bc0f1910b9cfa7dd5aa93b9c5d889c6fbacd037d
[linux-3.10.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 #include "ext4_extents.h"
45 #include "xattr.h"
46
47 #include <trace/events/ext4.h>
48
49 /*
50  * used by extent splitting.
51  */
52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
53                                         due to ENOSPC */
54 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
55 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
56
57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
59
60 static __le32 ext4_extent_block_csum(struct inode *inode,
61                                      struct ext4_extent_header *eh)
62 {
63         struct ext4_inode_info *ei = EXT4_I(inode);
64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
65         __u32 csum;
66
67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
68                            EXT4_EXTENT_TAIL_OFFSET(eh));
69         return cpu_to_le32(csum);
70 }
71
72 static int ext4_extent_block_csum_verify(struct inode *inode,
73                                          struct ext4_extent_header *eh)
74 {
75         struct ext4_extent_tail *et;
76
77         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
78                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
79                 return 1;
80
81         et = find_ext4_extent_tail(eh);
82         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
83                 return 0;
84         return 1;
85 }
86
87 static void ext4_extent_block_csum_set(struct inode *inode,
88                                        struct ext4_extent_header *eh)
89 {
90         struct ext4_extent_tail *et;
91
92         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
93                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
94                 return;
95
96         et = find_ext4_extent_tail(eh);
97         et->et_checksum = ext4_extent_block_csum(inode, eh);
98 }
99
100 static int ext4_split_extent(handle_t *handle,
101                                 struct inode *inode,
102                                 struct ext4_ext_path *path,
103                                 struct ext4_map_blocks *map,
104                                 int split_flag,
105                                 int flags);
106
107 static int ext4_split_extent_at(handle_t *handle,
108                              struct inode *inode,
109                              struct ext4_ext_path *path,
110                              ext4_lblk_t split,
111                              int split_flag,
112                              int flags);
113
114 static int ext4_find_delayed_extent(struct inode *inode,
115                                     struct extent_status *newes);
116
117 static int ext4_ext_truncate_extend_restart(handle_t *handle,
118                                             struct inode *inode,
119                                             int needed)
120 {
121         int err;
122
123         if (!ext4_handle_valid(handle))
124                 return 0;
125         if (handle->h_buffer_credits > needed)
126                 return 0;
127         err = ext4_journal_extend(handle, needed);
128         if (err <= 0)
129                 return err;
130         err = ext4_truncate_restart_trans(handle, inode, needed);
131         if (err == 0)
132                 err = -EAGAIN;
133
134         return err;
135 }
136
137 /*
138  * could return:
139  *  - EROFS
140  *  - ENOMEM
141  */
142 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
143                                 struct ext4_ext_path *path)
144 {
145         if (path->p_bh) {
146                 /* path points to block */
147                 return ext4_journal_get_write_access(handle, path->p_bh);
148         }
149         /* path points to leaf/index in inode body */
150         /* we use in-core data, no need to protect them */
151         return 0;
152 }
153
154 /*
155  * could return:
156  *  - EROFS
157  *  - ENOMEM
158  *  - EIO
159  */
160 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
161                      struct inode *inode, struct ext4_ext_path *path)
162 {
163         int err;
164         if (path->p_bh) {
165                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
166                 /* path points to block */
167                 err = __ext4_handle_dirty_metadata(where, line, handle,
168                                                    inode, path->p_bh);
169         } else {
170                 /* path points to leaf/index in inode body */
171                 err = ext4_mark_inode_dirty(handle, inode);
172         }
173         return err;
174 }
175
176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
177                               struct ext4_ext_path *path,
178                               ext4_lblk_t block)
179 {
180         if (path) {
181                 int depth = path->p_depth;
182                 struct ext4_extent *ex;
183
184                 /*
185                  * Try to predict block placement assuming that we are
186                  * filling in a file which will eventually be
187                  * non-sparse --- i.e., in the case of libbfd writing
188                  * an ELF object sections out-of-order but in a way
189                  * the eventually results in a contiguous object or
190                  * executable file, or some database extending a table
191                  * space file.  However, this is actually somewhat
192                  * non-ideal if we are writing a sparse file such as
193                  * qemu or KVM writing a raw image file that is going
194                  * to stay fairly sparse, since it will end up
195                  * fragmenting the file system's free space.  Maybe we
196                  * should have some hueristics or some way to allow
197                  * userspace to pass a hint to file system,
198                  * especially if the latter case turns out to be
199                  * common.
200                  */
201                 ex = path[depth].p_ext;
202                 if (ex) {
203                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
204                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
205
206                         if (block > ext_block)
207                                 return ext_pblk + (block - ext_block);
208                         else
209                                 return ext_pblk - (ext_block - block);
210                 }
211
212                 /* it looks like index is empty;
213                  * try to find starting block from index itself */
214                 if (path[depth].p_bh)
215                         return path[depth].p_bh->b_blocknr;
216         }
217
218         /* OK. use inode's group */
219         return ext4_inode_to_goal_block(inode);
220 }
221
222 /*
223  * Allocation for a meta data block
224  */
225 static ext4_fsblk_t
226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
227                         struct ext4_ext_path *path,
228                         struct ext4_extent *ex, int *err, unsigned int flags)
229 {
230         ext4_fsblk_t goal, newblock;
231
232         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
233         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
234                                         NULL, err);
235         return newblock;
236 }
237
238 static inline int ext4_ext_space_block(struct inode *inode, int check)
239 {
240         int size;
241
242         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
243                         / sizeof(struct ext4_extent);
244 #ifdef AGGRESSIVE_TEST
245         if (!check && size > 6)
246                 size = 6;
247 #endif
248         return size;
249 }
250
251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
252 {
253         int size;
254
255         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
256                         / sizeof(struct ext4_extent_idx);
257 #ifdef AGGRESSIVE_TEST
258         if (!check && size > 5)
259                 size = 5;
260 #endif
261         return size;
262 }
263
264 static inline int ext4_ext_space_root(struct inode *inode, int check)
265 {
266         int size;
267
268         size = sizeof(EXT4_I(inode)->i_data);
269         size -= sizeof(struct ext4_extent_header);
270         size /= sizeof(struct ext4_extent);
271 #ifdef AGGRESSIVE_TEST
272         if (!check && size > 3)
273                 size = 3;
274 #endif
275         return size;
276 }
277
278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
279 {
280         int size;
281
282         size = sizeof(EXT4_I(inode)->i_data);
283         size -= sizeof(struct ext4_extent_header);
284         size /= sizeof(struct ext4_extent_idx);
285 #ifdef AGGRESSIVE_TEST
286         if (!check && size > 4)
287                 size = 4;
288 #endif
289         return size;
290 }
291
292 /*
293  * Calculate the number of metadata blocks needed
294  * to allocate @blocks
295  * Worse case is one block per extent
296  */
297 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
298 {
299         struct ext4_inode_info *ei = EXT4_I(inode);
300         int idxs;
301
302         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
303                 / sizeof(struct ext4_extent_idx));
304
305         /*
306          * If the new delayed allocation block is contiguous with the
307          * previous da block, it can share index blocks with the
308          * previous block, so we only need to allocate a new index
309          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
310          * an additional index block, and at ldxs**3 blocks, yet
311          * another index blocks.
312          */
313         if (ei->i_da_metadata_calc_len &&
314             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
315                 int num = 0;
316
317                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
318                         num++;
319                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
320                         num++;
321                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
322                         num++;
323                         ei->i_da_metadata_calc_len = 0;
324                 } else
325                         ei->i_da_metadata_calc_len++;
326                 ei->i_da_metadata_calc_last_lblock++;
327                 return num;
328         }
329
330         /*
331          * In the worst case we need a new set of index blocks at
332          * every level of the inode's extent tree.
333          */
334         ei->i_da_metadata_calc_len = 1;
335         ei->i_da_metadata_calc_last_lblock = lblock;
336         return ext_depth(inode) + 1;
337 }
338
339 static int
340 ext4_ext_max_entries(struct inode *inode, int depth)
341 {
342         int max;
343
344         if (depth == ext_depth(inode)) {
345                 if (depth == 0)
346                         max = ext4_ext_space_root(inode, 1);
347                 else
348                         max = ext4_ext_space_root_idx(inode, 1);
349         } else {
350                 if (depth == 0)
351                         max = ext4_ext_space_block(inode, 1);
352                 else
353                         max = ext4_ext_space_block_idx(inode, 1);
354         }
355
356         return max;
357 }
358
359 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
360 {
361         ext4_fsblk_t block = ext4_ext_pblock(ext);
362         int len = ext4_ext_get_actual_len(ext);
363
364         if (len == 0)
365                 return 0;
366         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
367 }
368
369 static int ext4_valid_extent_idx(struct inode *inode,
370                                 struct ext4_extent_idx *ext_idx)
371 {
372         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
373
374         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
375 }
376
377 static int ext4_valid_extent_entries(struct inode *inode,
378                                 struct ext4_extent_header *eh,
379                                 int depth)
380 {
381         unsigned short entries;
382         if (eh->eh_entries == 0)
383                 return 1;
384
385         entries = le16_to_cpu(eh->eh_entries);
386
387         if (depth == 0) {
388                 /* leaf entries */
389                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
390                 while (entries) {
391                         if (!ext4_valid_extent(inode, ext))
392                                 return 0;
393                         ext++;
394                         entries--;
395                 }
396         } else {
397                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
398                 while (entries) {
399                         if (!ext4_valid_extent_idx(inode, ext_idx))
400                                 return 0;
401                         ext_idx++;
402                         entries--;
403                 }
404         }
405         return 1;
406 }
407
408 static int __ext4_ext_check(const char *function, unsigned int line,
409                             struct inode *inode, struct ext4_extent_header *eh,
410                             int depth)
411 {
412         const char *error_msg;
413         int max = 0;
414
415         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
416                 error_msg = "invalid magic";
417                 goto corrupted;
418         }
419         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
420                 error_msg = "unexpected eh_depth";
421                 goto corrupted;
422         }
423         if (unlikely(eh->eh_max == 0)) {
424                 error_msg = "invalid eh_max";
425                 goto corrupted;
426         }
427         max = ext4_ext_max_entries(inode, depth);
428         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
429                 error_msg = "too large eh_max";
430                 goto corrupted;
431         }
432         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
433                 error_msg = "invalid eh_entries";
434                 goto corrupted;
435         }
436         if (!ext4_valid_extent_entries(inode, eh, depth)) {
437                 error_msg = "invalid extent entries";
438                 goto corrupted;
439         }
440         /* Verify checksum on non-root extent tree nodes */
441         if (ext_depth(inode) != depth &&
442             !ext4_extent_block_csum_verify(inode, eh)) {
443                 error_msg = "extent tree corrupted";
444                 goto corrupted;
445         }
446         return 0;
447
448 corrupted:
449         ext4_error_inode(inode, function, line, 0,
450                         "bad header/extent: %s - magic %x, "
451                         "entries %u, max %u(%u), depth %u(%u)",
452                         error_msg, le16_to_cpu(eh->eh_magic),
453                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
454                         max, le16_to_cpu(eh->eh_depth), depth);
455
456         return -EIO;
457 }
458
459 #define ext4_ext_check(inode, eh, depth)        \
460         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
461
462 int ext4_ext_check_inode(struct inode *inode)
463 {
464         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
465 }
466
467 static int __ext4_ext_check_block(const char *function, unsigned int line,
468                                   struct inode *inode,
469                                   struct ext4_extent_header *eh,
470                                   int depth,
471                                   struct buffer_head *bh)
472 {
473         int ret;
474
475         if (buffer_verified(bh))
476                 return 0;
477         ret = ext4_ext_check(inode, eh, depth);
478         if (ret)
479                 return ret;
480         set_buffer_verified(bh);
481         return ret;
482 }
483
484 #define ext4_ext_check_block(inode, eh, depth, bh)      \
485         __ext4_ext_check_block(__func__, __LINE__, inode, eh, depth, bh)
486
487 #ifdef EXT_DEBUG
488 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
489 {
490         int k, l = path->p_depth;
491
492         ext_debug("path:");
493         for (k = 0; k <= l; k++, path++) {
494                 if (path->p_idx) {
495                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
