Merge branch 'for_linus' into for_linus_merged
[linux-2.6.git] / fs / ext4 / extents.c
1 /*
2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3  * Written by Alex Tomas <alex@clusterfs.com>
4  *
5  * Architecture independence:
6  *   Copyright (c) 2005, Bull S.A.
7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public Licens
19  * along with this program; if not, write to the Free Software
20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
21  */
22
23 /*
24  * Extents support for EXT4
25  *
26  * TODO:
27  *   - ext4*_error() should be used in some situations
28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
29  *   - smart tree reduction
30  */
31
32 #include <linux/fs.h>
33 #include <linux/time.h>
34 #include <linux/jbd2.h>
35 #include <linux/highuid.h>
36 #include <linux/pagemap.h>
37 #include <linux/quotaops.h>
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include <linux/falloc.h>
41 #include <asm/uaccess.h>
42 #include <linux/fiemap.h>
43 #include "ext4_jbd2.h"
44
45 #include <trace/events/ext4.h>
46
47 static int ext4_split_extent(handle_t *handle,
48                                 struct inode *inode,
49                                 struct ext4_ext_path *path,
50                                 struct ext4_map_blocks *map,
51                                 int split_flag,
52                                 int flags);
53
54 static int ext4_ext_truncate_extend_restart(handle_t *handle,
55                                             struct inode *inode,
56                                             int needed)
57 {
58         int err;
59
60         if (!ext4_handle_valid(handle))
61                 return 0;
62         if (handle->h_buffer_credits > needed)
63                 return 0;
64         err = ext4_journal_extend(handle, needed);
65         if (err <= 0)
66                 return err;
67         err = ext4_truncate_restart_trans(handle, inode, needed);
68         if (err == 0)
69                 err = -EAGAIN;
70
71         return err;
72 }
73
74 /*
75  * could return:
76  *  - EROFS
77  *  - ENOMEM
78  */
79 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
80                                 struct ext4_ext_path *path)
81 {
82         if (path->p_bh) {
83                 /* path points to block */
84                 return ext4_journal_get_write_access(handle, path->p_bh);
85         }
86         /* path points to leaf/index in inode body */
87         /* we use in-core data, no need to protect them */
88         return 0;
89 }
90
91 /*
92  * could return:
93  *  - EROFS
94  *  - ENOMEM
95  *  - EIO
96  */
97 #define ext4_ext_dirty(handle, inode, path) \
98                 __ext4_ext_dirty(__func__, __LINE__, (handle), (inode), (path))
99 static int __ext4_ext_dirty(const char *where, unsigned int line,
100                             handle_t *handle, struct inode *inode,
101                             struct ext4_ext_path *path)
102 {
103         int err;
104         if (path->p_bh) {
105                 /* path points to block */
106                 err = __ext4_handle_dirty_metadata(where, line, handle,
107                                                    inode, path->p_bh);
108         } else {
109                 /* path points to leaf/index in inode body */
110                 err = ext4_mark_inode_dirty(handle, inode);
111         }
112         return err;
113 }
114
115 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
116                               struct ext4_ext_path *path,
117                               ext4_lblk_t block)
118 {
119         if (path) {
120                 int depth = path->p_depth;
121                 struct ext4_extent *ex;
122
123                 /*
124                  * Try to predict block placement assuming that we are
125                  * filling in a file which will eventually be
126                  * non-sparse --- i.e., in the case of libbfd writing
127                  * an ELF object sections out-of-order but in a way
128                  * the eventually results in a contiguous object or
129                  * executable file, or some database extending a table
130                  * space file.  However, this is actually somewhat
131                  * non-ideal if we are writing a sparse file such as
132                  * qemu or KVM writing a raw image file that is going
133                  * to stay fairly sparse, since it will end up
134                  * fragmenting the file system's free space.  Maybe we
135                  * should have some hueristics or some way to allow
136                  * userspace to pass a hint to file system,
137                  * especially if the latter case turns out to be
138                  * common.
139                  */
140                 ex = path[depth].p_ext;
141                 if (ex) {
142                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
143                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
144
145                         if (block > ext_block)
146                                 return ext_pblk + (block - ext_block);
147                         else
148                                 return ext_pblk - (ext_block - block);
149                 }
150
151                 /* it looks like index is empty;
152                  * try to find starting block from index itself */
153                 if (path[depth].p_bh)
154                         return path[depth].p_bh->b_blocknr;
155         }
156
157         /* OK. use inode's group */
158         return ext4_inode_to_goal_block(inode);
159 }
160
161 /*
162  * Allocation for a meta data block
163  */
164 static ext4_fsblk_t
165 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
166                         struct ext4_ext_path *path,
167                         struct ext4_extent *ex, int *err, unsigned int flags)
168 {
169         ext4_fsblk_t goal, newblock;
170
171         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
172         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
173                                         NULL, err);
174         return newblock;
175 }
176
177 static inline int ext4_ext_space_block(struct inode *inode, int check)
178 {
179         int size;
180
181         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
182                         / sizeof(struct ext4_extent);
183 #ifdef AGGRESSIVE_TEST
184         if (!check && size > 6)
185                 size = 6;
186 #endif
187         return size;
188 }
189
190 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
191 {
192         int size;
193
194         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
195                         / sizeof(struct ext4_extent_idx);
196 #ifdef AGGRESSIVE_TEST
197         if (!check && size > 5)
198                 size = 5;
199 #endif
200         return size;
201 }
202
203 static inline int ext4_ext_space_root(struct inode *inode, int check)
204 {
205         int size;
206
207         size = sizeof(EXT4_I(inode)->i_data);
208         size -= sizeof(struct ext4_extent_header);
209         size /= sizeof(struct ext4_extent);
210 #ifdef AGGRESSIVE_TEST
211         if (!check && size > 3)
212                 size = 3;
213 #endif
214         return size;
215 }
216
217 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
218 {
219         int size;
220
221         size = sizeof(EXT4_I(inode)->i_data);
222         size -= sizeof(struct ext4_extent_header);
223         size /= sizeof(struct ext4_extent_idx);
224 #ifdef AGGRESSIVE_TEST
225         if (!check && size > 4)
226                 size = 4;
227 #endif
228         return size;
229 }
230
231 /*
232  * Calculate the number of metadata blocks needed
233  * to allocate @blocks
234  * Worse case is one block per extent
235  */
236 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
237 {
238         struct ext4_inode_info *ei = EXT4_I(inode);
239         int idxs;
240
241         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
242                 / sizeof(struct ext4_extent_idx));
243
244         /*
245          * If the new delayed allocation block is contiguous with the
246          * previous da block, it can share index blocks with the
247          * previous block, so we only need to allocate a new index
248          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
249          * an additional index block, and at ldxs**3 blocks, yet
250          * another index blocks.
251          */
252         if (ei->i_da_metadata_calc_len &&
253             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
254                 int num = 0;
255
256                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
257                         num++;
258                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
259                         num++;
260                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
261                         num++;
262                         ei->i_da_metadata_calc_len = 0;
263                 } else
264                         ei->i_da_metadata_calc_len++;
265                 ei->i_da_metadata_calc_last_lblock++;
266                 return num;
267         }
268
269         /*
270          * In the worst case we need a new set of index blocks at
271          * every level of the inode's extent tree.
272          */
273         ei->i_da_metadata_calc_len = 1;
274         ei->i_da_metadata_calc_last_lblock = lblock;
275         return ext_depth(inode) + 1;
276 }
277
278 static int
279 ext4_ext_max_entries(struct inode *inode, int depth)
280 {
281         int max;
282
283         if (depth == ext_depth(inode)) {
284                 if (depth == 0)
285                         max = ext4_ext_space_root(inode, 1);
286                 else
287                         max = ext4_ext_space_root_idx(inode, 1);
288         } else {
289                 if (depth == 0)
290                         max = ext4_ext_space_block(inode, 1);
291                 else
292                         max = ext4_ext_space_block_idx(inode, 1);
293         }
294
295         return max;
296 }
297
298 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
299 {
300         ext4_fsblk_t block = ext4_ext_pblock(ext);
301         int len = ext4_ext_get_actual_len(ext);
302
303         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
304 }
305
306 static int ext4_valid_extent_idx(struct inode *inode,
307                                 struct ext4_extent_idx *ext_idx)
308 {
309         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
310
311         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
312 }
313
314 static int ext4_valid_extent_entries(struct inode *inode,
315                                 struct ext4_extent_header *eh,
316                                 int depth)
317 {
318         unsigned short entries;
319         if (eh->eh_entries == 0)
320                 return 1;
321
322         entries = le16_to_cpu(eh->eh_entries);
323
324         if (depth == 0) {
325                 /* leaf entries */
326                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
327                 while (entries) {
328                         if (!ext4_valid_extent(inode, ext))
329                                 return 0;
330                         ext++;
331                         entries--;
332                 }
333         } else {
334                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
335                 while (entries) {
336                         if (!ext4_valid_extent_idx(inode, ext_idx))
337                                 return 0;
338                         ext_idx++;
339                         entries--;
340                 }
341         }
342         return 1;
343 }
344
345 static int __ext4_ext_check(const char *function, unsigned int line,
346                             struct inode *inode, struct ext4_extent_header *eh,
347                             int depth)
348 {
349         const char *error_msg;
350         int max = 0;
351
352         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
353                 error_msg = "invalid magic";
354                 goto corrupted;
355         }
356         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
357                 error_msg = "unexpected eh_depth";
358                 goto corrupted;
359         }
360         if (unlikely(eh->eh_max == 0)) {
361                 error_msg = "invalid eh_max";
362                 goto corrupted;
363         }
364         max = ext4_ext_max_entries(inode, depth);
365         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
366                 error_msg = "too large eh_max";
367                 goto corrupted;
368         }
369         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
370                 error_msg = "invalid eh_entries";
371                 goto corrupted;
372         }
373         if (!ext4_valid_extent_entries(inode, eh, depth)) {
374                 error_msg = "invalid extent entries";
375                 goto corrupted;
376         }
377         return 0;
378
379 corrupted:
380         ext4_error_inode(inode, function, line, 0,
381                         "bad header/extent: %s - magic %x, "
382                         "entries %u, max %u(%u), depth %u(%u)",
383                         error_msg, le16_to_cpu(eh->eh_magic),
384                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
385                         max, le16_to_cpu(eh->eh_depth), depth);
386
387         return -EIO;
388 }
389
390 #define ext4_ext_check(inode, eh, depth)        \
391         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
392
393 int ext4_ext_check_inode(struct inode *inode)
394 {
395         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
396 }
397
398 #ifdef EXT_DEBUG
399 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
400 {
401         int k, l = path->p_depth;
402
403         ext_debug("path:");
404         for (k = 0; k <= l; k++, path++) {
405                 if (path->p_idx) {
406                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
407                             ext4_idx_pblock(path->p_idx));
408                 } else if (path->p_ext) {
409                         ext_debug("  %d:[%d]%d:%llu ",
410                                   le32_to_cpu(path->p_ext->ee_block),
411                                   ext4_ext_is_uninitialized(path->p_ext),
412                                   ext4_ext_get_actual_len(path->p_ext),
413                                   ext4_ext_pblock(path->p_ext));
414                 } else
415                         ext_debug("  []");
416         }
417         ext_debug("\n");
418 }
419
420 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
421 {
422         int depth = ext_depth(inode);
423         struct ext4_extent_header *eh;
424         struct ext4_extent *ex;
425         int i;
426
427         if (!path)
428                 return;
429
430         eh = path[depth].p_hdr;
431         ex = EXT_FIRST_EXTENT(eh);
432
433         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
434
435         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
436                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
437                           ext4_ext_is_uninitialized(ex),
438                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
439         }
440         ext_debug("\n");
441 }
442
443 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
444                         ext4_fsblk_t newblock, int level)
445 {
446         int depth = ext_depth(inode);
447         struct ext4_extent *ex;
448
449         if (depth != level) {
450                 struct ext4_extent_idx *idx;
451                 idx = path[level].p_idx;
452                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
453                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
454                                         le32_to_cpu(idx->ei_block),
455                                         ext4_idx_pblock(idx),
456                                         newblock);
457                         idx++;
458                 }
459
460                 return;
461         }
462
463         ex = path[depth].p_ext;
464         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
465                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
466                                 le32_to_cpu(ex->ee_block),
467                                 ext4_ext_pblock(ex),
468                                 ext4_ext_is_uninitialized(ex),
469                                 ext4_ext_get_actual_len(ex),
470                                 newblock);
471                 ex++;
472         }
473 }
474
475 #else
476 #define ext4_ext_show_path(inode, path)
477 #define ext4_ext_show_leaf(inode, path)
478 #define ext4_ext_show_move(inode, path, newblock, level)
479 #endif
480
481 void ext4_ext_drop_refs(struct ext4_ext_path *path)
482 {
483         int depth = path->p_depth;
484         int i;
485
486         for (i = 0; i <= depth; i++, path++)
487                 if (path->p_bh) {
488                         brelse(path->p_bh);
489                         path->p_bh = NULL;
490                 }
491 }
492
493 /*
494  * ext4_ext_binsearch_idx:
495  * binary search for the closest index of the given block
496  * the header must be checked before calling this
497  */
498 static void
499 ext4_ext_binsearch_idx(struct inode *inode,
500                         struct ext4_ext_path *path, ext4_lblk_t block)
501 {
502         struct ext4_extent_header *eh = path->p_hdr;
503         struct ext4_extent_idx *r, *l, *m;
504
505
506         ext_debug("binsearch for %u(idx):  ", block);
507
508         l = EXT_FIRST_INDEX(eh) + 1;
509         r = EXT_LAST_INDEX(eh);
510         while (l <= r) {
511                 m = l + (r - l) / 2;
512                 if (block < le32_to_cpu(m->ei_block))
513                         r = m - 1;
514                 else
515                         l = m + 1;
516                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
517                                 m, le32_to_cpu(m->ei_block),
518                                 r, le32_to_cpu(r->ei_block));
519         }
520
521         path->p_idx = l - 1;
522         ext_debug("  -> %d->%lld ", le32_to_cpu(path->p_idx->ei_block),
523                   ext4_idx_pblock(path->p_idx));
524
525 #ifdef CHECK_BINSEARCH
526         {
527                 struct ext4_extent_idx *chix, *ix;
528                 int k;
529
530                 chix = ix = EXT_FIRST_INDEX(eh);
531                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
532                   if (k != 0 &&
533                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
534                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
535                                        "first=0x%p\n", k,
536                                        ix, EXT_FIRST_INDEX(eh));
537                                 printk(KERN_DEBUG "%u <= %u\n",
538                                        le32_to_cpu(ix->ei_block),
539                                        le32_to_cpu(ix[-1].ei_block));
540                         }
541                         BUG_ON(k && le32_to_cpu(ix->ei_block)
542                                            <= le32_to_cpu(ix[-1].ei_block));
543                         if (block < le32_to_cpu(ix->ei_block))
544                                 break;
545                         chix = ix;
546                 }
547                 BUG_ON(chix != path->p_idx);
548         }
549 #endif
550
551 }
552
553 /*
554  * ext4_ext_binsearch:
555  * binary search for closest extent of the given block
556  * the header must be checked before calling this
557  */
558 static void
559 ext4_ext_binsearch(struct inode *inode,
560                 struct ext4_ext_path *path, ext4_lblk_t block)
561 {
562         struct ext4_extent_header *eh = path->p_hdr;
563         struct ext4_extent *r, *l, *m;
564
565         if (eh->eh_entries == 0) {
566                 /*
567                  * this leaf is empty:
568                  * we get such a leaf in split/add case
569                  */
570                 return;
571         }
572
573         ext_debug("binsearch for %u:  ", block);
574
575         l = EXT_FIRST_EXTENT(eh) + 1;
576         r = EXT_LAST_EXTENT(eh);
577
578         while (l <= r) {
579                 m = l + (r - l) / 2;
580                 if (block < le32_to_cpu(m->ee_block))
581                         r = m - 1;
582                 else
583                         l = m + 1;
584                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
585                                 m, le32_to_cpu(m->ee_block),
586                                 r, le32_to_cpu(r->ee_block));
587         }
588
589         path->p_ext = l - 1;
590         ext_debug("  -> %d:%llu:[%d]%d ",
591                         le32_to_cpu(path->p_ext->ee_block),
592                         ext4_ext_pblock(path->p_ext),
593                         ext4_ext_is_uninitialized(path->p_ext),
594                         ext4_ext_get_actual_len(path->p_ext));
595
596 #ifdef CHECK_BINSEARCH
597         {
598                 struct ext4_extent *chex, *ex;
599                 int k;
600
601                 chex = ex = EXT_FIRST_EXTENT(eh);
602                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
603                         BUG_ON(k && le32_to_cpu(ex->ee_block)
604                                           <= le32_to_cpu(ex[-1].ee_block));
605                         if (block < le32_to_cpu(ex->ee_block))
606                                 break;
607                         chex = ex;
608                 }
609                 BUG_ON(chex != path->p_ext);
610         }
611 #endif
612
613 }
614
615 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
616 {
617         struct ext4_extent_header *eh;
618
619         eh = ext_inode_hdr(inode);
620         eh->eh_depth = 0;
621         eh->eh_entries = 0;
622         eh->eh_magic = EXT4_EXT_MAGIC;
623         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
624         ext4_mark_inode_dirty(handle, inode);
625         ext4_ext_invalidate_cache(inode);
626         return 0;
627 }
628
629 struct ext4_ext_path *
630 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
631                                         struct ext4_ext_path *path)
632 {
633         struct ext4_extent_header *eh;
634         struct buffer_head *bh;
635         short int depth, i, ppos = 0, alloc = 0;
636
637         eh = ext_inode_hdr(inode);
638         depth = ext_depth(inode);
639
640         /* account possible depth increase */
641         if (!path) {
642                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
643                                 GFP_NOFS);
644                 if (!path)
645                         return ERR_PTR(-ENOMEM);
646                 alloc = 1;
647         }
648         path[0].p_hdr = eh;
649         path[0].p_bh = NULL;
650
651         i = depth;
652         /* walk through the tree */
653         while (i) {
654                 int need_to_validate = 0;
655
656                 ext_debug("depth %d: num %d, max %d\n",
657                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
658
659                 ext4_ext_binsearch_idx(inode, path + ppos, block);
660                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
661                 path[ppos].p_depth = i;
662                 path[ppos].p_ext = NULL;
663
664                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
665                 if (unlikely(!bh))
666                         goto err;
667                 if (!bh_uptodate_or_lock(bh)) {
668                         trace_ext4_ext_load_extent(inode, block,
669                                                 path[ppos].p_block);
670                         if (bh_submit_read(bh) < 0) {
671                                 put_bh(bh);
672                                 goto err;
673                         }
674                         /* validate the extent entries */
675                         need_to_validate = 1;
676                 }
677                 eh = ext_block_hdr(bh);
678                 ppos++;
679                 if (unlikely(ppos > depth)) {
680                         put_bh(bh);
681                         EXT4_ERROR_INODE(inode,
682                                          "ppos %d > depth %d", ppos, depth);
683                         goto err;
684                 }
685                 path[ppos].p_bh = bh;
686                 path[ppos].p_hdr = eh;
687                 i--;
688
689                 if (need_to_validate && ext4_ext_check(inode, eh, i))
690                         goto err;
691         }
692
693         path[ppos].p_depth = i;
694         path[ppos].p_ext = NULL;
695         path[ppos].p_idx = NULL;
696
697         /* find extent */
698         ext4_ext_binsearch(inode, path + ppos, block);
699         /* if not an empty leaf */
700         if (path[ppos].p_ext)
701                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
702
703         ext4_ext_show_path(inode, path);
704
705         return path;
706
707 err:
708         ext4_ext_drop_refs(path);
709         if (alloc)
710                 kfree(path);
711         return ERR_PTR(-EIO);
712 }
713
714 /*
715  * ext4_ext_insert_index:
716  * insert new index [@logical;@ptr] into the block at @curp;
717  * check where to insert: before @curp or after @curp
718  */
719 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
720                                  struct ext4_ext_path *curp,
721                                  int logical, ext4_fsblk_t ptr)
722 {
723         struct ext4_extent_idx *ix;
724         int len, err;
725
726         err = ext4_ext_get_access(handle, inode, curp);
727         if (err)
728                 return err;
729
730         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
731                 EXT4_ERROR_INODE(inode,
732                                  "logical %d == ei_block %d!",
733                                  logical, le32_to_cpu(curp->p_idx->ei_block));
734                 return -EIO;
735         }
736
737         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
738                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
739                 EXT4_ERROR_INODE(inode,
740                                  "eh_entries %d >= eh_max %d!",
741                                  le16_to_cpu(curp->p_hdr->eh_entries),
742                                  le16_to_cpu(curp->p_hdr->eh_max));
743                 return -EIO;
744         }
745
746         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
747                 /* insert after */
748                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
749                 ix = curp->p_idx + 1;
750         } else {
751                 /* insert before */
752                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
753                 ix = curp->p_idx;
754         }
755
756         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
757         BUG_ON(len < 0);
758         if (len > 0) {
759                 ext_debug("insert new index %d: "
760                                 "move %d indices from 0x%p to 0x%p\n",
761                                 logical, len, ix, ix + 1);
762                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
763         }
764
765         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
766                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
767                 return -EIO;
768         }
769
770         ix->ei_block = cpu_to_le32(logical);
771         ext4_idx_store_pblock(ix, ptr);
772         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
773
774         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
775                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
776                 return -EIO;
777         }
778
779         err = ext4_ext_dirty(handle, inode, curp);
780         ext4_std_error(inode->i_sb, err);
781
782         return err;
783 }
784
785 /*
786  * ext4_ext_split:
787  * inserts new subtree into the path, using free index entry
788  * at depth @at:
789  * - allocates all needed blocks (new leaf and all intermediate index blocks)
790  * - makes decision where to split
791  * - moves remaining extents and index entries (right to the split point)
792  *   into the newly allocated blocks
793  * - initializes subtree
794  */
795 static int ext4_ext_split(handle_t *handle, struct inode *inode,
796                           unsigned int flags,
797                           struct ext4_ext_path *path,
798                           struct ext4_extent *newext, int at)
799 {
800         struct buffer_head *bh = NULL;
801         int depth = ext_depth(inode);
802         struct ext4_extent_header *neh;
803         struct ext4_extent_idx *fidx;
804         int i = at, k, m, a;
805         ext4_fsblk_t newblock, oldblock;
806         __le32 border;
807         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
808         int err = 0;
809
810         /* make decision: where to split? */
811         /* FIXME: now decision is simplest: at current extent */
812
813         /* if current leaf will be split, then we should use
814          * border from split point */
815         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
816                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
817                 return -EIO;
818         }
819         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
820                 border = path[depth].p_ext[1].ee_block;
821                 ext_debug("leaf will be split."
