                 mlog(ML_ERROR, "block offset is outside the allocated size: "
83                      "%llu\n", (unsigned long long)iblock);
84                 goto bail;
85         }
86
87         /* We don't use the page cache to create symlink data, so if
88          * need be, copy it over from the buffer cache. */
89         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
90                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
91                             iblock;
92                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
93                 if (!buffer_cache_bh) {
94                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
95                         goto bail;
96                 }
97
98                 /* we haven't locked out transactions, so a commit
99                  * could've happened. Since we've got a reference on
100                  * the bh, even if it commits while we're doing the
101                  * copy, the data is still good. */
102                 if (buffer_jbd(buffer_cache_bh)
103                     && ocfs2_inode_is_new(inode)) {
104                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
105                         if (!kaddr) {
106                                 mlog(ML_ERROR, "couldn't kmap!\n");
107                                 goto bail;
108                         }
109                         memcpy(kaddr + (bh_result->b_size * iblock),
110                                buffer_cache_bh->b_data,
111                                bh_result->b_size);
112                         kunmap_atomic(kaddr, KM_USER0);
113                         set_buffer_uptodate(bh_result);
114                 }
115                 brelse(buffer_cache_bh);
116         }
117
118         map_bh(bh_result, inode->i_sb,
119                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
120
121         err = 0;
122
123 bail:
124         brelse(bh);
125
126         mlog_exit(err);
127         return err;
128 }
129
130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
131                     struct buffer_head *bh_result, int create)
132 {
133         int err = 0;
134         unsigned int ext_flags;
135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
136         u64 p_blkno, count, past_eof;
137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
138
139         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
140                    (unsigned long long)iblock, bh_result, create);
141
142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
144                      inode, inode->i_ino);
145
146         if (S_ISLNK(inode->i_mode)) {
147                 /* this always does I/O for some reason. */
148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
149                 goto bail;
150         }
151
152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
153                                           &ext_flags);
154         if (err) {
155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
157                      (unsigned long long)p_blkno);
158                 goto bail;
159         }
160
161         if (max_blocks < count)
162                 count = max_blocks;
163
164         /*
165          * ocfs2 never allocates in this function - the only time we
166          * need to use BH_New is when we're extending i_size on a file
167          * system which doesn't support holes, in which case BH_New
168          * allows block_prepare_write() to zero.
169          *
170          * If we see this on a sparse file system, then a truncate has
171          * raced us and removed the cluster. In this case, we clear
172          * the buffers dirty and uptodate bits and let the buffer code
173          * ignore it as a hole.
174          */
175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
176                 clear_buffer_dirty(bh_result);
177                 clear_buffer_uptodate(bh_result);
178                 goto bail;
179         }
180
181         /* Treat the unwritten extent as a hole for zeroing purposes. */
182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
183                 map_bh(bh_result, inode->i_sb, p_blkno);
184
185         bh_result->b_size = count << inode->i_blkbits;
186
187         if (!ocfs2_sparse_alloc(osb)) {
188                 if (p_blkno == 0) {
189                         err = -EIO;
190                         mlog(ML_ERROR,
191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
192                              (unsigned long long)iblock,
193                              (unsigned long long)p_blkno,
194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
196                         dump_stack();
197                         goto bail;
198                 }
199
200                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
201                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
202                      (unsigned long long)past_eof);
203
204                 if (create && (iblock >= past_eof))
205                         set_buffer_new(bh_result);
206         }
207
208 bail:
209         if (err < 0)
210                 err = -EIO;
211
212         mlog_exit(err);
213         return err;
214 }
215
216 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
217                            struct buffer_head *di_bh)
218 {
219         void *kaddr;
220         loff_t size;
221         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
222
223         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
224                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
225                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
226                 return -EROFS;
227         }
228
229         size = i_size_read(inode);
230
231         if (size > PAGE_CACHE_SIZE ||
232             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
233                 ocfs2_error(inode->i_sb,
234                             "Inode %llu has with inline data has bad size: %Lu",
235                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
236                             (unsigned long long)size);
237                 return -EROFS;
238         }
239
240         kaddr = kmap_atomic(page, KM_USER0);
241         if (size)
242                 memcpy(kaddr, di->id2.i_data.id_data, size);
243         /* Clear the remaining part of the page */
244         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
245         flush_dcache_page(page);
246         kunmap_atomic(kaddr, KM_USER0);
247
248         SetPageUptodate(page);
249
250         return 0;
251 }
252
253 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
254 {
255         int ret;
256         struct buffer_head *di_bh = NULL;
257
258         BUG_ON(!PageLocked(page));
259         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
260
261         ret = ocfs2_read_inode_block(inode, &di_bh);
262         if (ret) {
263                 mlog_errno(ret);
264                 goto out;
265         }
266
267         ret = ocfs2_read_inline_data(inode, page, di_bh);
268 out:
269         unlock_page(page);
270
271         brelse(di_bh);
272         return ret;
273 }
274
275 static int ocfs2_readpage(struct file *file, struct page *page)
276 {
277         struct inode *inode = page->mapping->host;
278         struct ocfs2_inode_info *oi = OCFS2_I(inode);
279         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
280         int ret, unlock = 1;
281
282         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
283
284         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
285         if (ret != 0) {
286                 if (ret == AOP_TRUNCATED_PAGE)
287                         unlock = 0;
288                 mlog_errno(ret);
289                 goto out;
290         }
291
292         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
293                 ret = AOP_TRUNCATED_PAGE;
294                 goto out_inode_unlock;
295         }
296
297         /*
298          * i_size might have just been updated as we grabed the meta lock.  We
299          * might now be discovering a truncate that hit on another node.
300          * block_read_full_page->get_block freaks out if it is asked to read
301          * beyond the end of a file, so we check here.  Callers
302          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
303          * and notice that the page they just read isn't needed.
304          *
305          * XXX sys_readahead() seems to get that wrong?
306          */
307         if (start >= i_size_read(inode)) {
308                 zero_user(page, 0, PAGE_SIZE);
309                 SetPageUptodate(page);
310                 ret = 0;
311                 goto out_alloc;
312         }
313
314         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
315                 ret = ocfs2_readpage_inline(inode, page);
316         else
317                 ret = block_read_full_page(page, ocfs2_get_block);
318         unlock = 0;
319
320 out_alloc:
321         up_read(&OCFS2_I(inode)->ip_alloc_sem);
322 out_inode_unlock:
323         ocfs2_inode_unlock(inode, 0);
324 out:
325         if (unlock)
326                 unlock_page(page);
327         mlog_exit(ret);
328         return ret;
329 }
330
331 /*
332  * This is used only for read-ahead. Failures or difficult to handle
333  * situations are safe to ignore.
334  *
335  * Right now, we don't bother with BH_Boundary - in-inode extent lists
336  * are quite large (243 extents on 4k blocks), so most inodes don't
337  * grow out to a tree. If need be, detecting boundary extents could
338  * trivially be added in a future version of ocfs2_get_block().