496                             ext4_idx_pblock(path->p_idx));
497                 } else if (path->p_ext) {
498                         ext_debug("  %d:[%d]%d:%llu ",
499                                   le32_to_cpu(path->p_ext->ee_block),
500                                   ext4_ext_is_uninitialized(path->p_ext),
501                                   ext4_ext_get_actual_len(path->p_ext),
502                                   ext4_ext_pblock(path->p_ext));
503                 } else
504                         ext_debug("  []");
505         }
506         ext_debug("\n");
507 }
508
509 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
510 {
511         int depth = ext_depth(inode);
512         struct ext4_extent_header *eh;
513         struct ext4_extent *ex;
514         int i;
515
516         if (!path)
517                 return;
518
519         eh = path[depth].p_hdr;
520         ex = EXT_FIRST_EXTENT(eh);
521
522         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
523
524         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
525                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
526                           ext4_ext_is_uninitialized(ex),
527                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
528         }
529         ext_debug("\n");
530 }
531
532 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
533                         ext4_fsblk_t newblock, int level)
534 {
535         int depth = ext_depth(inode);
536         struct ext4_extent *ex;
537
538         if (depth != level) {
539                 struct ext4_extent_idx *idx;
540                 idx = path[level].p_idx;
541                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
542                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
543                                         le32_to_cpu(idx->ei_block),
544                                         ext4_idx_pblock(idx),
545                                         newblock);
546                         idx++;
547                 }
548
549                 return;
550         }
551
552         ex = path[depth].p_ext;
553         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
554                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
555                                 le32_to_cpu(ex->ee_block),
556                                 ext4_ext_pblock(ex),
557                                 ext4_ext_is_uninitialized(ex),
558                                 ext4_ext_get_actual_len(ex),
559                                 newblock);
560                 ex++;
561         }
562 }
563
564 #else
565 #define ext4_ext_show_path(inode, path)
566 #define ext4_ext_show_leaf(inode, path)
567 #define ext4_ext_show_move(inode, path, newblock, level)
568 #endif
569
570 void ext4_ext_drop_refs(struct ext4_ext_path *path)
571 {
572         int depth = path->p_depth;
573         int i;
574
575         for (i = 0; i <= depth; i++, path++)
576                 if (path->p_bh) {
577                         brelse(path->p_bh);
578                         path->p_bh = NULL;
579                 }
580 }
581
582 /*
583  * ext4_ext_binsearch_idx:
584  * binary search for the closest index of the given block
585  * the header must be checked before calling this
586  */
587 static void
588 ext4_ext_binsearch_idx(struct inode *inode,
589                         struct ext4_ext_path *path, ext4_lblk_t block)
590 {
591         struct ext4_extent_header *eh = path->p_hdr;
592         struct ext4_extent_idx *r, *l, *m;
593
594
595         ext_debug("binsearch for %u(idx):  ", block);
596
597         l = EXT_FIRST_INDEX(eh) + 1;
598         r = EXT_LAST_INDEX(eh);
599         while (l <= r) {
600                 m = l + (r - l) / 2;
601                 if (block < le32_to_cpu(m->ei_block))
602                         r = m - 1;
603                 else
604                         l = m + 1;
605                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
606                                 m, le32_to_cpu(m->ei_block),
607                                 r, le32_to_cpu(r->ei_block));
608         }
609
610         path->p_idx = l - 1;
611         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
612                   ext4_idx_pblock(path->p_idx));
613
614 #ifdef CHECK_BINSEARCH
615         {
616                 struct ext4_extent_idx *chix, *ix;
617                 int k;
618
619                 chix = ix = EXT_FIRST_INDEX(eh);
620                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
621                   if (k != 0 &&
622                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
623                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
624                                        "first=0x%p\n", k,
625                                        ix, EXT_FIRST_INDEX(eh));
626                                 printk(KERN_DEBUG "%u <= %u\n",
627                                        le32_to_cpu(ix->ei_block),
628                                        le32_to_cpu(ix[-1].ei_block));
629                         }
630                         BUG_ON(k && le32_to_cpu(ix->ei_block)
631                                            <= le32_to_cpu(ix[-1].ei_block));
632                         if (block < le32_to_cpu(ix->ei_block))
633                                 break;
634                         chix = ix;
635                 }
636                 BUG_ON(chix != path->p_idx);
637         }
638 #endif
639
640 }
641
642 /*
643  * ext4_ext_binsearch:
644  * binary search for closest extent of the given block
645  * the header must be checked before calling this
646  */
647 static void
648 ext4_ext_binsearch(struct inode *inode,
649                 struct ext4_ext_path *path, ext4_lblk_t block)
650 {
651         struct ext4_extent_header *eh = path->p_hdr;
652         struct ext4_extent *r, *l, *m;
653
654         if (eh->eh_entries == 0) {
655                 /*
656                  * this leaf is empty:
657                  * we get such a leaf in split/add case
658                  */
659                 return;
660         }
661
662         ext_debug("binsearch for %u:  ", block);
663
664         l = EXT_FIRST_EXTENT(eh) + 1;
665         r = EXT_LAST_EXTENT(eh);
666
667         while (l <= r) {
668                 m = l + (r - l) / 2;
669                 if (block < le32_to_cpu(m->ee_block))
670                         r = m - 1;
671                 else
672                         l = m + 1;
673                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
674                                 m, le32_to_cpu(m->ee_block),
675                                 r, le32_to_cpu(r->ee_block));
676         }
677
678         path->p_ext = l - 1;
679         ext_debug("  -> %d:%llu:[%d]%d ",
680                         le32_to_cpu(path->p_ext->ee_block),
681                         ext4_ext_pblock(path->p_ext),
682                         ext4_ext_is_uninitialized(path->p_ext),
683                         ext4_ext_get_actual_len(path->p_ext));
684
685 #ifdef CHECK_BINSEARCH
686         {
687                 struct ext4_extent *chex, *ex;
688                 int k;
689
690                 chex = ex = EXT_FIRST_EXTENT(eh);
691                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
692                         BUG_ON(k && le32_to_cpu(ex->ee_block)
693                                           <= le32_to_cpu(ex[-1].ee_block));
694                         if (block < le32_to_cpu(ex->ee_block))
695                                 break;
696                         chex = ex;
697                 }
698                 BUG_ON(chex != path->p_ext);
699         }
700 #endif
701
702 }
703
704 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
705 {
706         struct ext4_extent_header *eh;
707
708         eh = ext_inode_hdr(inode);
709         eh->eh_depth = 0;
710         eh->eh_entries = 0;
711         eh->eh_magic = EXT4_EXT_MAGIC;
712         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
713         ext4_mark_inode_dirty(handle, inode);
714         return 0;
715 }
716
717 struct ext4_ext_path *
718 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
719                                         struct ext4_ext_path *path)
720 {
721         struct ext4_extent_header *eh;
722         struct buffer_head *bh;
723         short int depth, i, ppos = 0, alloc = 0;
724         int ret;
725
726         eh = ext_inode_hdr(inode);
727         depth = ext_depth(inode);
728
729         /* account possible depth increase */
730         if (!path) {
731                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
732                                 GFP_NOFS);
733                 if (!path)
734                         return ERR_PTR(-ENOMEM);
735                 alloc = 1;
736         }
737         path[0].p_hdr = eh;
738         path[0].p_bh = NULL;
739
740         i = depth;
741         /* walk through the tree */
742         while (i) {
743                 ext_debug("depth %d: num %d, max %d\n",
744                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
745
746                 ext4_ext_binsearch_idx(inode, path + ppos, block);
747                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
748                 path[ppos].p_depth = i;
749                 path[ppos].p_ext = NULL;
750
751                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
752                 if (unlikely(!bh)) {
753                         ret = -ENOMEM;
754                         goto err;
755                 }
756                 if (!bh_uptodate_or_lock(bh)) {
757                         trace_ext4_ext_load_extent(inode, block,
758                                                 path[ppos].p_block);
759                         ret = bh_submit_read(bh);
760                         if (ret < 0) {
761                                 put_bh(bh);
762                                 goto err;
763                         }
764                 }
765                 eh = ext_block_hdr(bh);
766                 ppos++;
767                 if (unlikely(ppos > depth)) {
768                         put_bh(bh);
769                         EXT4_ERROR_INODE(inode,
770                                          "ppos %d > depth %d", ppos, depth);
771                         ret = -EIO;
772                         goto err;
773                 }
774                 path[ppos].p_bh = bh;
775                 path[ppos].p_hdr = eh;
776                 i--;
777
778                 ret = ext4_ext_check_block(inode, eh, i, bh);
779                 if (ret < 0)
780                         goto err;
781         }
782
783         path[ppos].p_depth = i;
784         path[ppos].p_ext = NULL;
785         path[ppos].p_idx = NULL;
786
787         /* find extent */
788         ext4_ext_binsearch(inode, path + ppos, block);
789         /* if not an empty leaf */
790         if (path[ppos].p_ext)
791                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
792
793         ext4_ext_show_path(inode, path);
794
795         return path;
796
797 err:
798         ext4_ext_drop_refs(path);
799         if (alloc)
800                 kfree(path);
801         return ERR_PTR(ret);
802 }
803
804 /*
805  * ext4_ext_insert_index:
806  * insert new index [@logical;@ptr] into the block at @curp;
807  * check where to insert: before @curp or after @curp
808  */
809 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
810                                  struct ext4_ext_path *curp,
811                                  int logical, ext4_fsblk_t ptr)
812 {
813         struct ext4_extent_idx *ix;
814         int len, err;
815
816         err = ext4_ext_get_access(handle, inode, curp);
817         if (err)
818                 return err;
819
820         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
821                 EXT4_ERROR_INODE(inode,
822                                  "logical %d == ei_block %d!",
823                                  logical, le32_to_cpu(curp->p_idx->ei_block));
824                 return -EIO;
825         }
826
827         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
828                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
829                 EXT4_ERROR_INODE(inode,
830                                  "eh_entries %d >= eh_max %d!",
831                                  le16_to_cpu(curp->p_hdr->eh_entries),
832                                  le16_to_cpu(curp->p_hdr->eh_max));
833                 return -EIO;
834         }
835
836         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
837                 /* insert after */
838                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
839                 ix = curp->p_idx + 1;
840         } else {
841                 /* insert before */
842                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
843                 ix = curp->p_idx;
844         }
845
846         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
847         BUG_ON(len < 0);
848         if (len > 0) {
849                 ext_debug("insert new index %d: "
850                                 "move %d indices from 0x%p to 0x%p\n",
851                                 logical, len, ix, ix + 1);
852                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
853         }
854
855         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
856                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
857                 return -EIO;
858         }
859
860         ix->ei_block = cpu_to_le32(logical);
861         ext4_idx_store_pblock(ix, ptr);
862         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
863
864         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
865                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
866                 return -EIO;
867         }
868
869         err = ext4_ext_dirty(handle, inode, curp);
870         ext4_std_error(inode->i_sb, err);
871
872         return err;
873 }
874
875 /*
876  * ext4_ext_split:
877  * inserts new subtree into the path, using free index entry
878  * at depth @at:
879  * - allocates all needed blocks (new leaf and all intermediate index blocks)
880  * - makes decision where to split
881  * - moves remaining extents and index entries (right to the split point)
882  *   into the newly allocated blocks
883  * - initializes subtree
884  */
885 static int ext4_ext_split(handle_t *handle, struct inode *inode,
886                           unsigned int flags,
887                           struct ext4_ext_path *path,
888                           struct ext4_extent *newext, int at)
889 {
890         struct buffer_head *bh = NULL;
891         int depth = ext_depth(inode);
892         struct ext4_extent_header *neh;
893         struct ext4_extent_idx *fidx;
894         int i = at, k, m, a;
895         ext4_fsblk_t newblock, oldblock;
896         __le32 border;
897         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
898         int err = 0;
899
900         /* make decision: where to split? */
901         /* FIXME: now decision is simplest: at current extent */
902
903         /* if current leaf will be split, then we should use
904          * border from split point */
905         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
906                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
907                 return -EIO;
908         }
909         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
910                 border = path[depth].p_ext[1].ee_block;
911                 ext_debug("leaf will be split."
912                                 " next leaf starts at %d\n",
913                                   le32_to_cpu(border));
914         } else {
915                 border = newext->ee_block;
916                 ext_debug("leaf will be added."
917                                 " next leaf starts at %d\n",
918                                 le32_to_cpu(border));
919         }
920
921         /*
922          * If error occurs, then we break processing
923          * and mark filesystem read-only. index won't
924          * be inserted and tree will be in consistent
925          * state. Next mount will repair buffers too.
926          */
927
928         /*
929          * Get array to track all allocated blocks.
930          * We need this to handle errors and free blocks
931          * upon them.
