GNU Linux-libre 4.19.211-gnu1
[releases.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 #include <linux/blkdev.h>
32 #include <linux/uio.h>
33
34 #include <cluster/masklog.h>
35
36 #include "ocfs2.h"
37
38 #include "alloc.h"
39 #include "aops.h"
40 #include "dlmglue.h"
41 #include "extent_map.h"
42 #include "file.h"
43 #include "inode.h"
44 #include "journal.h"
45 #include "suballoc.h"
46 #include "super.h"
47 #include "symlink.h"
48 #include "refcounttree.h"
49 #include "ocfs2_trace.h"
50
51 #include "buffer_head_io.h"
52 #include "dir.h"
53 #include "namei.h"
54 #include "sysfile.h"
55
56 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
57                                    struct buffer_head *bh_result, int create)
58 {
59         int err = -EIO;
60         int status;
61         struct ocfs2_dinode *fe = NULL;
62         struct buffer_head *bh = NULL;
63         struct buffer_head *buffer_cache_bh = NULL;
64         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
65         void *kaddr;
66
67         trace_ocfs2_symlink_get_block(
68                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
69                         (unsigned long long)iblock, bh_result, create);
70
71         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
72
73         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
74                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
75                      (unsigned long long)iblock);
76                 goto bail;
77         }
78
79         status = ocfs2_read_inode_block(inode, &bh);
80         if (status < 0) {
81                 mlog_errno(status);
82                 goto bail;
83         }
84         fe = (struct ocfs2_dinode *) bh->b_data;
85
86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
87                                                     le32_to_cpu(fe->i_clusters))) {
88                 err = -ENOMEM;
89                 mlog(ML_ERROR, "block offset is outside the allocated size: "
90                      "%llu\n", (unsigned long long)iblock);
91                 goto bail;
92         }
93
94         /* We don't use the page cache to create symlink data, so if
95          * need be, copy it over from the buffer cache. */
96         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
97                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
98                             iblock;
99                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
100                 if (!buffer_cache_bh) {
101                         err = -ENOMEM;
102                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
103                         goto bail;
104                 }
105
106                 /* we haven't locked out transactions, so a commit
107                  * could've happened. Since we've got a reference on
108                  * the bh, even if it commits while we're doing the
109                  * copy, the data is still good. */
110                 if (buffer_jbd(buffer_cache_bh)
111                     && ocfs2_inode_is_new(inode)) {
112                         kaddr = kmap_atomic(bh_result->b_page);
113                         if (!kaddr) {
114                                 mlog(ML_ERROR, "couldn't kmap!\n");
115                                 goto bail;
116                         }
117                         memcpy(kaddr + (bh_result->b_size * iblock),
118                                buffer_cache_bh->b_data,
119                                bh_result->b_size);
120                         kunmap_atomic(kaddr);
121                         set_buffer_uptodate(bh_result);
122                 }
123                 brelse(buffer_cache_bh);
124         }
125
126         map_bh(bh_result, inode->i_sb,
127                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
128
129         err = 0;
130
131 bail:
132         brelse(bh);
133
134         return err;
135 }
136
137 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
138                     struct buffer_head *bh_result, int create)
139 {
140         int ret = 0;
141         struct ocfs2_inode_info *oi = OCFS2_I(inode);
142
143         down_read(&oi->ip_alloc_sem);
144         ret = ocfs2_get_block(inode, iblock, bh_result, create);
145         up_read(&oi->ip_alloc_sem);
146
147         return ret;
148 }
149
150 int ocfs2_get_block(struct inode *inode, sector_t iblock,
151                     struct buffer_head *bh_result, int create)
152 {
153         int err = 0;
154         unsigned int ext_flags;
155         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
156         u64 p_blkno, count, past_eof;
157         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
158
159         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
160                               (unsigned long long)iblock, bh_result, create);
161
162         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
163                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
164                      inode, inode->i_ino);
165
166         if (S_ISLNK(inode->i_mode)) {
167                 /* this always does I/O for some reason. */
168                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
169                 goto bail;
170         }
171
172         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
173                                           &ext_flags);
174         if (err) {
175                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
176                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
177                      (unsigned long long)p_blkno);
178                 goto bail;
179         }
180
181         if (max_blocks < count)
182                 count = max_blocks;
183
184         /*
185          * ocfs2 never allocates in this function - the only time we
186          * need to use BH_New is when we're extending i_size on a file
187          * system which doesn't support holes, in which case BH_New
188          * allows __block_write_begin() to zero.
189          *
190          * If we see this on a sparse file system, then a truncate has
191          * raced us and removed the cluster. In this case, we clear
192          * the buffers dirty and uptodate bits and let the buffer code
193          * ignore it as a hole.
194          */
195         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
196                 clear_buffer_dirty(bh_result);
197                 clear_buffer_uptodate(bh_result);
198                 goto bail;
199         }
200
201         /* Treat the unwritten extent as a hole for zeroing purposes. */
202         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
203                 map_bh(bh_result, inode->i_sb, p_blkno);
204
205         bh_result->b_size = count << inode->i_blkbits;
206
207         if (!ocfs2_sparse_alloc(osb)) {
208                 if (p_blkno == 0) {
209                         err = -EIO;
210                         mlog(ML_ERROR,
211                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
212                              (unsigned long long)iblock,
213                              (unsigned long long)p_blkno,
214                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
215                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
216                         dump_stack();
217                         goto bail;
218                 }
219         }
220
221         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
222
223         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
224                                   (unsigned long long)past_eof);
225         if (create && (iblock >= past_eof))
226                 set_buffer_new(bh_result);
227
228 bail:
229         if (err < 0)
230                 err = -EIO;
231
232         return err;
233 }
234
235 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
236                            struct buffer_head *di_bh)
237 {
238         void *kaddr;
239         loff_t size;
240         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
241
242         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
243                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
244                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
245                 return -EROFS;
246         }
247
248         size = i_size_read(inode);
249
250         if (size > PAGE_SIZE ||
251             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
252                 ocfs2_error(inode->i_sb,
253                             "Inode %llu has with inline data has bad size: %Lu\n",
254                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
255                             (unsigned long long)size);
256                 return -EROFS;
257         }
258
259         kaddr = kmap_atomic(page);
260         if (size)
261                 memcpy(kaddr, di->id2.i_data.id_data, size);
262         /* Clear the remaining part of the page */
263         memset(kaddr + size, 0, PAGE_SIZE - size);
264         flush_dcache_page(page);
265         kunmap_atomic(kaddr);
266
267         SetPageUptodate(page);
268
269         return 0;
270 }
271
272 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
273 {
274         int ret;
275         struct buffer_head *di_bh = NULL;
276
277         BUG_ON(!PageLocked(page));
278         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
279
280         ret = ocfs2_read_inode_block(inode, &di_bh);
281         if (ret) {
282                 mlog_errno(ret);
283                 goto out;
284         }
285
286         ret = ocfs2_read_inline_data(inode, page, di_bh);
287 out:
288         unlock_page(page);
289
290         brelse(di_bh);
291         return ret;
292 }
293
294 static int ocfs2_readpage(struct file *file, struct page *page)
295 {
296         struct inode *inode = page->mapping->host;
297         struct ocfs2_inode_info *oi = OCFS2_I(inode);
298         loff_t start = (loff_t)page->index << PAGE_SHIFT;
299         int ret, unlock = 1;
300
301         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
302                              (page ? page->index : 0));
303
304         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
305         if (ret != 0) {
306                 if (ret == AOP_TRUNCATED_PAGE)
307                         unlock = 0;
308                 mlog_errno(ret);
309                 goto out;
310         }
311
312         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
313                 /*
314                  * Unlock the page and cycle ip_alloc_sem so that we don't
315                  * busyloop waiting for ip_alloc_sem to unlock
316                  */
317                 ret = AOP_TRUNCATED_PAGE;
318                 unlock_page(page);
319                 unlock = 0;
320                 down_read(&oi->ip_alloc_sem);
321                 up_read(&oi->ip_alloc_sem);
322                 goto out_inode_unlock;
323         }
324
325         /*
326          * i_size might have just been updated as we grabed the meta lock.  We
327          * might now be discovering a truncate that hit on another node.
328          * block_read_full_page->get_block freaks out if it is asked to read
329          * beyond the end of a file, so we check here.  Callers
330          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
331          * and notice that the page they just read isn't needed.
332          *
333          * XXX sys_readahead() seems to get that wrong?
334          */
335         if (start >= i_size_read(inode)) {
336                 zero_user(page, 0, PAGE_SIZE);
337                 SetPageUptodate(page);
338                 ret = 0;
339                 goto out_alloc;
340         }
341
342         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
343                 ret = ocfs2_readpage_inline(inode, page);
344         else
345                 ret = block_read_full_page(page, ocfs2_get_block);
346         unlock = 0;
347
348 out_alloc:
349         up_read(&oi->ip_alloc_sem);
350 out_inode_unlock:
351         ocfs2_inode_unlock(inode, 0);
352 out:
353         if (unlock)
354                 unlock_page(page);
355         return ret;
356 }
357
358 /*
359  * This is used only for read-ahead. Failures or difficult to handle
360  * situations are safe to ignore.