339  */
340 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
341                            struct list_head *pages, unsigned nr_pages)
342 {
343         int ret, err = -EIO;
344         struct inode *inode = mapping->host;
345         struct ocfs2_inode_info *oi = OCFS2_I(inode);
346         loff_t start;
347         struct page *last;
348
349         /*
350          * Use the nonblocking flag for the dlm code to avoid page
351          * lock inversion, but don't bother with retrying.
352          */
353         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
354         if (ret)
355                 return err;
356
357         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
358                 ocfs2_inode_unlock(inode, 0);
359                 return err;
360         }
361
362         /*
363          * Don't bother with inline-data. There isn't anything
364          * to read-ahead in that case anyway...
365          */
366         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
367                 goto out_unlock;
368
369         /*
370          * Check whether a remote node truncated this file - we just
371          * drop out in that case as it's not worth handling here.
372          */
373         last = list_entry(pages->prev, struct page, lru);
374         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
375         if (start >= i_size_read(inode))
376                 goto out_unlock;
377
378         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
379
380 out_unlock:
381         up_read(&oi->ip_alloc_sem);
382         ocfs2_inode_unlock(inode, 0);
383
384         return err;
385 }
386
387 /* Note: Because we don't support holes, our allocation has
388  * already happened (allocation writes zeros to the file data)
389  * so we don't have to worry about ordered writes in
390  * ocfs2_writepage.
391  *
392  * ->writepage is called during the process of invalidating the page cache
393  * during blocked lock processing.  It can't block on any cluster locks
394  * to during block mapping.  It's relying on the fact that the block
395  * mapping can't have disappeared under the dirty pages that it is
396  * being asked to write back.
397  */
398 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
399 {
400         int ret;
401
402         mlog_entry("(0x%p)\n", page);
403
404         ret = block_write_full_page(page, ocfs2_get_block, wbc);
405
406         mlog_exit(ret);
407
408         return ret;
409 }
410
411 /*
412  * This is called from ocfs2_write_zero_page() which has handled it's
413  * own cluster locking and has ensured allocation exists for those
414  * blocks to be written.
415  */
416 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
417                                unsigned from, unsigned to)
418 {
419         int ret;
420
421         ret = block_prepare_write(page, from, to, ocfs2_get_block);
422
423         return ret;
424 }
425
426 /* Taken from ext3. We don't necessarily need the full blown
427  * functionality yet, but IMHO it's better to cut and paste the whole
428  * thing so we can avoid introducing our own bugs (and easily pick up
429  * their fixes when they happen) --Mark */
430 int walk_page_buffers(  handle_t *handle,
431                         struct buffer_head *head,
432                         unsigned from,
433                         unsigned to,
434                         int *partial,
435                         int (*fn)(      handle_t *handle,
436                                         struct buffer_head *bh))
437 {
438         struct buffer_head *bh;
439         unsigned block_start, block_end;
440         unsigned blocksize = head->b_size;
441         int err, ret = 0;
442         struct buffer_head *next;
443
444         for (   bh = head, block_start = 0;
445                 ret == 0 && (bh != head || !block_start);
446                 block_start = block_end, bh = next)
447         {
448                 next = bh->b_this_page;
449                 block_end = block_start + blocksize;
450                 if (block_end <= from || block_start >= to) {
451                         if (partial && !buffer_uptodate(bh))
452                                 *partial = 1;
453                         continue;
454                 }
455                 err = (*fn)(handle, bh);
456                 if (!ret)
457                         ret = err;
458         }
459         return ret;
460 }
461
462 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
463                                                          struct page *page,
464                                                          unsigned from,
465                                                          unsigned to)
466 {
467         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
468         handle_t *handle;
469         int ret = 0;
470
471         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
472         if (IS_ERR(handle)) {
473                 ret = -ENOMEM;
474                 mlog_errno(ret);
475                 goto out;
476         }
477
478         if (ocfs2_should_order_data(inode)) {
479                 ret = ocfs2_jbd2_file_inode(handle, inode);
480                 if (ret < 0)
481                         mlog_errno(ret);
482         }
483 out:
484         if (ret) {
485                 if (!IS_ERR(handle))
486                         ocfs2_commit_trans(osb, handle);
487                 handle = ERR_PTR(ret);
488         }
489         return handle;
490 }
491
492 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
493 {
494         sector_t status;
495         u64 p_blkno = 0;
496         int err = 0;
497         struct inode *inode = mapping->host;
498
499         mlog_entry("(block = %llu)\n", (unsigned long long)block);
500
501         /* We don't need to lock journal system files, since they aren't
502          * accessed concurrently from multiple nodes.
503          */
504         if (!INODE_JOURNAL(inode)) {
505                 err = ocfs2_inode_lock(inode, NULL, 0);
506                 if (err) {
507                         if (err != -ENOENT)
508                                 mlog_errno(err);
509                         goto bail;
510                 }
511                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
512         }
513
514         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
515                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
516                                                   NULL);
517
518         if (!INODE_JOURNAL(inode)) {
519                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
520                 ocfs2_inode_unlock(inode, 0);
521         }
522
523         if (err) {
524                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
525                      (unsigned long long)block);
526                 mlog_errno(err);
527                 goto bail;
528         }
529
530 bail:
531         status = err ? 0 : p_blkno;
532
533         mlog_exit((int)status);
534
535         return status;
536 }
537
538 /*
539  * TODO: Make this into a generic get_blocks function.
540  *
541  * From do_direct_io in direct-io.c:
542  *  "So what we do is to permit the ->get_blocks function to populate
543  *   bh.b_size with the size of IO which is permitted at this offset and
544  *   this i_blkbits."
545  *
546  * This function is called directly from get_more_blocks in direct-io.c.
547  *
548  * called like this: dio->get_blocks(dio->inode, fs_startblk,
549  *                                      fs_count, map_bh, dio->rw == WRITE);
550  */
551 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
552                                      struct buffer_head *bh_result, int create)
553 {
554         int ret;
555         u64 p_blkno, inode_blocks, contig_blocks;
556         unsigned int ext_flags;
557         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
558         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
559
560         /* This function won't even be called if the request isn't all
561          * nicely aligned and of the right size, so there's no need
562          * for us to check any of that. */
563
564         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
565
566         /*
567          * Any write past EOF is not allowed because we'd be extending.
568          */
569         if (create && (iblock + max_blocks) > inode_blocks) {
570                 ret = -EIO;
571                 goto bail;
572         }
573
574         /* This figures out the size of the next contiguous block, and
575          * our logical offset */
576         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
577                                           &contig_blocks, &ext_flags);
578         if (ret) {
579                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
580                      (unsigned long long)iblock);
581                 ret = -EIO;
582                 goto bail;
583         }
584
585         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
586                 ocfs2_error(inode->i_sb,
587                             "Inode %llu has a hole at block %llu\n",
588                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
589                             (unsigned long long)iblock);
590                 ret = -EROFS;
591                 goto bail;
592         }
593
594         /* We should already CoW the refcounted extent. */
595         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
596         /*
597          * get_more_blocks() expects us to describe a hole by clearing
598          * the mapped bit on bh_result().