932          */
933         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
934         if (!ablocks)
935                 return -ENOMEM;
936
937         /* allocate all needed blocks */
938         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
939         for (a = 0; a < depth - at; a++) {
940                 newblock = ext4_ext_new_meta_block(handle, inode, path,
941                                                    newext, &err, flags);
942                 if (newblock == 0)
943                         goto cleanup;
944                 ablocks[a] = newblock;
945         }
946
947         /* initialize new leaf */
948         newblock = ablocks[--a];
949         if (unlikely(newblock == 0)) {
950                 EXT4_ERROR_INODE(inode, "newblock == 0!");
951                 err = -EIO;
952                 goto cleanup;
953         }
954         bh = sb_getblk(inode->i_sb, newblock);
955         if (unlikely(!bh)) {
956                 err = -ENOMEM;
957                 goto cleanup;
958         }
959         lock_buffer(bh);
960
961         err = ext4_journal_get_create_access(handle, bh);
962         if (err)
963                 goto cleanup;
964
965         neh = ext_block_hdr(bh);
966         neh->eh_entries = 0;
967         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
968         neh->eh_magic = EXT4_EXT_MAGIC;
969         neh->eh_depth = 0;
970
971         /* move remainder of path[depth] to the new leaf */
972         if (unlikely(path[depth].p_hdr->eh_entries !=
973                      path[depth].p_hdr->eh_max)) {
974                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
975                                  path[depth].p_hdr->eh_entries,
976                                  path[depth].p_hdr->eh_max);
977                 err = -EIO;
978                 goto cleanup;
979         }
980         /* start copy from next extent */
981         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
982         ext4_ext_show_move(inode, path, newblock, depth);
983         if (m) {
984                 struct ext4_extent *ex;
985                 ex = EXT_FIRST_EXTENT(neh);
986                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
987                 le16_add_cpu(&neh->eh_entries, m);
988         }
989
990         ext4_extent_block_csum_set(inode, neh);
991         set_buffer_uptodate(bh);
992         unlock_buffer(bh);
993
994         err = ext4_handle_dirty_metadata(handle, inode, bh);
995         if (err)
996                 goto cleanup;
997         brelse(bh);
998         bh = NULL;
999
1000         /* correct old leaf */
1001         if (m) {
1002                 err = ext4_ext_get_access(handle, inode, path + depth);
1003                 if (err)
1004                         goto cleanup;
1005                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1006                 err = ext4_ext_dirty(handle, inode, path + depth);
1007                 if (err)
1008                         goto cleanup;
1009
1010         }
1011
1012         /* create intermediate indexes */
1013         k = depth - at - 1;
1014         if (unlikely(k < 0)) {
1015                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1016                 err = -EIO;
1017                 goto cleanup;
1018         }
1019         if (k)
1020                 ext_debug("create %d intermediate indices\n", k);
1021         /* insert new index into current index block */
1022         /* current depth stored in i var */
1023         i = depth - 1;
1024         while (k--) {
1025                 oldblock = newblock;
1026                 newblock = ablocks[--a];
1027                 bh = sb_getblk(inode->i_sb, newblock);
1028                 if (unlikely(!bh)) {
1029                         err = -ENOMEM;
1030                         goto cleanup;
1031                 }
1032                 lock_buffer(bh);
1033
1034                 err = ext4_journal_get_create_access(handle, bh);
1035                 if (err)
1036                         goto cleanup;
1037
1038                 neh = ext_block_hdr(bh);
1039                 neh->eh_entries = cpu_to_le16(1);
1040                 neh->eh_magic = EXT4_EXT_MAGIC;
1041                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1042                 neh->eh_depth = cpu_to_le16(depth - i);
1043                 fidx = EXT_FIRST_INDEX(neh);
1044                 fidx->ei_block = border;
1045                 ext4_idx_store_pblock(fidx, oldblock);
1046
1047                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1048                                 i, newblock, le32_to_cpu(border), oldblock);
1049
1050                 /* move remainder of path[i] to the new index block */
1051                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1052                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1053                         EXT4_ERROR_INODE(inode,
1054                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1055                                          le32_to_cpu(path[i].p_ext->ee_block));
1056                         err = -EIO;
1057                         goto cleanup;
1058                 }
1059                 /* start copy indexes */
1060                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1061                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1062                                 EXT_MAX_INDEX(path[i].p_hdr));
1063                 ext4_ext_show_move(inode, path, newblock, i);
1064                 if (m) {
1065                         memmove(++fidx, path[i].p_idx,
1066                                 sizeof(struct ext4_extent_idx) * m);
1067                         le16_add_cpu(&neh->eh_entries, m);
1068                 }
1069                 ext4_extent_block_csum_set(inode, neh);
1070                 set_buffer_uptodate(bh);
1071                 unlock_buffer(bh);
1072
1073                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1074                 if (err)
1075                         goto cleanup;
1076                 brelse(bh);
1077                 bh = NULL;
1078
1079                 /* correct old index */
1080                 if (m) {
1081                         err = ext4_ext_get_access(handle, inode, path + i);
1082                         if (err)
1083                                 goto cleanup;
1084                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1085                         err = ext4_ext_dirty(handle, inode, path + i);
1086                         if (err)
1087                                 goto cleanup;
1088                 }
1089
1090                 i--;
1091         }
1092
1093         /* insert new index */
1094         err = ext4_ext_insert_index(handle, inode, path + at,
1095                                     le32_to_cpu(border), newblock);
1096
1097 cleanup:
1098         if (bh) {
1099                 if (buffer_locked(bh))
1100                         unlock_buffer(bh);
1101                 brelse(bh);
1102         }
1103
1104         if (err) {
1105                 /* free all allocated blocks in error case */
1106                 for (i = 0; i < depth; i++) {
1107                         if (!ablocks[i])
1108                                 continue;
1109                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1110                                          EXT4_FREE_BLOCKS_METADATA);
1111                 }
1112         }
1113         kfree(ablocks);
1114
1115         return err;
1116 }
1117
1118 /*
1119  * ext4_ext_grow_indepth:
1120  * implements tree growing procedure:
1121  * - allocates new block
1122  * - moves top-level data (index block or leaf) into the new block
1123  * - initializes new top-level, creating index that points to the
1124  *   just created block
1125  */
1126 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1127                                  unsigned int flags,
1128                                  struct ext4_extent *newext)
1129 {
1130         struct ext4_extent_header *neh;
1131         struct buffer_head *bh;
1132         ext4_fsblk_t newblock;
1133         int err = 0;
1134
1135         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1136                 newext, &err, flags);
1137         if (newblock == 0)
1138                 return err;
1139
1140         bh = sb_getblk(inode->i_sb, newblock);
1141         if (unlikely(!bh))
1142                 return -ENOMEM;
1143         lock_buffer(bh);
1144
1145         err = ext4_journal_get_create_access(handle, bh);
1146         if (err) {
1147                 unlock_buffer(bh);
1148                 goto out;
1149         }
1150
1151         /* move top-level index/leaf into new block */
1152         memmove(bh->b_data, EXT4_I(inode)->i_data,
1153                 sizeof(EXT4_I(inode)->i_data));
1154
1155         /* set size of new block */
1156         neh = ext_block_hdr(bh);
1157         /* old root could have indexes or leaves
1158          * so calculate e_max right way */
1159         if (ext_depth(inode))
1160                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1161         else
1162                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1163         neh->eh_magic = EXT4_EXT_MAGIC;
1164         ext4_extent_block_csum_set(inode, neh);
1165         set_buffer_uptodate(bh);
1166         unlock_buffer(bh);
1167
1168         err = ext4_handle_dirty_metadata(handle, inode, bh);
1169         if (err)
1170                 goto out;
1171
1172         /* Update top-level index: num,max,pointer */
1173         neh = ext_inode_hdr(inode);
1174         neh->eh_entries = cpu_to_le16(1);
1175         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1176         if (neh->eh_depth == 0) {
1177                 /* Root extent block becomes index block */
1178                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1179                 EXT_FIRST_INDEX(neh)->ei_block =
1180                         EXT_FIRST_EXTENT(neh)->ee_block;
1181         }
1182         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1183                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1184                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1185                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1186
1187         le16_add_cpu(&neh->eh_depth, 1);
1188         ext4_mark_inode_dirty(handle, inode);
1189 out:
1190         brelse(bh);
1191
1192         return err;
1193 }
1194
1195 /*
1196  * ext4_ext_create_new_leaf:
1197  * finds empty index and adds new leaf.
1198  * if no free index is found, then it requests in-depth growing.
1199  */
1200 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1201                                     unsigned int flags,
1202                                     struct ext4_ext_path *path,
1203                                     struct ext4_extent *newext)
1204 {
1205         struct ext4_ext_path *curp;
1206         int depth, i, err = 0;
1207
1208 repeat:
1209         i = depth = ext_depth(inode);
1210
1211         /* walk up to the tree and look for free index entry */
1212         curp = path + depth;
1213         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1214                 i--;
1215                 curp--;
1216         }
1217
1218         /* we use already allocated block for index block,
1219          * so subsequent data blocks should be contiguous */
1220         if (EXT_HAS_FREE_INDEX(curp)) {
1221                 /* if we found index with free entry, then use that
1222                  * entry: create all needed subtree and add new leaf */
1223                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1224                 if (err)
1225                         goto out;
1226
1227                 /* refill path */
1228                 ext4_ext_drop_refs(path);
1229                 path = ext4_ext_find_extent(inode,
1230                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1231                                     path);
1232                 if (IS_ERR(path))
1233                         err = PTR_ERR(path);
1234         } else {
1235                 /* tree is full, time to grow in depth */
1236                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1237                 if (err)
1238                         goto out;
1239
1240                 /* refill path */
1241                 ext4_ext_drop_refs(path);
1242                 path = ext4_ext_find_extent(inode,
1243                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1244                                     path);
1245                 if (IS_ERR(path)) {
1246                         err = PTR_ERR(path);
1247                         goto out;
1248                 }
1249
1250                 /*
1251                  * only first (depth 0 -> 1) produces free space;
1252                  * in all other cases we have to split the grown tree
1253                  */
1254                 depth = ext_depth(inode);
1255                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1256                         /* now we need to split */
1257                         goto repeat;
1258                 }
1259         }
1260
1261 out:
1262         return err;
1263 }
1264
1265 /*
1266  * search the closest allocated block to the left for *logical
1267  * and returns it at @logical + it's physical address at @phys
1268  * if *logical is the smallest allocated block, the function
1269  * returns 0 at @phys
1270  * return value contains 0 (success) or error code
1271  */
1272 static int ext4_ext_search_left(struct inode *inode,
1273                                 struct ext4_ext_path *path,
1274                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1275 {
1276         struct ext4_extent_idx *ix;
1277         struct ext4_extent *ex;
1278         int depth, ee_len;
1279
1280         if (unlikely(path == NULL)) {
1281                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1282                 return -EIO;
1283         }
1284         depth = path->p_depth;
1285         *phys = 0;
1286
1287         if (depth == 0 && path->p_ext == NULL)
1288                 return 0;
1289
1290         /* usually extent in the path covers blocks smaller
1291          * then *logical, but it can be that extent is the
1292          * first one in the file */
1293
1294         ex = path[depth].p_ext;
1295         ee_len = ext4_ext_get_actual_len(ex);
1296         if (*logical < le32_to_cpu(ex->ee_block)) {
1297                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1298                         EXT4_ERROR_INODE(inode,
1299                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1300                                          *logical, le32_to_cpu(ex->ee_block));
1301                         return -EIO;
1302                 }
1303                 while (--depth >= 0) {
1304                         ix = path[depth].p_idx;
1305                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1306                                 EXT4_ERROR_INODE(inode,
1307                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1308                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1309                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1310                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1311                                   depth);
1312                                 return -EIO;
1313                         }
1314                 }
1315                 return 0;
1316         }
1317
1318         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1319                 EXT4_ERROR_INODE(inode,
1320                                  "logical %d < ee_block %d + ee_len %d!",
1321                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1322                 return -EIO;
1323         }
1324
1325         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1326         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1327         return 0;
1328 }
1329
1330 /*
1331  * search the closest allocated block to the right for *logical
1332  * and returns it at @logical + it's physical address at @phys
1333  * if *logical is the largest allocated block, the function
1334  * returns 0 at @phys
1335  * return value contains 0 (success) or error code
1336  */
1337 static int ext4_ext_search_right(struct inode *inode,
1338                                  struct ext4_ext_path *path,
1339                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1340                                  struct ext4_extent **ret_ex)
1341 {
1342         struct buffer_head *bh = NULL;
1343         struct ext4_extent_header *eh;
1344         struct ext4_extent_idx *ix;
1345         struct ext4_extent *ex;
1346         ext4_fsblk_t block;
1347         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1348         int ee_len;
1349
1350         if (unlikely(path == NULL)) {
1351                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1352                 return -EIO;
1353         }
1354         depth = path->p_depth;
1355         *phys = 0;
1356
1357         if (depth == 0 && path->p_ext == NULL)
1358                 return 0;
1359
1360         /* usually extent in the path covers blocks smaller
1361          * then *logical, but it can be that extent is the
1362          * first one in the file */
1363
1364         ex = path[depth].p_ext;
1365         ee_len = ext4_ext_get_actual_len(ex);
1366         if (*logical < le32_to_cpu(ex->ee_block)) {
1367                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1368                         EXT4_ERROR_INODE(inode,
1369                                          "first_extent(path[%d].p_hdr) != ex",
1370                                          depth);
1371                         return -EIO;
1372                 }
1373                 while (--depth >= 0) {
1374                         ix = path[depth].p_idx;
1375                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1376                                 EXT4_ERROR_INODE(inode,
1377                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1378                                                  *logical);
1379                                 return -EIO;
1380                         }
1381                 }
1382                 goto found_extent;
1383         }
1384
1385         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1386                 EXT4_ERROR_INODE(inode,
1387                                  "logical %d < ee_block %d + ee_len %d!",
1388                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1389                 return -EIO;
1390         }
1391
1392         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1393                 /* next allocated block in this leaf */
1394                 ex++;
1395                 goto found_extent;
1396         }
1397
1398         /* go up and search for index to the right */
1399         while (--depth >= 0) {
1400                 ix = path[depth].p_idx;
1401                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1402                         goto got_index;
1403         }
1404
1405         /* we've gone up to the root and found no index to the right */
1406         return 0;
1407
1408 got_index:
1409         /* we've found index to the right, let's
1410          * follow it and find the closest allocated
1411          * block to the right */
1412         ix++;
1413         block = ext4_idx_pblock(ix);
1414         while (++depth < path->p_depth) {
1415                 bh = sb_bread(inode->i_sb, block);
1416                 if (bh == NULL)
1417                         return -EIO;
1418                 eh = ext_block_hdr(bh);
1419                 /* subtract from p_depth to get proper eh_depth */
1420                 if (ext4_ext_check_block(inode, eh,
1421                                          path->p_depth - depth, bh)) {
1422                         put_bh(bh);
1423                         return -EIO;
1424                 }
1425                 ix = EXT_FIRST_INDEX(eh);
1426                 block = ext4_idx_pblock(ix);
1427                 put_bh(bh);
1428         }
1429
1430         bh = sb_bread(inode->i_sb, block);
1431         if (bh == NULL)
1432                 return -EIO;
1433         eh = ext_block_hdr(bh);
1434         if (ext4_ext_check_block(inode, eh, path->p_depth - depth, bh)) {
1435                 put_bh(bh);
1436                 return -EIO;
1437         }
1438         ex = EXT_FIRST_EXTENT(eh);
1439 found_extent:
1440         *logical = le32_to_cpu(ex->ee_block);
1441         *phys = ext4_ext_pblock(ex);
1442         *ret_ex = ex;
1443         if (bh)
1444                 put_bh(bh);
1445         return 0;
1446 }
1447
1448 /*
1449  * ext4_ext_next_allocated_block:
1450  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1451  * NOTE: it considers block number from index entry as
1452  * allocated block. Thus, index entries have to be consistent
1453  * with leaves.