361  *
362  * Right now, we don't bother with BH_Boundary - in-inode extent lists
363  * are quite large (243 extents on 4k blocks), so most inodes don't
364  * grow out to a tree. If need be, detecting boundary extents could
365  * trivially be added in a future version of ocfs2_get_block().
366  */
367 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
368                            struct list_head *pages, unsigned nr_pages)
369 {
370         int ret, err = -EIO;
371         struct inode *inode = mapping->host;
372         struct ocfs2_inode_info *oi = OCFS2_I(inode);
373         loff_t start;
374         struct page *last;
375
376         /*
377          * Use the nonblocking flag for the dlm code to avoid page
378          * lock inversion, but don't bother with retrying.
379          */
380         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
381         if (ret)
382                 return err;
383
384         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
385                 ocfs2_inode_unlock(inode, 0);
386                 return err;
387         }
388
389         /*
390          * Don't bother with inline-data. There isn't anything
391          * to read-ahead in that case anyway...
392          */
393         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
394                 goto out_unlock;
395
396         /*
397          * Check whether a remote node truncated this file - we just
398          * drop out in that case as it's not worth handling here.
399          */
400         last = list_entry(pages->prev, struct page, lru);
401         start = (loff_t)last->index << PAGE_SHIFT;
402         if (start >= i_size_read(inode))
403                 goto out_unlock;
404
405         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
406
407 out_unlock:
408         up_read(&oi->ip_alloc_sem);
409         ocfs2_inode_unlock(inode, 0);
410
411         return err;
412 }
413
414 /* Note: Because we don't support holes, our allocation has
415  * already happened (allocation writes zeros to the file data)
416  * so we don't have to worry about ordered writes in
417  * ocfs2_writepage.
418  *
419  * ->writepage is called during the process of invalidating the page cache
420  * during blocked lock processing.  It can't block on any cluster locks
421  * to during block mapping.  It's relying on the fact that the block
422  * mapping can't have disappeared under the dirty pages that it is
423  * being asked to write back.
424  */
425 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
426 {
427         trace_ocfs2_writepage(
428                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
429                 page->index);
430
431         return block_write_full_page(page, ocfs2_get_block, wbc);
432 }
433
434 /* Taken from ext3. We don't necessarily need the full blown
435  * functionality yet, but IMHO it's better to cut and paste the whole
436  * thing so we can avoid introducing our own bugs (and easily pick up
437  * their fixes when they happen) --Mark */
438 int walk_page_buffers(  handle_t *handle,
439                         struct buffer_head *head,
440                         unsigned from,
441                         unsigned to,
442                         int *partial,
443                         int (*fn)(      handle_t *handle,
444                                         struct buffer_head *bh))
445 {
446         struct buffer_head *bh;
447         unsigned block_start, block_end;
448         unsigned blocksize = head->b_size;
449         int err, ret = 0;
450         struct buffer_head *next;
451
452         for (   bh = head, block_start = 0;
453                 ret == 0 && (bh != head || !block_start);
454                 block_start = block_end, bh = next)
455         {
456                 next = bh->b_this_page;
457                 block_end = block_start + blocksize;
458                 if (block_end <= from || block_start >= to) {
459                         if (partial && !buffer_uptodate(bh))
460                                 *partial = 1;
461                         continue;
462                 }
463                 err = (*fn)(handle, bh);
464                 if (!ret)
465                         ret = err;
466         }
467         return ret;
468 }
469
470 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
471 {
472         sector_t status;
473         u64 p_blkno = 0;
474         int err = 0;
475         struct inode *inode = mapping->host;
476
477         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
478                          (unsigned long long)block);
479
480         /*
481          * The swap code (ab-)uses ->bmap to get a block mapping and then
482          * bypasseÑ• the file system for actual I/O.  We really can't allow
483          * that on refcounted inodes, so we have to skip out here.  And yes,
484          * 0 is the magic code for a bmap error..
485          */
486         if (ocfs2_is_refcount_inode(inode))
487                 return 0;
488
489         /* We don't need to lock journal system files, since they aren't
490          * accessed concurrently from multiple nodes.
491          */
492         if (!INODE_JOURNAL(inode)) {
493                 err = ocfs2_inode_lock(inode, NULL, 0);
494                 if (err) {
495                         if (err != -ENOENT)
496                                 mlog_errno(err);
497                         goto bail;
498                 }
499                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
500         }
501
502         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
503                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
504                                                   NULL);
505
506         if (!INODE_JOURNAL(inode)) {
507                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
508                 ocfs2_inode_unlock(inode, 0);
509         }
510
511         if (err) {
512                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
513                      (unsigned long long)block);
514                 mlog_errno(err);
515                 goto bail;
516         }
517
518 bail:
519         status = err ? 0 : p_blkno;
520
521         return status;
522 }
523
524 static int ocfs2_releasepage(struct page *page, gfp_t wait)
525 {
526         if (!page_has_buffers(page))
527                 return 0;
528         return try_to_free_buffers(page);
529 }
530
531 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
532                                             u32 cpos,
533                                             unsigned int *start,
534                                             unsigned int *end)
535 {
536         unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
537
538         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
539                 unsigned int cpp;
540
541                 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
542
543                 cluster_start = cpos % cpp;
544                 cluster_start = cluster_start << osb->s_clustersize_bits;
545
546                 cluster_end = cluster_start + osb->s_clustersize;
547         }
548
549         BUG_ON(cluster_start > PAGE_SIZE);
550         BUG_ON(cluster_end > PAGE_SIZE);
551
552         if (start)
553                 *start = cluster_start;
554         if (end)
555                 *end = cluster_end;
556 }
557
558 /*
559  * 'from' and 'to' are the region in the page to avoid zeroing.
560  *
561  * If pagesize > clustersize, this function will avoid zeroing outside
562  * of the cluster boundary.
563  *
564  * from == to == 0 is code for "zero the entire cluster region"
565  */
566 static void ocfs2_clear_page_regions(struct page *page,
567                                      struct ocfs2_super *osb, u32 cpos,
568                                      unsigned from, unsigned to)
569 {
570         void *kaddr;
571         unsigned int cluster_start, cluster_end;
572
573         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
574
575         kaddr = kmap_atomic(page);
576
577         if (from || to) {
578                 if (from > cluster_start)
579                         memset(kaddr + cluster_start, 0, from - cluster_start);
580                 if (to < cluster_end)
581                         memset(kaddr + to, 0, cluster_end - to);
582         } else {
583                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
584         }
585
586         kunmap_atomic(kaddr);
587 }
588
589 /*
590  * Nonsparse file systems fully allocate before we get to the write
591  * code. This prevents ocfs2_write() from tagging the write as an
592  * allocating one, which means ocfs2_map_page_blocks() might try to
593  * read-in the blocks at the tail of our file. Avoid reading them by
594  * testing i_size against each block offset.
595  */
596 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
597                                  unsigned int block_start)
598 {
599         u64 offset = page_offset(page) + block_start;
600
601         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
602                 return 1;
603
604         if (i_size_read(inode) > offset)
605                 return 1;
606
607         return 0;
608 }
609
610 /*
611  * Some of this taken from __block_write_begin(). We already have our
612  * mapping by now though, and the entire write will be allocating or
613  * it won't, so not much need to use BH_New.
614  *
615  * This will also skip zeroing, which is handled externally.
616  */
617 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
618                           struct inode *inode, unsigned int from,
619                           unsigned int to, int new)
620 {
621         int ret = 0;
622         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
623         unsigned int block_end, block_start;
624         unsigned int bsize = i_blocksize(inode);
625
626         if (!page_has_buffers(page))
627                 create_empty_buffers(page, bsize, 0);
628
629         head = page_buffers(page);
630         for (bh = head, block_start = 0; bh != head || !block_start;
631              bh = bh->b_this_page, block_start += bsize) {
632                 block_end = block_start + bsize;
633
634                 clear_buffer_new(bh);
635
636                 /*
637                  * Ignore blocks outside of our i/o range -
638                  * they may belong to unallocated clusters.
639                  */
640                 if (block_start >= to || block_end <= from) {
641                         if (PageUptodate(page))
642                                 set_buffer_uptodate(bh);
643                         continue;
644                 }
645
646                 /*
647                  * For an allocating write with cluster size >= page
648                  * size, we always write the entire page.
649                  */
650                 if (new)
651                         set_buffer_new(bh);
652
653                 if (!buffer_mapped(bh)) {
654                         map_bh(bh, inode->i_sb, *p_blkno);
655                         clean_bdev_bh_alias(bh);
656                 }
657
658                 if (PageUptodate(page)) {
659                         if (!buffer_uptodate(bh))
660                                 set_buffer_uptodate(bh);
661                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
662                            !buffer_new(bh) &&
663                            ocfs2_should_read_blk(inode, page, block_start) &&
664                            (block_start < from || block_end > to)) {
665                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
666                         *wait_bh++=bh;
667                 }
668
669                 *p_blkno = *p_blkno + 1;
670         }
671
672         /*
673          * If we issued read requests - let them complete.