599          *
600          * Consider an unwritten extent as a hole.
601          */
602         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
603                 map_bh(bh_result, inode->i_sb, p_blkno);
604         else {
605                 /*
606                  * ocfs2_prepare_inode_for_write() should have caught
607                  * the case where we'd be filling a hole and triggered
608                  * a buffered write instead.
609                  */
610                 if (create) {
611                         ret = -EIO;
612                         mlog_errno(ret);
613                         goto bail;
614                 }
615
616                 clear_buffer_mapped(bh_result);
617         }
618
619         /* make sure we don't map more than max_blocks blocks here as
620            that's all the kernel will handle at this point. */
621         if (max_blocks < contig_blocks)
622                 contig_blocks = max_blocks;
623         bh_result->b_size = contig_blocks << blocksize_bits;
624 bail:
625         return ret;
626 }
627
628 /* 
629  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
630  * particularly interested in the aio/dio case.  Like the core uses
631  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
632  * truncation on another.
633  */
634 static void ocfs2_dio_end_io(struct kiocb *iocb,
635                              loff_t offset,
636                              ssize_t bytes,
637                              void *private)
638 {
639         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
640         int level;
641
642         /* this io's submitter should not have unlocked this before we could */
643         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
644
645         ocfs2_iocb_clear_rw_locked(iocb);
646
647         level = ocfs2_iocb_rw_locked_level(iocb);
648         if (!level)
649                 up_read(&inode->i_alloc_sem);
650         ocfs2_rw_unlock(inode, level);
651 }
652
653 /*
654  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
655  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
656  * do journalled data.
657  */
658 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
659 {
660         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
661
662         jbd2_journal_invalidatepage(journal, page, offset);
663 }
664
665 static int ocfs2_releasepage(struct page *page, gfp_t wait)
666 {
667         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
668
669         if (!page_has_buffers(page))
670                 return 0;
671         return jbd2_journal_try_to_free_buffers(journal, page, wait);
672 }
673
674 static ssize_t ocfs2_direct_IO(int rw,
675                                struct kiocb *iocb,
676                                const struct iovec *iov,
677                                loff_t offset,
678                                unsigned long nr_segs)
679 {
680         struct file *file = iocb->ki_filp;
681         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
682         int ret;
683
684         mlog_entry_void();
685
686         /*
687          * Fallback to buffered I/O if we see an inode without
688          * extents.
689          */
690         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
691                 return 0;
692
693         /* Fallback to buffered I/O if we are appending. */
694         if (i_size_read(inode) <= offset)
695                 return 0;
696
697         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
698                                             inode->i_sb->s_bdev, iov, offset,
699                                             nr_segs, 
700                                             ocfs2_direct_IO_get_blocks,
701                                             ocfs2_dio_end_io);
702
703         mlog_exit(ret);
704         return ret;
705 }
706
707 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
708                                             u32 cpos,
709                                             unsigned int *start,
710                                             unsigned int *end)
711 {
712         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
713
714         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
715                 unsigned int cpp;
716
717                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
718
719                 cluster_start = cpos % cpp;
720                 cluster_start = cluster_start << osb->s_clustersize_bits;
721
722                 cluster_end = cluster_start + osb->s_clustersize;
723         }
724
725         BUG_ON(cluster_start > PAGE_SIZE);
726         BUG_ON(cluster_end > PAGE_SIZE);
727
728         if (start)
729                 *start = cluster_start;
730         if (end)
731                 *end = cluster_end;
732 }
733
734 /*
735  * 'from' and 'to' are the region in the page to avoid zeroing.
736  *
737  * If pagesize > clustersize, this function will avoid zeroing outside
738  * of the cluster boundary.
739  *
740  * from == to == 0 is code for "zero the entire cluster region"
741  */
742 static void ocfs2_clear_page_regions(struct page *page,
743                                      struct ocfs2_super *osb, u32 cpos,
744                                      unsigned from, unsigned to)
745 {
746         void *kaddr;
747         unsigned int cluster_start, cluster_end;
748
749         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
750
751         kaddr = kmap_atomic(page, KM_USER0);
752
753         if (from || to) {
754                 if (from > cluster_start)
755                         memset(kaddr + cluster_start, 0, from - cluster_start);
756                 if (to < cluster_end)
757                         memset(kaddr + to, 0, cluster_end - to);
758         } else {
759                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
760         }
761
762         kunmap_atomic(kaddr, KM_USER0);
763 }
764
765 /*
766  * Nonsparse file systems fully allocate before we get to the write
767  * code. This prevents ocfs2_write() from tagging the write as an
768  * allocating one, which means ocfs2_map_page_blocks() might try to
769  * read-in the blocks at the tail of our file. Avoid reading them by
770  * testing i_size against each block offset.
771  */
772 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
773                                  unsigned int block_start)
774 {
775         u64 offset = page_offset(page) + block_start;
776
777         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
778                 return 1;
779
780         if (i_size_read(inode) > offset)
781                 return 1;
782
783         return 0;
784 }
785
786 /*
787  * Some of this taken from block_prepare_write(). We already have our
788  * mapping by now though, and the entire write will be allocating or
789  * it won't, so not much need to use BH_New.
790  *
791  * This will also skip zeroing, which is handled externally.
792  */
793 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
794                           struct inode *inode, unsigned int from,
795                           unsigned int to, int new)
796 {
797         int ret = 0;
798         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
799         unsigned int block_end, block_start;
800         unsigned int bsize = 1 << inode->i_blkbits;
801
802         if (!page_has_buffers(page))
803                 create_empty_buffers(page, bsize, 0);
804
805         head = page_buffers(page);
806         for (bh = head, block_start = 0; bh != head || !block_start;
807              bh = bh->b_this_page, block_start += bsize) {
808                 block_end = block_start + bsize;
809
810                 clear_buffer_new(bh);
811
812                 /*
813                  * Ignore blocks outside of our i/o range -
814                  * they may belong to unallocated clusters.
815                  */
816                 if (block_start >= to || block_end <= from) {
817                         if (PageUptodate(page))
818                                 set_buffer_uptodate(bh);
819                         continue;
820                 }
821
822                 /*
823                  * For an allocating write with cluster size >= page
824                  * size, we always write the entire page.
825                  */
826                 if (new)
827                         set_buffer_new(bh);
828
829                 if (!buffer_mapped(bh)) {
830                         map_bh(bh, inode->i_sb, *p_blkno);
831                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
832                 }
833
834                 if (PageUptodate(page)) {
835                         if (!buffer_uptodate(bh))
836                                 set_buffer_uptodate(bh);
837                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
838                            !buffer_new(bh) &&
839                            ocfs2_should_read_blk(inode, page, block_start) &&
840                            (block_start < from || block_end > to)) {
841                         ll_rw_block(READ, 1, &bh);
842                         *wait_bh++=bh;
843                 }
844
845                 *p_blkno = *p_blkno + 1;
846         }
847
848         /*
849          * If we issued read requests - let them complete.