1454  */
1455 static ext4_lblk_t
1456 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1457 {
1458         int depth;
1459
1460         BUG_ON(path == NULL);
1461         depth = path->p_depth;
1462
1463         if (depth == 0 && path->p_ext == NULL)
1464                 return EXT_MAX_BLOCKS;
1465
1466         while (depth >= 0) {
1467                 if (depth == path->p_depth) {
1468                         /* leaf */
1469                         if (path[depth].p_ext &&
1470                                 path[depth].p_ext !=
1471                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1472                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1473                 } else {
1474                         /* index */
1475                         if (path[depth].p_idx !=
1476                                         EXT_LAST_INDEX(path[depth].p_hdr))
1477                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1478                 }
1479                 depth--;
1480         }
1481
1482         return EXT_MAX_BLOCKS;
1483 }
1484
1485 /*
1486  * ext4_ext_next_leaf_block:
1487  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1488  */
1489 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1490 {
1491         int depth;
1492
1493         BUG_ON(path == NULL);
1494         depth = path->p_depth;
1495
1496         /* zero-tree has no leaf blocks at all */
1497         if (depth == 0)
1498                 return EXT_MAX_BLOCKS;
1499
1500         /* go to index block */
1501         depth--;
1502
1503         while (depth >= 0) {
1504                 if (path[depth].p_idx !=
1505                                 EXT_LAST_INDEX(path[depth].p_hdr))
1506                         return (ext4_lblk_t)
1507                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1508                 depth--;
1509         }
1510
1511         return EXT_MAX_BLOCKS;
1512 }
1513
1514 /*
1515  * ext4_ext_correct_indexes:
1516  * if leaf gets modified and modified extent is first in the leaf,
1517  * then we have to correct all indexes above.
1518  * TODO: do we need to correct tree in all cases?
1519  */
1520 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1521                                 struct ext4_ext_path *path)
1522 {
1523         struct ext4_extent_header *eh;
1524         int depth = ext_depth(inode);
1525         struct ext4_extent *ex;
1526         __le32 border;
1527         int k, err = 0;
1528
1529         eh = path[depth].p_hdr;
1530         ex = path[depth].p_ext;
1531
1532         if (unlikely(ex == NULL || eh == NULL)) {
1533                 EXT4_ERROR_INODE(inode,
1534                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1535                 return -EIO;
1536         }
1537
1538         if (depth == 0) {
1539                 /* there is no tree at all */
1540                 return 0;
1541         }
1542
1543         if (ex != EXT_FIRST_EXTENT(eh)) {
1544                 /* we correct tree if first leaf got modified only */
1545                 return 0;
1546         }
1547
1548         /*
1549          * TODO: we need correction if border is smaller than current one
1550          */
1551         k = depth - 1;
1552         border = path[depth].p_ext->ee_block;
1553         err = ext4_ext_get_access(handle, inode, path + k);
1554         if (err)
1555                 return err;
1556         path[k].p_idx->ei_block = border;
1557         err = ext4_ext_dirty(handle, inode, path + k);
1558         if (err)
1559                 return err;
1560
1561         while (k--) {
1562                 /* change all left-side indexes */
1563                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1564                         break;
1565                 err = ext4_ext_get_access(handle, inode, path + k);
1566                 if (err)
1567                         break;
1568                 path[k].p_idx->ei_block = border;
1569                 err = ext4_ext_dirty(handle, inode, path + k);
1570                 if (err)
1571                         break;
1572         }
1573
1574         return err;
1575 }
1576
1577 int
1578 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1579                                 struct ext4_extent *ex2)
1580 {
1581         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1582
1583         /*
1584          * Make sure that both extents are initialized. We don't merge
1585          * uninitialized extents so that we can be sure that end_io code has
1586          * the extent that was written properly split out and conversion to
1587          * initialized is trivial.
1588          */
1589         if (ext4_ext_is_uninitialized(ex1) || ext4_ext_is_uninitialized(ex2))
1590                 return 0;
1591
1592         if (ext4_ext_is_uninitialized(ex1))
1593                 max_len = EXT_UNINIT_MAX_LEN;
1594         else
1595                 max_len = EXT_INIT_MAX_LEN;
1596
1597         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1598         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1599
1600         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1601                         le32_to_cpu(ex2->ee_block))
1602                 return 0;
1603
1604         /*
1605          * To allow future support for preallocated extents to be added
1606          * as an RO_COMPAT feature, refuse to merge to extents if
1607          * this can result in the top bit of ee_len being set.
1608          */
1609         if (ext1_ee_len + ext2_ee_len > max_len)
1610                 return 0;
1611 #ifdef AGGRESSIVE_TEST
1612         if (ext1_ee_len >= 4)
1613                 return 0;
1614 #endif
1615
1616         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1617                 return 1;
1618         return 0;
1619 }
1620
1621 /*
1622  * This function tries to merge the "ex" extent to the next extent in the tree.
1623  * It always tries to merge towards right. If you want to merge towards
1624  * left, pass "ex - 1" as argument instead of "ex".
1625  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1626  * 1 if they got merged.
1627  */
1628 static int ext4_ext_try_to_merge_right(struct inode *inode,
1629                                  struct ext4_ext_path *path,
1630                                  struct ext4_extent *ex)
1631 {
1632         struct ext4_extent_header *eh;
1633         unsigned int depth, len;
1634         int merge_done = 0;
1635         int uninitialized = 0;
1636
1637         depth = ext_depth(inode);
1638         BUG_ON(path[depth].p_hdr == NULL);
1639         eh = path[depth].p_hdr;
1640
1641         while (ex < EXT_LAST_EXTENT(eh)) {
1642                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1643                         break;
1644                 /* merge with next extent! */
1645                 if (ext4_ext_is_uninitialized(ex))
1646                         uninitialized = 1;
1647                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1648                                 + ext4_ext_get_actual_len(ex + 1));
1649                 if (uninitialized)
1650                         ext4_ext_mark_uninitialized(ex);
1651
1652                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1653                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1654                                 * sizeof(struct ext4_extent);
1655                         memmove(ex + 1, ex + 2, len);
1656                 }
1657                 le16_add_cpu(&eh->eh_entries, -1);
1658                 merge_done = 1;
1659                 WARN_ON(eh->eh_entries == 0);
1660                 if (!eh->eh_entries)
1661                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1662         }
1663
1664         return merge_done;
1665 }
1666
1667 /*
1668  * This function does a very simple check to see if we can collapse
1669  * an extent tree with a single extent tree leaf block into the inode.
1670  */
1671 static void ext4_ext_try_to_merge_up(handle_t *handle,
1672                                      struct inode *inode,
1673                                      struct ext4_ext_path *path)
1674 {
1675         size_t s;
1676         unsigned max_root = ext4_ext_space_root(inode, 0);
1677         ext4_fsblk_t blk;
1678
1679         if ((path[0].p_depth != 1) ||
1680             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1681             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1682                 return;
1683
1684         /*
1685          * We need to modify the block allocation bitmap and the block
1686          * group descriptor to release the extent tree block.  If we
1687          * can't get the journal credits, give up.
1688          */
1689         if (ext4_journal_extend(handle, 2))
1690                 return;
1691
1692         /*
1693          * Copy the extent data up to the inode
1694          */
1695         blk = ext4_idx_pblock(path[0].p_idx);
1696         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1697                 sizeof(struct ext4_extent_idx);
1698         s += sizeof(struct ext4_extent_header);
1699
1700         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1701         path[0].p_depth = 0;
1702         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1703                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1704         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1705
1706         brelse(path[1].p_bh);
1707         ext4_free_blocks(handle, inode, NULL, blk, 1,
1708                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1709 }
1710
1711 /*
1712  * This function tries to merge the @ex extent to neighbours in the tree.
1713  * return 1 if merge left else 0.
1714  */
1715 static void ext4_ext_try_to_merge(handle_t *handle,
1716                                   struct inode *inode,
1717                                   struct ext4_ext_path *path,
1718                                   struct ext4_extent *ex) {
1719         struct ext4_extent_header *eh;
1720         unsigned int depth;
1721         int merge_done = 0;
1722
1723         depth = ext_depth(inode);
1724         BUG_ON(path[depth].p_hdr == NULL);
1725         eh = path[depth].p_hdr;
1726
1727         if (ex > EXT_FIRST_EXTENT(eh))
1728                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1729
1730         if (!merge_done)
1731                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1732
1733         ext4_ext_try_to_merge_up(handle, inode, path);
1734 }
1735
1736 /*
1737  * check if a portion of the "newext" extent overlaps with an
1738  * existing extent.
1739  *
1740  * If there is an overlap discovered, it updates the length of the newext
1741  * such that there will be no overlap, and then returns 1.
1742  * If there is no overlap found, it returns 0.