822                                 " next leaf starts at %d\n",
823                                   le32_to_cpu(border));
824         } else {
825                 border = newext->ee_block;
826                 ext_debug("leaf will be added."
827                                 " next leaf starts at %d\n",
828                                 le32_to_cpu(border));
829         }
830
831         /*
832          * If error occurs, then we break processing
833          * and mark filesystem read-only. index won't
834          * be inserted and tree will be in consistent
835          * state. Next mount will repair buffers too.
836          */
837
838         /*
839          * Get array to track all allocated blocks.
840          * We need this to handle errors and free blocks
841          * upon them.
842          */
843         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
844         if (!ablocks)
845                 return -ENOMEM;
846
847         /* allocate all needed blocks */
848         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
849         for (a = 0; a < depth - at; a++) {
850                 newblock = ext4_ext_new_meta_block(handle, inode, path,
851                                                    newext, &err, flags);
852                 if (newblock == 0)
853                         goto cleanup;
854                 ablocks[a] = newblock;
855         }
856
857         /* initialize new leaf */
858         newblock = ablocks[--a];
859         if (unlikely(newblock == 0)) {
860                 EXT4_ERROR_INODE(inode, "newblock == 0!");
861                 err = -EIO;
862                 goto cleanup;
863         }
864         bh = sb_getblk(inode->i_sb, newblock);
865         if (!bh) {
866                 err = -EIO;
867                 goto cleanup;
868         }
869         lock_buffer(bh);
870
871         err = ext4_journal_get_create_access(handle, bh);
872         if (err)
873                 goto cleanup;
874
875         neh = ext_block_hdr(bh);
876         neh->eh_entries = 0;
877         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
878         neh->eh_magic = EXT4_EXT_MAGIC;
879         neh->eh_depth = 0;
880
881         /* move remainder of path[depth] to the new leaf */
882         if (unlikely(path[depth].p_hdr->eh_entries !=
883                      path[depth].p_hdr->eh_max)) {
884                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
885                                  path[depth].p_hdr->eh_entries,
886                                  path[depth].p_hdr->eh_max);
887                 err = -EIO;
888                 goto cleanup;
889         }
890         /* start copy from next extent */
891         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
892         ext4_ext_show_move(inode, path, newblock, depth);
893         if (m) {
894                 struct ext4_extent *ex;
895                 ex = EXT_FIRST_EXTENT(neh);
896                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
897                 le16_add_cpu(&neh->eh_entries, m);
898         }
899
900         set_buffer_uptodate(bh);
901         unlock_buffer(bh);
902
903         err = ext4_handle_dirty_metadata(handle, inode, bh);
904         if (err)
905                 goto cleanup;
906         brelse(bh);
907         bh = NULL;
908
909         /* correct old leaf */
910         if (m) {
911                 err = ext4_ext_get_access(handle, inode, path + depth);
912                 if (err)
913                         goto cleanup;
914                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
915                 err = ext4_ext_dirty(handle, inode, path + depth);
916                 if (err)
917                         goto cleanup;
918
919         }
920
921         /* create intermediate indexes */
922         k = depth - at - 1;
923         if (unlikely(k < 0)) {
924                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
925                 err = -EIO;
926                 goto cleanup;
927         }
928         if (k)
929                 ext_debug("create %d intermediate indices\n", k);
930         /* insert new index into current index block */
931         /* current depth stored in i var */
932         i = depth - 1;
933         while (k--) {
934                 oldblock = newblock;
935                 newblock = ablocks[--a];
936                 bh = sb_getblk(inode->i_sb, newblock);
937                 if (!bh) {
938                         err = -EIO;
939                         goto cleanup;
940                 }
941                 lock_buffer(bh);
942
943                 err = ext4_journal_get_create_access(handle, bh);
944                 if (err)
945                         goto cleanup;
946
947                 neh = ext_block_hdr(bh);
948                 neh->eh_entries = cpu_to_le16(1);
949                 neh->eh_magic = EXT4_EXT_MAGIC;
950                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
951                 neh->eh_depth = cpu_to_le16(depth - i);
952                 fidx = EXT_FIRST_INDEX(neh);
953                 fidx->ei_block = border;
954                 ext4_idx_store_pblock(fidx, oldblock);
955
956                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
957                                 i, newblock, le32_to_cpu(border), oldblock);
958
959                 /* move remainder of path[i] to the new index block */
960                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
961                                         EXT_LAST_INDEX(path[i].p_hdr))) {
962                         EXT4_ERROR_INODE(inode,
963                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
964                                          le32_to_cpu(path[i].p_ext->ee_block));
965                         err = -EIO;
966                         goto cleanup;
967                 }
968                 /* start copy indexes */
969                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
970                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
971                                 EXT_MAX_INDEX(path[i].p_hdr));
972                 ext4_ext_show_move(inode, path, newblock, i);
973                 if (m) {
974                         memmove(++fidx, path[i].p_idx,
975                                 sizeof(struct ext4_extent_idx) * m);
976                         le16_add_cpu(&neh->eh_entries, m);
977                 }
978                 set_buffer_uptodate(bh);
979                 unlock_buffer(bh);
980
981                 err = ext4_handle_dirty_metadata(handle, inode, bh);
982                 if (err)
983                         goto cleanup;
984                 brelse(bh);
985                 bh = NULL;
986
987                 /* correct old index */
988                 if (m) {
989                         err = ext4_ext_get_access(handle, inode, path + i);
990                         if (err)
991                                 goto cleanup;
992                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
993                         err = ext4_ext_dirty(handle, inode, path + i);
994                         if (err)
995                                 goto cleanup;
996                 }
997
998                 i--;
999         }
1000
1001         /* insert new index */
1002         err = ext4_ext_insert_index(handle, inode, path + at,
1003                                     le32_to_cpu(border), newblock);
1004
1005 cleanup:
1006         if (bh) {
1007                 if (buffer_locked(bh))
1008                         unlock_buffer(bh);
1009                 brelse(bh);
1010         }
1011
1012         if (err) {
1013                 /* free all allocated blocks in error case */
1014                 for (i = 0; i < depth; i++) {
1015                         if (!ablocks[i])
1016                                 continue;
1017                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1018                                          EXT4_FREE_BLOCKS_METADATA);
1019                 }
1020         }
1021         kfree(ablocks);
1022
1023         return err;
1024 }
1025
1026 /*
1027  * ext4_ext_grow_indepth:
1028  * implements tree growing procedure:
1029  * - allocates new block
1030  * - moves top-level data (index block or leaf) into the new block
1031  * - initializes new top-level, creating index that points to the
1032  *   just created block
1033  */
1034 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1035                                  unsigned int flags,
1036                                  struct ext4_extent *newext)
1037 {
1038         struct ext4_extent_header *neh;
1039         struct buffer_head *bh;
1040         ext4_fsblk_t newblock;
1041         int err = 0;
1042
1043         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1044                 newext, &err, flags);
1045         if (newblock == 0)
1046                 return err;
1047
1048         bh = sb_getblk(inode->i_sb, newblock);
1049         if (!bh) {
1050                 err = -EIO;
1051                 ext4_std_error(inode->i_sb, err);
1052                 return err;
1053         }
1054         lock_buffer(bh);
1055
1056         err = ext4_journal_get_create_access(handle, bh);
1057         if (err) {
1058                 unlock_buffer(bh);
1059                 goto out;
1060         }
1061
1062         /* move top-level index/leaf into new block */
1063         memmove(bh->b_data, EXT4_I(inode)->i_data,
1064                 sizeof(EXT4_I(inode)->i_data));
1065
1066         /* set size of new block */
1067         neh = ext_block_hdr(bh);
1068         /* old root could have indexes or leaves
1069          * so calculate e_max right way */
1070         if (ext_depth(inode))
1071                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1072         else
1073                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1074         neh->eh_magic = EXT4_EXT_MAGIC;
1075         set_buffer_uptodate(bh);
1076         unlock_buffer(bh);
1077
1078         err = ext4_handle_dirty_metadata(handle, inode, bh);
1079         if (err)
1080                 goto out;
1081
1082         /* Update top-level index: num,max,pointer */
1083         neh = ext_inode_hdr(inode);
1084         neh->eh_entries = cpu_to_le16(1);
1085         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1086         if (neh->eh_depth == 0) {
1087                 /* Root extent block becomes index block */
1088                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1089                 EXT_FIRST_INDEX(neh)->ei_block =
1090                         EXT_FIRST_EXTENT(neh)->ee_block;
1091         }
1092         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1093                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1094                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1095                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1096
1097         neh->eh_depth = cpu_to_le16(le16_to_cpu(neh->eh_depth) + 1);
1098         ext4_mark_inode_dirty(handle, inode);
1099 out:
1100         brelse(bh);
1101
1102         return err;
1103 }
1104
1105 /*
1106  * ext4_ext_create_new_leaf:
1107  * finds empty index and adds new leaf.
1108  * if no free index is found, then it requests in-depth growing.
1109  */
1110 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1111                                     unsigned int flags,
1112                                     struct ext4_ext_path *path,
1113                                     struct ext4_extent *newext)
1114 {
1115         struct ext4_ext_path *curp;
1116         int depth, i, err = 0;
1117
1118 repeat:
1119         i = depth = ext_depth(inode);
1120
1121         /* walk up to the tree and look for free index entry */
1122         curp = path + depth;
1123         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1124                 i--;
1125                 curp--;
1126         }
1127
1128         /* we use already allocated block for index block,
1129          * so subsequent data blocks should be contiguous */
1130         if (EXT_HAS_FREE_INDEX(curp)) {
1131                 /* if we found index with free entry, then use that
1132                  * entry: create all needed subtree and add new leaf */
1133                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1134                 if (err)
1135                         goto out;
1136
1137                 /* refill path */
1138                 ext4_ext_drop_refs(path);
1139                 path = ext4_ext_find_extent(inode,
1140                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1141                                     path);
1142                 if (IS_ERR(path))
1143                         err = PTR_ERR(path);
1144         } else {
1145                 /* tree is full, time to grow in depth */
1146                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1147                 if (err)
1148                         goto out;
1149
1150                 /* refill path */
1151                 ext4_ext_drop_refs(path);
1152                 path = ext4_ext_find_extent(inode,
1153                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1154                                     path);
1155                 if (IS_ERR(path)) {
1156                         err = PTR_ERR(path);
1157                         goto out;
1158                 }
1159
1160                 /*
1161                  * only first (depth 0 -> 1) produces free space;
1162                  * in all other cases we have to split the grown tree
1163                  */
1164                 depth = ext_depth(inode);
1165                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1166                         /* now we need to split */
1167                         goto repeat;
1168                 }
1169         }
1170
1171 out:
1172         return err;
1173 }
1174
1175 /*
1176  * search the closest allocated block to the left for *logical
1177  * and returns it at @logical + it's physical address at @phys
1178  * if *logical is the smallest allocated block, the function
1179  * returns 0 at @phys
1180  * return value contains 0 (success) or error code
1181  */
1182 static int ext4_ext_search_left(struct inode *inode,
1183                                 struct ext4_ext_path *path,
1184                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1185 {
1186         struct ext4_extent_idx *ix;
1187         struct ext4_extent *ex;
1188         int depth, ee_len;
1189
1190         if (unlikely(path == NULL)) {
1191                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1192                 return -EIO;
1193         }
1194         depth = path->p_depth;
1195         *phys = 0;
1196
1197         if (depth == 0 && path->p_ext == NULL)
1198                 return 0;
1199
1200         /* usually extent in the path covers blocks smaller
1201          * then *logical, but it can be that extent is the
1202          * first one in the file */
1203
1204         ex = path[depth].p_ext;
1205         ee_len = ext4_ext_get_actual_len(ex);
1206         if (*logical < le32_to_cpu(ex->ee_block)) {
1207                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1208                         EXT4_ERROR_INODE(inode,
1209                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1210                                          *logical, le32_to_cpu(ex->ee_block));
1211                         return -EIO;
1212                 }
1213                 while (--depth >= 0) {
1214                         ix = path[depth].p_idx;
1215                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1216                                 EXT4_ERROR_INODE(inode,
1217                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1218                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1219                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1220                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1221                                   depth);
1222                                 return -EIO;
1223                         }
1224                 }
1225                 return 0;
1226         }
1227
1228         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1229                 EXT4_ERROR_INODE(inode,
1230                                  "logical %d < ee_block %d + ee_len %d!",
1231                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1232                 return -EIO;
1233         }
1234
1235         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1236         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1237         return 0;
1238 }
1239
1240 /*
1241  * search the closest allocated block to the right for *logical
1242  * and returns it at @logical + it's physical address at @phys
1243  * if *logical is the largest allocated block, the function
1244  * returns 0 at @phys
1245  * return value contains 0 (success) or error code
1246  */
1247 static int ext4_ext_search_right(struct inode *inode,
1248                                  struct ext4_ext_path *path,
1249                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1250                                  struct ext4_extent **ret_ex)
1251 {
1252         struct buffer_head *bh = NULL;
1253         struct ext4_extent_header *eh;
1254         struct ext4_extent_idx *ix;
1255         struct ext4_extent *ex;
1256         ext4_fsblk_t block;
1257         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1258         int ee_len;
1259
1260         if (unlikely(path == NULL)) {
1261                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1262                 return -EIO;
1263         }
1264         depth = path->p_depth;
1265         *phys = 0;
1266
1267         if (depth == 0 && path->p_ext == NULL)
1268                 return 0;
1269
1270         /* usually extent in the path covers blocks smaller
1271          * then *logical, but it can be that extent is the
1272          * first one in the file */
1273
1274         ex = path[depth].p_ext;
1275         ee_len = ext4_ext_get_actual_len(ex);
1276         if (*logical < le32_to_cpu(ex->ee_block)) {
1277                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1278                         EXT4_ERROR_INODE(inode,
1279                                          "first_extent(path[%d].p_hdr) != ex",
1280                                          depth);
1281                         return -EIO;
1282                 }
1283                 while (--depth >= 0) {
1284                         ix = path[depth].p_idx;
1285                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1286                                 EXT4_ERROR_INODE(inode,
1287                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1288                                                  *logical);
1289                                 return -EIO;
1290                         }
1291                 }
1292                 goto found_extent;
1293         }
1294
1295         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1296                 EXT4_ERROR_INODE(inode,
1297                                  "logical %d < ee_block %d + ee_len %d!",
1298                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1299                 return -EIO;
1300         }
1301
1302         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1303                 /* next allocated block in this leaf */
1304                 ex++;
1305                 goto found_extent;
1306         }
1307
1308         /* go up and search for index to the right */
1309         while (--depth >= 0) {
1310                 ix = path[depth].p_idx;
1311                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1312                         goto got_index;
1313         }
1314
1315         /* we've gone up to the root and found no index to the right */
1316         return 0;
1317
1318 got_index:
1319         /* we've found index to the right, let's
1320          * follow it and find the closest allocated
1321          * block to the right */
1322         ix++;
1323         block = ext4_idx_pblock(ix);
1324         while (++depth < path->p_depth) {
1325                 bh = sb_bread(inode->i_sb, block);
1326                 if (bh == NULL)
1327                         return -EIO;
1328                 eh = ext_block_hdr(bh);
1329                 /* subtract from p_depth to get proper eh_depth */
1330                 if (ext4_ext_check(inode, eh, path->p_depth - depth)) {
1331                         put_bh(bh);
1332                         return -EIO;
1333                 }
1334                 ix = EXT_FIRST_INDEX(eh);
1335                 block = ext4_idx_pblock(ix);
1336                 put_bh(bh);
1337         }
1338
1339         bh = sb_bread(inode->i_sb, block);
1340         if (bh == NULL)
1341                 return -EIO;
1342         eh = ext_block_hdr(bh);
1343         if (ext4_ext_check(inode, eh, path->p_depth - depth)) {
1344                 put_bh(bh);
1345                 return -EIO;
1346         }
1347         ex = EXT_FIRST_EXTENT(eh);
1348 found_extent:
1349         *logical = le32_to_cpu(ex->ee_block);
1350         *phys = ext4_ext_pblock(ex);
1351         *ret_ex = ex;
1352         if (bh)
1353                 put_bh(bh);
1354         return 0;
1355 }
1356
1357 /*
1358  * ext4_ext_next_allocated_block:
1359  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1360  * NOTE: it considers block number from index entry as
1361  * allocated block. Thus, index entries have to be consistent
1362  * with leaves.
1363  */
1364 static ext4_lblk_t
1365 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1366 {
1367         int depth;
1368
1369         BUG_ON(path == NULL);
1370         depth = path->p_depth;
1371
1372         if (depth == 0 && path->p_ext == NULL)
1373                 return EXT_MAX_BLOCKS;
1374
1375         while (depth >= 0) {
1376                 if (depth == path->p_depth) {
1377                         /* leaf */
1378                         if (path[depth].p_ext &&
1379                                 path[depth].p_ext !=
1380                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1381                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1382                 } else {
1383                         /* index */
1384                         if (path[depth].p_idx !=
1385                                         EXT_LAST_INDEX(path[depth].p_hdr))
1386                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1387                 }
1388                 depth--;
1389         }
1390
1391         return EXT_MAX_BLOCKS;
1392 }
1393
1394 /*
1395  * ext4_ext_next_leaf_block:
1396  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1397  */
1398 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1399 {
1400         int depth;
1401
1402         BUG_ON(path == NULL);
1403         depth = path->p_depth;
1404
1405         /* zero-tree has no leaf blocks at all */
1406         if (depth == 0)
1407                 return EXT_MAX_BLOCKS;
1408
1409         /* go to index block */
1410         depth--;
1411
1412         while (depth >= 0) {
1413                 if (path[depth].p_idx !=
1414                                 EXT_LAST_INDEX(path[depth].p_hdr))
1415                         return (ext4_lblk_t)
1416                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1417                 depth--;
1418         }
1419
1420         return EXT_MAX_BLOCKS;
1421 }
1422
1423 /*
1424  * ext4_ext_correct_indexes:
1425  * if leaf gets modified and modified extent is first in the leaf,
1426  * then we have to correct all indexes above.