674          */
675         while(wait_bh > wait) {
676                 wait_on_buffer(*--wait_bh);
677                 if (!buffer_uptodate(*wait_bh))
678                         ret = -EIO;
679         }
680
681         if (ret == 0 || !new)
682                 return ret;
683
684         /*
685          * If we get -EIO above, zero out any newly allocated blocks
686          * to avoid exposing stale data.
687          */
688         bh = head;
689         block_start = 0;
690         do {
691                 block_end = block_start + bsize;
692                 if (block_end <= from)
693                         goto next_bh;
694                 if (block_start >= to)
695                         break;
696
697                 zero_user(page, block_start, bh->b_size);
698                 set_buffer_uptodate(bh);
699                 mark_buffer_dirty(bh);
700
701 next_bh:
702                 block_start = block_end;
703                 bh = bh->b_this_page;
704         } while (bh != head);
705
706         return ret;
707 }
708
709 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
710 #define OCFS2_MAX_CTXT_PAGES    1
711 #else
712 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
713 #endif
714
715 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
716
717 struct ocfs2_unwritten_extent {
718         struct list_head        ue_node;
719         struct list_head        ue_ip_node;
720         u32                     ue_cpos;
721         u32                     ue_phys;
722 };
723
724 /*
725  * Describe the state of a single cluster to be written to.
726  */
727 struct ocfs2_write_cluster_desc {
728         u32             c_cpos;
729         u32             c_phys;
730         /*
731          * Give this a unique field because c_phys eventually gets
732          * filled.
733          */
734         unsigned        c_new;
735         unsigned        c_clear_unwritten;
736         unsigned        c_needs_zero;
737 };
738
739 struct ocfs2_write_ctxt {
740         /* Logical cluster position / len of write */
741         u32                             w_cpos;
742         u32                             w_clen;
743
744         /* First cluster allocated in a nonsparse extend */
745         u32                             w_first_new_cpos;
746
747         /* Type of caller. Must be one of buffer, mmap, direct.  */
748         ocfs2_write_type_t              w_type;
749
750         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
751
752         /*
753          * This is true if page_size > cluster_size.
754          *
755          * It triggers a set of special cases during write which might
756          * have to deal with allocating writes to partial pages.
757          */
758         unsigned int                    w_large_pages;
759
760         /*
761          * Pages involved in this write.
762          *
763          * w_target_page is the page being written to by the user.
764          *
765          * w_pages is an array of pages which always contains
766          * w_target_page, and in the case of an allocating write with
767          * page_size < cluster size, it will contain zero'd and mapped
768          * pages adjacent to w_target_page which need to be written
769          * out in so that future reads from that region will get
770          * zero's.
771          */
772         unsigned int                    w_num_pages;
773         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
774         struct page                     *w_target_page;
775
776         /*
777          * w_target_locked is used for page_mkwrite path indicating no unlocking
778          * against w_target_page in ocfs2_write_end_nolock.
779          */
780         unsigned int                    w_target_locked:1;
781
782         /*
783          * ocfs2_write_end() uses this to know what the real range to
784          * write in the target should be.
785          */
786         unsigned int                    w_target_from;
787         unsigned int                    w_target_to;
788
789         /*
790          * We could use journal_current_handle() but this is cleaner,
791          * IMHO -Mark
792          */
793         handle_t                        *w_handle;
794
795         struct buffer_head              *w_di_bh;
796
797         struct ocfs2_cached_dealloc_ctxt w_dealloc;
798
799         struct list_head                w_unwritten_list;
800         unsigned int                    w_unwritten_count;
801 };
802
803 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
804 {
805         int i;
806
807         for(i = 0; i < num_pages; i++) {
808                 if (pages[i]) {
809                         unlock_page(pages[i]);
810                         mark_page_accessed(pages[i]);
811                         put_page(pages[i]);
812                 }
813         }
814 }
815
816 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
817 {
818         int i;
819
820         /*
821          * w_target_locked is only set to true in the page_mkwrite() case.
822          * The intent is to allow us to lock the target page from write_begin()
823          * to write_end(). The caller must hold a ref on w_target_page.
824          */
825         if (wc->w_target_locked) {
826                 BUG_ON(!wc->w_target_page);
827                 for (i = 0; i < wc->w_num_pages; i++) {
828                         if (wc->w_target_page == wc->w_pages[i]) {
829                                 wc->w_pages[i] = NULL;
830                                 break;
831                         }
832                 }
833                 mark_page_accessed(wc->w_target_page);
834                 put_page(wc->w_target_page);
835         }
836         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
837 }
838
839 static void ocfs2_free_unwritten_list(struct inode *inode,
840                                  struct list_head *head)
841 {
842         struct ocfs2_inode_info *oi = OCFS2_I(inode);
843         struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
844
845         list_for_each_entry_safe(ue, tmp, head, ue_node) {
846                 list_del(&ue->ue_node);
847                 spin_lock(&oi->ip_lock);
848                 list_del(&ue->ue_ip_node);
849                 spin_unlock(&oi->ip_lock);
850                 kfree(ue);
851         }
852 }
853
854 static void ocfs2_free_write_ctxt(struct inode *inode,
855                                   struct ocfs2_write_ctxt *wc)
856 {
857         ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
858         ocfs2_unlock_pages(wc);
859         brelse(wc->w_di_bh);
860         kfree(wc);
861 }
862
863 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
864                                   struct ocfs2_super *osb, loff_t pos,
865                                   unsigned len, ocfs2_write_type_t type,
866                                   struct buffer_head *di_bh)
867 {
868         u32 cend;
869         struct ocfs2_write_ctxt *wc;
870
871         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
872         if (!wc)
873                 return -ENOMEM;
874
875         wc->w_cpos = pos >> osb->s_clustersize_bits;
876         wc->w_first_new_cpos = UINT_MAX;
877         cend = (pos + len - 1) >> osb->s_clustersize_bits;
878         wc->w_clen = cend - wc->w_cpos + 1;
879         get_bh(di_bh);
880         wc->w_di_bh = di_bh;
881         wc->w_type = type;
882
883         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
884                 wc->w_large_pages = 1;
885         else
886                 wc->w_large_pages = 0;
887
888         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
889         INIT_LIST_HEAD(&wc->w_unwritten_list);
890
891         *wcp = wc;
892
893         return 0;
894 }
895
896 /*
897  * If a page has any new buffers, zero them out here, and mark them uptodate
898  * and dirty so they'll be written out (in order to prevent uninitialised
899  * block data from leaking). And clear the new bit.
900  */
901 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
902 {
903         unsigned int block_start, block_end;
904         struct buffer_head *head, *bh;
905
906         BUG_ON(!PageLocked(page));
907         if (!page_has_buffers(page))
908                 return;
909
910         bh = head = page_buffers(page);
911         block_start = 0;
912         do {
913                 block_end = block_start + bh->b_size;
914
915                 if (buffer_new(bh)) {
916                         if (block_end > from && block_start < to) {
917                                 if (!PageUptodate(page)) {
918                                         unsigned start, end;
919
920                                         start = max(from, block_start);
921                                         end = min(to, block_end);
922
923                                         zero_user_segment(page, start, end);
924                                         set_buffer_uptodate(bh);
925                                 }
926
927                                 clear_buffer_new(bh);
928                                 mark_buffer_dirty(bh);
929                         }
930                 }
931
932                 block_start = block_end;
933                 bh = bh->b_this_page;
934         } while (bh != head);
935 }
936
937 /*
938  * Only called when we have a failure during allocating write to write
939  * zero's to the newly allocated region.
940  */
941 static void ocfs2_write_failure(struct inode *inode,
942                                 struct ocfs2_write_ctxt *wc,
943                                 loff_t user_pos, unsigned user_len)
944 {
945         int i;
946         unsigned from = user_pos & (PAGE_SIZE - 1),
947                 to = user_pos + user_len;
948         struct page *tmppage;
949
950         if (wc->w_target_page)
951                 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
952
953         for(i = 0; i < wc->w_num_pages; i++) {
954                 tmppage = wc->w_pages[i];
955
956                 if (tmppage && page_has_buffers(tmppage)) {
957                         if (ocfs2_should_order_data(inode))
958                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
959
960                         block_commit_write(tmppage, from, to);
961                 }
962         }
963 }
964
965 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
966                                         struct ocfs2_write_ctxt *wc,
967                                         struct page *page, u32 cpos,
968                                         loff_t user_pos, unsigned user_len,
969                                         int new)
970 {
971         int ret;
972         unsigned int map_from = 0, map_to = 0;
973         unsigned int cluster_start, cluster_end;
974         unsigned int user_data_from = 0, user_data_to = 0;
975
976         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
977                                         &cluster_start, &cluster_end);
978
979         /* treat the write as new if the a hole/lseek spanned across
980          * the page boundary.