850          */
851         while(wait_bh > wait) {
852                 wait_on_buffer(*--wait_bh);
853                 if (!buffer_uptodate(*wait_bh))
854                         ret = -EIO;
855         }
856
857         if (ret == 0 || !new)
858                 return ret;
859
860         /*
861          * If we get -EIO above, zero out any newly allocated blocks
862          * to avoid exposing stale data.
863          */
864         bh = head;
865         block_start = 0;
866         do {
867                 block_end = block_start + bsize;
868                 if (block_end <= from)
869                         goto next_bh;
870                 if (block_start >= to)
871                         break;
872
873                 zero_user(page, block_start, bh->b_size);
874                 set_buffer_uptodate(bh);
875                 mark_buffer_dirty(bh);
876
877 next_bh:
878                 block_start = block_end;
879                 bh = bh->b_this_page;
880         } while (bh != head);
881
882         return ret;
883 }
884
885 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
886 #define OCFS2_MAX_CTXT_PAGES    1
887 #else
888 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
889 #endif
890
891 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
892
893 /*
894  * Describe the state of a single cluster to be written to.
895  */
896 struct ocfs2_write_cluster_desc {
897         u32             c_cpos;
898         u32             c_phys;
899         /*
900          * Give this a unique field because c_phys eventually gets
901          * filled.
902          */
903         unsigned        c_new;
904         unsigned        c_unwritten;
905         unsigned        c_needs_zero;
906 };
907
908 struct ocfs2_write_ctxt {
909         /* Logical cluster position / len of write */
910         u32                             w_cpos;
911         u32                             w_clen;
912
913         /* First cluster allocated in a nonsparse extend */
914         u32                             w_first_new_cpos;
915
916         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
917
918         /*
919          * This is true if page_size > cluster_size.
920          *
921          * It triggers a set of special cases during write which might
922          * have to deal with allocating writes to partial pages.
923          */
924         unsigned int                    w_large_pages;
925
926         /*
927          * Pages involved in this write.
928          *
929          * w_target_page is the page being written to by the user.
930          *
931          * w_pages is an array of pages which always contains
932          * w_target_page, and in the case of an allocating write with
933          * page_size < cluster size, it will contain zero'd and mapped
934          * pages adjacent to w_target_page which need to be written
935          * out in so that future reads from that region will get
936          * zero's.
937          */
938         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
939         unsigned int                    w_num_pages;
940         struct page                     *w_target_page;
941
942         /*
943          * ocfs2_write_end() uses this to know what the real range to
944          * write in the target should be.
945          */
946         unsigned int                    w_target_from;
947         unsigned int                    w_target_to;
948
949         /*
950          * We could use journal_current_handle() but this is cleaner,
951          * IMHO -Mark
952          */
953         handle_t                        *w_handle;
954
955         struct buffer_head              *w_di_bh;
956
957         struct ocfs2_cached_dealloc_ctxt w_dealloc;
958 };
959
960 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
961 {
962         int i;
963
964         for(i = 0; i < num_pages; i++) {
965                 if (pages[i]) {
966                         unlock_page(pages[i]);
967                         mark_page_accessed(pages[i]);
968                         page_cache_release(pages[i]);
969                 }
970         }
971 }
972
973 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
974 {
975         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
976
977         brelse(wc->w_di_bh);
978         kfree(wc);
979 }
980
981 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
982                                   struct ocfs2_super *osb, loff_t pos,
983                                   unsigned len, struct buffer_head *di_bh)
984 {
985         u32 cend;
986         struct ocfs2_write_ctxt *wc;
987
988         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
989         if (!wc)
990                 return -ENOMEM;
991
992         wc->w_cpos = pos >> osb->s_clustersize_bits;
993         wc->w_first_new_cpos = UINT_MAX;
994         cend = (pos + len - 1) >> osb->s_clustersize_bits;
995         wc->w_clen = cend - wc->w_cpos + 1;
996         get_bh(di_bh);
997         wc->w_di_bh = di_bh;
998
999         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1000                 wc->w_large_pages = 1;
1001         else
1002                 wc->w_large_pages = 0;
1003
1004         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1005
1006         *wcp = wc;
1007
1008         return 0;
1009 }
1010
1011 /*
1012  * If a page has any new buffers, zero them out here, and mark them uptodate
1013  * and dirty so they'll be written out (in order to prevent uninitialised
1014  * block data from leaking). And clear the new bit.
1015  */
1016 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1017 {
1018         unsigned int block_start, block_end;
1019         struct buffer_head *head, *bh;
1020
1021         BUG_ON(!PageLocked(page));
1022         if (!page_has_buffers(page))
1023                 return;
1024
1025         bh = head = page_buffers(page);
1026         block_start = 0;
1027         do {
1028                 block_end = block_start + bh->b_size;
1029
1030                 if (buffer_new(bh)) {
1031                         if (block_end > from && block_start < to) {
1032                                 if (!PageUptodate(page)) {
1033                                         unsigned start, end;
1034
1035                                         start = max(from, block_start);
1036                                         end = min(to, block_end);
1037
1038                                         zero_user_segment(page, start, end);
1039                                         set_buffer_uptodate(bh);
1040                                 }
1041
1042                                 clear_buffer_new(bh);
1043                                 mark_buffer_dirty(bh);
1044                         }
1045                 }
1046
1047                 block_start = block_end;
1048                 bh = bh->b_this_page;
1049         } while (bh != head);
1050 }
1051
1052 /*
1053  * Only called when we have a failure during allocating write to write
1054  * zero's to the newly allocated region.
1055  */
1056 static void ocfs2_write_failure(struct inode *inode,
1057                                 struct ocfs2_write_ctxt *wc,
1058                                 loff_t user_pos, unsigned user_len)
1059 {
1060         int i;
1061         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1062                 to = user_pos + user_len;
1063         struct page *tmppage;
1064
1065         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1066
1067         for(i = 0; i < wc->w_num_pages; i++) {
1068                 tmppage = wc->w_pages[i];
1069
1070                 if (page_has_buffers(tmppage)) {
1071                         if (ocfs2_should_order_data(inode))
1072                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1073
1074                         block_commit_write(tmppage, from, to);
1075                 }
1076         }
1077 }
1078
1079 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1080                                         struct ocfs2_write_ctxt *wc,
1081                                         struct page *page, u32 cpos,
1082                                         loff_t user_pos, unsigned user_len,
1083                                         int new)
1084 {
1085         int ret;
1086         unsigned int map_from = 0, map_to = 0;
1087         unsigned int cluster_start, cluster_end;
1088         unsigned int user_data_from = 0, user_data_to = 0;
1089
1090         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1091                                         &cluster_start, &cluster_end);
1092
1093         if (page == wc->w_target_page) {
1094                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1095                 map_to = map_from + user_len;
1096
1097                 if (new)
1098                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1099                                                     cluster_start, cluster_end,
1100                                                     new);
1101                 else
1102                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1103                                                     map_from, map_to, new);
1104                 if (ret) {
1105                         mlog_errno(ret);
1106                         goto out;
1107                 }
1108
1109                 user_data_from = map_from;
1110                 user_data_to = map_to;
1111                 if (new) {
1112                         map_from = cluster_start;
1113                         map_to = cluster_end;
1114                 }
1115         } else {
1116                 /*
1117                  * If we haven't allocated the new page yet, we
1118                  * shouldn't be writing it out without copying user
1119                  * data. This is likely a math error from the caller.