1743  */
1744 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1745                                            struct inode *inode,
1746                                            struct ext4_extent *newext,
1747                                            struct ext4_ext_path *path)
1748 {
1749         ext4_lblk_t b1, b2;
1750         unsigned int depth, len1;
1751         unsigned int ret = 0;
1752
1753         b1 = le32_to_cpu(newext->ee_block);
1754         len1 = ext4_ext_get_actual_len(newext);
1755         depth = ext_depth(inode);
1756         if (!path[depth].p_ext)
1757                 goto out;
1758         b2 = le32_to_cpu(path[depth].p_ext->ee_block);
1759         b2 &= ~(sbi->s_cluster_ratio - 1);
1760
1761         /*
1762          * get the next allocated block if the extent in the path
1763          * is before the requested block(s)
1764          */
1765         if (b2 < b1) {
1766                 b2 = ext4_ext_next_allocated_block(path);
1767                 if (b2 == EXT_MAX_BLOCKS)
1768                         goto out;
1769                 b2 &= ~(sbi->s_cluster_ratio - 1);
1770         }
1771
1772         /* check for wrap through zero on extent logical start block*/
1773         if (b1 + len1 < b1) {
1774                 len1 = EXT_MAX_BLOCKS - b1;
1775                 newext->ee_len = cpu_to_le16(len1);
1776                 ret = 1;
1777         }
1778
1779         /* check for overlap */
1780         if (b1 + len1 > b2) {
1781                 newext->ee_len = cpu_to_le16(b2 - b1);
1782                 ret = 1;
1783         }
1784 out:
1785         return ret;
1786 }
1787
1788 /*
1789  * ext4_ext_insert_extent:
1790  * tries to merge requsted extent into the existing extent or
1791  * inserts requested extent as new one into the tree,
1792  * creating new leaf in the no-space case.
1793  */
1794 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1795                                 struct ext4_ext_path *path,
1796                                 struct ext4_extent *newext, int flag)
1797 {
1798         struct ext4_extent_header *eh;
1799         struct ext4_extent *ex, *fex;
1800         struct ext4_extent *nearex; /* nearest extent */
1801         struct ext4_ext_path *npath = NULL;
1802         int depth, len, err;
1803         ext4_lblk_t next;
1804         unsigned uninitialized = 0;
1805         int flags = 0;
1806
1807         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1808                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1809                 return -EIO;
1810         }
1811         depth = ext_depth(inode);
1812         ex = path[depth].p_ext;
1813         eh = path[depth].p_hdr;
1814         if (unlikely(path[depth].p_hdr == NULL)) {
1815                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1816                 return -EIO;
1817         }
1818
1819         /* try to insert block into found extent and return */
1820         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)) {
1821
1822                 /*
1823                  * Try to see whether we should rather test the extent on
1824                  * right from ex, or from the left of ex. This is because
1825                  * ext4_ext_find_extent() can return either extent on the
1826                  * left, or on the right from the searched position. This
1827                  * will make merging more effective.
1828                  */
1829                 if (ex < EXT_LAST_EXTENT(eh) &&
1830                     (le32_to_cpu(ex->ee_block) +
1831                     ext4_ext_get_actual_len(ex) <
1832                     le32_to_cpu(newext->ee_block))) {
1833                         ex += 1;
1834                         goto prepend;
1835                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1836                            (le32_to_cpu(newext->ee_block) +
1837                            ext4_ext_get_actual_len(newext) <
1838                            le32_to_cpu(ex->ee_block)))
1839                         ex -= 1;
1840
1841                 /* Try to append newex to the ex */
1842                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1843                         ext_debug("append [%d]%d block to %u:[%d]%d"
1844                                   "(from %llu)\n",
1845                                   ext4_ext_is_uninitialized(newext),
1846                                   ext4_ext_get_actual_len(newext),
1847                                   le32_to_cpu(ex->ee_block),
1848                                   ext4_ext_is_uninitialized(ex),
1849                                   ext4_ext_get_actual_len(ex),
1850                                   ext4_ext_pblock(ex));
1851                         err = ext4_ext_get_access(handle, inode,
1852                                                   path + depth);
1853                         if (err)
1854                                 return err;
1855
1856                         /*
1857                          * ext4_can_extents_be_merged should have checked
1858                          * that either both extents are uninitialized, or
1859                          * both aren't. Thus we need to check only one of
1860                          * them here.
1861                          */
1862                         if (ext4_ext_is_uninitialized(ex))
1863                                 uninitialized = 1;
1864                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1865                                         + ext4_ext_get_actual_len(newext));
1866                         if (uninitialized)
1867                                 ext4_ext_mark_uninitialized(ex);
1868                         eh = path[depth].p_hdr;
1869                         nearex = ex;
1870                         goto merge;
1871                 }
1872
1873 prepend:
1874                 /* Try to prepend newex to the ex */
1875                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1876                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
1877                                   "(from %llu)\n",
1878                                   le32_to_cpu(newext->ee_block),
1879                                   ext4_ext_is_uninitialized(newext),
1880                                   ext4_ext_get_actual_len(newext),
1881                                   le32_to_cpu(ex->ee_block),
1882                                   ext4_ext_is_uninitialized(ex),
1883                                   ext4_ext_get_actual_len(ex),
1884                                   ext4_ext_pblock(ex));
1885                         err = ext4_ext_get_access(handle, inode,
1886                                                   path + depth);
1887                         if (err)
1888                                 return err;
1889
1890                         /*
1891                          * ext4_can_extents_be_merged should have checked
1892                          * that either both extents are uninitialized, or
1893                          * both aren't. Thus we need to check only one of
1894                          * them here.
1895                          */
1896                         if (ext4_ext_is_uninitialized(ex))
1897                                 uninitialized = 1;
1898                         ex->ee_block = newext->ee_block;
1899                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
1900                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1901                                         + ext4_ext_get_actual_len(newext));
1902                         if (uninitialized)
1903                                 ext4_ext_mark_uninitialized(ex);
1904                         eh = path[depth].p_hdr;
1905                         nearex = ex;
1906                         goto merge;
1907                 }
1908         }
1909
1910         depth = ext_depth(inode);
1911         eh = path[depth].p_hdr;
1912         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1913                 goto has_space;
1914
1915         /* probably next leaf has space for us? */
1916         fex = EXT_LAST_EXTENT(eh);
1917         next = EXT_MAX_BLOCKS;
1918         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1919                 next = ext4_ext_next_leaf_block(path);
1920         if (next != EXT_MAX_BLOCKS) {
1921                 ext_debug("next leaf block - %u\n", next);
1922                 BUG_ON(npath != NULL);
1923                 npath = ext4_ext_find_extent(inode, next, NULL);
1924                 if (IS_ERR(npath))
1925                         return PTR_ERR(npath);
1926                 BUG_ON(npath->p_depth != path->p_depth);
1927                 eh = npath[depth].p_hdr;
1928                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1929                         ext_debug("next leaf isn't full(%d)\n",
1930                                   le16_to_cpu(eh->eh_entries));
1931                         path = npath;
1932                         goto has_space;
1933                 }
1934                 ext_debug("next leaf has no free space(%d,%d)\n",
1935                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1936         }
1937
1938         /*
1939          * There is no free space in the found leaf.
1940          * We're gonna add a new leaf in the tree.
1941          */
1942         if (flag & EXT4_GET_BLOCKS_METADATA_NOFAIL)
1943                 flags = EXT4_MB_USE_RESERVED;
1944         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1945         if (err)
1946                 goto cleanup;
1947         depth = ext_depth(inode);
1948         eh = path[depth].p_hdr;
1949
1950 has_space:
1951         nearex = path[depth].p_ext;
1952
1953         err = ext4_ext_get_access(handle, inode, path + depth);
1954         if (err)
1955                 goto cleanup;
1956
1957         if (!nearex) {
1958                 /* there is no extent in this leaf, create first one */
1959                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1960                                 le32_to_cpu(newext->ee_block),
1961                                 ext4_ext_pblock(newext),
1962                                 ext4_ext_is_uninitialized(newext),
1963                                 ext4_ext_get_actual_len(newext));
1964                 nearex = EXT_FIRST_EXTENT(eh);
1965         } else {
1966                 if (le32_to_cpu(newext->ee_block)
1967                            > le32_to_cpu(nearex->ee_block)) {
1968                         /* Insert after */
1969                         ext_debug("insert %u:%llu:[%d]%d before: "
1970                                         "nearest %p\n",
1971                                         le32_to_cpu(newext->ee_block),
1972                                         ext4_ext_pblock(newext),
1973                                         ext4_ext_is_uninitialized(newext),
1974                                         ext4_ext_get_actual_len(newext),
1975                                         nearex);
1976                         nearex++;
1977                 } else {
1978                         /* Insert before */
1979                         BUG_ON(newext->ee_block == nearex->ee_block);
1980                         ext_debug("insert %u:%llu:[%d]%d after: "
1981                                         "nearest %p\n",
1982                                         le32_to_cpu(newext->ee_block),
1983                                         ext4_ext_pblock(newext),
1984                                         ext4_ext_is_uninitialized(newext),
1985                                         ext4_ext_get_actual_len(newext),
1986                                         nearex);
1987                 }
1988                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
1989                 if (len > 0) {
1990                         ext_debug("insert %u:%llu:[%d]%d: "
1991                                         "move %d extents from 0x%p to 0x%p\n",
1992                                         le32_to_cpu(newext->ee_block),
1993                                         ext4_ext_pblock(newext),
1994                                         ext4_ext_is_uninitialized(newext),
1995                                         ext4_ext_get_actual_len(newext),
1996                                         len, nearex, nearex + 1);
1997                         memmove(nearex + 1, nearex,
1998                                 len * sizeof(struct ext4_extent));
1999                 }
2000         }
2001
2002         le16_add_cpu(&eh->eh_entries, 1);
2003         path[depth].p_ext = nearex;
2004         nearex->ee_block = newext->ee_block;
2005         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2006         nearex->ee_len = newext->ee_len;
2007
2008 merge:
2009         /* try to merge extents */
2010         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
2011                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2012
2013
2014         /* time to correct all indexes above */
2015         err = ext4_ext_correct_indexes(handle, inode, path);
2016         if (err)
2017                 goto cleanup;
2018
2019         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2020
2021 cleanup:
2022         if (npath) {
2023                 ext4_ext_drop_refs(npath);
2024                 kfree(npath);
2025         }
2026         return err;
2027 }
2028
2029 static int ext4_fill_fiemap_extents(struct inode *inode,
2030                                     ext4_lblk_t block, ext4_lblk_t num,
2031                                     struct fiemap_extent_info *fieinfo)
2032 {
2033         struct ext4_ext_path *path = NULL;
2034         struct ext4_extent *ex;
2035         struct extent_status es;
2036         ext4_lblk_t next, next_del, start = 0, end = 0;
2037         ext4_lblk_t last = block + num;
2038         int exists, depth = 0, err = 0;
2039         unsigned int flags = 0;
2040         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2041
2042         while (block < last && block != EXT_MAX_BLOCKS) {
2043                 num = last - block;
2044                 /* find extent for this block */
2045                 down_read(&EXT4_I(inode)->i_data_sem);
2046
2047                 if (path && ext_depth(inode) != depth) {
2048                         /* depth was changed. we have to realloc path */
2049                         kfree(path);
2050                         path = NULL;
2051                 }
2052
2053                 path = ext4_ext_find_extent(inode, block, path);
2054                 if (IS_ERR(path)) {
2055                         up_read(&EXT4_I(inode)->i_data_sem);
2056                         err = PTR_ERR(path);
2057                         path = NULL;
2058                         break;
2059                 }
2060
2061                 depth = ext_depth(inode);
2062                 if (unlikely(path[depth].p_hdr == NULL)) {
2063                         up_read(&EXT4_I(inode)->i_data_sem);
2064                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2065                         err = -EIO;
2066                         break;
2067                 }
2068                 ex = path[depth].p_ext;
2069                 next = ext4_ext_next_allocated_block(path);
2070                 ext4_ext_drop_refs(path);
2071
2072                 flags = 0;
2073                 exists = 0;
2074                 if (!ex) {
2075                         /* there is no extent yet, so try to allocate
2076                          * all requested space */
2077                         start = block;
2078                         end = block + num;
2079                 } else if (le32_to_cpu(ex->ee_block) > block) {
2080                         /* need to allocate space before found extent */
2081                         start = block;
2082                         end = le32_to_cpu(ex->ee_block);
2083                         if (block + num < end)
2084                                 end = block + num;
2085                 } else if (block >= le32_to_cpu(ex->ee_block)
2086                                         + ext4_ext_get_actual_len(ex)) {
2087                         /* need to allocate space after found extent */
2088                         start = block;
2089                         end = block + num;
2090                         if (end >= next)
2091                                 end = next;
2092                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2093                         /*
2094                          * some part of requested space is covered
2095                          * by found extent
2096                          */
2097                         start = block;
2098                         end = le32_to_cpu(ex->ee_block)
2099                                 + ext4_ext_get_actual_len(ex);
2100                         if (block + num < end)
2101                                 end = block + num;
2102                         exists = 1;
2103                 } else {
2104                         BUG();
2105                 }
2106                 BUG_ON(end <= start);
2107
2108                 if (!exists) {
2109                         es.es_lblk = start;
2110                         es.es_len = end - start;
2111                         es.es_pblk = 0;
2112                 } else {
2113                         es.es_lblk = le32_to_cpu(ex->ee_block);
2114                         es.es_len = ext4_ext_get_actual_len(ex);
2115                         es.es_pblk = ext4_ext_pblock(ex);
2116                         if (ext4_ext_is_uninitialized(ex))
2117                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2118                 }
2119
2120                 /*
2121                  * Find delayed extent and update es accordingly. We call
2122                  * it even in !exists case to find out whether es is the
2123                  * last existing extent or not.