1427  * TODO: do we need to correct tree in all cases?
1428  */
1429 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1430                                 struct ext4_ext_path *path)
1431 {
1432         struct ext4_extent_header *eh;
1433         int depth = ext_depth(inode);
1434         struct ext4_extent *ex;
1435         __le32 border;
1436         int k, err = 0;
1437
1438         eh = path[depth].p_hdr;
1439         ex = path[depth].p_ext;
1440
1441         if (unlikely(ex == NULL || eh == NULL)) {
1442                 EXT4_ERROR_INODE(inode,
1443                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1444                 return -EIO;
1445         }
1446
1447         if (depth == 0) {
1448                 /* there is no tree at all */
1449                 return 0;
1450         }
1451
1452         if (ex != EXT_FIRST_EXTENT(eh)) {
1453                 /* we correct tree if first leaf got modified only */
1454                 return 0;
1455         }
1456
1457         /*
1458          * TODO: we need correction if border is smaller than current one
1459          */
1460         k = depth - 1;
1461         border = path[depth].p_ext->ee_block;
1462         err = ext4_ext_get_access(handle, inode, path + k);
1463         if (err)
1464                 return err;
1465         path[k].p_idx->ei_block = border;
1466         err = ext4_ext_dirty(handle, inode, path + k);
1467         if (err)
1468                 return err;
1469
1470         while (k--) {
1471                 /* change all left-side indexes */
1472                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1473                         break;
1474                 err = ext4_ext_get_access(handle, inode, path + k);
1475                 if (err)
1476                         break;
1477                 path[k].p_idx->ei_block = border;
1478                 err = ext4_ext_dirty(handle, inode, path + k);
1479                 if (err)
1480                         break;
1481         }
1482
1483         return err;
1484 }
1485
1486 int
1487 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1488                                 struct ext4_extent *ex2)
1489 {
1490         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1491
1492         /*
1493          * Make sure that either both extents are uninitialized, or
1494          * both are _not_.
1495          */
1496         if (ext4_ext_is_uninitialized(ex1) ^ ext4_ext_is_uninitialized(ex2))
1497                 return 0;
1498
1499         if (ext4_ext_is_uninitialized(ex1))
1500                 max_len = EXT_UNINIT_MAX_LEN;
1501         else
1502                 max_len = EXT_INIT_MAX_LEN;
1503
1504         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1505         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1506
1507         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1508                         le32_to_cpu(ex2->ee_block))
1509                 return 0;
1510
1511         /*
1512          * To allow future support for preallocated extents to be added
1513          * as an RO_COMPAT feature, refuse to merge to extents if
1514          * this can result in the top bit of ee_len being set.
1515          */
1516         if (ext1_ee_len + ext2_ee_len > max_len)
1517                 return 0;
1518 #ifdef AGGRESSIVE_TEST
1519         if (ext1_ee_len >= 4)
1520                 return 0;
1521 #endif
1522
1523         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1524                 return 1;
1525         return 0;
1526 }
1527
1528 /*
1529  * This function tries to merge the "ex" extent to the next extent in the tree.
1530  * It always tries to merge towards right. If you want to merge towards
1531  * left, pass "ex - 1" as argument instead of "ex".
1532  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1533  * 1 if they got merged.
1534  */
1535 static int ext4_ext_try_to_merge_right(struct inode *inode,
1536                                  struct ext4_ext_path *path,
1537                                  struct ext4_extent *ex)
1538 {
1539         struct ext4_extent_header *eh;
1540         unsigned int depth, len;
1541         int merge_done = 0;
1542         int uninitialized = 0;
1543
1544         depth = ext_depth(inode);
1545         BUG_ON(path[depth].p_hdr == NULL);
1546         eh = path[depth].p_hdr;
1547
1548         while (ex < EXT_LAST_EXTENT(eh)) {
1549                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1550                         break;
1551                 /* merge with next extent! */
1552                 if (ext4_ext_is_uninitialized(ex))
1553                         uninitialized = 1;
1554                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1555                                 + ext4_ext_get_actual_len(ex + 1));
1556                 if (uninitialized)
1557                         ext4_ext_mark_uninitialized(ex);
1558
1559                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1560                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1561                                 * sizeof(struct ext4_extent);
1562                         memmove(ex + 1, ex + 2, len);
1563                 }
1564                 le16_add_cpu(&eh->eh_entries, -1);
1565                 merge_done = 1;
1566                 WARN_ON(eh->eh_entries == 0);
1567                 if (!eh->eh_entries)
1568                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1569         }
1570
1571         return merge_done;
1572 }
1573
1574 /*
1575  * This function tries to merge the @ex extent to neighbours in the tree.
1576  * return 1 if merge left else 0.
1577  */
1578 static int ext4_ext_try_to_merge(struct inode *inode,
1579                                   struct ext4_ext_path *path,
1580                                   struct ext4_extent *ex) {
1581         struct ext4_extent_header *eh;
1582         unsigned int depth;
1583         int merge_done = 0;
1584         int ret = 0;
1585
1586         depth = ext_depth(inode);
1587         BUG_ON(path[depth].p_hdr == NULL);
1588         eh = path[depth].p_hdr;
1589
1590         if (ex > EXT_FIRST_EXTENT(eh))
1591                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1592
1593         if (!merge_done)
1594                 ret = ext4_ext_try_to_merge_right(inode, path, ex);
1595
1596         return ret;
1597 }
1598
1599 /*
1600  * check if a portion of the "newext" extent overlaps with an
1601  * existing extent.
1602  *
1603  * If there is an overlap discovered, it updates the length of the newext
1604  * such that there will be no overlap, and then returns 1.
1605  * If there is no overlap found, it returns 0.
1606  */
1607 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1608                                            struct inode *inode,
1609                                            struct ext4_extent *newext,
1610                                            struct ext4_ext_path *path)
1611 {
1612         ext4_lblk_t b1, b2;
1613         unsigned int depth, len1;
1614         unsigned int ret = 0;
1615
1616         b1 = le32_to_cpu(newext->ee_block);
1617         len1 = ext4_ext_get_actual_len(newext);
1618         depth = ext_depth(inode);
1619         if (!path[depth].p_ext)
1620                 goto out;
1621         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1622         b2 &= ~(sbi->s_cluster_ratio - 1);
1623
1624         /*
1625          * get the next allocated block if the extent in the path
1626          * is before the requested block(s)
1627          */
1628         if (b2 < b1) {
1629                 b2 = ext4_ext_next_allocated_block(path);
1630                 if (b2 == EXT_MAX_BLOCKS)
1631                         goto out;
1632                 b2 &= ~(sbi->s_cluster_ratio - 1);
1633         }
1634
1635         /* check for wrap through zero on extent logical start block*/
1636         if (b1 + len1 < b1) {
1637                 len1 = EXT_MAX_BLOCKS - b1;
1638                 newext->ee_len = cpu_to_le16(len1);
1639                 ret = 1;
1640         }
1641
1642         /* check for overlap */
1643         if (b1 + len1 > b2) {
1644                 newext->ee_len = cpu_to_le16(b2 - b1);
1645                 ret = 1;
1646         }
1647 out:
1648         return ret;
1649 }
1650
1651 /*
1652  * ext4_ext_insert_extent:
1653  * tries to merge requsted extent into the existing extent or
1654  * inserts requested extent as new one into the tree,
1655  * creating new leaf in the no-space case.
1656  */
1657 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1658                                 struct ext4_ext_path *path,
1659                                 struct ext4_extent *newext, int flag)
1660 {
1661         struct ext4_extent_header *eh;
1662         struct ext4_extent *ex, *fex;
1663         struct ext4_extent *nearex; /* nearest extent */
1664         struct ext4_ext_path *npath = NULL;
1665         int depth, len, err;
1666         ext4_lblk_t next;
1667         unsigned uninitialized = 0;
1668         int flags = 0;
1669
1670         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1671                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1672                 return -EIO;
1673         }
1674         depth = ext_depth(inode);
1675         ex = path[depth].p_ext;
1676         if (unlikely(path[depth].p_hdr == NULL)) {
1677                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1678                 return -EIO;
1679         }
1680
1681         /* try to insert block into found extent and return */
1682         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)
1683                 && ext4_can_extents_be_merged(inode, ex, newext)) {
1684                 ext_debug("append [%d]%d block to %u:[%d]%d (from %llu)\n",
1685                           ext4_ext_is_uninitialized(newext),
1686                           ext4_ext_get_actual_len(newext),
1687                           le32_to_cpu(ex->ee_block),
1688                           ext4_ext_is_uninitialized(ex),
1689                           ext4_ext_get_actual_len(ex),
1690                           ext4_ext_pblock(ex));
1691                 err = ext4_ext_get_access(handle, inode, path + depth);
1692                 if (err)
1693                         return err;
1694
1695                 /*
1696                  * ext4_can_extents_be_merged should have checked that either
1697                  * both extents are uninitialized, or both aren't. Thus we
1698                  * need to check only one of them here.
1699                  */
1700                 if (ext4_ext_is_uninitialized(ex))
1701                         uninitialized = 1;
1702                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1703                                         + ext4_ext_get_actual_len(newext));
1704                 if (uninitialized)
1705                         ext4_ext_mark_uninitialized(ex);
1706                 eh = path[depth].p_hdr;
1707                 nearex = ex;
1708                 goto merge;
1709         }
1710
1711         depth = ext_depth(inode);
1712         eh = path[depth].p_hdr;
1713         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1714                 goto has_space;
1715
1716         /* probably next leaf has space for us? */
1717         fex = EXT_LAST_EXTENT(eh);
1718         next = EXT_MAX_BLOCKS;
1719         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1720                 next = ext4_ext_next_leaf_block(path);
1721         if (next != EXT_MAX_BLOCKS) {
1722                 ext_debug("next leaf block - %u\n", next);
1723                 BUG_ON(npath != NULL);
1724                 npath = ext4_ext_find_extent(inode, next, NULL);
1725                 if (IS_ERR(npath))
1726                         return PTR_ERR(npath);
1727                 BUG_ON(npath->p_depth != path->p_depth);
1728                 eh = npath[depth].p_hdr;
1729                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1730                         ext_debug("next leaf isn't full(%d)\n",
1731                                   le16_to_cpu(eh->eh_entries));
1732                         path = npath;
1733                         goto has_space;
1734                 }
1735                 ext_debug("next leaf has no free space(%d,%d)\n",
1736                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1737         }
1738
1739         /*
1740          * There is no free space in the found leaf.
1741          * We're gonna add a new leaf in the tree.
1742          */
1743         if (flag & EXT4_GET_BLOCKS_PUNCH_OUT_EXT)
1744                 flags = EXT4_MB_USE_ROOT_BLOCKS;
1745         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1746         if (err)
1747                 goto cleanup;
1748         depth = ext_depth(inode);
1749         eh = path[depth].p_hdr;
1750
1751 has_space:
1752         nearex = path[depth].p_ext;
1753
1754         err = ext4_ext_get_access(handle, inode, path + depth);
1755         if (err)
1756                 goto cleanup;
1757
1758         if (!nearex) {
1759                 /* there is no extent in this leaf, create first one */
1760                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1761                                 le32_to_cpu(newext->ee_block),
1762                                 ext4_ext_pblock(newext),
1763                                 ext4_ext_is_uninitialized(newext),
1764                                 ext4_ext_get_actual_len(newext));
1765                 nearex = EXT_FIRST_EXTENT(eh);
1766         } else {
1767                 if (le32_to_cpu(newext->ee_block)
1768                            > le32_to_cpu(nearex->ee_block)) {
1769                         /* Insert after */
1770                         ext_debug("insert %u:%llu:[%d]%d before: "
1771                                         "nearest %p\n",
1772                                         le32_to_cpu(newext->ee_block),
1773                                         ext4_ext_pblock(newext),
1774                                         ext4_ext_is_uninitialized(newext),
1775                                         ext4_ext_get_actual_len(newext),
1776                                         nearex);
1777                         nearex++;
1778                 } else {
1779                         /* Insert before */
1780                         BUG_ON(newext->ee_block == nearex->ee_block);
1781                         ext_debug("insert %u:%llu:[%d]%d after: "
1782                                         "nearest %p\n",
1783                                         le32_to_cpu(newext->ee_block),
1784                                         ext4_ext_pblock(newext),
1785                                         ext4_ext_is_uninitialized(newext),
1786                                         ext4_ext_get_actual_len(newext),
1787                                         nearex);
1788                 }
1789                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1790                 if (len > 0) {
1791                         ext_debug("insert %u:%llu:[%d]%d: "
1792                                         "move %d extents from 0x%p to 0x%p\n",
1793                                         le32_to_cpu(newext->ee_block),
1794                                         ext4_ext_pblock(newext),
1795                                         ext4_ext_is_uninitialized(newext),
1796                                         ext4_ext_get_actual_len(newext),
1797                                         len, nearex, nearex + 1);
1798                         memmove(nearex + 1, nearex,
1799                                 len * sizeof(struct ext4_extent));
1800                 }
1801         }
1802
1803         le16_add_cpu(&eh->eh_entries, 1);
1804         path[depth].p_ext = nearex;
1805         nearex->ee_block = newext->ee_block;
1806         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
1807         nearex->ee_len = newext->ee_len;
1808
1809 merge:
1810         /* try to merge extents to the right */
1811         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
1812                 ext4_ext_try_to_merge(inode, path, nearex);
1813
1814         /* try to merge extents to the left */
1815
1816         /* time to correct all indexes above */
1817         err = ext4_ext_correct_indexes(handle, inode, path);
1818         if (err)
1819                 goto cleanup;
1820
1821         err = ext4_ext_dirty(handle, inode, path + depth);
1822
1823 cleanup:
1824         if (npath) {
1825                 ext4_ext_drop_refs(npath);
1826                 kfree(npath);
1827         }
1828         ext4_ext_invalidate_cache(inode);
1829         return err;
1830 }
1831
1832 static int ext4_ext_walk_space(struct inode *inode, ext4_lblk_t block,
1833                                ext4_lblk_t num, ext_prepare_callback func,
1834                                void *cbdata)
1835 {
1836         struct ext4_ext_path *path = NULL;
1837         struct ext4_ext_cache cbex;
1838         struct ext4_extent *ex;
1839         ext4_lblk_t next, start = 0, end = 0;
1840         ext4_lblk_t last = block + num;
1841         int depth, exists, err = 0;
1842
1843         BUG_ON(func == NULL);
1844         BUG_ON(inode == NULL);
1845
1846         while (block < last && block != EXT_MAX_BLOCKS) {
1847                 num = last - block;
1848                 /* find extent for this block */
1849                 down_read(&EXT4_I(inode)->i_data_sem);
1850                 path = ext4_ext_find_extent(inode, block, path);
1851                 up_read(&EXT4_I(inode)->i_data_sem);
1852                 if (IS_ERR(path)) {
1853                         err = PTR_ERR(path);
1854                         path = NULL;
1855                         break;
1856                 }
1857
1858                 depth = ext_depth(inode);
1859                 if (unlikely(path[depth].p_hdr == NULL)) {
1860                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1861                         err = -EIO;
1862                         break;
1863                 }
1864                 ex = path[depth].p_ext;
1865                 next = ext4_ext_next_allocated_block(path);
1866
1867                 exists = 0;
1868                 if (!ex) {
1869                         /* there is no extent yet, so try to allocate
1870                          * all requested space */
1871                         start = block;
1872                         end = block + num;
1873                 } else if (le32_to_cpu(ex->ee_block) > block) {
1874                         /* need to allocate space before found extent */
1875                         start = block;
1876                         end = le32_to_cpu(ex->ee_block);
1877                         if (block + num < end)
1878                                 end = block + num;
1879                 } else if (block >= le32_to_cpu(ex->ee_block)
1880                                         + ext4_ext_get_actual_len(ex)) {
1881                         /* need to allocate space after found extent */
1882                         start = block;
1883                         end = block + num;
1884                         if (end >= next)
1885                                 end = next;
1886                 } else if (block >= le32_to_cpu(ex->ee_block)) {
1887                         /*
1888                          * some part of requested space is covered
1889                          * by found extent
1890                          */
1891                         start = block;
1892                         end = le32_to_cpu(ex->ee_block)
1893                                 + ext4_ext_get_actual_len(ex);
1894                         if (block + num < end)
1895                                 end = block + num;
1896                         exists = 1;
1897                 } else {
1898                         BUG();
1899                 }
1900                 BUG_ON(end <= start);
1901
1902                 if (!exists) {
1903                         cbex.ec_block = start;
1904                         cbex.ec_len = end - start;
1905                         cbex.ec_start = 0;
1906                 } else {
1907                         cbex.ec_block = le32_to_cpu(ex->ee_block);
1908                         cbex.ec_len = ext4_ext_get_actual_len(ex);
1909                         cbex.ec_start = ext4_ext_pblock(ex);
1910                 }
1911
1912                 if (unlikely(cbex.ec_len == 0)) {
1913                         EXT4_ERROR_INODE(inode, "cbex.ec_len == 0");
1914                         err = -EIO;
1915                         break;
1916                 }
1917                 err = func(inode, next, &cbex, ex, cbdata);
1918                 ext4_ext_drop_refs(path);
1919
1920                 if (err < 0)
1921                         break;
1922
1923                 if (err == EXT_REPEAT)
1924                         continue;
1925                 else if (err == EXT_BREAK) {
1926                         err = 0;
1927                         break;
1928                 }
1929
1930                 if (ext_depth(inode) != depth) {
1931                         /* depth was changed. we have to realloc path */
1932                         kfree(path);
1933                         path = NULL;
1934                 }
1935
1936                 block = cbex.ec_block + cbex.ec_len;
1937         }
1938
1939         if (path) {
1940                 ext4_ext_drop_refs(path);
1941                 kfree(path);
1942         }
1943
1944         return err;
1945 }
1946
1947 static void
1948 ext4_ext_put_in_cache(struct inode *inode, ext4_lblk_t block,
1949                         __u32 len, ext4_fsblk_t start)
1950 {
1951         struct ext4_ext_cache *cex;
1952         BUG_ON(len == 0);
1953         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1954         trace_ext4_ext_put_in_cache(inode, block, len, start);
1955         cex = &EXT4_I(inode)->i_cached_extent;
1956         cex->ec_block = block;
1957         cex->ec_len = len;
1958         cex->ec_start = start;
1959         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1960 }
1961
1962 /*
1963  * ext4_ext_put_gap_in_cache:
1964  * calculate boundaries of the gap that the requested block fits into
1965  * and cache this gap
1966  */
1967 static void
1968 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
1969                                 ext4_lblk_t block)
1970 {
1971         int depth = ext_depth(inode);
1972         unsigned long len;
1973         ext4_lblk_t lblock;
1974         struct ext4_extent *ex;
1975
1976         ex = path[depth].p_ext;
1977         if (ex == NULL) {
1978                 /* there is no extent yet, so gap is [0;-] */
1979                 lblock = 0;
1980                 len = EXT_MAX_BLOCKS;
1981                 ext_debug("cache gap(whole file):");
1982         } else if (block < le32_to_cpu(ex->ee_block)) {
1983                 lblock = block;
1984                 len = le32_to_cpu(ex->ee_block) - block;
1985                 ext_debug("cache gap(before): %u [%u:%u]",
1986                                 block,
1987                                 le32_to_cpu(ex->ee_block),
1988                                  ext4_ext_get_actual_len(ex));
1989         } else if (block >= le32_to_cpu(ex->ee_block)
1990                         + ext4_ext_get_actual_len(ex)) {
1991                 ext4_lblk_t next;
1992                 lblock = le32_to_cpu(ex->ee_block)
1993                         + ext4_ext_get_actual_len(ex);
1994
1995                 next = ext4_ext_next_allocated_block(path);
1996                 ext_debug("cache gap(after): [%u:%u] %u",
1997                                 le32_to_cpu(ex->ee_block),
1998                                 ext4_ext_get_actual_len(ex),
1999                                 block);
2000                 BUG_ON(next == lblock);
2001                 len = next - lblock;
2002         } else {
2003                 lblock = len = 0;
2004                 BUG();
2005         }
2006
2007         ext_debug(" -> %u:%lu\n", lblock, len);
2008         ext4_ext_put_in_cache(inode, lblock, len, 0);
2009 }
2010
2011 /*
2012  * ext4_ext_check_cache()
2013  * Checks to see if the given block is in the cache.