1120                  */
1121                 BUG_ON(!new);
1122
1123                 map_from = cluster_start;
1124                 map_to = cluster_end;
1125
1126                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1127                                             cluster_start, cluster_end, new);
1128                 if (ret) {
1129                         mlog_errno(ret);
1130                         goto out;
1131                 }
1132         }
1133
1134         /*
1135          * Parts of newly allocated pages need to be zero'd.
1136          *
1137          * Above, we have also rewritten 'to' and 'from' - as far as
1138          * the rest of the function is concerned, the entire cluster
1139          * range inside of a page needs to be written.
1140          *
1141          * We can skip this if the page is up to date - it's already
1142          * been zero'd from being read in as a hole.
1143          */
1144         if (new && !PageUptodate(page))
1145                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1146                                          cpos, user_data_from, user_data_to);
1147
1148         flush_dcache_page(page);
1149
1150 out:
1151         return ret;
1152 }
1153
1154 /*
1155  * This function will only grab one clusters worth of pages.
1156  */
1157 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1158                                       struct ocfs2_write_ctxt *wc,
1159                                       u32 cpos, loff_t user_pos, int new,
1160                                       struct page *mmap_page)
1161 {
1162         int ret = 0, i;
1163         unsigned long start, target_index, index;
1164         struct inode *inode = mapping->host;
1165
1166         target_index = user_pos >> PAGE_CACHE_SHIFT;
1167
1168         /*
1169          * Figure out how many pages we'll be manipulating here. For
1170          * non allocating write, we just change the one
1171          * page. Otherwise, we'll need a whole clusters worth.
1172          */
1173         if (new) {
1174                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1175                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1176         } else {
1177                 wc->w_num_pages = 1;
1178                 start = target_index;
1179         }
1180
1181         for(i = 0; i < wc->w_num_pages; i++) {
1182                 index = start + i;
1183
1184                 if (index == target_index && mmap_page) {
1185                         /*
1186                          * ocfs2_pagemkwrite() is a little different
1187                          * and wants us to directly use the page
1188                          * passed in.
1189                          */
1190                         lock_page(mmap_page);
1191
1192                         if (mmap_page->mapping != mapping) {
1193                                 unlock_page(mmap_page);
1194                                 /*
1195                                  * Sanity check - the locking in
1196                                  * ocfs2_pagemkwrite() should ensure
1197                                  * that this code doesn't trigger.
1198                                  */
1199                                 ret = -EINVAL;
1200                                 mlog_errno(ret);
1201                                 goto out;
1202                         }
1203
1204                         page_cache_get(mmap_page);
1205                         wc->w_pages[i] = mmap_page;
1206                 } else {
1207                         wc->w_pages[i] = find_or_create_page(mapping, index,
1208                                                              GFP_NOFS);
1209                         if (!wc->w_pages[i]) {
1210                                 ret = -ENOMEM;
1211                                 mlog_errno(ret);
1212                                 goto out;
1213                         }
1214                 }
1215
1216                 if (index == target_index)
1217                         wc->w_target_page = wc->w_pages[i];
1218         }
1219 out:
1220         return ret;
1221 }
1222
1223 /*
1224  * Prepare a single cluster for write one cluster into the file.
1225  */
1226 static int ocfs2_write_cluster(struct address_space *mapping,
1227                                u32 phys, unsigned int unwritten,
1228                                unsigned int should_zero,
1229                                struct ocfs2_alloc_context *data_ac,
1230                                struct ocfs2_alloc_context *meta_ac,
1231                                struct ocfs2_write_ctxt *wc, u32 cpos,
1232                                loff_t user_pos, unsigned user_len)
1233 {
1234         int ret, i, new;
1235         u64 v_blkno, p_blkno;
1236         struct inode *inode = mapping->host;
1237         struct ocfs2_extent_tree et;
1238
1239         new = phys == 0 ? 1 : 0;
1240         if (new) {
1241                 u32 tmp_pos;
1242
1243                 /*
1244                  * This is safe to call with the page locks - it won't take
1245                  * any additional semaphores or cluster locks.
1246                  */
1247                 tmp_pos = cpos;
1248                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1249                                            &tmp_pos, 1, 0, wc->w_di_bh,
1250                                            wc->w_handle, data_ac,
1251                                            meta_ac, NULL);
1252                 /*
1253                  * This shouldn't happen because we must have already
1254                  * calculated the correct meta data allocation required. The
1255                  * internal tree allocation code should know how to increase
1256                  * transaction credits itself.
1257                  *
1258                  * If need be, we could handle -EAGAIN for a
1259                  * RESTART_TRANS here.
1260                  */
1261                 mlog_bug_on_msg(ret == -EAGAIN,
1262                                 "Inode %llu: EAGAIN return during allocation.\n",
1263                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1264                 if (ret < 0) {
1265                         mlog_errno(ret);
1266                         goto out;
1267                 }
1268         } else if (unwritten) {
1269                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1270                                               wc->w_di_bh);
1271                 ret = ocfs2_mark_extent_written(inode, &et,
1272                                                 wc->w_handle, cpos, 1, phys,
1273                                                 meta_ac, &wc->w_dealloc);
1274                 if (ret < 0) {
1275                         mlog_errno(ret);
1276                         goto out;
1277                 }
1278         }
1279
1280         if (should_zero)
1281                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1282         else
1283                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1284
1285         /*
1286          * The only reason this should fail is due to an inability to
1287          * find the extent added.
1288          */
1289         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1290                                           NULL);
1291         if (ret < 0) {
1292                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1293                             "at logical block %llu",
1294                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1295                             (unsigned long long)v_blkno);
1296                 goto out;
1297         }
1298
1299         BUG_ON(p_blkno == 0);
1300
1301         for(i = 0; i < wc->w_num_pages; i++) {
1302                 int tmpret;
1303
1304                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1305                                                       wc->w_pages[i], cpos,
1306                                                       user_pos, user_len,
1307                                                       should_zero);
1308                 if (tmpret) {
1309                         mlog_errno(tmpret);
1310                         if (ret == 0)
1311                                 ret = tmpret;
1312                 }
1313         }
1314
1315         /*
1316          * We only have cleanup to do in case of allocating write.