2124                  */
2125                 next_del = ext4_find_delayed_extent(inode, &es);
2126                 if (!exists && next_del) {
2127                         exists = 1;
2128                         flags |= FIEMAP_EXTENT_DELALLOC;
2129                 }
2130                 up_read(&EXT4_I(inode)->i_data_sem);
2131
2132                 if (unlikely(es.es_len == 0)) {
2133                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2134                         err = -EIO;
2135                         break;
2136                 }
2137
2138                 /*
2139                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2140                  * we need to check next == EXT_MAX_BLOCKS because it is
2141                  * possible that an extent is with unwritten and delayed
2142                  * status due to when an extent is delayed allocated and
2143                  * is allocated by fallocate status tree will track both of
2144                  * them in a extent.
2145                  *
2146                  * So we could return a unwritten and delayed extent, and
2147                  * its block is equal to 'next'.
2148                  */
2149                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2150                         flags |= FIEMAP_EXTENT_LAST;
2151                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2152                                      next != EXT_MAX_BLOCKS)) {
2153                                 EXT4_ERROR_INODE(inode,
2154                                                  "next extent == %u, next "
2155                                                  "delalloc extent = %u",
2156                                                  next, next_del);
2157                                 err = -EIO;
2158                                 break;
2159                         }
2160                 }
2161
2162                 if (exists) {
2163                         err = fiemap_fill_next_extent(fieinfo,
2164                                 (__u64)es.es_lblk << blksize_bits,
2165                                 (__u64)es.es_pblk << blksize_bits,
2166                                 (__u64)es.es_len << blksize_bits,
2167                                 flags);
2168                         if (err < 0)
2169                                 break;
2170                         if (err == 1) {
2171                                 err = 0;
2172                                 break;
2173                         }
2174                 }
2175
2176                 block = es.es_lblk + es.es_len;
2177         }
2178
2179         if (path) {
2180                 ext4_ext_drop_refs(path);
2181                 kfree(path);
2182         }
2183
2184         return err;
2185 }
2186
2187 /*
2188  * ext4_ext_put_gap_in_cache:
2189  * calculate boundaries of the gap that the requested block fits into
2190  * and cache this gap
2191  */
2192 static void
2193 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2194                                 ext4_lblk_t block)
2195 {
2196         int depth = ext_depth(inode);
2197         unsigned long len;
2198         ext4_lblk_t lblock;
2199         struct ext4_extent *ex;
2200
2201         ex = path[depth].p_ext;
2202         if (ex == NULL) {
2203                 /*
2204                  * there is no extent yet, so gap is [0;-] and we
2205                  * don't cache it
2206                  */
2207                 ext_debug("cache gap(whole file):");
2208         } else if (block < le32_to_cpu(ex->ee_block)) {
2209                 lblock = block;
2210                 len = le32_to_cpu(ex->ee_block) - block;
2211                 ext_debug("cache gap(before): %u [%u:%u]",
2212                                 block,
2213                                 le32_to_cpu(ex->ee_block),
2214                                  ext4_ext_get_actual_len(ex));
2215                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2216                         ext4_es_insert_extent(inode, lblock, len, ~0,
2217                                               EXTENT_STATUS_HOLE);
2218         } else if (block >= le32_to_cpu(ex->ee_block)
2219                         + ext4_ext_get_actual_len(ex)) {
2220                 ext4_lblk_t next;
2221                 lblock = le32_to_cpu(ex->ee_block)
2222                         + ext4_ext_get_actual_len(ex);
2223
2224                 next = ext4_ext_next_allocated_block(path);
2225                 ext_debug("cache gap(after): [%u:%u] %u",
2226                                 le32_to_cpu(ex->ee_block),
2227                                 ext4_ext_get_actual_len(ex),
2228                                 block);
2229                 BUG_ON(next == lblock);
2230                 len = next - lblock;
2231                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2232                         ext4_es_insert_extent(inode, lblock, len, ~0,
2233                                               EXTENT_STATUS_HOLE);
2234         } else {
2235                 lblock = len = 0;
2236                 BUG();
2237         }
2238
2239         ext_debug(" -> %u:%lu\n", lblock, len);
2240 }
2241
2242 /*
2243  * ext4_ext_rm_idx:
2244  * removes index from the index block.
2245  */
2246 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2247                         struct ext4_ext_path *path, int depth)
2248 {
2249         int err;
2250         ext4_fsblk_t leaf;
2251
2252         /* free index block */
2253         depth--;
2254         path = path + depth;
2255         leaf = ext4_idx_pblock(path->p_idx);
2256         if (unlikely(path->p_hdr->eh_entries == 0)) {
2257                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2258                 return -EIO;
2259         }
2260         err = ext4_ext_get_access(handle, inode, path);
2261         if (err)
2262                 return err;
2263
2264         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2265                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2266                 len *= sizeof(struct ext4_extent_idx);
2267                 memmove(path->p_idx, path->p_idx + 1, len);
2268         }
2269
2270         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2271         err = ext4_ext_dirty(handle, inode, path);
2272         if (err)
2273                 return err;
2274         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2275         trace_ext4_ext_rm_idx(inode, leaf);
2276
2277         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2278                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2279
2280         while (--depth >= 0) {
2281                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2282                         break;
2283                 path--;
2284                 err = ext4_ext_get_access(handle, inode, path);
2285                 if (err)
2286                         break;
2287                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2288                 err = ext4_ext_dirty(handle, inode, path);
2289                 if (err)
2290                         break;
2291         }
2292         return err;
2293 }
2294
2295 /*
2296  * ext4_ext_calc_credits_for_single_extent:
2297  * This routine returns max. credits that needed to insert an extent
2298  * to the extent tree.
2299  * When pass the actual path, the caller should calculate credits
2300  * under i_data_sem.
2301  */
2302 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2303                                                 struct ext4_ext_path *path)
2304 {
2305         if (path) {
2306                 int depth = ext_depth(inode);
2307                 int ret = 0;
2308
2309                 /* probably there is space in leaf? */
2310                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2311                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2312
2313                         /*
2314                          *  There are some space in the leaf tree, no
2315                          *  need to account for leaf block credit
2316                          *
2317                          *  bitmaps and block group descriptor blocks
2318                          *  and other metadata blocks still need to be
2319                          *  accounted.
2320                          */
2321                         /* 1 bitmap, 1 block group descriptor */
2322                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2323                         return ret;
2324                 }
2325         }
2326
2327         return ext4_chunk_trans_blocks(inode, nrblocks);
2328 }
2329
2330 /*
2331  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2332  *
2333  * if nrblocks are fit in a single extent (chunk flag is 1), then
2334  * in the worse case, each tree level index/leaf need to be changed
2335  * if the tree split due to insert a new extent, then the old tree
2336  * index/leaf need to be updated too
2337  *
2338  * If the nrblocks are discontiguous, they could cause
2339  * the whole tree split more than once, but this is really rare.
2340  */
2341 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2342 {
2343         int index;
2344         int depth;
2345
2346         /* If we are converting the inline data, only one is needed here. */
2347         if (ext4_has_inline_data(inode))
2348                 return 1;
2349
2350         depth = ext_depth(inode);
2351
2352         if (chunk)
2353                 index = depth * 2;
2354         else
2355                 index = depth * 3;
2356
2357         return index;
2358 }
2359
2360 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2361                               struct ext4_extent *ex,
2362                               ext4_fsblk_t *partial_cluster,
2363                               ext4_lblk_t from, ext4_lblk_t to)
2364 {
2365         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2366         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2367         ext4_fsblk_t pblk;
2368         int flags = 0;
2369
2370         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2371                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2372         else if (ext4_should_journal_data(inode))
2373                 flags |= EXT4_FREE_BLOCKS_FORGET;
2374
2375         /*
2376          * For bigalloc file systems, we never free a partial cluster
2377          * at the beginning of the extent.  Instead, we make a note
2378          * that we tried freeing the cluster, and check to see if we
2379          * need to free it on a subsequent call to ext4_remove_blocks,
2380          * or at the end of the ext4_truncate() operation.
2381          */
2382         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2383
2384         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2385         /*
2386          * If we have a partial cluster, and it's different from the
2387          * cluster of the last block, we need to explicitly free the
2388          * partial cluster here.
2389          */
2390         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2391         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2392                 ext4_free_blocks(handle, inode, NULL,
2393                                  EXT4_C2B(sbi, *partial_cluster),
2394                                  sbi->s_cluster_ratio, flags);
2395                 *partial_cluster = 0;
2396         }
2397
2398 #ifdef EXTENTS_STATS
2399         {
2400                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2401                 spin_lock(&sbi->s_ext_stats_lock);
2402                 sbi->s_ext_blocks += ee_len;
2403                 sbi->s_ext_extents++;
2404                 if (ee_len < sbi->s_ext_min)
2405                         sbi->s_ext_min = ee_len;
2406                 if (ee_len > sbi->s_ext_max)
2407                         sbi->s_ext_max = ee_len;
2408                 if (ext_depth(inode) > sbi->s_depth_max)
2409                         sbi->s_depth_max = ext_depth(inode);
2410                 spin_unlock(&sbi->s_ext_stats_lock);
2411         }
2412 #endif
2413         if (from >= le32_to_cpu(ex->ee_block)
2414             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2415                 /* tail removal */
2416                 ext4_lblk_t num;
2417
2418                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2419                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2420                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2421                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2422                 /*
2423                  * If the block range to be freed didn't start at the
2424                  * beginning of a cluster, and we removed the entire
2425                  * extent, save the partial cluster here, since we
2426                  * might need to delete if we determine that the
2427                  * truncate operation has removed all of the blocks in
2428                  * the cluster.
2429                  */
2430                 if (pblk & (sbi->s_cluster_ratio - 1) &&
2431                     (ee_len == num))
2432                         *partial_cluster = EXT4_B2C(sbi, pblk);
2433                 else
2434                         *partial_cluster = 0;
2435         } else if (from == le32_to_cpu(ex->ee_block)
2436                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2437                 /* head removal */
2438                 ext4_lblk_t num;
2439                 ext4_fsblk_t start;
2440
2441                 num = to - from;
2442                 start = ext4_ext_pblock(ex);
2443
2444                 ext_debug("free first %u blocks starting %llu\n", num, start);
2445                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2446
2447         } else {
2448                 printk(KERN_INFO "strange request: removal(2) "
2449                                 "%u-%u from %u:%u\n",
2450                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2451         }
2452         return 0;
2453 }
2454
2455
2456 /*
2457  * ext4_ext_rm_leaf() Removes the extents associated with the
2458  * blocks appearing between "start" and "end", and splits the extents
2459  * if "start" and "end" appear in the same extent
2460  *
2461  * @handle: The journal handle
2462  * @inode:  The files inode
2463  * @path:   The path to the leaf
2464  * @start:  The first block to remove
2465  * @end:   The last block to remove
2466  */
2467 static int
2468 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2469                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2470                  ext4_lblk_t start, ext4_lblk_t end)
2471 {
2472         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2473         int err = 0, correct_index = 0;
2474         int depth = ext_depth(inode), credits;
2475         struct ext4_extent_header *eh;
2476         ext4_lblk_t a, b;
2477         unsigned num;
2478         ext4_lblk_t ex_ee_block;
2479         unsigned short ex_ee_len;
2480         unsigned uninitialized = 0;
2481         struct ext4_extent *ex;
2482
2483         /* the header must be checked already in ext4_ext_remove_space() */
2484         ext_debug("truncate since %u in leaf to %u\n", start, end);
2485         if (!path[depth].p_hdr)
2486                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2487         eh = path[depth].p_hdr;
2488         if (unlikely(path[depth].p_hdr == NULL)) {
2489                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2490                 return -EIO;
2491         }
2492         /* find where to start removing */
2493         ex = EXT_LAST_EXTENT(eh);
2494
2495         ex_ee_block = le32_to_cpu(ex->ee_block);
2496         ex_ee_len = ext4_ext_get_actual_len(ex);
2497
2498         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2499
2500         while (ex >= EXT_FIRST_EXTENT(eh) &&
2501                         ex_ee_block + ex_ee_len > start) {
2502
2503                 if (ext4_ext_is_uninitialized(ex))
2504                         uninitialized = 1;
2505                 else
2506                         uninitialized = 0;
2507
2508                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2509                          uninitialized, ex_ee_len);
2510                 path[depth].p_ext = ex;
2511
2512                 a = ex_ee_block > start ? ex_ee_block : start;
2513                 b = ex_ee_block+ex_ee_len - 1 < end ?