981          */
982         new = new | ((i_size_read(inode) <= page_offset(page)) &&
983                         (page_offset(page) <= user_pos));
984
985         if (page == wc->w_target_page) {
986                 map_from = user_pos & (PAGE_SIZE - 1);
987                 map_to = map_from + user_len;
988
989                 if (new)
990                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
991                                                     cluster_start, cluster_end,
992                                                     new);
993                 else
994                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
995                                                     map_from, map_to, new);
996                 if (ret) {
997                         mlog_errno(ret);
998                         goto out;
999                 }
1000
1001                 user_data_from = map_from;
1002                 user_data_to = map_to;
1003                 if (new) {
1004                         map_from = cluster_start;
1005                         map_to = cluster_end;
1006                 }
1007         } else {
1008                 /*
1009                  * If we haven't allocated the new page yet, we
1010                  * shouldn't be writing it out without copying user
1011                  * data. This is likely a math error from the caller.
1012                  */
1013                 BUG_ON(!new);
1014
1015                 map_from = cluster_start;
1016                 map_to = cluster_end;
1017
1018                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1019                                             cluster_start, cluster_end, new);
1020                 if (ret) {
1021                         mlog_errno(ret);
1022                         goto out;
1023                 }
1024         }
1025
1026         /*
1027          * Parts of newly allocated pages need to be zero'd.
1028          *
1029          * Above, we have also rewritten 'to' and 'from' - as far as
1030          * the rest of the function is concerned, the entire cluster
1031          * range inside of a page needs to be written.
1032          *
1033          * We can skip this if the page is up to date - it's already
1034          * been zero'd from being read in as a hole.
1035          */
1036         if (new && !PageUptodate(page))
1037                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1038                                          cpos, user_data_from, user_data_to);
1039
1040         flush_dcache_page(page);
1041
1042 out:
1043         return ret;
1044 }
1045
1046 /*
1047  * This function will only grab one clusters worth of pages.
1048  */
1049 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1050                                       struct ocfs2_write_ctxt *wc,
1051                                       u32 cpos, loff_t user_pos,
1052                                       unsigned user_len, int new,
1053                                       struct page *mmap_page)
1054 {
1055         int ret = 0, i;
1056         unsigned long start, target_index, end_index, index;
1057         struct inode *inode = mapping->host;
1058         loff_t last_byte;
1059
1060         target_index = user_pos >> PAGE_SHIFT;
1061
1062         /*
1063          * Figure out how many pages we'll be manipulating here. For
1064          * non allocating write, we just change the one
1065          * page. Otherwise, we'll need a whole clusters worth.  If we're
1066          * writing past i_size, we only need enough pages to cover the
1067          * last page of the write.
1068          */
1069         if (new) {
1070                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1071                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1072                 /*
1073                  * We need the index *past* the last page we could possibly
1074                  * touch.  This is the page past the end of the write or
1075                  * i_size, whichever is greater.
1076                  */
1077                 last_byte = max(user_pos + user_len, i_size_read(inode));
1078                 BUG_ON(last_byte < 1);
1079                 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1080                 if ((start + wc->w_num_pages) > end_index)
1081                         wc->w_num_pages = end_index - start;
1082         } else {
1083                 wc->w_num_pages = 1;
1084                 start = target_index;
1085         }
1086         end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1087
1088         for(i = 0; i < wc->w_num_pages; i++) {
1089                 index = start + i;
1090
1091                 if (index >= target_index && index <= end_index &&
1092                     wc->w_type == OCFS2_WRITE_MMAP) {
1093                         /*
1094                          * ocfs2_pagemkwrite() is a little different
1095                          * and wants us to directly use the page
1096                          * passed in.
1097                          */
1098                         lock_page(mmap_page);
1099
1100                         /* Exit and let the caller retry */
1101                         if (mmap_page->mapping != mapping) {
1102                                 WARN_ON(mmap_page->mapping);
1103                                 unlock_page(mmap_page);
1104                                 ret = -EAGAIN;
1105                                 goto out;
1106                         }
1107
1108                         get_page(mmap_page);
1109                         wc->w_pages[i] = mmap_page;
1110                         wc->w_target_locked = true;
1111                 } else if (index >= target_index && index <= end_index &&
1112                            wc->w_type == OCFS2_WRITE_DIRECT) {
1113                         /* Direct write has no mapping page. */
1114                         wc->w_pages[i] = NULL;
1115                         continue;
1116                 } else {
1117                         wc->w_pages[i] = find_or_create_page(mapping, index,
1118                                                              GFP_NOFS);
1119                         if (!wc->w_pages[i]) {
1120                                 ret = -ENOMEM;
1121                                 mlog_errno(ret);
1122                                 goto out;
1123                         }
1124                 }
1125                 wait_for_stable_page(wc->w_pages[i]);
1126
1127                 if (index == target_index)
1128                         wc->w_target_page = wc->w_pages[i];
1129         }
1130 out:
1131         if (ret)
1132                 wc->w_target_locked = false;
1133         return ret;
1134 }
1135
1136 /*
1137  * Prepare a single cluster for write one cluster into the file.
1138  */
1139 static int ocfs2_write_cluster(struct address_space *mapping,
1140                                u32 *phys, unsigned int new,
1141                                unsigned int clear_unwritten,
1142                                unsigned int should_zero,
1143                                struct ocfs2_alloc_context *data_ac,
1144                                struct ocfs2_alloc_context *meta_ac,
1145                                struct ocfs2_write_ctxt *wc, u32 cpos,
1146                                loff_t user_pos, unsigned user_len)
1147 {
1148         int ret, i;
1149         u64 p_blkno;
1150         struct inode *inode = mapping->host;
1151         struct ocfs2_extent_tree et;
1152         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1153
1154         if (new) {
1155                 u32 tmp_pos;
1156
1157                 /*
1158                  * This is safe to call with the page locks - it won't take
1159                  * any additional semaphores or cluster locks.
1160                  */
1161                 tmp_pos = cpos;
1162                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1163                                            &tmp_pos, 1, !clear_unwritten,
1164                                            wc->w_di_bh, wc->w_handle,
1165                                            data_ac, meta_ac, NULL);
1166                 /*
1167                  * This shouldn't happen because we must have already
1168                  * calculated the correct meta data allocation required. The
1169                  * internal tree allocation code should know how to increase
1170                  * transaction credits itself.
1171                  *
1172                  * If need be, we could handle -EAGAIN for a
1173                  * RESTART_TRANS here.
1174                  */
1175                 mlog_bug_on_msg(ret == -EAGAIN,
1176                                 "Inode %llu: EAGAIN return during allocation.\n",
1177                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1178                 if (ret < 0) {
1179                         mlog_errno(ret);
1180                         goto out;
1181                 }
1182         } else if (clear_unwritten) {
1183                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1184                                               wc->w_di_bh);
1185                 ret = ocfs2_mark_extent_written(inode, &et,
1186                                                 wc->w_handle, cpos, 1, *phys,
1187                                                 meta_ac, &wc->w_dealloc);
1188                 if (ret < 0) {
1189                         mlog_errno(ret);
1190                         goto out;
1191                 }
1192         }
1193
1194         /*
1195          * The only reason this should fail is due to an inability to
1196          * find the extent added.
1197          */
1198         ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1199         if (ret < 0) {
1200                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1201                             "at logical cluster %u",
1202                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1203                 goto out;
1204         }
1205
1206         BUG_ON(*phys == 0);
1207
1208         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1209         if (!should_zero)
1210                 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1211
1212         for(i = 0; i < wc->w_num_pages; i++) {
1213                 int tmpret;
1214
1215                 /* This is the direct io target page. */
1216                 if (wc->w_pages[i] == NULL) {
1217                         p_blkno++;
1218                         continue;
1219                 }
1220
1221                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1222                                                       wc->w_pages[i], cpos,
1223                                                       user_pos, user_len,
1224                                                       should_zero);
1225                 if (tmpret) {
1226                         mlog_errno(tmpret);
1227                         if (ret == 0)
1228                                 ret = tmpret;
1229                 }
1230         }
1231
1232         /*
1233          * We only have cleanup to do in case of allocating write.
1234          */
1235         if (ret && new)
1236                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1237
1238 out:
1239
1240         return ret;
1241 }
1242
1243 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1244                                        struct ocfs2_alloc_context *data_ac,
1245                                        struct ocfs2_alloc_context *meta_ac,
1246                                        struct ocfs2_write_ctxt *wc,
1247                                        loff_t pos, unsigned len)
1248 {
1249         int ret, i;
1250         loff_t cluster_off;
1251         unsigned int local_len = len;
1252         struct ocfs2_write_cluster_desc *desc;
1253         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1254
1255         for (i = 0; i < wc->w_clen; i++) {
1256                 desc = &wc->w_desc[i];
1257
1258                 /*
1259                  * We have to make sure that the total write passed in
1260                  * doesn't extend past a single cluster.