1317          */
1318         if (ret && new)
1319                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1320
1321 out:
1322
1323         return ret;
1324 }
1325
1326 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1327                                        struct ocfs2_alloc_context *data_ac,
1328                                        struct ocfs2_alloc_context *meta_ac,
1329                                        struct ocfs2_write_ctxt *wc,
1330                                        loff_t pos, unsigned len)
1331 {
1332         int ret, i;
1333         loff_t cluster_off;
1334         unsigned int local_len = len;
1335         struct ocfs2_write_cluster_desc *desc;
1336         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1337
1338         for (i = 0; i < wc->w_clen; i++) {
1339                 desc = &wc->w_desc[i];
1340
1341                 /*
1342                  * We have to make sure that the total write passed in
1343                  * doesn't extend past a single cluster.
1344                  */
1345                 local_len = len;
1346                 cluster_off = pos & (osb->s_clustersize - 1);
1347                 if ((cluster_off + local_len) > osb->s_clustersize)
1348                         local_len = osb->s_clustersize - cluster_off;
1349
1350                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1351                                           desc->c_unwritten,
1352                                           desc->c_needs_zero,
1353                                           data_ac, meta_ac,
1354                                           wc, desc->c_cpos, pos, local_len);
1355                 if (ret) {
1356                         mlog_errno(ret);
1357                         goto out;
1358                 }
1359
1360                 len -= local_len;
1361                 pos += local_len;
1362         }
1363
1364         ret = 0;
1365 out:
1366         return ret;
1367 }
1368
1369 /*
1370  * ocfs2_write_end() wants to know which parts of the target page it
1371  * should complete the write on. It's easiest to compute them ahead of
1372  * time when a more complete view of the write is available.
1373  */
1374 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1375                                         struct ocfs2_write_ctxt *wc,
1376                                         loff_t pos, unsigned len, int alloc)
1377 {
1378         struct ocfs2_write_cluster_desc *desc;
1379
1380         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1381         wc->w_target_to = wc->w_target_from + len;
1382
1383         if (alloc == 0)
1384                 return;
1385
1386         /*
1387          * Allocating write - we may have different boundaries based
1388          * on page size and cluster size.
1389          *
1390          * NOTE: We can no longer compute one value from the other as
1391          * the actual write length and user provided length may be
1392          * different.
1393          */
1394
1395         if (wc->w_large_pages) {
1396                 /*
1397                  * We only care about the 1st and last cluster within
1398                  * our range and whether they should be zero'd or not. Either
1399                  * value may be extended out to the start/end of a
1400                  * newly allocated cluster.
1401                  */
1402                 desc = &wc->w_desc[0];
1403                 if (desc->c_needs_zero)
1404                         ocfs2_figure_cluster_boundaries(osb,
1405                                                         desc->c_cpos,
1406                                                         &wc->w_target_from,
1407                                                         NULL);
1408
1409                 desc = &wc->w_desc[wc->w_clen - 1];
1410                 if (desc->c_needs_zero)
1411                         ocfs2_figure_cluster_boundaries(osb,
1412                                                         desc->c_cpos,
1413                                                         NULL,
1414                                                         &wc->w_target_to);
1415         } else {
1416                 wc->w_target_from = 0;
1417                 wc->w_target_to = PAGE_CACHE_SIZE;
1418         }
1419 }
1420
1421 /*
1422  * Populate each single-cluster write descriptor in the write context
1423  * with information about the i/o to be done.
1424  *
1425  * Returns the number of clusters that will have to be allocated, as
1426  * well as a worst case estimate of the number of extent records that
1427  * would have to be created during a write to an unwritten region.
1428  */
1429 static int ocfs2_populate_write_desc(struct inode *inode,
1430                                      struct ocfs2_write_ctxt *wc,
1431                                      unsigned int *clusters_to_alloc,
1432                                      unsigned int *extents_to_split)
1433 {
1434         int ret;
1435         struct ocfs2_write_cluster_desc *desc;
1436         unsigned int num_clusters = 0;
1437         unsigned int ext_flags = 0;
1438         u32 phys = 0;
1439         int i;
1440
1441         *clusters_to_alloc = 0;
1442         *extents_to_split = 0;
1443
1444         for (i = 0; i < wc->w_clen; i++) {
1445                 desc = &wc->w_desc[i];
1446                 desc->c_cpos = wc->w_cpos + i;
1447
1448                 if (num_clusters == 0) {
1449                         /*
1450                          * Need to look up the next extent record.
1451                          */
1452                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1453                                                  &num_clusters, &ext_flags);
1454                         if (ret) {
1455                                 mlog_errno(ret);
1456                                 goto out;
1457                         }
1458
1459                         /* We should already CoW the refcountd extent. */
1460                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1461
1462                         /*
1463                          * Assume worst case - that we're writing in
1464                          * the middle of the extent.
1465                          *
1466                          * We can assume that the write proceeds from
1467                          * left to right, in which case the extent
1468                          * insert code is smart enough to coalesce the
1469                          * next splits into the previous records created.
1470                          */
1471                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1472                                 *extents_to_split = *extents_to_split + 2;
1473                 } else if (phys) {
1474                         /*
1475                          * Only increment phys if it doesn't describe
1476                          * a hole.
1477                          */
1478                         phys++;
1479                 }
1480
1481                 /*
1482                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1483                  * file that got extended.  w_first_new_cpos tells us
1484                  * where the newly allocated clusters are so we can
1485                  * zero them.
1486                  */
1487                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1488                         BUG_ON(phys == 0);
1489                         desc->c_needs_zero = 1;
1490                 }
1491
1492                 desc->c_phys = phys;
1493                 if (phys == 0) {
1494                         desc->c_new = 1;
1495                         desc->c_needs_zero = 1;
1496                         *clusters_to_alloc = *clusters_to_alloc + 1;
1497                 }
1498
1499                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1500                         desc->c_unwritten = 1;
1501                         desc->c_needs_zero = 1;
1502                 }
1503
1504                 num_clusters--;
1505         }
1506
1507         ret = 0;
1508 out:
1509         return ret;
1510 }
1511
1512 static int ocfs2_write_begin_inline(struct address_space *mapping,
1513                                     struct inode *inode,
1514                                     struct ocfs2_write_ctxt *wc)
1515 {
1516         int ret;
1517         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1518         struct page *page;
1519         handle_t *handle;
1520         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1521
1522         page = find_or_create_page(mapping, 0, GFP_NOFS);
1523         if (!page) {
1524                 ret = -ENOMEM;
1525                 mlog_errno(ret);
1526                 goto out;
1527         }
1528         /*
1529          * If we don't set w_num_pages then this page won't get unlocked
1530          * and freed on cleanup of the write context.