2014  * If it is, the cached extent is stored in the given
2015  * cache extent pointer.  If the cached extent is a hole,
2016  * this routine should be used instead of
2017  * ext4_ext_in_cache if the calling function needs to
2018  * know the size of the hole.
2019  *
2020  * @inode: The files inode
2021  * @block: The block to look for in the cache
2022  * @ex:    Pointer where the cached extent will be stored
2023  *         if it contains block
2024  *
2025  * Return 0 if cache is invalid; 1 if the cache is valid
2026  */
2027 static int ext4_ext_check_cache(struct inode *inode, ext4_lblk_t block,
2028         struct ext4_ext_cache *ex){
2029         struct ext4_ext_cache *cex;
2030         struct ext4_sb_info *sbi;
2031         int ret = 0;
2032
2033         /*
2034          * We borrow i_block_reservation_lock to protect i_cached_extent
2035          */
2036         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
2037         cex = &EXT4_I(inode)->i_cached_extent;
2038         sbi = EXT4_SB(inode->i_sb);
2039
2040         /* has cache valid data? */
2041         if (cex->ec_len == 0)
2042                 goto errout;
2043
2044         if (in_range(block, cex->ec_block, cex->ec_len)) {
2045                 memcpy(ex, cex, sizeof(struct ext4_ext_cache));
2046                 ext_debug("%u cached by %u:%u:%llu\n",
2047                                 block,
2048                                 cex->ec_block, cex->ec_len, cex->ec_start);
2049                 ret = 1;
2050         }
2051 errout:
2052         if (!ret)
2053                 sbi->extent_cache_misses++;
2054         else
2055                 sbi->extent_cache_hits++;
2056         trace_ext4_ext_in_cache(inode, block, ret);
2057         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
2058         return ret;
2059 }
2060
2061 /*
2062  * ext4_ext_in_cache()
2063  * Checks to see if the given block is in the cache.
2064  * If it is, the cached extent is stored in the given
2065  * extent pointer.
2066  *
2067  * @inode: The files inode
2068  * @block: The block to look for in the cache
2069  * @ex:    Pointer where the cached extent will be stored
2070  *         if it contains block
2071  *
2072  * Return 0 if cache is invalid; 1 if the cache is valid
2073  */
2074 static int
2075 ext4_ext_in_cache(struct inode *inode, ext4_lblk_t block,
2076                         struct ext4_extent *ex)
2077 {
2078         struct ext4_ext_cache cex;
2079         int ret = 0;
2080
2081         if (ext4_ext_check_cache(inode, block, &cex)) {
2082                 ex->ee_block = cpu_to_le32(cex.ec_block);
2083                 ext4_ext_store_pblock(ex, cex.ec_start);
2084                 ex->ee_len = cpu_to_le16(cex.ec_len);
2085                 ret = 1;
2086         }
2087
2088         return ret;
2089 }
2090
2091
2092 /*
2093  * ext4_ext_rm_idx:
2094  * removes index from the index block.
2095  */
2096 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2097                         struct ext4_ext_path *path)
2098 {
2099         int err;
2100         ext4_fsblk_t leaf;
2101
2102         /* free index block */
2103         path--;
2104         leaf = ext4_idx_pblock(path->p_idx);
2105         if (unlikely(path->p_hdr->eh_entries == 0)) {
2106                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2107                 return -EIO;
2108         }
2109         err = ext4_ext_get_access(handle, inode, path);
2110         if (err)
2111                 return err;
2112
2113         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2114                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2115                 len *= sizeof(struct ext4_extent_idx);
2116                 memmove(path->p_idx, path->p_idx + 1, len);
2117         }
2118
2119         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2120         err = ext4_ext_dirty(handle, inode, path);
2121         if (err)
2122                 return err;
2123         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2124         trace_ext4_ext_rm_idx(inode, leaf);
2125
2126         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2127                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2128         return err;
2129 }
2130
2131 /*
2132  * ext4_ext_calc_credits_for_single_extent:
2133  * This routine returns max. credits that needed to insert an extent
2134  * to the extent tree.
2135  * When pass the actual path, the caller should calculate credits
2136  * under i_data_sem.
2137  */
2138 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2139                                                 struct ext4_ext_path *path)
2140 {
2141         if (path) {
2142                 int depth = ext_depth(inode);
2143                 int ret = 0;
2144
2145                 /* probably there is space in leaf? */
2146                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2147                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2148
2149                         /*
2150                          *  There are some space in the leaf tree, no
2151                          *  need to account for leaf block credit
2152                          *
2153                          *  bitmaps and block group descriptor blocks
2154                          *  and other metadata blocks still need to be
2155                          *  accounted.
2156                          */
2157                         /* 1 bitmap, 1 block group descriptor */
2158                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2159                         return ret;
2160                 }
2161         }
2162
2163         return ext4_chunk_trans_blocks(inode, nrblocks);
2164 }
2165
2166 /*
2167  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2168  *
2169  * if nrblocks are fit in a single extent (chunk flag is 1), then
2170  * in the worse case, each tree level index/leaf need to be changed
2171  * if the tree split due to insert a new extent, then the old tree
2172  * index/leaf need to be updated too
2173  *
2174  * If the nrblocks are discontiguous, they could cause
2175  * the whole tree split more than once, but this is really rare.
2176  */
2177 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2178 {
2179         int index;
2180         int depth = ext_depth(inode);
2181
2182         if (chunk)
2183                 index = depth * 2;
2184         else
2185                 index = depth * 3;
2186
2187         return index;
2188 }
2189
2190 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2191                               struct ext4_extent *ex,
2192                               ext4_fsblk_t *partial_cluster,
2193                               ext4_lblk_t from, ext4_lblk_t to)
2194 {
2195         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2196         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2197         ext4_fsblk_t pblk;
2198         int flags = EXT4_FREE_BLOCKS_FORGET;
2199
2200         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2201                 flags |= EXT4_FREE_BLOCKS_METADATA;
2202         /*
2203          * For bigalloc file systems, we never free a partial cluster
2204          * at the beginning of the extent.  Instead, we make a note
2205          * that we tried freeing the cluster, and check to see if we
2206          * need to free it on a subsequent call to ext4_remove_blocks,
2207          * or at the end of the ext4_truncate() operation.
2208          */
2209         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2210
2211         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2212         /*
2213          * If we have a partial cluster, and it's different from the
2214          * cluster of the last block, we need to explicitly free the
2215          * partial cluster here.
2216          */
2217         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2218         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2219                 ext4_free_blocks(handle, inode, NULL,
2220                                  EXT4_C2B(sbi, *partial_cluster),
2221                                  sbi->s_cluster_ratio, flags);
2222                 *partial_cluster = 0;
2223         }
2224
2225 #ifdef EXTENTS_STATS
2226         {
2227                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2228                 spin_lock(&sbi->s_ext_stats_lock);
2229                 sbi->s_ext_blocks += ee_len;
2230                 sbi->s_ext_extents++;
2231                 if (ee_len < sbi->s_ext_min)
2232                         sbi->s_ext_min = ee_len;
2233                 if (ee_len > sbi->s_ext_max)
2234                         sbi->s_ext_max = ee_len;
2235                 if (ext_depth(inode) > sbi->s_depth_max)
2236                         sbi->s_depth_max = ext_depth(inode);
2237                 spin_unlock(&sbi->s_ext_stats_lock);
2238         }
2239 #endif
2240         if (from >= le32_to_cpu(ex->ee_block)
2241             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2242                 /* tail removal */
2243                 ext4_lblk_t num;
2244
2245                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2246                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2247                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2248                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2249                 /*
2250                  * If the block range to be freed didn't start at the
2251                  * beginning of a cluster, and we removed the entire
2252                  * extent, save the partial cluster here, since we
2253                  * might need to delete if we determine that the
2254                  * truncate operation has removed all of the blocks in
2255                  * the cluster.
2256                  */
2257                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2258                     (ee_len == num))
2259                         *partial_cluster = EXT4_B2C(sbi, pblk);
2260                 else
2261                         *partial_cluster = 0;
2262         } else if (from == le32_to_cpu(ex->ee_block)
2263                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2264                 /* head removal */
2265                 ext4_lblk_t num;
2266                 ext4_fsblk_t start;
2267
2268                 num = to - from;
2269                 start = ext4_ext_pblock(ex);
2270
2271                 ext_debug("free first %u blocks starting %llu\n", num, start);
2272                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2273
2274         } else {
2275                 printk(KERN_INFO "strange request: removal(2) "
2276                                 "%u-%u from %u:%u\n",
2277                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2278         }
2279         return 0;
2280 }
2281
2282
2283 /*
2284  * ext4_ext_rm_leaf() Removes the extents associated with the
2285  * blocks appearing between "start" and "end", and splits the extents
2286  * if "start" and "end" appear in the same extent
2287  *
2288  * @handle: The journal handle
2289  * @inode:  The files inode
2290  * @path:   The path to the leaf
2291  * @start:  The first block to remove
2292  * @end:   The last block to remove
2293  */
2294 static int
2295 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2296                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2297                  ext4_lblk_t start, ext4_lblk_t end)
2298 {
2299         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2300         int err = 0, correct_index = 0;
2301         int depth = ext_depth(inode), credits;
2302         struct ext4_extent_header *eh;
2303         ext4_lblk_t a, b;
2304         unsigned num;
2305         ext4_lblk_t ex_ee_block;
2306         unsigned short ex_ee_len;
2307         unsigned uninitialized = 0;
2308         struct ext4_extent *ex;
2309
2310         /* the header must be checked already in ext4_ext_remove_space() */
2311         ext_debug("truncate since %u in leaf\n", start);
2312         if (!path[depth].p_hdr)
2313                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2314         eh = path[depth].p_hdr;
2315         if (unlikely(path[depth].p_hdr == NULL)) {
2316                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2317                 return -EIO;
2318         }
2319         /* find where to start removing */
2320         ex = EXT_LAST_EXTENT(eh);
2321
2322         ex_ee_block = le32_to_cpu(ex->ee_block);
2323         ex_ee_len = ext4_ext_get_actual_len(ex);
2324
2325         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2326
2327         while (ex >= EXT_FIRST_EXTENT(eh) &&
2328                         ex_ee_block + ex_ee_len > start) {
2329
2330                 if (ext4_ext_is_uninitialized(ex))
2331                         uninitialized = 1;
2332                 else
2333                         uninitialized = 0;
2334
2335                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2336                          uninitialized, ex_ee_len);
2337                 path[depth].p_ext = ex;
2338
2339                 a = ex_ee_block > start ? ex_ee_block : start;
2340                 b = ex_ee_block+ex_ee_len - 1 < end ?
2341                         ex_ee_block+ex_ee_len - 1 : end;
2342
2343                 ext_debug("  border %u:%u\n", a, b);
2344
2345                 /* If this extent is beyond the end of the hole, skip it */
2346                 if (end <= ex_ee_block) {
2347                         ex--;
2348                         ex_ee_block = le32_to_cpu(ex->ee_block);
2349                         ex_ee_len = ext4_ext_get_actual_len(ex);
2350                         continue;
2351                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2352                         EXT4_ERROR_INODE(inode,"  bad truncate %u:%u\n",
2353                                          start, end);
2354                         err = -EIO;
2355                         goto out;
2356                 } else if (a != ex_ee_block) {
2357                         /* remove tail of the extent */
2358                         num = a - ex_ee_block;
2359                 } else {
2360                         /* remove whole extent: excellent! */
2361                         num = 0;
2362                 }
2363                 /*
2364                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2365                  * descriptor) for each block group; assume two block
2366                  * groups plus ex_ee_len/blocks_per_block_group for
2367                  * the worst case
2368                  */
2369                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2370                 if (ex == EXT_FIRST_EXTENT(eh)) {
2371                         correct_index = 1;
2372                         credits += (ext_depth(inode)) + 1;
2373                 }
2374                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2375
2376                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2377                 if (err)
2378                         goto out;
2379
2380                 err = ext4_ext_get_access(handle, inode, path + depth);
2381                 if (err)
2382                         goto out;
2383
2384                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2385                                          a, b);
2386                 if (err)
2387                         goto out;
2388
2389                 if (num == 0)
2390                         /* this extent is removed; mark slot entirely unused */
2391                         ext4_ext_store_pblock(ex, 0);
2392
2393                 ex->ee_len = cpu_to_le16(num);
2394                 /*
2395                  * Do not mark uninitialized if all the blocks in the
2396                  * extent have been removed.
2397                  */
2398                 if (uninitialized && num)
2399                         ext4_ext_mark_uninitialized(ex);
2400                 /*
2401                  * If the extent was completely released,
2402                  * we need to remove it from the leaf
2403                  */
2404                 if (num == 0) {
2405                         if (end != EXT_MAX_BLOCKS - 1) {
2406                                 /*
2407                                  * For hole punching, we need to scoot all the
2408                                  * extents up when an extent is removed so that
2409                                  * we dont have blank extents in the middle
2410                                  */
2411                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2412                                         sizeof(struct ext4_extent));
2413
2414                                 /* Now get rid of the one at the end */
2415                                 memset(EXT_LAST_EXTENT(eh), 0,
2416                                         sizeof(struct ext4_extent));
2417                         }
2418                         le16_add_cpu(&eh->eh_entries, -1);
2419                 } else
2420                         *partial_cluster = 0;
2421
2422                 err = ext4_ext_dirty(handle, inode, path + depth);
2423                 if (err)
2424                         goto out;
2425
2426                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2427                                 ext4_ext_pblock(ex));
2428                 ex--;
2429                 ex_ee_block = le32_to_cpu(ex->ee_block);
2430                 ex_ee_len = ext4_ext_get_actual_len(ex);
2431         }
2432
2433         if (correct_index && eh->eh_entries)
2434                 err = ext4_ext_correct_indexes(handle, inode, path);
2435
2436         /*
2437          * If there is still a entry in the leaf node, check to see if
2438          * it references the partial cluster.  This is the only place
2439          * where it could; if it doesn't, we can free the cluster.
2440          */
2441         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2442             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2443              *partial_cluster)) {
2444                 int flags = EXT4_FREE_BLOCKS_FORGET;
2445
2446                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2447                         flags |= EXT4_FREE_BLOCKS_METADATA;
2448
2449                 ext4_free_blocks(handle, inode, NULL,
2450                                  EXT4_C2B(sbi, *partial_cluster),
2451                                  sbi->s_cluster_ratio, flags);
2452                 *partial_cluster = 0;
2453         }
2454
2455         /* if this leaf is free, then we should
2456          * remove it from index block above */
2457         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2458                 err = ext4_ext_rm_idx(handle, inode, path + depth);
2459
2460 out:
2461         return err;
2462 }
2463
2464 /*
2465  * ext4_ext_more_to_rm:
2466  * returns 1 if current index has to be freed (even partial)
2467  */
2468 static int
2469 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2470 {
2471         BUG_ON(path->p_idx == NULL);
2472
2473         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2474                 return 0;
2475
2476         /*
2477          * if truncate on deeper level happened, it wasn't partial,
2478          * so we have to consider current index for truncation
2479          */
2480         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2481                 return 0;
2482         return 1;
2483 }
2484
2485 static int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start)
2486 {
2487         struct super_block *sb = inode->i_sb;
2488         int depth = ext_depth(inode);
2489         struct ext4_ext_path *path;
2490         ext4_fsblk_t partial_cluster = 0;
2491         handle_t *handle;
2492         int i, err;
2493
2494         ext_debug("truncate since %u\n", start);
2495
2496         /* probably first extent we're gonna free will be last in block */
2497         handle = ext4_journal_start(inode, depth + 1);
2498         if (IS_ERR(handle))
2499                 return PTR_ERR(handle);
2500
2501 again:
2502         ext4_ext_invalidate_cache(inode);
2503
2504         trace_ext4_ext_remove_space(inode, start, depth);
2505
2506         /*
2507          * We start scanning from right side, freeing all the blocks
2508          * after i_size and walking into the tree depth-wise.
2509          */
2510         depth = ext_depth(inode);
2511         path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1), GFP_NOFS);
2512         if (path == NULL) {
2513                 ext4_journal_stop(handle);
2514                 return -ENOMEM;
2515         }
2516         path[0].p_depth = depth;
2517         path[0].p_hdr = ext_inode_hdr(inode);
2518         if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2519                 err = -EIO;
2520                 goto out;
2521         }
2522         i = err = 0;
2523
2524         while (i >= 0 && err == 0) {
2525                 if (i == depth) {
2526                         /* this is leaf block */
2527                         err = ext4_ext_rm_leaf(handle, inode, path,
2528                                                &partial_cluster, start,
2529                                                EXT_MAX_BLOCKS - 1);
2530                         /* root level has p_bh == NULL, brelse() eats this */
2531                         brelse(path[i].p_bh);
2532                         path[i].p_bh = NULL;
2533                         i--;
2534                         continue;
2535                 }
2536
2537                 /* this is index block */
2538                 if (!path[i].p_hdr) {
2539                         ext_debug("initialize header\n");
2540                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2541                 }
2542
2543                 if (!path[i].p_idx) {
2544                         /* this level hasn't been touched yet */
2545                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2546                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2547                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2548                                   path[i].p_hdr,
2549                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2550                 } else {
2551                         /* we were already here, see at next index */
2552                         path[i].p_idx--;
2553                 }
2554
2555                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2556                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2557                                 path[i].p_idx);
2558                 if (ext4_ext_more_to_rm(path + i)) {
2559                         struct buffer_head *bh;
2560                         /* go to the next level */
2561                         ext_debug("move to level %d (block %llu)\n",
2562                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2563                         memset(path + i + 1, 0, sizeof(*path));
2564                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2565                         if (!bh) {
2566                                 /* should we reset i_size? */
2567                                 err = -EIO;
2568                                 break;
2569                         }
2570                         if (WARN_ON(i + 1 > depth)) {
2571                                 err = -EIO;
2572                                 break;
2573                         }
2574                         if (ext4_ext_check(inode, ext_block_hdr(bh),
2575                                                         depth - i - 1)) {
2576                                 err = -EIO;
2577                                 break;
2578                         }
2579                         path[i + 1].p_bh = bh;
2580
2581                         /* save actual number of indexes since this
2582                          * number is changed at the next iteration */
2583                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2584                         i++;
2585                 } else {
2586                         /* we finished processing this index, go up */
2587                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2588                                 /* index is empty, remove it;
2589                                  * handle must be already prepared by the
2590                                  * truncatei_leaf() */
2591                                 err = ext4_ext_rm_idx(handle, inode, path + i);
2592                         }
2593                         /* root level has p_bh == NULL, brelse() eats this */
2594                         brelse(path[i].p_bh);
2595                         path[i].p_bh = NULL;
2596                         i--;
2597                         ext_debug("return to level %d\n", i);
2598                 }
2599         }
2600
2601         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2602                         path->p_hdr->eh_entries);
2603
2604         /* If we still have something in the partial cluster and we have removed
2605          * even the first extent, then we should free the blocks in the partial
2606          * cluster as well. */
2607         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2608                 int flags = EXT4_FREE_BLOCKS_FORGET;
2609
2610                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2611                         flags |= EXT4_FREE_BLOCKS_METADATA;
2612
2613                 ext4_free_blocks(handle, inode, NULL,
2614                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2615                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2616                 partial_cluster = 0;
2617         }
2618
2619         /* TODO: flexible tree reduction should be here */
2620         if (path->p_hdr->eh_entries == 0) {
2621                 /*
2622                  * truncate to zero freed all the tree,
2623                  * so we need to correct eh_depth
2624                  */
2625                 err = ext4_ext_get_access(handle, inode, path);
2626                 if (err == 0) {
2627                         ext_inode_hdr(inode)->eh_depth = 0;
2628                         ext_inode_hdr(inode)->eh_max =
2629                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2630                         err = ext4_ext_dirty(handle, inode, path);
2631                 }
2632         }
2633 out:
2634         ext4_ext_drop_refs(path);
2635         kfree(path);
2636         if (err == -EAGAIN)
2637                 goto again;
2638         ext4_journal_stop(handle);
2639
2640         return err;
2641 }
2642
2643 /*
2644  * called at mount time
2645  */
2646 void ext4_ext_init(struct super_block *sb)
2647 {
2648         /*
2649          * possible initialization would be here
2650          */
2651
2652         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2653 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2654                 printk(KERN_INFO "EXT4-fs: file extents enabled");
2655 #ifdef AGGRESSIVE_TEST
2656                 printk(", aggressive tests");
2657 #endif
2658 #ifdef CHECK_BINSEARCH
2659                 printk(", check binsearch");
2660 #endif
2661 #ifdef EXTENTS_STATS
2662                 printk(", stats");
2663 #endif
2664                 printk("\n");
2665 #endif
2666 #ifdef EXTENTS_STATS
2667                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2668                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2669                 EXT4_SB(sb)->s_ext_max = 0;
2670 #endif
2671         }
2672 }
2673
2674 /*
2675  * called at umount time
2676  */
2677 void ext4_ext_release(struct super_block *sb)
2678 {
2679         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2680                 return;
2681
2682 #ifdef EXTENTS_STATS
2683         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2684                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2685                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2686                         sbi->s_ext_blocks, sbi->s_ext_extents,
2687                         sbi->s_ext_blocks / sbi->s_ext_extents);
2688                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2689                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2690         }
2691 #endif
2692 }
2693
2694 /* FIXME!! we need to try to merge to left or right after zero-out  */
2695 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2696 {
2697         ext4_fsblk_t ee_pblock;
2698         unsigned int ee_len;
2699         int ret;
2700
2701         ee_len    = ext4_ext_get_actual_len(ex);
2702         ee_pblock = ext4_ext_pblock(ex);
2703
2704         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2705         if (ret > 0)
2706                 ret = 0;
2707
2708         return ret;
2709 }
2710
2711 /*
2712  * used by extent splitting.