1261                  */
1262                 local_len = len;
1263                 cluster_off = pos & (osb->s_clustersize - 1);
1264                 if ((cluster_off + local_len) > osb->s_clustersize)
1265                         local_len = osb->s_clustersize - cluster_off;
1266
1267                 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1268                                           desc->c_new,
1269                                           desc->c_clear_unwritten,
1270                                           desc->c_needs_zero,
1271                                           data_ac, meta_ac,
1272                                           wc, desc->c_cpos, pos, local_len);
1273                 if (ret) {
1274                         mlog_errno(ret);
1275                         goto out;
1276                 }
1277
1278                 len -= local_len;
1279                 pos += local_len;
1280         }
1281
1282         ret = 0;
1283 out:
1284         return ret;
1285 }
1286
1287 /*
1288  * ocfs2_write_end() wants to know which parts of the target page it
1289  * should complete the write on. It's easiest to compute them ahead of
1290  * time when a more complete view of the write is available.
1291  */
1292 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1293                                         struct ocfs2_write_ctxt *wc,
1294                                         loff_t pos, unsigned len, int alloc)
1295 {
1296         struct ocfs2_write_cluster_desc *desc;
1297
1298         wc->w_target_from = pos & (PAGE_SIZE - 1);
1299         wc->w_target_to = wc->w_target_from + len;
1300
1301         if (alloc == 0)
1302                 return;
1303
1304         /*
1305          * Allocating write - we may have different boundaries based
1306          * on page size and cluster size.
1307          *
1308          * NOTE: We can no longer compute one value from the other as
1309          * the actual write length and user provided length may be
1310          * different.
1311          */
1312
1313         if (wc->w_large_pages) {
1314                 /*
1315                  * We only care about the 1st and last cluster within
1316                  * our range and whether they should be zero'd or not. Either
1317                  * value may be extended out to the start/end of a
1318                  * newly allocated cluster.
1319                  */
1320                 desc = &wc->w_desc[0];
1321                 if (desc->c_needs_zero)
1322                         ocfs2_figure_cluster_boundaries(osb,
1323                                                         desc->c_cpos,
1324                                                         &wc->w_target_from,
1325                                                         NULL);
1326
1327                 desc = &wc->w_desc[wc->w_clen - 1];
1328                 if (desc->c_needs_zero)
1329                         ocfs2_figure_cluster_boundaries(osb,
1330                                                         desc->c_cpos,
1331                                                         NULL,
1332                                                         &wc->w_target_to);
1333         } else {
1334                 wc->w_target_from = 0;
1335                 wc->w_target_to = PAGE_SIZE;
1336         }
1337 }
1338
1339 /*
1340  * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1341  * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1342  * by the direct io procedure.
1343  * If this is a new extent that allocated by direct io, we should mark it in
1344  * the ip_unwritten_list.
1345  */
1346 static int ocfs2_unwritten_check(struct inode *inode,
1347                                  struct ocfs2_write_ctxt *wc,
1348                                  struct ocfs2_write_cluster_desc *desc)
1349 {
1350         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1351         struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1352         int ret = 0;
1353
1354         if (!desc->c_needs_zero)
1355                 return 0;
1356
1357 retry:
1358         spin_lock(&oi->ip_lock);
1359         /* Needs not to zero no metter buffer or direct. The one who is zero
1360          * the cluster is doing zero. And he will clear unwritten after all
1361          * cluster io finished. */
1362         list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1363                 if (desc->c_cpos == ue->ue_cpos) {
1364                         BUG_ON(desc->c_new);
1365                         desc->c_needs_zero = 0;
1366                         desc->c_clear_unwritten = 0;
1367                         goto unlock;
1368                 }
1369         }
1370
1371         if (wc->w_type != OCFS2_WRITE_DIRECT)
1372                 goto unlock;
1373
1374         if (new == NULL) {
1375                 spin_unlock(&oi->ip_lock);
1376                 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1377                              GFP_NOFS);
1378                 if (new == NULL) {
1379                         ret = -ENOMEM;
1380                         goto out;
1381                 }
1382                 goto retry;
1383         }
1384         /* This direct write will doing zero. */
1385         new->ue_cpos = desc->c_cpos;
1386         new->ue_phys = desc->c_phys;
1387         desc->c_clear_unwritten = 0;
1388         list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1389         list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1390         wc->w_unwritten_count++;
1391         new = NULL;
1392 unlock:
1393         spin_unlock(&oi->ip_lock);
1394 out:
1395         if (new)
1396                 kfree(new);
1397         return ret;
1398 }
1399
1400 /*
1401  * Populate each single-cluster write descriptor in the write context
1402  * with information about the i/o to be done.
1403  *
1404  * Returns the number of clusters that will have to be allocated, as
1405  * well as a worst case estimate of the number of extent records that
1406  * would have to be created during a write to an unwritten region.
1407  */
1408 static int ocfs2_populate_write_desc(struct inode *inode,
1409                                      struct ocfs2_write_ctxt *wc,
1410                                      unsigned int *clusters_to_alloc,
1411                                      unsigned int *extents_to_split)
1412 {
1413         int ret;
1414         struct ocfs2_write_cluster_desc *desc;
1415         unsigned int num_clusters = 0;
1416         unsigned int ext_flags = 0;
1417         u32 phys = 0;
1418         int i;
1419
1420         *clusters_to_alloc = 0;
1421         *extents_to_split = 0;
1422
1423         for (i = 0; i < wc->w_clen; i++) {
1424                 desc = &wc->w_desc[i];
1425                 desc->c_cpos = wc->w_cpos + i;
1426
1427                 if (num_clusters == 0) {
1428                         /*
1429                          * Need to look up the next extent record.
1430                          */
1431                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1432                                                  &num_clusters, &ext_flags);
1433                         if (ret) {
1434                                 mlog_errno(ret);
1435                                 goto out;
1436                         }
1437
1438                         /* We should already CoW the refcountd extent. */
1439                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1440
1441                         /*
1442                          * Assume worst case - that we're writing in
1443                          * the middle of the extent.
1444                          *
1445                          * We can assume that the write proceeds from
1446                          * left to right, in which case the extent
1447                          * insert code is smart enough to coalesce the
1448                          * next splits into the previous records created.
1449                          */
1450                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1451                                 *extents_to_split = *extents_to_split + 2;
1452                 } else if (phys) {
1453                         /*
1454                          * Only increment phys if it doesn't describe
1455                          * a hole.
1456                          */
1457                         phys++;
1458                 }
1459
1460                 /*
1461                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1462                  * file that got extended.  w_first_new_cpos tells us
1463                  * where the newly allocated clusters are so we can
1464                  * zero them.
1465                  */
1466                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1467                         BUG_ON(phys == 0);
1468                         desc->c_needs_zero = 1;
1469                 }
1470
1471                 desc->c_phys = phys;
1472                 if (phys == 0) {
1473                         desc->c_new = 1;
1474                         desc->c_needs_zero = 1;
1475                         desc->c_clear_unwritten = 1;
1476                         *clusters_to_alloc = *clusters_to_alloc + 1;
1477                 }
1478
1479                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1480                         desc->c_clear_unwritten = 1;
1481                         desc->c_needs_zero = 1;
1482                 }
1483
1484                 ret = ocfs2_unwritten_check(inode, wc, desc);
1485                 if (ret) {
1486                         mlog_errno(ret);
1487                         goto out;
1488                 }
1489
1490                 num_clusters--;
1491         }
1492
1493         ret = 0;
1494 out:
1495         return ret;
1496 }
1497
1498 static int ocfs2_write_begin_inline(struct address_space *mapping,
1499                                     struct inode *inode,
1500                                     struct ocfs2_write_ctxt *wc)
1501 {
1502         int ret;
1503         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1504         struct page *page;
1505         handle_t *handle;
1506         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1507
1508         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1509         if (IS_ERR(handle)) {
1510                 ret = PTR_ERR(handle);
1511                 mlog_errno(ret);
1512                 goto out;
1513         }
1514
1515         page = find_or_create_page(mapping, 0, GFP_NOFS);
1516         if (!page) {
1517                 ocfs2_commit_trans(osb, handle);
1518                 ret = -ENOMEM;
1519                 mlog_errno(ret);
1520                 goto out;
1521         }
1522         /*
1523          * If we don't set w_num_pages then this page won't get unlocked
1524          * and freed on cleanup of the write context.
1525          */
1526         wc->w_pages[0] = wc->w_target_page = page;
1527         wc->w_num_pages = 1;
1528
1529         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1530                                       OCFS2_JOURNAL_ACCESS_WRITE);
1531         if (ret) {
1532                 ocfs2_commit_trans(osb, handle);
1533
1534                 mlog_errno(ret);
1535                 goto out;
1536         }
1537
1538         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1539                 ocfs2_set_inode_data_inline(inode, di);
1540
1541         if (!PageUptodate(page)) {
1542                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1543                 if (ret) {
1544                         ocfs2_commit_trans(osb, handle);
1545
1546                         goto out;
1547                 }
1548         }
1549
1550         wc->w_handle = handle;
1551 out:
1552         return ret;
1553 }
1554
1555 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1556 {
1557         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1558
1559         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1560                 return 1;
1561         return 0;
1562 }
1563
1564 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1565                                           struct inode *inode, loff_t pos,
1566                                           unsigned len, struct page *mmap_page,
1567                                           struct ocfs2_write_ctxt *wc)
1568 {
1569         int ret, written = 0;
1570         loff_t end = pos + len;
1571         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1572         struct ocfs2_dinode *di = NULL;
1573
1574         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1575                                              len, (unsigned long long)pos,
1576                                              oi->ip_dyn_features);
1577
1578         /*
1579          * Handle inodes which already have inline data 1st.