1531          */
1532         wc->w_pages[0] = wc->w_target_page = page;
1533         wc->w_num_pages = 1;
1534
1535         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1536         if (IS_ERR(handle)) {
1537                 ret = PTR_ERR(handle);
1538                 mlog_errno(ret);
1539                 goto out;
1540         }
1541
1542         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1543                                       OCFS2_JOURNAL_ACCESS_WRITE);
1544         if (ret) {
1545                 ocfs2_commit_trans(osb, handle);
1546
1547                 mlog_errno(ret);
1548                 goto out;
1549         }
1550
1551         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1552                 ocfs2_set_inode_data_inline(inode, di);
1553
1554         if (!PageUptodate(page)) {
1555                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1556                 if (ret) {
1557                         ocfs2_commit_trans(osb, handle);
1558
1559                         goto out;
1560                 }
1561         }
1562
1563         wc->w_handle = handle;
1564 out:
1565         return ret;
1566 }
1567
1568 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1569 {
1570         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1571
1572         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1573                 return 1;
1574         return 0;
1575 }
1576
1577 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1578                                           struct inode *inode, loff_t pos,
1579                                           unsigned len, struct page *mmap_page,
1580                                           struct ocfs2_write_ctxt *wc)
1581 {
1582         int ret, written = 0;
1583         loff_t end = pos + len;
1584         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1585         struct ocfs2_dinode *di = NULL;
1586
1587         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1588              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1589              oi->ip_dyn_features);
1590
1591         /*
1592          * Handle inodes which already have inline data 1st.
1593          */
1594         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1595                 if (mmap_page == NULL &&
1596                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1597                         goto do_inline_write;
1598
1599                 /*
1600                  * The write won't fit - we have to give this inode an
1601                  * inline extent list now.
1602                  */
1603                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1604                 if (ret)
1605                         mlog_errno(ret);
1606                 goto out;
1607         }
1608
1609         /*
1610          * Check whether the inode can accept inline data.
1611          */
1612         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1613                 return 0;
1614
1615         /*
1616          * Check whether the write can fit.
1617          */
1618         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1619         if (mmap_page ||
1620             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1621                 return 0;
1622
1623 do_inline_write:
1624         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1625         if (ret) {
1626                 mlog_errno(ret);
1627                 goto out;
1628         }
1629
1630         /*
1631          * This signals to the caller that the data can be written
1632          * inline.
1633          */
1634         written = 1;
1635 out:
1636         return written ? written : ret;
1637 }
1638
1639 /*
1640  * This function only does anything for file systems which can't
1641  * handle sparse files.
1642  *
1643  * What we want to do here is fill in any hole between the current end
1644  * of allocation and the end of our write. That way the rest of the
1645  * write path can treat it as an non-allocating write, which has no
1646  * special case code for sparse/nonsparse files.
1647  */
1648 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1649                                         unsigned len,
1650                                         struct ocfs2_write_ctxt *wc)
1651 {
1652         int ret;
1653         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1654         loff_t newsize = pos + len;
1655
1656         if (ocfs2_sparse_alloc(osb))
1657                 return 0;
1658
1659         if (newsize <= i_size_read(inode))
1660                 return 0;
1661
1662         ret = ocfs2_extend_no_holes(inode, newsize, pos);
1663         if (ret)
1664                 mlog_errno(ret);
1665
1666         wc->w_first_new_cpos =
1667                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1668
1669         return ret;
1670 }
1671
1672 int ocfs2_write_begin_nolock(struct address_space *mapping,
1673                              loff_t pos, unsigned len, unsigned flags,
1674                              struct page **pagep, void **fsdata,
1675                              struct buffer_head *di_bh, struct page *mmap_page)
1676 {
1677         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1678         unsigned int clusters_to_alloc, extents_to_split;
1679         struct ocfs2_write_ctxt *wc;
1680         struct inode *inode = mapping->host;
1681         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1682         struct ocfs2_dinode *di;
1683         struct ocfs2_alloc_context *data_ac = NULL;
1684         struct ocfs2_alloc_context *meta_ac = NULL;
1685         handle_t *handle;
1686         struct ocfs2_extent_tree et;
1687
1688         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1689         if (ret) {
1690                 mlog_errno(ret);
1691                 return ret;
1692         }
1693
1694         if (ocfs2_supports_inline_data(osb)) {
1695                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1696                                                      mmap_page, wc);
1697                 if (ret == 1) {
1698                         ret = 0;
1699                         goto success;
1700                 }
1701                 if (ret < 0) {
1702                         mlog_errno(ret);
1703                         goto out;
1704                 }
1705         }
1706
1707         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1708         if (ret) {
1709                 mlog_errno(ret);
1710                 goto out;
1711         }
1712
1713         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1714         if (ret < 0) {
1715                 mlog_errno(ret);
1716                 goto out;
1717         } else if (ret == 1) {
1718                 ret = ocfs2_refcount_cow(inode, di_bh,
1719                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1720                 if (ret) {
1721                         mlog_errno(ret);
1722                         goto out;
1723                 }
1724         }
1725
1726         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1727                                         &extents_to_split);
1728         if (ret) {
1729                 mlog_errno(ret);
1730                 goto out;
1731         }
1732
1733         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1734
1735         /*
1736          * We set w_target_from, w_target_to here so that
1737          * ocfs2_write_end() knows which range in the target page to
1738          * write out. An allocation requires that we write the entire
1739          * cluster range.
1740          */
1741         if (clusters_to_alloc || extents_to_split) {
1742                 /*
1743                  * XXX: We are stretching the limits of
1744                  * ocfs2_lock_allocators(). It greatly over-estimates
1745                  * the work to be done.
1746                  */
1747                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1748                      " clusters_to_add = %u, extents_to_split = %u\n",
1749                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1750                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1751                      clusters_to_alloc, extents_to_split);
1752
1753                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1754                                               wc->w_di_bh);
1755                 ret = ocfs2_lock_allocators(inode, &et,
1756                                             clusters_to_alloc, extents_to_split,
1757                                             &data_ac, &meta_ac);
1758                 if (ret) {
1759                         mlog_errno(ret);
1760                         goto out;
1761                 }
1762
1763                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1764                                                     &di->id2.i_list,
1765                                                     clusters_to_alloc);
1766
1767         }
1768
1769         /*
1770          * We have to zero sparse allocated clusters, unwritten extent clusters,
1771          * and non-sparse clusters we just extended.  For non-sparse writes,
1772          * we know zeros will only be needed in the first and/or last cluster.
1773          */
1774         if (clusters_to_alloc || extents_to_split ||
1775             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1776                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1777                 cluster_of_pages = 1;
1778         else
1779                 cluster_of_pages = 0;
1780
1781         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1782
1783         handle = ocfs2_start_trans(osb, credits);
1784         if (IS_ERR(handle)) {
1785                 ret = PTR_ERR(handle);
1786                 mlog_errno(ret);
1787                 goto out;
1788         }
1789
1790         wc->w_handle = handle;
1791
1792         if (clusters_to_alloc && vfs_dq_alloc_space_nodirty(inode,
1793                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc))) {
1794                 ret = -EDQUOT;
1795                 goto out_commit;
1796         }
1797         /*
1798          * We don't want this to fail in ocfs2_write_end(), so do it
1799          * here.