2713  */
2714 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
2715                                         due to ENOSPC */
2716 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
2717 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
2718
2719 /*
2720  * ext4_split_extent_at() splits an extent at given block.
2721  *
2722  * @handle: the journal handle
2723  * @inode: the file inode
2724  * @path: the path to the extent
2725  * @split: the logical block where the extent is splitted.
2726  * @split_flags: indicates if the extent could be zeroout if split fails, and
2727  *               the states(init or uninit) of new extents.
2728  * @flags: flags used to insert new extent to extent tree.
2729  *
2730  *
2731  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2732  * of which are deterimined by split_flag.
2733  *
2734  * There are two cases:
2735  *  a> the extent are splitted into two extent.
2736  *  b> split is not needed, and just mark the extent.
2737  *
2738  * return 0 on success.
2739  */
2740 static int ext4_split_extent_at(handle_t *handle,
2741                              struct inode *inode,
2742                              struct ext4_ext_path *path,
2743                              ext4_lblk_t split,
2744                              int split_flag,
2745                              int flags)
2746 {
2747         ext4_fsblk_t newblock;
2748         ext4_lblk_t ee_block;
2749         struct ext4_extent *ex, newex, orig_ex;
2750         struct ext4_extent *ex2 = NULL;
2751         unsigned int ee_len, depth;
2752         int err = 0;
2753
2754         ext_debug("ext4_split_extents_at: inode %lu, logical"
2755                 "block %llu\n", inode->i_ino, (unsigned long long)split);
2756
2757         ext4_ext_show_leaf(inode, path);
2758
2759         depth = ext_depth(inode);
2760         ex = path[depth].p_ext;
2761         ee_block = le32_to_cpu(ex->ee_block);
2762         ee_len = ext4_ext_get_actual_len(ex);
2763         newblock = split - ee_block + ext4_ext_pblock(ex);
2764
2765         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
2766
2767         err = ext4_ext_get_access(handle, inode, path + depth);
2768         if (err)
2769                 goto out;
2770
2771         if (split == ee_block) {
2772                 /*
2773                  * case b: block @split is the block that the extent begins with
2774                  * then we just change the state of the extent, and splitting
2775                  * is not needed.
2776                  */
2777                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2778                         ext4_ext_mark_uninitialized(ex);
2779                 else
2780                         ext4_ext_mark_initialized(ex);
2781
2782                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
2783                         ext4_ext_try_to_merge(inode, path, ex);
2784
2785                 err = ext4_ext_dirty(handle, inode, path + depth);
2786                 goto out;
2787         }
2788
2789         /* case a */
2790         memcpy(&orig_ex, ex, sizeof(orig_ex));
2791         ex->ee_len = cpu_to_le16(split - ee_block);
2792         if (split_flag & EXT4_EXT_MARK_UNINIT1)
2793                 ext4_ext_mark_uninitialized(ex);
2794
2795         /*
2796          * path may lead to new leaf, not to original leaf any more
2797          * after ext4_ext_insert_extent() returns,
2798          */
2799         err = ext4_ext_dirty(handle, inode, path + depth);
2800         if (err)
2801                 goto fix_extent_len;
2802
2803         ex2 = &newex;
2804         ex2->ee_block = cpu_to_le32(split);
2805         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
2806         ext4_ext_store_pblock(ex2, newblock);
2807         if (split_flag & EXT4_EXT_MARK_UNINIT2)
2808                 ext4_ext_mark_uninitialized(ex2);
2809
2810         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
2811         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
2812                 err = ext4_ext_zeroout(inode, &orig_ex);
2813                 if (err)
2814                         goto fix_extent_len;
2815                 /* update the extent length and mark as initialized */
2816                 ex->ee_len = cpu_to_le32(ee_len);
2817                 ext4_ext_try_to_merge(inode, path, ex);
2818                 err = ext4_ext_dirty(handle, inode, path + depth);
2819                 goto out;
2820         } else if (err)
2821                 goto fix_extent_len;
2822
2823 out:
2824         ext4_ext_show_leaf(inode, path);
2825         return err;
2826
2827 fix_extent_len:
2828         ex->ee_len = orig_ex.ee_len;
2829         ext4_ext_dirty(handle, inode, path + depth);
2830         return err;
2831 }
2832
2833 /*
2834  * ext4_split_extents() splits an extent and mark extent which is covered
2835  * by @map as split_flags indicates
2836  *
2837  * It may result in splitting the extent into multiple extents (upto three)
2838  * There are three possibilities:
2839  *   a> There is no split required
2840  *   b> Splits in two extents: Split is happening at either end of the extent
2841  *   c> Splits in three extents: Somone is splitting in middle of the extent
2842  *
2843  */
2844 static int ext4_split_extent(handle_t *handle,
2845                               struct inode *inode,
2846                               struct ext4_ext_path *path,
2847                               struct ext4_map_blocks *map,
2848                               int split_flag,
2849                               int flags)
2850 {
2851         ext4_lblk_t ee_block;
2852         struct ext4_extent *ex;
2853         unsigned int ee_len, depth;
2854         int err = 0;
2855         int uninitialized;
2856         int split_flag1, flags1;
2857
2858         depth = ext_depth(inode);
2859         ex = path[depth].p_ext;
2860         ee_block = le32_to_cpu(ex->ee_block);
2861         ee_len = ext4_ext_get_actual_len(ex);
2862         uninitialized = ext4_ext_is_uninitialized(ex);
2863
2864         if (map->m_lblk + map->m_len < ee_block + ee_len) {
2865                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT ?
2866                               EXT4_EXT_MAY_ZEROOUT : 0;
2867                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
2868                 if (uninitialized)
2869                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
2870                                        EXT4_EXT_MARK_UNINIT2;
2871                 err = ext4_split_extent_at(handle, inode, path,
2872                                 map->m_lblk + map->m_len, split_flag1, flags1);
2873                 if (err)
2874                         goto out;
2875         }
2876
2877         ext4_ext_drop_refs(path);
2878         path = ext4_ext_find_extent(inode, map->m_lblk, path);
2879         if (IS_ERR(path))
2880                 return PTR_ERR(path);
2881
2882         if (map->m_lblk >= ee_block) {
2883                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT ?
2884                               EXT4_EXT_MAY_ZEROOUT : 0;
2885                 if (uninitialized)
2886                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
2887                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
2888                         split_flag1 |= EXT4_EXT_MARK_UNINIT2;
2889                 err = ext4_split_extent_at(handle, inode, path,
2890                                 map->m_lblk, split_flag1, flags);
2891                 if (err)
2892                         goto out;
2893         }
2894
2895         ext4_ext_show_leaf(inode, path);
2896 out:
2897         return err ? err : map->m_len;
2898 }
2899
2900 #define EXT4_EXT_ZERO_LEN 7
2901 /*
2902  * This function is called by ext4_ext_map_blocks() if someone tries to write
2903  * to an uninitialized extent. It may result in splitting the uninitialized
2904  * extent into multiple extents (up to three - one initialized and two
2905  * uninitialized).
2906  * There are three possibilities:
2907  *   a> There is no split required: Entire extent should be initialized
2908  *   b> Splits in two extents: Write is happening at either end of the extent
2909  *   c> Splits in three extents: Somone is writing in middle of the extent
2910  *
2911  * Pre-conditions:
2912  *  - The extent pointed to by 'path' is uninitialized.
2913  *  - The extent pointed to by 'path' contains a superset
2914  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
2915  *
2916  * Post-conditions on success:
2917  *  - the returned value is the number of blocks beyond map->l_lblk
2918  *    that are allocated and initialized.
2919  *    It is guaranteed to be >= map->m_len.
2920  */
2921 static int ext4_ext_convert_to_initialized(handle_t *handle,
2922                                            struct inode *inode,
2923                                            struct ext4_map_blocks *map,
2924                                            struct ext4_ext_path *path)
2925 {
2926         struct ext4_extent_header *eh;
2927         struct ext4_map_blocks split_map;
2928         struct ext4_extent zero_ex;
2929         struct ext4_extent *ex;
2930         ext4_lblk_t ee_block, eof_block;
2931         unsigned int ee_len, depth;
2932         int allocated;
2933         int err = 0;
2934         int split_flag = 0;
2935
2936         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
2937                 "block %llu, max_blocks %u\n", inode->i_ino,
2938                 (unsigned long long)map->m_lblk, map->m_len);
2939
2940         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
2941                 inode->i_sb->s_blocksize_bits;
2942         if (eof_block < map->m_lblk + map->m_len)
2943                 eof_block = map->m_lblk + map->m_len;
2944
2945         depth = ext_depth(inode);
2946         eh = path[depth].p_hdr;
2947         ex = path[depth].p_ext;
2948         ee_block = le32_to_cpu(ex->ee_block);
2949         ee_len = ext4_ext_get_actual_len(ex);
2950         allocated = ee_len - (map->m_lblk - ee_block);
2951
2952         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
2953
2954         /* Pre-conditions */
2955         BUG_ON(!ext4_ext_is_uninitialized(ex));
2956         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
2957
2958         /*
2959          * Attempt to transfer newly initialized blocks from the currently
2960          * uninitialized extent to its left neighbor. This is much cheaper
2961          * than an insertion followed by a merge as those involve costly
2962          * memmove() calls. This is the common case in steady state for
2963          * workloads doing fallocate(FALLOC_FL_KEEP_SIZE) followed by append
2964          * writes.
2965          *
2966          * Limitations of the current logic:
2967          *  - L1: we only deal with writes at the start of the extent.
2968          *    The approach could be extended to writes at the end
2969          *    of the extent but this scenario was deemed less common.
2970          *  - L2: we do not deal with writes covering the whole extent.
2971          *    This would require removing the extent if the transfer
2972          *    is possible.
2973          *  - L3: we only attempt to merge with an extent stored in the
2974          *    same extent tree node.
2975          */
2976         if ((map->m_lblk == ee_block) &&        /*L1*/
2977                 (map->m_len < ee_len) &&        /*L2*/
2978                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L3*/
2979                 struct ext4_extent *prev_ex;
2980                 ext4_lblk_t prev_lblk;
2981                 ext4_fsblk_t prev_pblk, ee_pblk;
2982                 unsigned int prev_len, write_len;
2983
2984                 prev_ex = ex - 1;
2985                 prev_lblk = le32_to_cpu(prev_ex->ee_block);
2986                 prev_len = ext4_ext_get_actual_len(prev_ex);
2987                 prev_pblk = ext4_ext_pblock(prev_ex);
2988                 ee_pblk = ext4_ext_pblock(ex);
2989                 write_len = map->m_len;
2990
2991                 /*
2992                  * A transfer of blocks from 'ex' to 'prev_ex' is allowed
2993                  * upon those conditions:
2994                  * - C1: prev_ex is initialized,
2995                  * - C2: prev_ex is logically abutting ex,
2996                  * - C3: prev_ex is physically abutting ex,
2997                  * - C4: prev_ex can receive the additional blocks without
2998                  *   overflowing the (initialized) length limit.
2999                  */
3000                 if ((!ext4_ext_is_uninitialized(prev_ex)) &&            /*C1*/
3001                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3002                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3003                         (prev_len < (EXT_INIT_MAX_LEN - write_len))) {  /*C4*/
3004                         err = ext4_ext_get_access(handle, inode, path + depth);
3005                         if (err)
3006                                 goto out;
3007
3008                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3009                                 map, ex, prev_ex);
3010
3011                         /* Shift the start of ex by 'write_len' blocks */
3012                         ex->ee_block = cpu_to_le32(ee_block + write_len);
3013                         ext4_ext_store_pblock(ex, ee_pblk + write_len);
3014                         ex->ee_len = cpu_to_le16(ee_len - write_len);
3015                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3016
3017                         /* Extend prev_ex by 'write_len' blocks */
3018                         prev_ex->ee_len = cpu_to_le16(prev_len + write_len);
3019
3020                         /* Mark the block containing both extents as dirty */
3021                         ext4_ext_dirty(handle, inode, path + depth);
3022
3023                         /* Update path to point to the right extent */
3024                         path[depth].p_ext = prev_ex;
3025
3026                         /* Result: number of initialized blocks past m_lblk */
3027                         allocated = write_len;
3028                         goto out;
3029                 }
3030         }
3031
3032         WARN_ON(map->m_lblk < ee_block);
3033         /*
3034          * It is safe to convert extent to initialized via explicit
3035          * zeroout only if extent is fully insde i_size or new_size.
3036          */
3037         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3038
3039         /* If extent has less than 2*EXT4_EXT_ZERO_LEN zerout directly */
3040         if (ee_len <= 2*EXT4_EXT_ZERO_LEN &&
3041             (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3042                 err = ext4_ext_zeroout(inode, ex);
3043                 if (err)
3044                         goto out;
3045
3046                 err = ext4_ext_get_access(handle, inode, path + depth);
3047                 if (err)
3048                         goto out;
3049                 ext4_ext_mark_initialized(ex);
3050                 ext4_ext_try_to_merge(inode, path, ex);
3051                 err = ext4_ext_dirty(handle, inode, path + depth);
3052                 goto out;
3053         }
3054
3055         /*
3056          * four cases:
3057          * 1. split the extent into three extents.
3058          * 2. split the extent into two extents, zeroout the first half.
3059          * 3. split the extent into two extents, zeroout the second half.
3060          * 4. split the extent into two extents with out zeroout.
3061          */
3062         split_map.m_lblk = map->m_lblk;
3063         split_map.m_len = map->m_len;
3064
3065         if (allocated > map->m_len) {
3066                 if (allocated <= EXT4_EXT_ZERO_LEN &&
3067                     (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3068                         /* case 3 */
3069                         zero_ex.ee_block =
3070                                          cpu_to_le32(map->m_lblk);
3071                         zero_ex.ee_len = cpu_to_le16(allocated);
3072                         ext4_ext_store_pblock(&zero_ex,
3073                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3074                         err = ext4_ext_zeroout(inode, &zero_ex);
3075                         if (err)
3076                                 goto out;
3077                         split_map.m_lblk = map->m_lblk;
3078                         split_map.m_len = allocated;
3079                 } else if ((map->m_lblk - ee_block + map->m_len <
3080                            EXT4_EXT_ZERO_LEN) &&
3081                            (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3082                         /* case 2 */
3083                         if (map->m_lblk != ee_block) {
3084                                 zero_ex.ee_block = ex->ee_block;
3085                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3086                                                         ee_block);
3087                                 ext4_ext_store_pblock(&zero_ex,
3088                                                       ext4_ext_pblock(ex));
3089                                 err = ext4_ext_zeroout(inode, &zero_ex);
3090                                 if (err)
3091                                         goto out;
3092                         }
3093
3094                         split_map.m_lblk = ee_block;
3095                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3096                         allocated = map->m_len;
3097                 }
3098         }
3099
3100         allocated = ext4_split_extent(handle, inode, path,
3101                                        &split_map, split_flag, 0);
3102         if (allocated < 0)
3103                 err = allocated;
3104
3105 out:
3106         return err ? err : allocated;
3107 }
3108
3109 /*
3110  * This function is called by ext4_ext_map_blocks() from
3111  * ext4_get_blocks_dio_write() when DIO to write
3112  * to an uninitialized extent.
3113  *
3114  * Writing to an uninitialized extent may result in splitting the uninitialized
3115  * extent into multiple /initialized uninitialized extents (up to three)
3116  * There are three possibilities:
3117  *   a> There is no split required: Entire extent should be uninitialized
3118  *   b> Splits in two extents: Write is happening at either end of the extent
3119  *   c> Splits in three extents: Somone is writing in middle of the extent
3120  *
3121  * One of more index blocks maybe needed if the extent tree grow after
3122  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3123  * complete, we need to split the uninitialized extent before DIO submit
3124  * the IO. The uninitialized extent called at this time will be split
3125  * into three uninitialized extent(at most). After IO complete, the part
3126  * being filled will be convert to initialized by the end_io callback function
3127  * via ext4_convert_unwritten_extents().
3128  *
3129  * Returns the size of uninitialized extent to be written on success.
3130  */
3131 static int ext4_split_unwritten_extents(handle_t *handle,
3132                                         struct inode *inode,
3133                                         struct ext4_map_blocks *map,
3134                                         struct ext4_ext_path *path,
3135                                         int flags)
3136 {
3137         ext4_lblk_t eof_block;
3138         ext4_lblk_t ee_block;
3139         struct ext4_extent *ex;
3140         unsigned int ee_len;
3141         int split_flag = 0, depth;
3142
3143         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3144                 "block %llu, max_blocks %u\n", inode->i_ino,
3145                 (unsigned long long)map->m_lblk, map->m_len);
3146
3147         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3148                 inode->i_sb->s_blocksize_bits;
3149         if (eof_block < map->m_lblk + map->m_len)
3150                 eof_block = map->m_lblk + map->m_len;
3151         /*
3152          * It is safe to convert extent to initialized via explicit
3153          * zeroout only if extent is fully insde i_size or new_size.