1580          */
1581         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1582                 if (mmap_page == NULL &&
1583                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1584                         goto do_inline_write;
1585
1586                 /*
1587                  * The write won't fit - we have to give this inode an
1588                  * inline extent list now.
1589                  */
1590                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1591                 if (ret)
1592                         mlog_errno(ret);
1593                 goto out;
1594         }
1595
1596         /*
1597          * Check whether the inode can accept inline data.
1598          */
1599         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1600                 return 0;
1601
1602         /*
1603          * Check whether the write can fit.
1604          */
1605         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1606         if (mmap_page ||
1607             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1608                 return 0;
1609
1610 do_inline_write:
1611         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1612         if (ret) {
1613                 mlog_errno(ret);
1614                 goto out;
1615         }
1616
1617         /*
1618          * This signals to the caller that the data can be written
1619          * inline.
1620          */
1621         written = 1;
1622 out:
1623         return written ? written : ret;
1624 }
1625
1626 /*
1627  * This function only does anything for file systems which can't
1628  * handle sparse files.
1629  *
1630  * What we want to do here is fill in any hole between the current end
1631  * of allocation and the end of our write. That way the rest of the
1632  * write path can treat it as an non-allocating write, which has no
1633  * special case code for sparse/nonsparse files.
1634  */
1635 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1636                                         struct buffer_head *di_bh,
1637                                         loff_t pos, unsigned len,
1638                                         struct ocfs2_write_ctxt *wc)
1639 {
1640         int ret;
1641         loff_t newsize = pos + len;
1642
1643         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1644
1645         if (newsize <= i_size_read(inode))
1646                 return 0;
1647
1648         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1649         if (ret)
1650                 mlog_errno(ret);
1651
1652         /* There is no wc if this is call from direct. */
1653         if (wc)
1654                 wc->w_first_new_cpos =
1655                         ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1656
1657         return ret;
1658 }
1659
1660 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1661                            loff_t pos)
1662 {
1663         int ret = 0;
1664
1665         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1666         if (pos > i_size_read(inode))
1667                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1668
1669         return ret;
1670 }
1671
1672 int ocfs2_write_begin_nolock(struct address_space *mapping,
1673                              loff_t pos, unsigned len, ocfs2_write_type_t type,
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, clusters_need = 0;
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         int try_free = 1, ret1;
1688
1689 try_again:
1690         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1691         if (ret) {
1692                 mlog_errno(ret);
1693                 return ret;
1694         }
1695
1696         if (ocfs2_supports_inline_data(osb)) {
1697                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1698                                                      mmap_page, wc);
1699                 if (ret == 1) {
1700                         ret = 0;
1701                         goto success;
1702                 }
1703                 if (ret < 0) {
1704                         mlog_errno(ret);
1705                         goto out;
1706                 }
1707         }
1708
1709         /* Direct io change i_size late, should not zero tail here. */
1710         if (type != OCFS2_WRITE_DIRECT) {
1711                 if (ocfs2_sparse_alloc(osb))
1712                         ret = ocfs2_zero_tail(inode, di_bh, pos);
1713                 else
1714                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1715                                                            len, wc);
1716                 if (ret) {
1717                         mlog_errno(ret);
1718                         goto out;
1719                 }
1720         }
1721
1722         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1723         if (ret < 0) {
1724                 mlog_errno(ret);
1725                 goto out;
1726         } else if (ret == 1) {
1727                 clusters_need = wc->w_clen;
1728                 ret = ocfs2_refcount_cow(inode, di_bh,
1729                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1730                 if (ret) {
1731                         mlog_errno(ret);
1732                         goto out;
1733                 }
1734         }
1735
1736         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1737                                         &extents_to_split);
1738         if (ret) {
1739                 mlog_errno(ret);
1740                 goto out;
1741         }
1742         clusters_need += clusters_to_alloc;
1743
1744         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1745
1746         trace_ocfs2_write_begin_nolock(
1747                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1748                         (long long)i_size_read(inode),
1749                         le32_to_cpu(di->i_clusters),
1750                         pos, len, type, mmap_page,
1751                         clusters_to_alloc, extents_to_split);
1752
1753         /*
1754          * We set w_target_from, w_target_to here so that
1755          * ocfs2_write_end() knows which range in the target page to
1756          * write out. An allocation requires that we write the entire
1757          * cluster range.
1758          */
1759         if (clusters_to_alloc || extents_to_split) {
1760                 /*
1761                  * XXX: We are stretching the limits of
1762                  * ocfs2_lock_allocators(). It greatly over-estimates
1763                  * the work to be done.
1764                  */
1765                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1766                                               wc->w_di_bh);
1767                 ret = ocfs2_lock_allocators(inode, &et,
1768                                             clusters_to_alloc, extents_to_split,
1769                                             &data_ac, &meta_ac);
1770                 if (ret) {
1771                         mlog_errno(ret);
1772                         goto out;
1773                 }
1774
1775                 if (data_ac)
1776                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1777
1778                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1779                                                     &di->id2.i_list);
1780         } else if (type == OCFS2_WRITE_DIRECT)
1781                 /* direct write needs not to start trans if no extents alloc. */
1782                 goto success;
1783
1784         /*
1785          * We have to zero sparse allocated clusters, unwritten extent clusters,
1786          * and non-sparse clusters we just extended.  For non-sparse writes,
1787          * we know zeros will only be needed in the first and/or last cluster.
1788          */
1789         if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1790                            wc->w_desc[wc->w_clen - 1].c_needs_zero))
1791                 cluster_of_pages = 1;
1792         else
1793                 cluster_of_pages = 0;
1794
1795         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1796
1797         handle = ocfs2_start_trans(osb, credits);
1798         if (IS_ERR(handle)) {
1799                 ret = PTR_ERR(handle);
1800                 mlog_errno(ret);
1801                 goto out;
1802         }
1803
1804         wc->w_handle = handle;
1805
1806         if (clusters_to_alloc) {
1807                 ret = dquot_alloc_space_nodirty(inode,
1808                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1809                 if (ret)
1810                         goto out_commit;
1811         }
1812
1813         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1814                                       OCFS2_JOURNAL_ACCESS_WRITE);
1815         if (ret) {
1816                 mlog_errno(ret);
1817                 goto out_quota;
1818         }
1819
1820         /*
1821          * Fill our page array first. That way we've grabbed enough so
1822          * that we can zero and flush if we error after adding the
1823          * extent.
1824          */
1825         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1826                                          cluster_of_pages, mmap_page);
1827         if (ret && ret != -EAGAIN) {
1828                 mlog_errno(ret);
1829                 goto out_quota;
1830         }
1831
1832         /*
1833          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1834          * the target page. In this case, we exit with no error and no target
1835          * page. This will trigger the caller, page_mkwrite(), to re-try
1836          * the operation.
1837          */
1838         if (ret == -EAGAIN) {
1839                 BUG_ON(wc->w_target_page);
1840                 ret = 0;
1841                 goto out_quota;
1842         }
1843
1844         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1845                                           len);
1846         if (ret) {
1847                 mlog_errno(ret);
1848                 goto out_quota;
1849         }
1850
1851         if (data_ac)
1852                 ocfs2_free_alloc_context(data_ac);
1853         if (meta_ac)
1854                 ocfs2_free_alloc_context(meta_ac);
1855
1856 success:
1857         if (pagep)
1858                 *pagep = wc->w_target_page;
1859         *fsdata = wc;
1860         return 0;
1861 out_quota:
1862         if (clusters_to_alloc)
1863                 dquot_free_space(inode,
1864                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1865 out_commit:
1866         ocfs2_commit_trans(osb, handle);
1867
1868 out:
1869         /*
1870          * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1871          * even in case of error here like ENOSPC and ENOMEM. So, we need
1872          * to unlock the target page manually to prevent deadlocks when
1873          * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1874          * to VM code.
1875          */
1876         if (wc->w_target_locked)
1877                 unlock_page(mmap_page);
1878
1879         ocfs2_free_write_ctxt(inode, wc);
1880
1881         if (data_ac) {
1882                 ocfs2_free_alloc_context(data_ac);
1883                 data_ac = NULL;
1884         }
1885         if (meta_ac) {
1886                 ocfs2_free_alloc_context(meta_ac);
1887                 meta_ac = NULL;
1888         }
1889
1890         if (ret == -ENOSPC && try_free) {
1891                 /*
1892                  * Try to free some truncate log so that we can have enough
1893                  * clusters to allocate.