2514                         ex_ee_block+ex_ee_len - 1 : end;
2515
2516                 ext_debug("  border %u:%u\n", a, b);
2517
2518                 /* If this extent is beyond the end of the hole, skip it */
2519                 if (end < ex_ee_block) {
2520                         ex--;
2521                         ex_ee_block = le32_to_cpu(ex->ee_block);
2522                         ex_ee_len = ext4_ext_get_actual_len(ex);
2523                         continue;
2524                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2525                         EXT4_ERROR_INODE(inode,
2526                                          "can not handle truncate %u:%u "
2527                                          "on extent %u:%u",
2528                                          start, end, ex_ee_block,
2529                                          ex_ee_block + ex_ee_len - 1);
2530                         err = -EIO;
2531                         goto out;
2532                 } else if (a != ex_ee_block) {
2533                         /* remove tail of the extent */
2534                         num = a - ex_ee_block;
2535                 } else {
2536                         /* remove whole extent: excellent! */
2537                         num = 0;
2538                 }
2539                 /*
2540                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2541                  * descriptor) for each block group; assume two block
2542                  * groups plus ex_ee_len/blocks_per_block_group for
2543                  * the worst case
2544                  */
2545                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2546                 if (ex == EXT_FIRST_EXTENT(eh)) {
2547                         correct_index = 1;
2548                         credits += (ext_depth(inode)) + 1;
2549                 }
2550                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2551
2552                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2553                 if (err)
2554                         goto out;
2555
2556                 err = ext4_ext_get_access(handle, inode, path + depth);
2557                 if (err)
2558                         goto out;
2559
2560                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2561                                          a, b);
2562                 if (err)
2563                         goto out;
2564
2565                 if (num == 0)
2566                         /* this extent is removed; mark slot entirely unused */
2567                         ext4_ext_store_pblock(ex, 0);
2568
2569                 ex->ee_len = cpu_to_le16(num);
2570                 /*
2571                  * Do not mark uninitialized if all the blocks in the
2572                  * extent have been removed.
2573                  */
2574                 if (uninitialized && num)
2575                         ext4_ext_mark_uninitialized(ex);
2576                 /*
2577                  * If the extent was completely released,
2578                  * we need to remove it from the leaf
2579                  */
2580                 if (num == 0) {
2581                         if (end != EXT_MAX_BLOCKS - 1) {
2582                                 /*
2583                                  * For hole punching, we need to scoot all the
2584                                  * extents up when an extent is removed so that
2585                                  * we dont have blank extents in the middle
2586                                  */
2587                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2588                                         sizeof(struct ext4_extent));
2589
2590                                 /* Now get rid of the one at the end */
2591                                 memset(EXT_LAST_EXTENT(eh), 0,
2592                                         sizeof(struct ext4_extent));
2593                         }
2594                         le16_add_cpu(&eh->eh_entries, -1);
2595                 } else
2596                         *partial_cluster = 0;
2597
2598                 err = ext4_ext_dirty(handle, inode, path + depth);
2599                 if (err)
2600                         goto out;
2601
2602                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2603                                 ext4_ext_pblock(ex));
2604                 ex--;
2605                 ex_ee_block = le32_to_cpu(ex->ee_block);
2606                 ex_ee_len = ext4_ext_get_actual_len(ex);
2607         }
2608
2609         if (correct_index && eh->eh_entries)
2610                 err = ext4_ext_correct_indexes(handle, inode, path);
2611
2612         /*
2613          * If there is still a entry in the leaf node, check to see if
2614          * it references the partial cluster.  This is the only place
2615          * where it could; if it doesn't, we can free the cluster.
2616          */
2617         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2618             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2619              *partial_cluster)) {
2620                 int flags = EXT4_FREE_BLOCKS_FORGET;
2621
2622                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2623                         flags |= EXT4_FREE_BLOCKS_METADATA;
2624
2625                 ext4_free_blocks(handle, inode, NULL,
2626                                  EXT4_C2B(sbi, *partial_cluster),
2627                                  sbi->s_cluster_ratio, flags);
2628                 *partial_cluster = 0;
2629         }
2630
2631         /* if this leaf is free, then we should
2632          * remove it from index block above */
2633         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2634                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2635
2636 out:
2637         return err;
2638 }
2639
2640 /*
2641  * ext4_ext_more_to_rm:
2642  * returns 1 if current index has to be freed (even partial)
2643  */
2644 static int
2645 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2646 {
2647         BUG_ON(path->p_idx == NULL);
2648
2649         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2650                 return 0;
2651
2652         /*
2653          * if truncate on deeper level happened, it wasn't partial,
2654          * so we have to consider current index for truncation
2655          */
2656         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2657                 return 0;
2658         return 1;
2659 }
2660
2661 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2662                           ext4_lblk_t end)
2663 {
2664         struct super_block *sb = inode->i_sb;
2665         int depth = ext_depth(inode);
2666         struct ext4_ext_path *path = NULL;
2667         ext4_fsblk_t partial_cluster = 0;
2668         handle_t *handle;
2669         int i = 0, err = 0;
2670
2671         ext_debug("truncate since %u to %u\n", start, end);
2672
2673         /* probably first extent we're gonna free will be last in block */
2674         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2675         if (IS_ERR(handle))
2676                 return PTR_ERR(handle);
2677
2678 again:
2679         trace_ext4_ext_remove_space(inode, start, depth);
2680
2681         /*
2682          * Check if we are removing extents inside the extent tree. If that
2683          * is the case, we are going to punch a hole inside the extent tree
2684          * so we have to check whether we need to split the extent covering
2685          * the last block to remove so we can easily remove the part of it
2686          * in ext4_ext_rm_leaf().
2687          */
2688         if (end < EXT_MAX_BLOCKS - 1) {
2689                 struct ext4_extent *ex;
2690                 ext4_lblk_t ee_block;
2691
2692                 /* find extent for this block */
2693                 path = ext4_ext_find_extent(inode, end, NULL);
2694                 if (IS_ERR(path)) {
2695                         ext4_journal_stop(handle);
2696                         return PTR_ERR(path);
2697                 }
2698                 depth = ext_depth(inode);
2699                 /* Leaf not may not exist only if inode has no blocks at all */
2700                 ex = path[depth].p_ext;
2701                 if (!ex) {
2702                         if (depth) {
2703                                 EXT4_ERROR_INODE(inode,
2704                                                  "path[%d].p_hdr == NULL",
2705                                                  depth);
2706                                 err = -EIO;
2707                         }
2708                         goto out;
2709                 }
2710
2711                 ee_block = le32_to_cpu(ex->ee_block);
2712
2713                 /*
2714                  * See if the last block is inside the extent, if so split
2715                  * the extent at 'end' block so we can easily remove the
2716                  * tail of the first part of the split extent in
2717                  * ext4_ext_rm_leaf().
2718                  */
2719                 if (end >= ee_block &&
2720                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2721                         int split_flag = 0;
2722
2723                         if (ext4_ext_is_uninitialized(ex))
2724                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2725                                              EXT4_EXT_MARK_UNINIT2;
2726
2727                         /*
2728                          * Split the extent in two so that 'end' is the last
2729                          * block in the first new extent. Also we should not
2730                          * fail removing space due to ENOSPC so try to use
2731                          * reserved block if that happens.
2732                          */
2733                         err = ext4_split_extent_at(handle, inode, path,
2734                                         end + 1, split_flag,
2735                                         EXT4_GET_BLOCKS_PRE_IO |
2736                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
2737
2738                         if (err < 0)
2739                                 goto out;
2740                 }
2741         }
2742         /*
2743          * We start scanning from right side, freeing all the blocks
2744          * after i_size and walking into the tree depth-wise.
2745          */
2746         depth = ext_depth(inode);
2747         if (path) {
2748                 int k = i = depth;
2749                 while (--k > 0)
2750                         path[k].p_block =
2751                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2752         } else {
2753                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2754                                GFP_NOFS);
2755                 if (path == NULL) {
2756                         ext4_journal_stop(handle);
2757                         return -ENOMEM;
2758                 }
2759                 path[0].p_depth = depth;
2760                 path[0].p_hdr = ext_inode_hdr(inode);
2761                 i = 0;
2762
2763                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2764                         err = -EIO;
2765                         goto out;
2766                 }
2767         }
2768         err = 0;
2769
2770         while (i >= 0 && err == 0) {
2771                 if (i == depth) {
2772                         /* this is leaf block */
2773                         err = ext4_ext_rm_leaf(handle, inode, path,
2774                                                &partial_cluster, start,
2775                                                end);
2776                         /* root level has p_bh == NULL, brelse() eats this */
2777                         brelse(path[i].p_bh);
2778                         path[i].p_bh = NULL;
2779                         i--;
2780                         continue;
2781                 }
2782
2783                 /* this is index block */
2784                 if (!path[i].p_hdr) {
2785                         ext_debug("initialize header\n");
2786                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2787                 }
2788
2789                 if (!path[i].p_idx) {
2790                         /* this level hasn't been touched yet */
2791                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2792                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2793                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2794                                   path[i].p_hdr,
2795                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2796                 } else {
2797                         /* we were already here, see at next index */
2798                         path[i].p_idx--;
2799                 }
2800
2801                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2802                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2803                                 path[i].p_idx);
2804                 if (ext4_ext_more_to_rm(path + i)) {
2805                         struct buffer_head *bh;
2806                         /* go to the next level */
2807                         ext_debug("move to level %d (block %llu)\n",
2808                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2809                         memset(path + i + 1, 0, sizeof(*path));
2810                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2811                         if (!bh) {
2812                                 /* should we reset i_size? */
2813                                 err = -EIO;
2814                                 break;
2815                         }
2816                         if (WARN_ON(i + 1 > depth)) {
2817                                 err = -EIO;
2818                                 break;
2819                         }
2820                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2821                                                         depth - i - 1, bh)) {
2822                                 err = -EIO;
2823                                 break;
2824                         }
2825                         path[i + 1].p_bh = bh;
2826
2827                         /* save actual number of indexes since this
2828                          * number is changed at the next iteration */
2829                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2830                         i++;
2831                 } else {
2832                         /* we finished processing this index, go up */
2833                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2834                                 /* index is empty, remove it;
2835                                  * handle must be already prepared by the
2836                                  * truncatei_leaf() */
2837                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2838                         }
2839                         /* root level has p_bh == NULL, brelse() eats this */
2840                         brelse(path[i].p_bh);
2841                         path[i].p_bh = NULL;
2842                         i--;
2843                         ext_debug("return to level %d\n", i);
2844                 }
2845         }
2846
2847         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2848                         path->p_hdr->eh_entries);
2849
2850         /* If we still have something in the partial cluster and we have removed
2851          * even the first extent, then we should free the blocks in the partial
2852          * cluster as well. */
2853         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2854                 int flags = EXT4_FREE_BLOCKS_FORGET;
2855
2856                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2857                         flags |= EXT4_FREE_BLOCKS_METADATA;
2858
2859                 ext4_free_blocks(handle, inode, NULL,
2860                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2861                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2862                 partial_cluster = 0;
2863         }
2864
2865         /* TODO: flexible tree reduction should be here */
2866         if (path->p_hdr->eh_entries == 0) {
2867                 /*
2868                  * truncate to zero freed all the tree,
2869                  * so we need to correct eh_depth
2870                  */
2871                 err = ext4_ext_get_access(handle, inode, path);
2872                 if (err == 0) {
2873                         ext_inode_hdr(inode)->eh_depth = 0;
2874                         ext_inode_hdr(inode)->eh_max =
2875                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2876                         err = ext4_ext_dirty(handle, inode, path);
2877                 }
2878         }
2879 out:
2880         ext4_ext_drop_refs(path);
2881         kfree(path);
2882         if (err == -EAGAIN) {
2883                 path = NULL;
2884                 goto again;
2885         }
2886         ext4_journal_stop(handle);
2887
2888         return err;
2889 }
2890
2891 /*
2892  * called at mount time
2893  */
2894 void ext4_ext_init(struct super_block *sb)
2895 {
2896         /*
2897          * possible initialization would be here
2898          */
2899
2900         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2901 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2902                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2903 #ifdef AGGRESSIVE_TEST
2904                        ", aggressive tests"
2905 #endif
2906 #ifdef CHECK_BINSEARCH
2907                        ", check binsearch"
2908 #endif
2909 #ifdef EXTENTS_STATS
2910                        ", stats"
2911 #endif
2912                        "\n");
2913 #endif
2914 #ifdef EXTENTS_STATS
2915                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2916                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2917                 EXT4_SB(sb)->s_ext_max = 0;
2918 #endif
2919         }
2920 }
2921
2922 /*
2923  * called at umount time
2924  */
2925 void ext4_ext_release(struct super_block *sb)
2926 {
2927         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2928                 return;
2929
2930 #ifdef EXTENTS_STATS
2931         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2932                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2933                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2934                         sbi->s_ext_blocks, sbi->s_ext_extents,
2935                         sbi->s_ext_blocks / sbi->s_ext_extents);
2936                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2937                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2938         }
2939 #endif
2940 }
2941
2942 /* FIXME!! we need to try to merge to left or right after zero-out  */
2943 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2944 {
2945         ext4_fsblk_t ee_pblock;
2946         unsigned int ee_len;
2947         int ret;
2948
2949         ee_len    = ext4_ext_get_actual_len(ex);
2950         ee_pblock = ext4_ext_pblock(ex);
2951
2952         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2953         if (ret > 0)
2954                 ret = 0;
2955
2956         return ret;
2957 }
2958
2959 /*
2960  * ext4_split_extent_at() splits an extent at given block.