3154          */
3155         depth = ext_depth(inode);
3156         ex = path[depth].p_ext;
3157         ee_block = le32_to_cpu(ex->ee_block);
3158         ee_len = ext4_ext_get_actual_len(ex);
3159
3160         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3161         split_flag |= EXT4_EXT_MARK_UNINIT2;
3162
3163         flags |= EXT4_GET_BLOCKS_PRE_IO;
3164         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3165 }
3166
3167 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3168                                               struct inode *inode,
3169                                               struct ext4_ext_path *path)
3170 {
3171         struct ext4_extent *ex;
3172         int depth;
3173         int err = 0;
3174
3175         depth = ext_depth(inode);
3176         ex = path[depth].p_ext;
3177
3178         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3179                 "block %llu, max_blocks %u\n", inode->i_ino,
3180                 (unsigned long long)le32_to_cpu(ex->ee_block),
3181                 ext4_ext_get_actual_len(ex));
3182
3183         err = ext4_ext_get_access(handle, inode, path + depth);
3184         if (err)
3185                 goto out;
3186         /* first mark the extent as initialized */
3187         ext4_ext_mark_initialized(ex);
3188
3189         /* note: ext4_ext_correct_indexes() isn't needed here because
3190          * borders are not changed
3191          */
3192         ext4_ext_try_to_merge(inode, path, ex);
3193
3194         /* Mark modified extent as dirty */
3195         err = ext4_ext_dirty(handle, inode, path + depth);
3196 out:
3197         ext4_ext_show_leaf(inode, path);
3198         return err;
3199 }
3200
3201 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3202                         sector_t block, int count)
3203 {
3204         int i;
3205         for (i = 0; i < count; i++)
3206                 unmap_underlying_metadata(bdev, block + i);
3207 }
3208
3209 /*
3210  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3211  */
3212 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3213                               ext4_lblk_t lblk,
3214                               struct ext4_ext_path *path,
3215                               unsigned int len)
3216 {
3217         int i, depth;
3218         struct ext4_extent_header *eh;
3219         struct ext4_extent *last_ex;
3220
3221         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3222                 return 0;
3223
3224         depth = ext_depth(inode);
3225         eh = path[depth].p_hdr;
3226
3227         if (unlikely(!eh->eh_entries)) {
3228                 EXT4_ERROR_INODE(inode, "eh->eh_entries == 0 and "
3229                                  "EOFBLOCKS_FL set");
3230                 return -EIO;
3231         }
3232         last_ex = EXT_LAST_EXTENT(eh);
3233         /*
3234          * We should clear the EOFBLOCKS_FL flag if we are writing the
3235          * last block in the last extent in the file.  We test this by
3236          * first checking to see if the caller to
3237          * ext4_ext_get_blocks() was interested in the last block (or
3238          * a block beyond the last block) in the current extent.  If
3239          * this turns out to be false, we can bail out from this
3240          * function immediately.
3241          */
3242         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3243             ext4_ext_get_actual_len(last_ex))
3244                 return 0;
3245         /*
3246          * If the caller does appear to be planning to write at or
3247          * beyond the end of the current extent, we then test to see
3248          * if the current extent is the last extent in the file, by
3249          * checking to make sure it was reached via the rightmost node
3250          * at each level of the tree.
3251          */
3252         for (i = depth-1; i >= 0; i--)
3253                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3254                         return 0;
3255         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3256         return ext4_mark_inode_dirty(handle, inode);
3257 }
3258
3259 /**
3260  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3261  *
3262  * Goes through the buffer heads in the range [lblk_start, lblk_end] and returns
3263  * whether there are any buffers marked for delayed allocation. It returns '1'
3264  * on the first delalloc'ed buffer head found. If no buffer head in the given
3265  * range is marked for delalloc, it returns 0.
3266  * lblk_start should always be <= lblk_end.
3267  * search_hint_reverse is to indicate that searching in reverse from lblk_end to
3268  * lblk_start might be more efficient (i.e., we will likely hit the delalloc'ed
3269  * block sooner). This is useful when blocks are truncated sequentially from
3270  * lblk_start towards lblk_end.
3271  */
3272 static int ext4_find_delalloc_range(struct inode *inode,
3273                                     ext4_lblk_t lblk_start,
3274                                     ext4_lblk_t lblk_end,
3275                                     int search_hint_reverse)
3276 {
3277         struct address_space *mapping = inode->i_mapping;
3278         struct buffer_head *head, *bh = NULL;
3279         struct page *page;
3280         ext4_lblk_t i, pg_lblk;
3281         pgoff_t index;
3282
3283         if (!test_opt(inode->i_sb, DELALLOC))
3284                 return 0;
3285
3286         /* reverse search wont work if fs block size is less than page size */
3287         if (inode->i_blkbits < PAGE_CACHE_SHIFT)
3288                 search_hint_reverse = 0;
3289
3290         if (search_hint_reverse)
3291                 i = lblk_end;
3292         else
3293                 i = lblk_start;
3294
3295         index = i >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
3296
3297         while ((i >= lblk_start) && (i <= lblk_end)) {
3298                 page = find_get_page(mapping, index);
3299                 if (!page)
3300                         goto nextpage;
3301
3302                 if (!page_has_buffers(page))
3303                         goto nextpage;
3304
3305                 head = page_buffers(page);
3306                 if (!head)
3307                         goto nextpage;
3308
3309                 bh = head;
3310                 pg_lblk = index << (PAGE_CACHE_SHIFT -
3311                                                 inode->i_blkbits);
3312                 do {
3313                         if (unlikely(pg_lblk < lblk_start)) {
3314                                 /*
3315                                  * This is possible when fs block size is less
3316                                  * than page size and our cluster starts/ends in
3317                                  * middle of the page. So we need to skip the
3318                                  * initial few blocks till we reach the 'lblk'
3319                                  */
3320                                 pg_lblk++;
3321                                 continue;
3322                         }
3323
3324                         /* Check if the buffer is delayed allocated and that it
3325                          * is not yet mapped. (when da-buffers are mapped during
3326                          * their writeout, their da_mapped bit is set.)
3327                          */
3328                         if (buffer_delay(bh) && !buffer_da_mapped(bh)) {
3329                                 page_cache_release(page);
3330                                 trace_ext4_find_delalloc_range(inode,
3331                                                 lblk_start, lblk_end,
3332                                                 search_hint_reverse,
3333                                                 1, i);
3334                                 return 1;
3335                         }
3336                         if (search_hint_reverse)
3337                                 i--;
3338                         else
3339                                 i++;
3340                 } while ((i >= lblk_start) && (i <= lblk_end) &&
3341                                 ((bh = bh->b_this_page) != head));
3342 nextpage:
3343                 if (page)
3344                         page_cache_release(page);
3345                 /*
3346                  * Move to next page. 'i' will be the first lblk in the next
3347                  * page.
3348                  */
3349                 if (search_hint_reverse)
3350                         index--;
3351                 else
3352                         index++;
3353                 i = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
3354         }
3355
3356         trace_ext4_find_delalloc_range(inode, lblk_start, lblk_end,
3357                                         search_hint_reverse, 0, 0);
3358         return 0;
3359 }
3360
3361 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk,
3362                                int search_hint_reverse)
3363 {
3364         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3365         ext4_lblk_t lblk_start, lblk_end;
3366         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3367         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3368
3369         return ext4_find_delalloc_range(inode, lblk_start, lblk_end,
3370                                         search_hint_reverse);
3371 }
3372
3373 /**
3374  * Determines how many complete clusters (out of those specified by the 'map')
3375  * are under delalloc and were reserved quota for.
3376  * This function is called when we are writing out the blocks that were
3377  * originally written with their allocation delayed, but then the space was
3378  * allocated using fallocate() before the delayed allocation could be resolved.
3379  * The cases to look for are:
3380  * ('=' indicated delayed allocated blocks
3381  *  '-' indicates non-delayed allocated blocks)
3382  * (a) partial clusters towards beginning and/or end outside of allocated range
3383  *     are not delalloc'ed.
3384  *      Ex:
3385  *      |----c---=|====c====|====c====|===-c----|
3386  *               |++++++ allocated ++++++|
3387  *      ==> 4 complete clusters in above example
3388  *
3389  * (b) partial cluster (outside of allocated range) towards either end is
3390  *     marked for delayed allocation. In this case, we will exclude that
3391  *     cluster.
3392  *      Ex:
3393  *      |----====c========|========c========|
3394  *           |++++++ allocated ++++++|
3395  *      ==> 1 complete clusters in above example
3396  *
3397  *      Ex:
3398  *      |================c================|
3399  *            |++++++ allocated ++++++|
3400  *      ==> 0 complete clusters in above example
3401  *
3402  * The ext4_da_update_reserve_space will be called only if we
3403  * determine here that there were some "entire" clusters that span
3404  * this 'allocated' range.
3405  * In the non-bigalloc case, this function will just end up returning num_blks
3406  * without ever calling ext4_find_delalloc_range.
3407  */
3408 static unsigned int
3409 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3410                            unsigned int num_blks)
3411 {
3412         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3413         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3414         ext4_lblk_t lblk_from, lblk_to, c_offset;
3415         unsigned int allocated_clusters = 0;
3416
3417         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3418         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3419
3420         /* max possible clusters for this allocation */
3421         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3422
3423         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3424
3425         /* Check towards left side */
3426         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3427         if (c_offset) {
3428                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3429                 lblk_to = lblk_from + c_offset - 1;
3430
3431                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to, 0))
3432                         allocated_clusters--;
3433         }
3434
3435         /* Now check towards right. */
3436         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3437         if (allocated_clusters && c_offset) {
3438                 lblk_from = lblk_start + num_blks;
3439                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3440
3441                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to, 0))
3442                         allocated_clusters--;
3443         }
3444
3445         return allocated_clusters;
3446 }
3447
3448 static int
3449 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3450                         struct ext4_map_blocks *map,
3451                         struct ext4_ext_path *path, int flags,
3452                         unsigned int allocated, ext4_fsblk_t newblock)
3453 {
3454         int ret = 0;
3455         int err = 0;
3456         ext4_io_end_t *io = EXT4_I(inode)->cur_aio_dio;
3457
3458         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3459                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3460                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3461                   flags, allocated);
3462         ext4_ext_show_leaf(inode, path);
3463
3464         trace_ext4_ext_handle_uninitialized_extents(inode, map, allocated,
3465                                                     newblock);
3466
3467         /* get_block() before submit the IO, split the extent */
3468         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3469                 ret = ext4_split_unwritten_extents(handle, inode, map,
3470                                                    path, flags);
3471                 /*
3472                  * Flag the inode(non aio case) or end_io struct (aio case)
3473                  * that this IO needs to conversion to written when IO is
3474                  * completed
3475                  */
3476                 if (io)
3477                         ext4_set_io_unwritten_flag(inode, io);
3478                 else
3479                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3480                 if (ext4_should_dioread_nolock(inode))
3481                         map->m_flags |= EXT4_MAP_UNINIT;
3482                 goto out;
3483         }
3484         /* IO end_io complete, convert the filled extent to written */
3485         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3486                 ret = ext4_convert_unwritten_extents_endio(handle, inode,
3487                                                         path);
3488                 if (ret >= 0) {
3489                         ext4_update_inode_fsync_trans(handle, inode, 1);
3490                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3491                                                  path, map->m_len);
3492                 } else
3493                         err = ret;
3494                 goto out2;
3495         }
3496         /* buffered IO case */
3497         /*
3498          * repeat fallocate creation request
3499          * we already have an unwritten extent
3500          */
3501         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT)
3502                 goto map_out;
3503
3504         /* buffered READ or buffered write_begin() lookup */
3505         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3506                 /*
3507                  * We have blocks reserved already.  We
3508                  * return allocated blocks so that delalloc
3509                  * won't do block reservation for us.  But
3510                  * the buffer head will be unmapped so that
3511                  * a read from the block returns 0s.
3512                  */
3513                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3514                 goto out1;
3515         }
3516
3517         /* buffered write, writepage time, convert*/
3518         ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
3519         if (ret >= 0)
3520                 ext4_update_inode_fsync_trans(handle, inode, 1);
3521 out:
3522         if (ret <= 0) {
3523                 err = ret;
3524                 goto out2;
3525         } else
3526                 allocated = ret;
3527         map->m_flags |= EXT4_MAP_NEW;
3528         /*
3529          * if we allocated more blocks than requested
3530          * we need to make sure we unmap the extra block
3531          * allocated. The actual needed block will get
3532          * unmapped later when we find the buffer_head marked
3533          * new.
3534          */
3535         if (allocated > map->m_len) {
3536                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3537                                         newblock + map->m_len,
3538                                         allocated - map->m_len);
3539                 allocated = map->m_len;
3540         }
3541
3542         /*
3543          * If we have done fallocate with the offset that is already
3544          * delayed allocated, we would have block reservation
3545          * and quota reservation done in the delayed write path.
3546          * But fallocate would have already updated quota and block
3547          * count for this offset. So cancel these reservation
3548          */
3549         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3550                 unsigned int reserved_clusters;
3551                 reserved_clusters = get_reserved_cluster_alloc(inode,
3552                                 map->m_lblk, map->m_len);
3553                 if (reserved_clusters)
3554                         ext4_da_update_reserve_space(inode,
3555                                                      reserved_clusters,
3556                                                      0);
3557         }
3558
3559 map_out:
3560         map->m_flags |= EXT4_MAP_MAPPED;
3561         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3562                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3563                                          map->m_len);
3564                 if (err < 0)
3565                         goto out2;
3566         }
3567 out1:
3568         if (allocated > map->m_len)
3569                 allocated = map->m_len;
3570         ext4_ext_show_leaf(inode, path);
3571         map->m_pblk = newblock;
3572         map->m_len = allocated;
3573 out2:
3574         if (path) {
3575                 ext4_ext_drop_refs(path);
3576                 kfree(path);
3577         }
3578         return err ? err : allocated;
3579 }
3580
3581 /*
3582  * get_implied_cluster_alloc - check to see if the requested
3583  * allocation (in the map structure) overlaps with a cluster already
3584  * allocated in an extent.
3585  *      @sb     The filesystem superblock structure
3586  *      @map    The requested lblk->pblk mapping
3587  *      @ex     The extent structure which might contain an implied
3588  *                      cluster allocation
3589  *
3590  * This function is called by ext4_ext_map_blocks() after we failed to
3591  * find blocks that were already in the inode's extent tree.  Hence,
3592  * we know that the beginning of the requested region cannot overlap
3593  * the extent from the inode's extent tree.  There are three cases we
3594  * want to catch.  The first is this case:
3595  *
3596  *               |--- cluster # N--|
3597  *    |--- extent ---|  |---- requested region ---|
3598  *                      |==========|
3599  *
3600  * The second case that we need to test for is this one:
3601  *
3602  *   |--------- cluster # N ----------------|
3603  *         |--- requested region --|   |------- extent ----|
3604  *         |=======================|
3605  *
3606  * The third case is when the requested region lies between two extents
3607  * within the same cluster:
3608  *          |------------- cluster # N-------------|
3609  * |----- ex -----|                  |---- ex_right ----|
3610  *                  |------ requested region ------|
3611  *                  |================|
3612  *
3613  * In each of the above cases, we need to set the map->m_pblk and
3614  * map->m_len so it corresponds to the return the extent labelled as
3615  * "|====|" from cluster #N, since it is already in use for data in
3616  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3617  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3618  * as a new "allocated" block region.  Otherwise, we will return 0 and
3619  * ext4_ext_map_blocks() will then allocate one or more new clusters
3620  * by calling ext4_mb_new_blocks().
3621  */
3622 static int get_implied_cluster_alloc(struct super_block *sb,
3623                                      struct ext4_map_blocks *map,
3624                                      struct ext4_extent *ex,
3625                                      struct ext4_ext_path *path)
3626 {
3627         struct ext4_sb_info *sbi = EXT4_SB(sb);
3628         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3629         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3630         ext4_lblk_t rr_cluster_start;
3631         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3632         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3633         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3634
3635         /* The extent passed in that we are trying to match */
3636         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3637         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3638
3639         /* The requested region passed into ext4_map_blocks() */
3640         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3641
3642         if ((rr_cluster_start == ex_cluster_end) ||
3643             (rr_cluster_start == ex_cluster_start)) {
3644                 if (rr_cluster_start == ex_cluster_end)
3645                         ee_start += ee_len - 1;
3646                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3647                         c_offset;
3648                 map->m_len = min(map->m_len,
3649                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3650                 /*
3651                  * Check for and handle this case:
3652                  *
3653                  *   |--------- cluster # N-------------|
3654                  *                     |------- extent ----|
3655                  *         |--- requested region ---|
3656                  *         |===========|
3657                  */
3658
3659                 if (map->m_lblk < ee_block)
3660                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3661
3662                 /*
3663                  * Check for the case where there is already another allocated
3664                  * block to the right of 'ex' but before the end of the cluster.
3665                  *
3666                  *          |------------- cluster # N-------------|
3667                  * |----- ex -----|                  |---- ex_right ----|
3668                  *                  |------ requested region ------|
3669                  *                  |================|
3670                  */
3671                 if (map->m_lblk > ee_block) {
3672                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3673                         map->m_len = min(map->m_len, next - map->m_lblk);
3674                 }
3675
3676                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3677                 return 1;
3678         }
3679
3680         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3681         return 0;
3682 }
3683
3684
3685 /*
3686  * Block allocation/map/preallocation routine for extents based files
3687  *
3688  *
3689  * Need to be called with
3690  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
3691  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
3692  *
3693  * return > 0, number of of blocks already mapped/allocated
3694  *          if create == 0 and these are pre-allocated blocks
3695  *              buffer head is unmapped
3696  *          otherwise blocks are mapped
3697  *
3698  * return = 0, if plain look up failed (blocks have not been allocated)
3699  *          buffer head is unmapped
3700  *
3701  * return < 0, error case.
3702  */
3703 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
3704                         struct ext4_map_blocks *map, int flags)
3705 {
3706         struct ext4_ext_path *path = NULL;
3707         struct ext4_extent newex, *ex, *ex2;
3708         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3709         ext4_fsblk_t newblock = 0;
3710         int free_on_err = 0, err = 0, depth, ret;
3711         unsigned int allocated = 0, offset = 0;
3712         unsigned int allocated_clusters = 0;
3713         unsigned int punched_out = 0;
3714         unsigned int result = 0;
3715         struct ext4_allocation_request ar;
3716         ext4_io_end_t *io = EXT4_I(inode)->cur_aio_dio;
3717         ext4_lblk_t cluster_offset;
3718
3719         ext_debug("blocks %u/%u requested for inode %lu\n",
3720                   map->m_lblk, map->m_len, inode->i_ino);
3721         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
3722
3723         /* check in cache */
3724         if (!(flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) &&
3725                 ext4_ext_in_cache(inode, map->m_lblk, &newex)) {
3726                 if (!newex.ee_start_lo && !newex.ee_start_hi) {
3727                         if ((sbi->s_cluster_ratio > 1) &&
3728                             ext4_find_delalloc_cluster(inode, map->m_lblk, 0))
3729                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3730
3731                         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3732                                 /*
3733                                  * block isn't allocated yet and
3734                                  * user doesn't want to allocate it
3735                                  */
3736                                 goto out2;
3737                         }
3738                         /* we should allocate requested block */
3739                 } else {
3740                         /* block is already allocated */
3741                         if (sbi->s_cluster_ratio > 1)
3742                                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3743                         newblock = map->m_lblk
3744                                    - le32_to_cpu(newex.ee_block)
3745                                    + ext4_ext_pblock(&newex);
3746                         /* number of remaining blocks in the extent */
3747                         allocated = ext4_ext_get_actual_len(&newex) -
3748                                 (map->m_lblk - le32_to_cpu(newex.ee_block));
3749                         goto out;
3750                 }
3751         }
3752
3753         /* find extent for this block */
3754         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
3755         if (IS_ERR(path)) {
3756                 err = PTR_ERR(path);
3757                 path = NULL;
3758                 goto out2;
3759         }
3760
3761         depth = ext_depth(inode);
3762
3763         /*
3764          * consistent leaf must not be empty;
3765          * this situation is possible, though, _during_ tree modification;
3766          * this is why assert can't be put in ext4_ext_find_extent()
3767          */
3768         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
3769                 EXT4_ERROR_INODE(inode, "bad extent address "
3770                                  "lblock: %lu, depth: %d pblock %lld",
3771                                  (unsigned long) map->m_lblk, depth,
3772                                  path[depth].p_block);
3773                 err = -EIO;
3774                 goto out2;
3775         }
3776
3777         ex = path[depth].p_ext;
3778         if (ex) {
3779                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3780                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3781                 unsigned short ee_len;
3782
3783                 /*
3784                  * Uninitialized extents are treated as holes, except that
3785                  * we split out initialized portions during a write.