1894                  */
1895                 try_free = 0;
1896
1897                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1898                 if (ret1 == 1)
1899                         goto try_again;
1900
1901                 if (ret1 < 0)
1902                         mlog_errno(ret1);
1903         }
1904
1905         return ret;
1906 }
1907
1908 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1909                              loff_t pos, unsigned len, unsigned flags,
1910                              struct page **pagep, void **fsdata)
1911 {
1912         int ret;
1913         struct buffer_head *di_bh = NULL;
1914         struct inode *inode = mapping->host;
1915
1916         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1917         if (ret) {
1918                 mlog_errno(ret);
1919                 return ret;
1920         }
1921
1922         /*
1923          * Take alloc sem here to prevent concurrent lookups. That way
1924          * the mapping, zeroing and tree manipulation within
1925          * ocfs2_write() will be safe against ->readpage(). This
1926          * should also serve to lock out allocation from a shared
1927          * writeable region.
1928          */
1929         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1930
1931         ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1932                                        pagep, fsdata, di_bh, NULL);
1933         if (ret) {
1934                 mlog_errno(ret);
1935                 goto out_fail;
1936         }
1937
1938         brelse(di_bh);
1939
1940         return 0;
1941
1942 out_fail:
1943         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1944
1945         brelse(di_bh);
1946         ocfs2_inode_unlock(inode, 1);
1947
1948         return ret;
1949 }
1950
1951 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1952                                    unsigned len, unsigned *copied,
1953                                    struct ocfs2_dinode *di,
1954                                    struct ocfs2_write_ctxt *wc)
1955 {
1956         void *kaddr;
1957
1958         if (unlikely(*copied < len)) {
1959                 if (!PageUptodate(wc->w_target_page)) {
1960                         *copied = 0;
1961                         return;
1962                 }
1963         }
1964
1965         kaddr = kmap_atomic(wc->w_target_page);
1966         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1967         kunmap_atomic(kaddr);
1968
1969         trace_ocfs2_write_end_inline(
1970              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1971              (unsigned long long)pos, *copied,
1972              le16_to_cpu(di->id2.i_data.id_count),
1973              le16_to_cpu(di->i_dyn_features));
1974 }
1975
1976 int ocfs2_write_end_nolock(struct address_space *mapping,
1977                            loff_t pos, unsigned len, unsigned copied, void *fsdata)
1978 {
1979         int i, ret;
1980         unsigned from, to, start = pos & (PAGE_SIZE - 1);
1981         struct inode *inode = mapping->host;
1982         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1983         struct ocfs2_write_ctxt *wc = fsdata;
1984         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1985         handle_t *handle = wc->w_handle;
1986         struct page *tmppage;
1987
1988         BUG_ON(!list_empty(&wc->w_unwritten_list));
1989
1990         if (handle) {
1991                 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1992                                 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1993                 if (ret) {
1994                         copied = ret;
1995                         mlog_errno(ret);
1996                         goto out;
1997                 }
1998         }
1999
2000         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2001                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2002                 goto out_write_size;
2003         }
2004
2005         if (unlikely(copied < len) && wc->w_target_page) {
2006                 if (!PageUptodate(wc->w_target_page))
2007                         copied = 0;
2008
2009                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2010                                        start+len);
2011         }
2012         if (wc->w_target_page)
2013                 flush_dcache_page(wc->w_target_page);
2014
2015         for(i = 0; i < wc->w_num_pages; i++) {
2016                 tmppage = wc->w_pages[i];
2017
2018                 /* This is the direct io target page. */
2019                 if (tmppage == NULL)
2020                         continue;
2021
2022                 if (tmppage == wc->w_target_page) {
2023                         from = wc->w_target_from;
2024                         to = wc->w_target_to;
2025
2026                         BUG_ON(from > PAGE_SIZE ||
2027                                to > PAGE_SIZE ||
2028                                to < from);
2029                 } else {
2030                         /*
2031                          * Pages adjacent to the target (if any) imply
2032                          * a hole-filling write in which case we want
2033                          * to flush their entire range.
2034                          */
2035                         from = 0;
2036                         to = PAGE_SIZE;
2037                 }
2038
2039                 if (page_has_buffers(tmppage)) {
2040                         if (handle && ocfs2_should_order_data(inode))
2041                                 ocfs2_jbd2_file_inode(handle, inode);
2042                         block_commit_write(tmppage, from, to);
2043                 }
2044         }
2045
2046 out_write_size:
2047         /* Direct io do not update i_size here. */
2048         if (wc->w_type != OCFS2_WRITE_DIRECT) {
2049                 pos += copied;
2050                 if (pos > i_size_read(inode)) {
2051                         i_size_write(inode, pos);
2052                         mark_inode_dirty(inode);
2053                 }
2054                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2055                 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2056                 inode->i_mtime = inode->i_ctime = current_time(inode);
2057                 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2058                 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2059                 if (handle)
2060                         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2061         }
2062         if (handle)
2063                 ocfs2_journal_dirty(handle, wc->w_di_bh);
2064
2065 out:
2066         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2067          * lock, or it will cause a deadlock since journal commit threads holds
2068          * this lock and will ask for the page lock when flushing the data.
2069          * put it here to preserve the unlock order.
2070          */
2071         ocfs2_unlock_pages(wc);
2072
2073         if (handle)
2074                 ocfs2_commit_trans(osb, handle);
2075
2076         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2077
2078         brelse(wc->w_di_bh);
2079         kfree(wc);
2080
2081         return copied;
2082 }
2083
2084 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2085                            loff_t pos, unsigned len, unsigned copied,
2086                            struct page *page, void *fsdata)
2087 {
2088         int ret;
2089         struct inode *inode = mapping->host;
2090
2091         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2092
2093         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2094         ocfs2_inode_unlock(inode, 1);
2095
2096         return ret;
2097 }
2098
2099 struct ocfs2_dio_write_ctxt {
2100         struct list_head        dw_zero_list;
2101         unsigned                dw_zero_count;
2102         int                     dw_orphaned;
2103         pid_t                   dw_writer_pid;
2104 };
2105
2106 static struct ocfs2_dio_write_ctxt *
2107 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2108 {
2109         struct ocfs2_dio_write_ctxt *dwc = NULL;
2110
2111         if (bh->b_private)
2112                 return bh->b_private;
2113
2114         dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2115         if (dwc == NULL)
2116                 return NULL;
2117         INIT_LIST_HEAD(&dwc->dw_zero_list);
2118         dwc->dw_zero_count = 0;
2119         dwc->dw_orphaned = 0;
2120         dwc->dw_writer_pid = task_pid_nr(current);
2121         bh->b_private = dwc;
2122         *alloc = 1;
2123
2124         return dwc;
2125 }
2126
2127 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2128                                      struct ocfs2_dio_write_ctxt *dwc)
2129 {
2130         ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2131         kfree(dwc);
2132 }
2133
2134 /*
2135  * TODO: Make this into a generic get_blocks function.
2136  *
2137  * From do_direct_io in direct-io.c:
2138  *  "So what we do is to permit the ->get_blocks function to populate
2139  *   bh.b_size with the size of IO which is permitted at this offset and
2140  *   this i_blkbits."
2141  *
2142  * This function is called directly from get_more_blocks in direct-io.c.
2143  *
2144  * called like this: dio->get_blocks(dio->inode, fs_startblk,
2145  *                                      fs_count, map_bh, dio->rw == WRITE);
2146  */
2147 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2148                                struct buffer_head *bh_result, int create)
2149 {
2150         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2151         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2152         struct ocfs2_write_ctxt *wc;
2153         struct ocfs2_write_cluster_desc *desc = NULL;
2154         struct ocfs2_dio_write_ctxt *dwc = NULL;
2155         struct buffer_head *di_bh = NULL;
2156         u64 p_blkno;
2157         unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2158         loff_t pos = iblock << i_blkbits;
2159         sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2160         unsigned len, total_len = bh_result->b_size;
2161         int ret = 0, first_get_block = 0;
2162
2163         len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2164         len = min(total_len, len);
2165
2166         /*
2167          * bh_result->b_size is count in get_more_blocks according to write
2168          * "pos" and "end", we need map twice to return different buffer state:
2169          * 1. area in file size, not set NEW;
2170          * 2. area out file size, set  NEW.
2171          *
2172          *                 iblock    endblk
2173          * |--------|---------|---------|---------
2174          * |<-------area in file------->|
2175          */
2176
2177         if ((iblock <= endblk) &&
2178             ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2179                 len = (endblk - iblock + 1) << i_blkbits;
2180
2181         mlog(0, "get block of %lu at %llu:%u req %u\n",
2182                         inode->i_ino, pos, len, total_len);
2183
2184         /*
2185          * Because we need to change file size in ocfs2_dio_end_io_write(), or
2186          * we may need to add it to orphan dir. So can not fall to fast path
2187          * while file size will be changed.
2188          */
2189         if (pos + total_len <= i_size_read(inode)) {
2190
2191                 /* This is the fast path for re-write. */
2192                 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2193                 if (buffer_mapped(bh_result) &&
2194                     !buffer_new(bh_result) &&
2195                     ret == 0)
2196                         goto out;
2197
2198                 /* Clear state set by ocfs2_get_block. */
2199                 bh_result->b_state = 0;
2200         }
2201
2202         dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2203         if (unlikely(dwc == NULL)) {
2204                 ret = -ENOMEM;
2205                 mlog_errno(ret);
2206                 goto out;
2207         }
2208
2209         if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2210             ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2211             !dwc->dw_orphaned) {
2212                 /*
2213                  * when we are going to alloc extents beyond file size, add the
2214                  * inode to orphan dir, so we can recall those spaces when
2215                  * system crashed during write.