1800          */
1801         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1802                                       OCFS2_JOURNAL_ACCESS_WRITE);
1803         if (ret) {
1804                 mlog_errno(ret);
1805                 goto out_quota;
1806         }
1807
1808         /*
1809          * Fill our page array first. That way we've grabbed enough so
1810          * that we can zero and flush if we error after adding the
1811          * extent.
1812          */
1813         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1814                                          cluster_of_pages, mmap_page);
1815         if (ret) {
1816                 mlog_errno(ret);
1817                 goto out_quota;
1818         }
1819
1820         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1821                                           len);
1822         if (ret) {
1823                 mlog_errno(ret);
1824                 goto out_quota;
1825         }
1826
1827         if (data_ac)
1828                 ocfs2_free_alloc_context(data_ac);
1829         if (meta_ac)
1830                 ocfs2_free_alloc_context(meta_ac);
1831
1832 success:
1833         *pagep = wc->w_target_page;
1834         *fsdata = wc;
1835         return 0;
1836 out_quota:
1837         if (clusters_to_alloc)
1838                 vfs_dq_free_space(inode,
1839                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1840 out_commit:
1841         ocfs2_commit_trans(osb, handle);
1842
1843 out:
1844         ocfs2_free_write_ctxt(wc);
1845
1846         if (data_ac)
1847                 ocfs2_free_alloc_context(data_ac);
1848         if (meta_ac)
1849                 ocfs2_free_alloc_context(meta_ac);
1850         return ret;
1851 }
1852
1853 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1854                              loff_t pos, unsigned len, unsigned flags,
1855                              struct page **pagep, void **fsdata)
1856 {
1857         int ret;
1858         struct buffer_head *di_bh = NULL;
1859         struct inode *inode = mapping->host;
1860
1861         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1862         if (ret) {
1863                 mlog_errno(ret);
1864                 return ret;
1865         }
1866
1867         /*
1868          * Take alloc sem here to prevent concurrent lookups. That way
1869          * the mapping, zeroing and tree manipulation within
1870          * ocfs2_write() will be safe against ->readpage(). This
1871          * should also serve to lock out allocation from a shared
1872          * writeable region.
1873          */
1874         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1875
1876         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1877                                        fsdata, di_bh, NULL);
1878         if (ret) {
1879                 mlog_errno(ret);
1880                 goto out_fail;
1881         }
1882
1883         brelse(di_bh);
1884
1885         return 0;
1886
1887 out_fail:
1888         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1889
1890         brelse(di_bh);
1891         ocfs2_inode_unlock(inode, 1);
1892
1893         return ret;
1894 }
1895
1896 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1897                                    unsigned len, unsigned *copied,
1898                                    struct ocfs2_dinode *di,
1899                                    struct ocfs2_write_ctxt *wc)
1900 {
1901         void *kaddr;
1902
1903         if (unlikely(*copied < len)) {
1904                 if (!PageUptodate(wc->w_target_page)) {
1905                         *copied = 0;
1906                         return;
1907                 }
1908         }
1909
1910         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1911         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1912         kunmap_atomic(kaddr, KM_USER0);
1913
1914         mlog(0, "Data written to inode at offset %llu. "
1915              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1916              (unsigned long long)pos, *copied,
1917              le16_to_cpu(di->id2.i_data.id_count),
1918              le16_to_cpu(di->i_dyn_features));
1919 }
1920
1921 int ocfs2_write_end_nolock(struct address_space *mapping,
1922                            loff_t pos, unsigned len, unsigned copied,
1923                            struct page *page, void *fsdata)
1924 {
1925         int i;
1926         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1927         struct inode *inode = mapping->host;
1928         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1929         struct ocfs2_write_ctxt *wc = fsdata;
1930         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1931         handle_t *handle = wc->w_handle;
1932         struct page *tmppage;
1933
1934         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1935                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1936                 goto out_write_size;
1937         }
1938
1939         if (unlikely(copied < len)) {
1940                 if (!PageUptodate(wc->w_target_page))
1941                         copied = 0;
1942
1943                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1944                                        start+len);
1945         }
1946         flush_dcache_page(wc->w_target_page);
1947
1948         for(i = 0; i < wc->w_num_pages; i++) {
1949                 tmppage = wc->w_pages[i];
1950
1951                 if (tmppage == wc->w_target_page) {
1952                         from = wc->w_target_from;
1953                         to = wc->w_target_to;
1954
1955                         BUG_ON(from > PAGE_CACHE_SIZE ||
1956                                to > PAGE_CACHE_SIZE ||
1957                                to < from);
1958                 } else {
1959                         /*
1960                          * Pages adjacent to the target (if any) imply
1961                          * a hole-filling write in which case we want
1962                          * to flush their entire range.
1963                          */
1964                         from = 0;
1965                         to = PAGE_CACHE_SIZE;
1966                 }
1967
1968                 if (page_has_buffers(tmppage)) {
1969                         if (ocfs2_should_order_data(inode))
1970                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1971                         block_commit_write(tmppage, from, to);
1972                 }
1973         }
1974
1975 out_write_size:
1976         pos += copied;
1977         if (pos > inode->i_size) {
1978                 i_size_write(inode, pos);
1979                 mark_inode_dirty(inode);
1980         }
1981         inode->i_blocks = ocfs2_inode_sector_count(inode);
1982         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1983         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1984         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1985         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1986         ocfs2_journal_dirty(handle, wc->w_di_bh);
1987
1988         ocfs2_commit_trans(osb, handle);
1989
1990         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1991
1992         ocfs2_free_write_ctxt(wc);
1993
1994         return copied;
1995 }
1996
1997 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1998                            loff_t pos, unsigned len, unsigned copied,
1999                            struct page *page, void *fsdata)
2000 {
2001         int ret;
2002         struct inode *inode = mapping->host;
2003
2004         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2005
2006         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2007         ocfs2_inode_unlock(inode, 1);
2008
2009         return ret;
2010 }
2011
2012 const struct address_space_operations ocfs2_aops = {
2013         .readpage               = ocfs2_readpage,
2014         .readpages              = ocfs2_readpages,
2015         .writepage              = ocfs2_writepage,
2016         .write_begin            = ocfs2_write_begin,
2017         .write_end              = ocfs2_write_end,
2018         .bmap                   = ocfs2_bmap,
2019         .sync_page              = block_sync_page,
2020         .direct_IO              = ocfs2_direct_IO,
2021         .invalidatepage         = ocfs2_invalidatepage,
2022         .releasepage            = ocfs2_releasepage,
2023         .migratepage            = buffer_migrate_page,
2024         .is_partially_uptodate  = block_is_partially_uptodate,
2025         .error_remove_page      = generic_error_remove_page,
2026 };