3786                  */
3787                 ee_len = ext4_ext_get_actual_len(ex);
3788
3789                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
3790
3791                 /* if found extent covers block, simply return it */
3792                 if (in_range(map->m_lblk, ee_block, ee_len)) {
3793                         struct ext4_map_blocks punch_map;
3794                         ext4_fsblk_t partial_cluster = 0;
3795
3796                         newblock = map->m_lblk - ee_block + ee_start;
3797                         /* number of remaining blocks in the extent */
3798                         allocated = ee_len - (map->m_lblk - ee_block);
3799                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
3800                                   ee_block, ee_len, newblock);
3801
3802                         if ((flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) == 0) {
3803                                 /*
3804                                  * Do not put uninitialized extent
3805                                  * in the cache
3806                                  */
3807                                 if (!ext4_ext_is_uninitialized(ex)) {
3808                                         ext4_ext_put_in_cache(inode, ee_block,
3809                                                 ee_len, ee_start);
3810                                         goto out;
3811                                 }
3812                                 ret = ext4_ext_handle_uninitialized_extents(
3813                                         handle, inode, map, path, flags,
3814                                         allocated, newblock);
3815                                 return ret;
3816                         }
3817
3818                         /*
3819                          * Punch out the map length, but only to the
3820                          * end of the extent
3821                          */
3822                         punched_out = allocated < map->m_len ?
3823                                 allocated : map->m_len;
3824
3825                         /*
3826                          * Sense extents need to be converted to
3827                          * uninitialized, they must fit in an
3828                          * uninitialized extent
3829                          */
3830                         if (punched_out > EXT_UNINIT_MAX_LEN)
3831                                 punched_out = EXT_UNINIT_MAX_LEN;
3832
3833                         punch_map.m_lblk = map->m_lblk;
3834                         punch_map.m_pblk = newblock;
3835                         punch_map.m_len = punched_out;
3836                         punch_map.m_flags = 0;
3837
3838                         /* Check to see if the extent needs to be split */
3839                         if (punch_map.m_len != ee_len ||
3840                                 punch_map.m_lblk != ee_block) {
3841
3842                                 ret = ext4_split_extent(handle, inode,
3843                                 path, &punch_map, 0,
3844                                 EXT4_GET_BLOCKS_PUNCH_OUT_EXT |
3845                                 EXT4_GET_BLOCKS_PRE_IO);
3846
3847                                 if (ret < 0) {
3848                                         err = ret;
3849                                         goto out2;
3850                                 }
3851                                 /*
3852                                  * find extent for the block at
3853                                  * the start of the hole
3854                                  */
3855                                 ext4_ext_drop_refs(path);
3856                                 kfree(path);
3857
3858                                 path = ext4_ext_find_extent(inode,
3859                                 map->m_lblk, NULL);
3860                                 if (IS_ERR(path)) {
3861                                         err = PTR_ERR(path);
3862                                         path = NULL;
3863                                         goto out2;
3864                                 }
3865
3866                                 depth = ext_depth(inode);
3867                                 ex = path[depth].p_ext;
3868                                 ee_len = ext4_ext_get_actual_len(ex);
3869                                 ee_block = le32_to_cpu(ex->ee_block);
3870                                 ee_start = ext4_ext_pblock(ex);
3871
3872                         }
3873
3874                         ext4_ext_mark_uninitialized(ex);
3875
3876                         ext4_ext_invalidate_cache(inode);
3877
3878                         err = ext4_ext_rm_leaf(handle, inode, path,
3879                                                &partial_cluster, map->m_lblk,
3880                                                map->m_lblk + punched_out);
3881
3882                         if (!err && path->p_hdr->eh_entries == 0) {
3883                                 /*
3884                                  * Punch hole freed all of this sub tree,
3885                                  * so we need to correct eh_depth
3886                                  */
3887                                 err = ext4_ext_get_access(handle, inode, path);
3888                                 if (err == 0) {
3889                                         ext_inode_hdr(inode)->eh_depth = 0;
3890                                         ext_inode_hdr(inode)->eh_max =
3891                                         cpu_to_le16(ext4_ext_space_root(
3892                                                 inode, 0));
3893
3894                                         err = ext4_ext_dirty(
3895                                                 handle, inode, path);
3896                                 }
3897                         }
3898
3899                         goto out2;
3900                 }
3901         }
3902
3903         if ((sbi->s_cluster_ratio > 1) &&
3904             ext4_find_delalloc_cluster(inode, map->m_lblk, 0))
3905                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3906
3907         /*
3908          * requested block isn't allocated yet;
3909          * we couldn't try to create block if create flag is zero
3910          */
3911         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3912                 /*
3913                  * put just found gap into cache to speed up
3914                  * subsequent requests
3915                  */
3916                 ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
3917                 goto out2;
3918         }
3919
3920         /*
3921          * Okay, we need to do block allocation.
3922          */
3923         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
3924         newex.ee_block = cpu_to_le32(map->m_lblk);
3925         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3926
3927         /*
3928          * If we are doing bigalloc, check to see if the extent returned
3929          * by ext4_ext_find_extent() implies a cluster we can use.
3930          */
3931         if (cluster_offset && ex &&
3932             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
3933                 ar.len = allocated = map->m_len;
3934                 newblock = map->m_pblk;
3935                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3936                 goto got_allocated_blocks;
3937         }
3938
3939         /* find neighbour allocated blocks */
3940         ar.lleft = map->m_lblk;
3941         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
3942         if (err)
3943                 goto out2;
3944         ar.lright = map->m_lblk;
3945         ex2 = NULL;
3946         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
3947         if (err)
3948                 goto out2;
3949
3950         /* Check if the extent after searching to the right implies a
3951          * cluster we can use. */
3952         if ((sbi->s_cluster_ratio > 1) && ex2 &&
3953             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
3954                 ar.len = allocated = map->m_len;
3955                 newblock = map->m_pblk;
3956                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
3957                 goto got_allocated_blocks;
3958         }
3959
3960         /*
3961          * See if request is beyond maximum number of blocks we can have in
3962          * a single extent. For an initialized extent this limit is
3963          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
3964          * EXT_UNINIT_MAX_LEN.
3965          */
3966         if (map->m_len > EXT_INIT_MAX_LEN &&
3967             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
3968                 map->m_len = EXT_INIT_MAX_LEN;
3969         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
3970                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
3971                 map->m_len = EXT_UNINIT_MAX_LEN;
3972
3973         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
3974         newex.ee_len = cpu_to_le16(map->m_len);
3975         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
3976         if (err)
3977                 allocated = ext4_ext_get_actual_len(&newex);
3978         else
3979                 allocated = map->m_len;
3980
3981         /* allocate new block */
3982         ar.inode = inode;
3983         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
3984         ar.logical = map->m_lblk;
3985         /*
3986          * We calculate the offset from the beginning of the cluster
3987          * for the logical block number, since when we allocate a
3988          * physical cluster, the physical block should start at the
3989          * same offset from the beginning of the cluster.  This is
3990          * needed so that future calls to get_implied_cluster_alloc()
3991          * work correctly.
3992          */
3993         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
3994         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
3995         ar.goal -= offset;
3996         ar.logical -= offset;
3997         if (S_ISREG(inode->i_mode))
3998                 ar.flags = EXT4_MB_HINT_DATA;
3999         else
4000                 /* disable in-core preallocation for non-regular files */
4001                 ar.flags = 0;
4002         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4003                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4004         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4005         if (!newblock)
4006                 goto out2;
4007         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4008                   ar.goal, newblock, allocated);
4009         free_on_err = 1;
4010         allocated_clusters = ar.len;
4011         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4012         if (ar.len > allocated)
4013                 ar.len = allocated;
4014
4015 got_allocated_blocks:
4016         /* try to insert new extent into found leaf and return */
4017         ext4_ext_store_pblock(&newex, newblock + offset);
4018         newex.ee_len = cpu_to_le16(ar.len);
4019         /* Mark uninitialized */
4020         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4021                 ext4_ext_mark_uninitialized(&newex);
4022                 /*
4023                  * io_end structure was created for every IO write to an
4024                  * uninitialized extent. To avoid unnecessary conversion,
4025                  * here we flag the IO that really needs the conversion.
4026                  * For non asycn direct IO case, flag the inode state
4027                  * that we need to perform conversion when IO is done.
4028                  */
4029                 if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
4030                         if (io)
4031                                 ext4_set_io_unwritten_flag(inode, io);
4032                         else
4033                                 ext4_set_inode_state(inode,
4034                                                      EXT4_STATE_DIO_UNWRITTEN);
4035                 }
4036                 if (ext4_should_dioread_nolock(inode))
4037                         map->m_flags |= EXT4_MAP_UNINIT;
4038         }
4039
4040         err = 0;
4041         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4042                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4043                                          path, ar.len);
4044         if (!err)
4045                 err = ext4_ext_insert_extent(handle, inode, path,
4046                                              &newex, flags);
4047         if (err && free_on_err) {
4048                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4049                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4050                 /* free data blocks we just allocated */
4051                 /* not a good idea to call discard here directly,
4052                  * but otherwise we'd need to call it every free() */
4053                 ext4_discard_preallocations(inode);
4054                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4055                                  ext4_ext_get_actual_len(&newex), fb_flags);
4056                 goto out2;
4057         }
4058
4059         /* previous routine could use block we allocated */
4060         newblock = ext4_ext_pblock(&newex);
4061         allocated = ext4_ext_get_actual_len(&newex);
4062         if (allocated > map->m_len)
4063                 allocated = map->m_len;
4064         map->m_flags |= EXT4_MAP_NEW;
4065
4066         /*
4067          * Update reserved blocks/metadata blocks after successful
4068          * block allocation which had been deferred till now.
4069          */
4070         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4071                 unsigned int reserved_clusters;
4072                 /*
4073                  * Check how many clusters we had reserved this allocated range
4074                  */
4075                 reserved_clusters = get_reserved_cluster_alloc(inode,
4076                                                 map->m_lblk, allocated);
4077                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4078                         if (reserved_clusters) {
4079                                 /*
4080                                  * We have clusters reserved for this range.
4081                                  * But since we are not doing actual allocation
4082                                  * and are simply using blocks from previously
4083                                  * allocated cluster, we should release the
4084                                  * reservation and not claim quota.
4085                                  */
4086                                 ext4_da_update_reserve_space(inode,
4087                                                 reserved_clusters, 0);
4088                         }
4089                 } else {
4090                         BUG_ON(allocated_clusters < reserved_clusters);
4091                         /* We will claim quota for all newly allocated blocks.*/
4092                         ext4_da_update_reserve_space(inode, allocated_clusters,
4093                                                         1);
4094                         if (reserved_clusters < allocated_clusters) {
4095                                 struct ext4_inode_info *ei = EXT4_I(inode);
4096                                 int reservation = allocated_clusters -
4097                                                   reserved_clusters;
4098                                 /*
4099                                  * It seems we claimed few clusters outside of
4100                                  * the range of this allocation. We should give
4101                                  * it back to the reservation pool. This can
4102                                  * happen in the following case:
4103                                  *
4104                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4105                                  *   cluster has 4 blocks. Thus, the clusters
4106                                  *   are [0-3],[4-7],[8-11]...
4107                                  * * First comes delayed allocation write for
4108                                  *   logical blocks 10 & 11. Since there were no
4109                                  *   previous delayed allocated blocks in the
4110                                  *   range [8-11], we would reserve 1 cluster
4111                                  *   for this write.
4112                                  * * Next comes write for logical blocks 3 to 8.
4113                                  *   In this case, we will reserve 2 clusters
4114                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4115                                  *   that range has a delayed allocated blocks.
4116                                  *   Thus total reserved clusters now becomes 3.
4117                                  * * Now, during the delayed allocation writeout
4118                                  *   time, we will first write blocks [3-8] and
4119                                  *   allocate 3 clusters for writing these
4120                                  *   blocks. Also, we would claim all these
4121                                  *   three clusters above.
4122                                  * * Now when we come here to writeout the
4123                                  *   blocks [10-11], we would expect to claim
4124                                  *   the reservation of 1 cluster we had made
4125                                  *   (and we would claim it since there are no
4126                                  *   more delayed allocated blocks in the range
4127                                  *   [8-11]. But our reserved cluster count had
4128                                  *   already gone to 0.
4129                                  *
4130                                  *   Thus, at the step 4 above when we determine
4131                                  *   that there are still some unwritten delayed
4132                                  *   allocated blocks outside of our current
4133                                  *   block range, we should increment the
4134                                  *   reserved clusters count so that when the
4135                                  *   remaining blocks finally gets written, we
4136                                  *   could claim them.
4137                                  */
4138                                 dquot_reserve_block(inode,
4139                                                 EXT4_C2B(sbi, reservation));
4140                                 spin_lock(&ei->i_block_reservation_lock);
4141                                 ei->i_reserved_data_blocks += reservation;
4142                                 spin_unlock(&ei->i_block_reservation_lock);
4143                         }
4144                 }
4145         }
4146
4147         /*
4148          * Cache the extent and update transaction to commit on fdatasync only
4149          * when it is _not_ an uninitialized extent.
4150          */
4151         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0) {
4152                 ext4_ext_put_in_cache(inode, map->m_lblk, allocated, newblock);
4153                 ext4_update_inode_fsync_trans(handle, inode, 1);
4154         } else
4155                 ext4_update_inode_fsync_trans(handle, inode, 0);
4156 out:
4157         if (allocated > map->m_len)
4158                 allocated = map->m_len;
4159         ext4_ext_show_leaf(inode, path);
4160         map->m_flags |= EXT4_MAP_MAPPED;
4161         map->m_pblk = newblock;
4162         map->m_len = allocated;
4163 out2:
4164         if (path) {
4165                 ext4_ext_drop_refs(path);
4166                 kfree(path);
4167         }
4168         result = (flags & EXT4_GET_BLOCKS_PUNCH_OUT_EXT) ?
4169                         punched_out : allocated;
4170
4171         trace_ext4_ext_map_blocks_exit(inode, map->m_lblk,
4172                 newblock, map->m_len, err ? err : result);
4173
4174         return err ? err : result;
4175 }
4176
4177 void ext4_ext_truncate(struct inode *inode)
4178 {
4179         struct address_space *mapping = inode->i_mapping;
4180         struct super_block *sb = inode->i_sb;
4181         ext4_lblk_t last_block;
4182         handle_t *handle;
4183         loff_t page_len;
4184         int err = 0;
4185
4186         /*
4187          * finish any pending end_io work so we won't run the risk of
4188          * converting any truncated blocks to initialized later
4189          */
4190         ext4_flush_completed_IO(inode);
4191
4192         /*
4193          * probably first extent we're gonna free will be last in block
4194          */
4195         err = ext4_writepage_trans_blocks(inode);
4196         handle = ext4_journal_start(inode, err);
4197         if (IS_ERR(handle))
4198                 return;
4199
4200         if (inode->i_size % PAGE_CACHE_SIZE != 0) {
4201                 page_len = PAGE_CACHE_SIZE -
4202                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4203
4204                 err = ext4_discard_partial_page_buffers(handle,
4205                         mapping, inode->i_size, page_len, 0);
4206
4207                 if (err)
4208                         goto out_stop;
4209         }
4210
4211         if (ext4_orphan_add(handle, inode))
4212                 goto out_stop;
4213
4214         down_write(&EXT4_I(inode)->i_data_sem);
4215         ext4_ext_invalidate_cache(inode);
4216
4217         ext4_discard_preallocations(inode);
4218
4219         /*
4220          * TODO: optimization is possible here.
4221          * Probably we need not scan at all,
4222          * because page truncation is enough.
4223          */
4224
4225         /* we have to know where to truncate from in crash case */
4226         EXT4_I(inode)->i_disksize = inode->i_size;
4227         ext4_mark_inode_dirty(handle, inode);
4228
4229         last_block = (inode->i_size + sb->s_blocksize - 1)
4230                         >> EXT4_BLOCK_SIZE_BITS(sb);
4231         err = ext4_ext_remove_space(inode, last_block);
4232
4233         /* In a multi-transaction truncate, we only make the final
4234          * transaction synchronous.
4235          */
4236         if (IS_SYNC(inode))
4237                 ext4_handle_sync(handle);
4238
4239         up_write(&EXT4_I(inode)->i_data_sem);
4240
4241 out_stop:
4242         /*
4243          * If this was a simple ftruncate() and the file will remain alive,
4244          * then we need to clear up the orphan record which we created above.
4245          * However, if this was a real unlink then we were called by
4246          * ext4_delete_inode(), and we allow that function to clean up the
4247          * orphan info for us.
4248          */
4249         if (inode->i_nlink)
4250                 ext4_orphan_del(handle, inode);
4251
4252         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4253         ext4_mark_inode_dirty(handle, inode);
4254         ext4_journal_stop(handle);
4255 }
4256
4257 static void ext4_falloc_update_inode(struct inode *inode,
4258                                 int mode, loff_t new_size, int update_ctime)
4259 {
4260         struct timespec now;
4261
4262         if (update_ctime) {
4263                 now = current_fs_time(inode->i_sb);
4264                 if (!timespec_equal(&inode->i_ctime, &now))
4265                         inode->i_ctime = now;
4266         }
4267         /*
4268          * Update only when preallocation was requested beyond
4269          * the file size.
4270          */
4271         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4272                 if (new_size > i_size_read(inode))
4273                         i_size_write(inode, new_size);
4274                 if (new_size > EXT4_I(inode)->i_disksize)
4275                         ext4_update_i_disksize(inode, new_size);
4276         } else {
4277                 /*
4278                  * Mark that we allocate beyond EOF so the subsequent truncate
4279                  * can proceed even if the new size is the same as i_size.
4280                  */
4281                 if (new_size > i_size_read(inode))
4282                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4283         }
4284
4285 }
4286
4287 /*
4288  * preallocate space for a file. This implements ext4's fallocate file
4289  * operation, which gets called from sys_fallocate system call.
4290  * For block-mapped files, posix_fallocate should fall back to the method
4291  * of writing zeroes to the required new blocks (the same behavior which is
4292  * expected for file systems which do not support fallocate() system call).
4293  */
4294 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4295 {
4296         struct inode *inode = file->f_path.dentry->d_inode;
4297         handle_t *handle;
4298         loff_t new_size;
4299         unsigned int max_blocks;
4300         int ret = 0;
4301         int ret2 = 0;
4302         int retries = 0;
4303         int flags;
4304         struct ext4_map_blocks map;
4305         unsigned int credits, blkbits = inode->i_blkbits;
4306
4307         /*
4308          * currently supporting (pre)allocate mode for extent-based
4309          * files _only_
4310          */
4311         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4312                 return -EOPNOTSUPP;
4313
4314         /* Return error if mode is not supported */
4315         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4316                 return -EOPNOTSUPP;
4317
4318         if (mode & FALLOC_FL_PUNCH_HOLE)
4319                 return ext4_punch_hole(file, offset, len);
4320
4321         trace_ext4_fallocate_enter(inode, offset, len, mode);
4322         map.m_lblk = offset >> blkbits;
4323         /*
4324          * We can't just convert len to max_blocks because
4325          * If blocksize = 4096 offset = 3072 and len = 2048
4326          */
4327         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4328                 - map.m_lblk;
4329         /*
4330          * credits to insert 1 extent into extent tree
4331          */
4332         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4333         mutex_lock(&inode->i_mutex);
4334         ret = inode_newsize_ok(inode, (len + offset));
4335         if (ret) {
4336                 mutex_unlock(&inode->i_mutex);
4337                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4338                 return ret;
4339         }
4340         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4341         if (mode & FALLOC_FL_KEEP_SIZE)
4342                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4343         /*
4344          * Don't normalize the request if it can fit in one extent so
4345          * that it doesn't get unnecessarily split into multiple
4346          * extents.