2216                  */
2217                 ret = ocfs2_add_inode_to_orphan(osb, inode);
2218                 if (ret < 0) {
2219                         mlog_errno(ret);
2220                         goto out;
2221                 }
2222                 dwc->dw_orphaned = 1;
2223         }
2224
2225         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2226         if (ret) {
2227                 mlog_errno(ret);
2228                 goto out;
2229         }
2230
2231         down_write(&oi->ip_alloc_sem);
2232
2233         if (first_get_block) {
2234                 if (ocfs2_sparse_alloc(osb))
2235                         ret = ocfs2_zero_tail(inode, di_bh, pos);
2236                 else
2237                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2238                                                            total_len, NULL);
2239                 if (ret < 0) {
2240                         mlog_errno(ret);
2241                         goto unlock;
2242                 }
2243         }
2244
2245         ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2246                                        OCFS2_WRITE_DIRECT, NULL,
2247                                        (void **)&wc, di_bh, NULL);
2248         if (ret) {
2249                 mlog_errno(ret);
2250                 goto unlock;
2251         }
2252
2253         desc = &wc->w_desc[0];
2254
2255         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2256         BUG_ON(p_blkno == 0);
2257         p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2258
2259         map_bh(bh_result, inode->i_sb, p_blkno);
2260         bh_result->b_size = len;
2261         if (desc->c_needs_zero)
2262                 set_buffer_new(bh_result);
2263
2264         if (iblock > endblk)
2265                 set_buffer_new(bh_result);
2266
2267         /* May sleep in end_io. It should not happen in a irq context. So defer
2268          * it to dio work queue. */
2269         set_buffer_defer_completion(bh_result);
2270
2271         if (!list_empty(&wc->w_unwritten_list)) {
2272                 struct ocfs2_unwritten_extent *ue = NULL;
2273
2274                 ue = list_first_entry(&wc->w_unwritten_list,
2275                                       struct ocfs2_unwritten_extent,
2276                                       ue_node);
2277                 BUG_ON(ue->ue_cpos != desc->c_cpos);
2278                 /* The physical address may be 0, fill it. */
2279                 ue->ue_phys = desc->c_phys;
2280
2281                 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2282                 dwc->dw_zero_count += wc->w_unwritten_count;
2283         }
2284
2285         ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2286         BUG_ON(ret != len);
2287         ret = 0;
2288 unlock:
2289         up_write(&oi->ip_alloc_sem);
2290         ocfs2_inode_unlock(inode, 1);
2291         brelse(di_bh);
2292 out:
2293         if (ret < 0)
2294                 ret = -EIO;
2295         return ret;
2296 }
2297
2298 static int ocfs2_dio_end_io_write(struct inode *inode,
2299                                   struct ocfs2_dio_write_ctxt *dwc,
2300                                   loff_t offset,
2301                                   ssize_t bytes)
2302 {
2303         struct ocfs2_cached_dealloc_ctxt dealloc;
2304         struct ocfs2_extent_tree et;
2305         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2306         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2307         struct ocfs2_unwritten_extent *ue = NULL;
2308         struct buffer_head *di_bh = NULL;
2309         struct ocfs2_dinode *di;
2310         struct ocfs2_alloc_context *data_ac = NULL;
2311         struct ocfs2_alloc_context *meta_ac = NULL;
2312         handle_t *handle = NULL;
2313         loff_t end = offset + bytes;
2314         int ret = 0, credits = 0;
2315
2316         ocfs2_init_dealloc_ctxt(&dealloc);
2317
2318         /* We do clear unwritten, delete orphan, change i_size here. If neither
2319          * of these happen, we can skip all this. */
2320         if (list_empty(&dwc->dw_zero_list) &&
2321             end <= i_size_read(inode) &&
2322             !dwc->dw_orphaned)
2323                 goto out;
2324
2325         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2326         if (ret < 0) {
2327                 mlog_errno(ret);
2328                 goto out;
2329         }
2330
2331         down_write(&oi->ip_alloc_sem);
2332
2333         /* Delete orphan before acquire i_mutex. */
2334         if (dwc->dw_orphaned) {
2335                 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2336
2337                 end = end > i_size_read(inode) ? end : 0;
2338
2339                 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2340                                 !!end, end);
2341                 if (ret < 0)
2342                         mlog_errno(ret);
2343         }
2344
2345         di = (struct ocfs2_dinode *)di_bh->b_data;
2346
2347         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2348
2349         /* Attach dealloc with extent tree in case that we may reuse extents
2350          * which are already unlinked from current extent tree due to extent
2351          * rotation and merging.
2352          */
2353         et.et_dealloc = &dealloc;
2354
2355         ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2356                                     &data_ac, &meta_ac);
2357         if (ret) {
2358                 mlog_errno(ret);
2359                 goto unlock;
2360         }
2361
2362         credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2363
2364         handle = ocfs2_start_trans(osb, credits);
2365         if (IS_ERR(handle)) {
2366                 ret = PTR_ERR(handle);
2367                 mlog_errno(ret);
2368                 goto unlock;
2369         }
2370         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2371                                       OCFS2_JOURNAL_ACCESS_WRITE);
2372         if (ret) {
2373                 mlog_errno(ret);
2374                 goto commit;
2375         }
2376
2377         list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2378                 ret = ocfs2_mark_extent_written(inode, &et, handle,
2379                                                 ue->ue_cpos, 1,
2380                                                 ue->ue_phys,
2381                                                 meta_ac, &dealloc);
2382                 if (ret < 0) {
2383                         mlog_errno(ret);
2384                         break;
2385                 }
2386         }
2387
2388         if (end > i_size_read(inode)) {
2389                 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2390                 if (ret < 0)
2391                         mlog_errno(ret);
2392         }
2393 commit:
2394         ocfs2_commit_trans(osb, handle);
2395 unlock:
2396         up_write(&oi->ip_alloc_sem);
2397         ocfs2_inode_unlock(inode, 1);
2398         brelse(di_bh);
2399 out:
2400         if (data_ac)
2401                 ocfs2_free_alloc_context(data_ac);
2402         if (meta_ac)
2403                 ocfs2_free_alloc_context(meta_ac);
2404         ocfs2_run_deallocs(osb, &dealloc);
2405         ocfs2_dio_free_write_ctx(inode, dwc);
2406
2407         return ret;
2408 }
2409
2410 /*
2411  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2412  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2413  * to protect io on one node from truncation on another.
2414  */
2415 static int ocfs2_dio_end_io(struct kiocb *iocb,
2416                             loff_t offset,
2417                             ssize_t bytes,
2418                             void *private)
2419 {
2420         struct inode *inode = file_inode(iocb->ki_filp);
2421         int level;
2422         int ret = 0;
2423
2424         /* this io's submitter should not have unlocked this before we could */
2425         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2426
2427         if (bytes <= 0)
2428                 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2429                                  (long long)bytes);
2430         if (private) {
2431                 if (bytes > 0)
2432                         ret = ocfs2_dio_end_io_write(inode, private, offset,
2433                                                      bytes);
2434                 else
2435                         ocfs2_dio_free_write_ctx(inode, private);
2436         }
2437
2438         ocfs2_iocb_clear_rw_locked(iocb);
2439
2440         level = ocfs2_iocb_rw_locked_level(iocb);
2441         ocfs2_rw_unlock(inode, level);
2442         return ret;
2443 }
2444
2445 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2446 {
2447         struct file *file = iocb->ki_filp;
2448         struct inode *inode = file->f_mapping->host;
2449         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2450         get_block_t *get_block;
2451
2452         /*
2453          * Fallback to buffered I/O if we see an inode without
2454          * extents.
2455          */
2456         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2457                 return 0;
2458
2459         /* Fallback to buffered I/O if we do not support append dio. */
2460         if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2461             !ocfs2_supports_append_dio(osb))
2462                 return 0;
2463
2464         if (iov_iter_rw(iter) == READ)
2465                 get_block = ocfs2_lock_get_block;
2466         else
2467                 get_block = ocfs2_dio_wr_get_block;
2468
2469         return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2470                                     iter, get_block,
2471                                     ocfs2_dio_end_io, NULL, 0);
2472 }
2473
2474 const struct address_space_operations ocfs2_aops = {
2475         .readpage               = ocfs2_readpage,
2476         .readpages              = ocfs2_readpages,
2477         .writepage              = ocfs2_writepage,
2478         .write_begin            = ocfs2_write_begin,
2479         .write_end              = ocfs2_write_end,
2480         .bmap                   = ocfs2_bmap,
2481         .direct_IO              = ocfs2_direct_IO,
2482         .invalidatepage         = block_invalidatepage,
2483         .releasepage            = ocfs2_releasepage,
2484         .migratepage            = buffer_migrate_page,
2485         .is_partially_uptodate  = block_is_partially_uptodate,
2486         .error_remove_page      = generic_error_remove_page,
2487 };