4347          */
4348         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4349                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4350 retry:
4351         while (ret >= 0 && ret < max_blocks) {
4352                 map.m_lblk = map.m_lblk + ret;
4353                 map.m_len = max_blocks = max_blocks - ret;
4354                 handle = ext4_journal_start(inode, credits);
4355                 if (IS_ERR(handle)) {
4356                         ret = PTR_ERR(handle);
4357                         break;
4358                 }
4359                 ret = ext4_map_blocks(handle, inode, &map, flags);
4360                 if (ret <= 0) {
4361 #ifdef EXT4FS_DEBUG
4362                         WARN_ON(ret <= 0);
4363                         printk(KERN_ERR "%s: ext4_ext_map_blocks "
4364                                     "returned error inode#%lu, block=%u, "
4365                                     "max_blocks=%u", __func__,
4366                                     inode->i_ino, map.m_lblk, max_blocks);
4367 #endif
4368                         ext4_mark_inode_dirty(handle, inode);
4369                         ret2 = ext4_journal_stop(handle);
4370                         break;
4371                 }
4372                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4373                                                 blkbits) >> blkbits))
4374                         new_size = offset + len;
4375                 else
4376                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4377
4378                 ext4_falloc_update_inode(inode, mode, new_size,
4379                                          (map.m_flags & EXT4_MAP_NEW));
4380                 ext4_mark_inode_dirty(handle, inode);
4381                 ret2 = ext4_journal_stop(handle);
4382                 if (ret2)
4383                         break;
4384         }
4385         if (ret == -ENOSPC &&
4386                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4387                 ret = 0;
4388                 goto retry;
4389         }
4390         mutex_unlock(&inode->i_mutex);
4391         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4392                                 ret > 0 ? ret2 : ret);
4393         return ret > 0 ? ret2 : ret;
4394 }
4395
4396 /*
4397  * This function convert a range of blocks to written extents
4398  * The caller of this function will pass the start offset and the size.
4399  * all unwritten extents within this range will be converted to
4400  * written extents.
4401  *
4402  * This function is called from the direct IO end io call back
4403  * function, to convert the fallocated extents after IO is completed.
4404  * Returns 0 on success.
4405  */
4406 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4407                                     ssize_t len)
4408 {
4409         handle_t *handle;
4410         unsigned int max_blocks;
4411         int ret = 0;
4412         int ret2 = 0;
4413         struct ext4_map_blocks map;
4414         unsigned int credits, blkbits = inode->i_blkbits;
4415
4416         map.m_lblk = offset >> blkbits;
4417         /*
4418          * We can't just convert len to max_blocks because
4419          * If blocksize = 4096 offset = 3072 and len = 2048
4420          */
4421         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4422                       map.m_lblk);
4423         /*
4424          * credits to insert 1 extent into extent tree
4425          */
4426         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4427         while (ret >= 0 && ret < max_blocks) {
4428                 map.m_lblk += ret;
4429                 map.m_len = (max_blocks -= ret);
4430                 handle = ext4_journal_start(inode, credits);
4431                 if (IS_ERR(handle)) {
4432                         ret = PTR_ERR(handle);
4433                         break;
4434                 }
4435                 ret = ext4_map_blocks(handle, inode, &map,
4436                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4437                 if (ret <= 0) {
4438                         WARN_ON(ret <= 0);
4439                         printk(KERN_ERR "%s: ext4_ext_map_blocks "
4440                                     "returned error inode#%lu, block=%u, "
4441                                     "max_blocks=%u", __func__,
4442                                     inode->i_ino, map.m_lblk, map.m_len);
4443                 }
4444                 ext4_mark_inode_dirty(handle, inode);
4445                 ret2 = ext4_journal_stop(handle);
4446                 if (ret <= 0 || ret2 )
4447                         break;
4448         }
4449         return ret > 0 ? ret2 : ret;
4450 }
4451
4452 /*
4453  * Callback function called for each extent to gather FIEMAP information.
4454  */
4455 static int ext4_ext_fiemap_cb(struct inode *inode, ext4_lblk_t next,
4456                        struct ext4_ext_cache *newex, struct ext4_extent *ex,
4457                        void *data)
4458 {
4459         __u64   logical;
4460         __u64   physical;
4461         __u64   length;
4462         __u32   flags = 0;
4463         int             ret = 0;
4464         struct fiemap_extent_info *fieinfo = data;
4465         unsigned char blksize_bits;
4466
4467         blksize_bits = inode->i_sb->s_blocksize_bits;
4468         logical = (__u64)newex->ec_block << blksize_bits;
4469
4470         if (newex->ec_start == 0) {
4471                 /*
4472                  * No extent in extent-tree contains block @newex->ec_start,
4473                  * then the block may stay in 1)a hole or 2)delayed-extent.
4474                  *
4475                  * Holes or delayed-extents are processed as follows.
4476                  * 1. lookup dirty pages with specified range in pagecache.
4477                  *    If no page is got, then there is no delayed-extent and
4478                  *    return with EXT_CONTINUE.
4479                  * 2. find the 1st mapped buffer,
4480                  * 3. check if the mapped buffer is both in the request range
4481                  *    and a delayed buffer. If not, there is no delayed-extent,
4482                  *    then return.
4483                  * 4. a delayed-extent is found, the extent will be collected.
4484                  */
4485                 ext4_lblk_t     end = 0;
4486                 pgoff_t         last_offset;
4487                 pgoff_t         offset;
4488                 pgoff_t         index;
4489                 pgoff_t         start_index = 0;
4490                 struct page     **pages = NULL;
4491                 struct buffer_head *bh = NULL;
4492                 struct buffer_head *head = NULL;
4493                 unsigned int nr_pages = PAGE_SIZE / sizeof(struct page *);
4494
4495                 pages = kmalloc(PAGE_SIZE, GFP_KERNEL);
4496                 if (pages == NULL)
4497                         return -ENOMEM;
4498
4499                 offset = logical >> PAGE_SHIFT;
4500 repeat:
4501                 last_offset = offset;
4502                 head = NULL;
4503                 ret = find_get_pages_tag(inode->i_mapping, &offset,
4504                                         PAGECACHE_TAG_DIRTY, nr_pages, pages);
4505
4506                 if (!(flags & FIEMAP_EXTENT_DELALLOC)) {
4507                         /* First time, try to find a mapped buffer. */
4508                         if (ret == 0) {
4509 out:
4510                                 for (index = 0; index < ret; index++)
4511                                         page_cache_release(pages[index]);
4512                                 /* just a hole. */
4513                                 kfree(pages);
4514                                 return EXT_CONTINUE;
4515                         }
4516                         index = 0;
4517
4518 next_page:
4519                         /* Try to find the 1st mapped buffer. */
4520                         end = ((__u64)pages[index]->index << PAGE_SHIFT) >>
4521                                   blksize_bits;
4522                         if (!page_has_buffers(pages[index]))
4523                                 goto out;
4524                         head = page_buffers(pages[index]);
4525                         if (!head)
4526                                 goto out;
4527
4528                         index++;
4529                         bh = head;
4530                         do {
4531                                 if (end >= newex->ec_block +
4532                                         newex->ec_len)
4533                                         /* The buffer is out of
4534                                          * the request range.
4535                                          */
4536                                         goto out;
4537
4538                                 if (buffer_mapped(bh) &&
4539                                     end >= newex->ec_block) {
4540                                         start_index = index - 1;
4541                                         /* get the 1st mapped buffer. */
4542                                         goto found_mapped_buffer;
4543                                 }
4544
4545                                 bh = bh->b_this_page;
4546                                 end++;
4547                         } while (bh != head);
4548
4549                         /* No mapped buffer in the range found in this page,
4550                          * We need to look up next page.
4551                          */
4552                         if (index >= ret) {
4553                                 /* There is no page left, but we need to limit
4554                                  * newex->ec_len.
4555                                  */
4556                                 newex->ec_len = end - newex->ec_block;
4557                                 goto out;
4558                         }
4559                         goto next_page;
4560                 } else {
4561                         /*Find contiguous delayed buffers. */
4562                         if (ret > 0 && pages[0]->index == last_offset)
4563                                 head = page_buffers(pages[0]);
4564                         bh = head;
4565                         index = 1;
4566                         start_index = 0;
4567                 }
4568
4569 found_mapped_buffer:
4570                 if (bh != NULL && buffer_delay(bh)) {
4571                         /* 1st or contiguous delayed buffer found. */
4572                         if (!(flags & FIEMAP_EXTENT_DELALLOC)) {
4573                                 /*
4574                                  * 1st delayed buffer found, record
4575                                  * the start of extent.
4576                                  */
4577                                 flags |= FIEMAP_EXTENT_DELALLOC;
4578                                 newex->ec_block = end;
4579                                 logical = (__u64)end << blksize_bits;
4580                         }
4581                         /* Find contiguous delayed buffers. */
4582                         do {
4583                                 if (!buffer_delay(bh))
4584                                         goto found_delayed_extent;
4585                                 bh = bh->b_this_page;
4586                                 end++;
4587                         } while (bh != head);
4588
4589                         for (; index < ret; index++) {
4590                                 if (!page_has_buffers(pages[index])) {
4591                                         bh = NULL;
4592                                         break;
4593                                 }
4594                                 head = page_buffers(pages[index]);
4595                                 if (!head) {
4596                                         bh = NULL;
4597                                         break;
4598                                 }
4599
4600                                 if (pages[index]->index !=
4601                                     pages[start_index]->index + index
4602                                     - start_index) {
4603                                         /* Blocks are not contiguous. */
4604                                         bh = NULL;
4605                                         break;
4606                                 }
4607                                 bh = head;
4608                                 do {
4609                                         if (!buffer_delay(bh))
4610                                                 /* Delayed-extent ends. */
4611                                                 goto found_delayed_extent;
4612                                         bh = bh->b_this_page;
4613                                         end++;
4614                                 } while (bh != head);
4615                         }
4616                 } else if (!(flags & FIEMAP_EXTENT_DELALLOC))
4617                         /* a hole found. */
4618                         goto out;
4619
4620 found_delayed_extent:
4621                 newex->ec_len = min(end - newex->ec_block,
4622                                                 (ext4_lblk_t)EXT_INIT_MAX_LEN);
4623                 if (ret == nr_pages && bh != NULL &&
4624                         newex->ec_len < EXT_INIT_MAX_LEN &&
4625                         buffer_delay(bh)) {
4626                         /* Have not collected an extent and continue. */
4627                         for (index = 0; index < ret; index++)
4628                                 page_cache_release(pages[index]);
4629                         goto repeat;
4630                 }
4631
4632                 for (index = 0; index < ret; index++)
4633                         page_cache_release(pages[index]);
4634                 kfree(pages);
4635         }
4636
4637         physical = (__u64)newex->ec_start << blksize_bits;
4638         length =   (__u64)newex->ec_len << blksize_bits;
4639
4640         if (ex && ext4_ext_is_uninitialized(ex))
4641                 flags |= FIEMAP_EXTENT_UNWRITTEN;
4642
4643         if (next == EXT_MAX_BLOCKS)
4644                 flags |= FIEMAP_EXTENT_LAST;
4645
4646         ret = fiemap_fill_next_extent(fieinfo, logical, physical,
4647                                         length, flags);
4648         if (ret < 0)
4649                 return ret;
4650         if (ret == 1)
4651                 return EXT_BREAK;
4652         return EXT_CONTINUE;
4653 }
4654 /* fiemap flags we can handle specified here */
4655 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4656
4657 static int ext4_xattr_fiemap(struct inode *inode,
4658                                 struct fiemap_extent_info *fieinfo)
4659 {
4660         __u64 physical = 0;
4661         __u64 length;
4662         __u32 flags = FIEMAP_EXTENT_LAST;
4663         int blockbits = inode->i_sb->s_blocksize_bits;
4664         int error = 0;
4665
4666         /* in-inode? */
4667         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4668                 struct ext4_iloc iloc;
4669                 int offset;     /* offset of xattr in inode */
4670
4671                 error = ext4_get_inode_loc(inode, &iloc);
4672                 if (error)
4673                         return error;
4674                 physical = iloc.bh->b_blocknr << blockbits;
4675                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4676                                 EXT4_I(inode)->i_extra_isize;
4677                 physical += offset;
4678                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4679                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4680                 brelse(iloc.bh);
4681         } else { /* external block */
4682                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4683                 length = inode->i_sb->s_blocksize;
4684         }
4685
4686         if (physical)
4687                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4688                                                 length, flags);
4689         return (error < 0 ? error : 0);
4690 }
4691
4692 /*
4693  * ext4_ext_punch_hole
4694  *
4695  * Punches a hole of "length" bytes in a file starting
4696  * at byte "offset"
4697  *
4698  * @inode:  The inode of the file to punch a hole in
4699  * @offset: The starting byte offset of the hole
4700  * @length: The length of the hole
4701  *
4702  * Returns the number of blocks removed or negative on err
4703  */
4704 int ext4_ext_punch_hole(struct file *file, loff_t offset, loff_t length)
4705 {
4706         struct inode *inode = file->f_path.dentry->d_inode;
4707         struct super_block *sb = inode->i_sb;
4708         struct ext4_ext_cache cache_ex;
4709         ext4_lblk_t first_block, last_block, num_blocks, iblock, max_blocks;
4710         struct address_space *mapping = inode->i_mapping;
4711         struct ext4_map_blocks map;
4712         handle_t *handle;
4713         loff_t first_page, last_page, page_len;
4714         loff_t first_page_offset, last_page_offset;
4715         int ret, credits, blocks_released, err = 0;
4716
4717         /* No need to punch hole beyond i_size */
4718         if (offset >= inode->i_size)
4719                 return 0;
4720
4721         /*
4722          * If the hole extends beyond i_size, set the hole
4723          * to end after the page that contains i_size
4724          */
4725         if (offset + length > inode->i_size) {
4726                 length = inode->i_size +
4727                    PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
4728                    offset;
4729         }
4730
4731         first_block = (offset + sb->s_blocksize - 1) >>
4732                 EXT4_BLOCK_SIZE_BITS(sb);
4733         last_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
4734
4735         first_page = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
4736         last_page = (offset + length) >> PAGE_CACHE_SHIFT;
4737
4738         first_page_offset = first_page << PAGE_CACHE_SHIFT;
4739         last_page_offset = last_page << PAGE_CACHE_SHIFT;
4740
4741         /*
4742          * Write out all dirty pages to avoid race conditions
4743          * Then release them.
4744          */
4745         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4746                 err = filemap_write_and_wait_range(mapping,
4747                         offset, offset + length - 1);
4748
4749                 if (err)
4750                         return err;
4751         }
4752
4753         /* Now release the pages */
4754         if (last_page_offset > first_page_offset) {
4755                 truncate_inode_pages_range(mapping, first_page_offset,
4756                                            last_page_offset-1);
4757         }
4758
4759         /* finish any pending end_io work */
4760         ext4_flush_completed_IO(inode);
4761
4762         credits = ext4_writepage_trans_blocks(inode);
4763         handle = ext4_journal_start(inode, credits);
4764         if (IS_ERR(handle))
4765                 return PTR_ERR(handle);
4766
4767         err = ext4_orphan_add(handle, inode);
4768         if (err)
4769                 goto out;
4770
4771         /*
4772          * Now we need to zero out the non-page-aligned data in the
4773          * pages at the start and tail of the hole, and unmap the buffer
4774          * heads for the block aligned regions of the page that were
4775          * completely zeroed.
4776          */
4777         if (first_page > last_page) {
4778                 /*
4779                  * If the file space being truncated is contained within a page
4780                  * just zero out and unmap the middle of that page
4781                  */
4782                 err = ext4_discard_partial_page_buffers(handle,
4783                         mapping, offset, length, 0);
4784
4785                 if (err)
4786                         goto out;
4787         } else {
4788                 /*
4789                  * zero out and unmap the partial page that contains
4790                  * the start of the hole
4791                  */
4792                 page_len  = first_page_offset - offset;
4793                 if (page_len > 0) {
4794                         err = ext4_discard_partial_page_buffers(handle, mapping,
4795                                                    offset, page_len, 0);
4796                         if (err)
4797                                 goto out;
4798                 }
4799
4800                 /*
4801                  * zero out and unmap the partial page that contains
4802                  * the end of the hole
4803                  */
4804                 page_len = offset + length - last_page_offset;
4805                 if (page_len > 0) {
4806                         err = ext4_discard_partial_page_buffers(handle, mapping,
4807                                         last_page_offset, page_len, 0);
4808                         if (err)
4809                                 goto out;
4810                 }
4811         }
4812
4813
4814         /*
4815          * If i_size is contained in the last page, we need to
4816          * unmap and zero the partial page after i_size
4817          */
4818         if (inode->i_size >> PAGE_CACHE_SHIFT == last_page &&
4819            inode->i_size % PAGE_CACHE_SIZE != 0) {
4820
4821                 page_len = PAGE_CACHE_SIZE -
4822                         (inode->i_size & (PAGE_CACHE_SIZE - 1));
4823
4824                 if (page_len > 0) {
4825                         err = ext4_discard_partial_page_buffers(handle,
4826                           mapping, inode->i_size, page_len, 0);
4827
4828                         if (err)
4829                                 goto out;
4830                 }
4831         }
4832
4833         /* If there are no blocks to remove, return now */
4834         if (first_block >= last_block)
4835                 goto out;
4836
4837         down_write(&EXT4_I(inode)->i_data_sem);
4838         ext4_ext_invalidate_cache(inode);
4839         ext4_discard_preallocations(inode);
4840
4841         /*
4842          * Loop over all the blocks and identify blocks
4843          * that need to be punched out
4844          */
4845         iblock = first_block;
4846         blocks_released = 0;
4847         while (iblock < last_block) {
4848                 max_blocks = last_block - iblock;
4849                 num_blocks = 1;
4850                 memset(&map, 0, sizeof(map));
4851                 map.m_lblk = iblock;
4852                 map.m_len = max_blocks;
4853                 ret = ext4_ext_map_blocks(handle, inode, &map,
4854                         EXT4_GET_BLOCKS_PUNCH_OUT_EXT);
4855
4856                 if (ret > 0) {
4857                         blocks_released += ret;
4858                         num_blocks = ret;
4859                 } else if (ret == 0) {
4860                         /*
4861                          * If map blocks could not find the block,
4862                          * then it is in a hole.  If the hole was
4863                          * not already cached, then map blocks should
4864                          * put it in the cache.  So we can get the hole
4865                          * out of the cache
4866                          */
4867                         memset(&cache_ex, 0, sizeof(cache_ex));
4868                         if ((ext4_ext_check_cache(inode, iblock, &cache_ex)) &&
4869                                 !cache_ex.ec_start) {
4870
4871                                 /* The hole is cached */
4872                                 num_blocks = cache_ex.ec_block +
4873                                 cache_ex.ec_len - iblock;
4874
4875                         } else {
4876                                 /* The block could not be identified */
4877                                 err = -EIO;
4878                                 break;
4879                         }
4880                 } else {
4881                         /* Map blocks error */
4882                         err = ret;
4883                         break;
4884                 }
4885
4886                 if (num_blocks == 0) {
4887                         /* This condition should never happen */
4888                         ext_debug("Block lookup failed");
4889                         err = -EIO;
4890                         break;
4891                 }
4892
4893                 iblock += num_blocks;
4894         }
4895
4896         if (blocks_released > 0) {
4897                 ext4_ext_invalidate_cache(inode);
4898                 ext4_discard_preallocations(inode);
4899         }
4900
4901         if (IS_SYNC(inode))
4902                 ext4_handle_sync(handle);
4903
4904         up_write(&EXT4_I(inode)->i_data_sem);
4905
4906 out:
4907         ext4_orphan_del(handle, inode);
4908         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4909         ext4_mark_inode_dirty(handle, inode);
4910         ext4_journal_stop(handle);
4911         return err;
4912 }
4913 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4914                 __u64 start, __u64 len)
4915 {
4916         ext4_lblk_t start_blk;
4917         int error = 0;
4918
4919         /* fallback to generic here if not in extents fmt */
4920         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4921                 return generic_block_fiemap(inode, fieinfo, start, len,
4922                         ext4_get_block);
4923
4924         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4925                 return -EBADR;
4926
4927         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4928                 error = ext4_xattr_fiemap(inode, fieinfo);
4929         } else {
4930                 ext4_lblk_t len_blks;
4931                 __u64 last_blk;
4932
4933                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4934                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4935                 if (last_blk >= EXT_MAX_BLOCKS)
4936                         last_blk = EXT_MAX_BLOCKS-1;
4937                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4938
4939                 /*
4940                  * Walk the extent tree gathering extent information.
4941                  * ext4_ext_fiemap_cb will push extents back to user.
4942                  */
4943                 error = ext4_ext_walk_space(inode, start_blk, len_blks,
4944                                           ext4_ext_fiemap_cb, fieinfo);
4945         }
4946
4947         return error;
4948 }