2961  *
2962  * @handle: the journal handle
2963  * @inode: the file inode
2964  * @path: the path to the extent
2965  * @split: the logical block where the extent is splitted.
2966  * @split_flags: indicates if the extent could be zeroout if split fails, and
2967  *               the states(init or uninit) of new extents.
2968  * @flags: flags used to insert new extent to extent tree.
2969  *
2970  *
2971  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
2972  * of which are deterimined by split_flag.
2973  *
2974  * There are two cases:
2975  *  a> the extent are splitted into two extent.
2976  *  b> split is not needed, and just mark the extent.
2977  *
2978  * return 0 on success.
2979  */
2980 static int ext4_split_extent_at(handle_t *handle,
2981                              struct inode *inode,
2982                              struct ext4_ext_path *path,
2983                              ext4_lblk_t split,
2984                              int split_flag,
2985                              int flags)
2986 {
2987         ext4_fsblk_t newblock;
2988         ext4_lblk_t ee_block;
2989         struct ext4_extent *ex, newex, orig_ex, zero_ex;
2990         struct ext4_extent *ex2 = NULL;
2991         unsigned int ee_len, depth;
2992         int err = 0;
2993
2994         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
2995                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
2996
2997         ext_debug("ext4_split_extents_at: inode %lu, logical"
2998                 "block %llu\n", inode->i_ino, (unsigned long long)split);
2999
3000         ext4_ext_show_leaf(inode, path);
3001
3002         depth = ext_depth(inode);
3003         ex = path[depth].p_ext;
3004         ee_block = le32_to_cpu(ex->ee_block);
3005         ee_len = ext4_ext_get_actual_len(ex);
3006         newblock = split - ee_block + ext4_ext_pblock(ex);
3007
3008         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3009         BUG_ON(!ext4_ext_is_uninitialized(ex) &&
3010                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3011                              EXT4_EXT_MARK_UNINIT1 |
3012                              EXT4_EXT_MARK_UNINIT2));
3013
3014         err = ext4_ext_get_access(handle, inode, path + depth);
3015         if (err)
3016                 goto out;
3017
3018         if (split == ee_block) {
3019                 /*
3020                  * case b: block @split is the block that the extent begins with
3021                  * then we just change the state of the extent, and splitting
3022                  * is not needed.
3023                  */
3024                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
3025                         ext4_ext_mark_uninitialized(ex);
3026                 else
3027                         ext4_ext_mark_initialized(ex);
3028
3029                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3030                         ext4_ext_try_to_merge(handle, inode, path, ex);
3031
3032                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3033                 goto out;
3034         }
3035
3036         /* case a */
3037         memcpy(&orig_ex, ex, sizeof(orig_ex));
3038         ex->ee_len = cpu_to_le16(split - ee_block);
3039         if (split_flag & EXT4_EXT_MARK_UNINIT1)
3040                 ext4_ext_mark_uninitialized(ex);
3041
3042         /*
3043          * path may lead to new leaf, not to original leaf any more
3044          * after ext4_ext_insert_extent() returns,
3045          */
3046         err = ext4_ext_dirty(handle, inode, path + depth);
3047         if (err)
3048                 goto fix_extent_len;
3049
3050         ex2 = &newex;
3051         ex2->ee_block = cpu_to_le32(split);
3052         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3053         ext4_ext_store_pblock(ex2, newblock);
3054         if (split_flag & EXT4_EXT_MARK_UNINIT2)
3055                 ext4_ext_mark_uninitialized(ex2);
3056
3057         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3058         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3059                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3060                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3061                                 err = ext4_ext_zeroout(inode, ex2);
3062                                 zero_ex.ee_block = ex2->ee_block;
3063                                 zero_ex.ee_len = cpu_to_le16(
3064                                                 ext4_ext_get_actual_len(ex2));
3065                                 ext4_ext_store_pblock(&zero_ex,
3066                                                       ext4_ext_pblock(ex2));
3067                         } else {
3068                                 err = ext4_ext_zeroout(inode, ex);
3069                                 zero_ex.ee_block = ex->ee_block;
3070                                 zero_ex.ee_len = cpu_to_le16(
3071                                                 ext4_ext_get_actual_len(ex));
3072                                 ext4_ext_store_pblock(&zero_ex,
3073                                                       ext4_ext_pblock(ex));
3074                         }
3075                 } else {
3076                         err = ext4_ext_zeroout(inode, &orig_ex);
3077                         zero_ex.ee_block = orig_ex.ee_block;
3078                         zero_ex.ee_len = cpu_to_le16(
3079                                                 ext4_ext_get_actual_len(&orig_ex));
3080                         ext4_ext_store_pblock(&zero_ex,
3081                                               ext4_ext_pblock(&orig_ex));
3082                 }
3083
3084                 if (err)
3085                         goto fix_extent_len;
3086                 /* update the extent length and mark as initialized */
3087                 ex->ee_len = cpu_to_le16(ee_len);
3088                 ext4_ext_try_to_merge(handle, inode, path, ex);
3089                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3090                 if (err)
3091                         goto fix_extent_len;
3092
3093                 /* update extent status tree */
3094                 err = ext4_es_zeroout(inode, &zero_ex);
3095
3096                 goto out;
3097         } else if (err)
3098                 goto fix_extent_len;
3099
3100 out:
3101         ext4_ext_show_leaf(inode, path);
3102         return err;
3103
3104 fix_extent_len:
3105         ex->ee_len = orig_ex.ee_len;
3106         ext4_ext_dirty(handle, inode, path + depth);
3107         return err;
3108 }
3109
3110 /*
3111  * ext4_split_extents() splits an extent and mark extent which is covered
3112  * by @map as split_flags indicates
3113  *
3114  * It may result in splitting the extent into multiple extents (upto three)
3115  * There are three possibilities:
3116  *   a> There is no split required
3117  *   b> Splits in two extents: Split is happening at either end of the extent
3118  *   c> Splits in three extents: Somone is splitting in middle of the extent
3119  *
3120  */
3121 static int ext4_split_extent(handle_t *handle,
3122                               struct inode *inode,
3123                               struct ext4_ext_path *path,
3124                               struct ext4_map_blocks *map,
3125                               int split_flag,
3126                               int flags)
3127 {
3128         ext4_lblk_t ee_block;
3129         struct ext4_extent *ex;
3130         unsigned int ee_len, depth;
3131         int err = 0;
3132         int uninitialized;
3133         int split_flag1, flags1;
3134         int allocated = map->m_len;
3135
3136         depth = ext_depth(inode);
3137         ex = path[depth].p_ext;
3138         ee_block = le32_to_cpu(ex->ee_block);
3139         ee_len = ext4_ext_get_actual_len(ex);
3140         uninitialized = ext4_ext_is_uninitialized(ex);
3141
3142         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3143                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3144                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3145                 if (uninitialized)
3146                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3147                                        EXT4_EXT_MARK_UNINIT2;
3148                 if (split_flag & EXT4_EXT_DATA_VALID2)
3149                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3150                 err = ext4_split_extent_at(handle, inode, path,
3151                                 map->m_lblk + map->m_len, split_flag1, flags1);
3152                 if (err)
3153                         goto out;
3154         } else {
3155                 allocated = ee_len - (map->m_lblk - ee_block);
3156         }
3157         /*
3158          * Update path is required because previous ext4_split_extent_at() may
3159          * result in split of original leaf or extent zeroout.
3160          */
3161         ext4_ext_drop_refs(path);
3162         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3163         if (IS_ERR(path))
3164                 return PTR_ERR(path);
3165         depth = ext_depth(inode);
3166         ex = path[depth].p_ext;
3167         uninitialized = ext4_ext_is_uninitialized(ex);
3168         split_flag1 = 0;
3169
3170         if (map->m_lblk >= ee_block) {
3171                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3172                 if (uninitialized) {
3173                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3174                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3175                                                      EXT4_EXT_MARK_UNINIT2);
3176                 }
3177                 err = ext4_split_extent_at(handle, inode, path,
3178                                 map->m_lblk, split_flag1, flags);
3179                 if (err)
3180                         goto out;
3181         }
3182
3183         ext4_ext_show_leaf(inode, path);
3184 out:
3185         return err ? err : allocated;
3186 }
3187
3188 /*
3189  * This function is called by ext4_ext_map_blocks() if someone tries to write
3190  * to an uninitialized extent. It may result in splitting the uninitialized
3191  * extent into multiple extents (up to three - one initialized and two
3192  * uninitialized).
3193  * There are three possibilities:
3194  *   a> There is no split required: Entire extent should be initialized
3195  *   b> Splits in two extents: Write is happening at either end of the extent
3196  *   c> Splits in three extents: Somone is writing in middle of the extent
3197  *
3198  * Pre-conditions:
3199  *  - The extent pointed to by 'path' is uninitialized.
3200  *  - The extent pointed to by 'path' contains a superset
3201  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3202  *
3203  * Post-conditions on success:
3204  *  - the returned value is the number of blocks beyond map->l_lblk
3205  *    that are allocated and initialized.
3206  *    It is guaranteed to be >= map->m_len.
3207  */
3208 static int ext4_ext_convert_to_initialized(handle_t *handle,
3209                                            struct inode *inode,
3210                                            struct ext4_map_blocks *map,
3211                                            struct ext4_ext_path *path,
3212                                            int flags)
3213 {
3214         struct ext4_sb_info *sbi;
3215         struct ext4_extent_header *eh;
3216         struct ext4_map_blocks split_map;
3217         struct ext4_extent zero_ex;
3218         struct ext4_extent *ex, *abut_ex;
3219         ext4_lblk_t ee_block, eof_block;
3220         unsigned int ee_len, depth, map_len = map->m_len;
3221         int allocated = 0, max_zeroout = 0;
3222         int err = 0;
3223         int split_flag = 0;
3224
3225         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3226                 "block %llu, max_blocks %u\n", inode->i_ino,
3227                 (unsigned long long)map->m_lblk, map_len);
3228
3229         sbi = EXT4_SB(inode->i_sb);
3230         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3231                 inode->i_sb->s_blocksize_bits;
3232         if (eof_block < map->m_lblk + map_len)
3233                 eof_block = map->m_lblk + map_len;
3234
3235         depth = ext_depth(inode);
3236         eh = path[depth].p_hdr;
3237         ex = path[depth].p_ext;
3238         ee_block = le32_to_cpu(ex->ee_block);
3239         ee_len = ext4_ext_get_actual_len(ex);
3240         zero_ex.ee_len = 0;
3241
3242         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3243
3244         /* Pre-conditions */
3245         BUG_ON(!ext4_ext_is_uninitialized(ex));
3246         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3247
3248         /*
3249          * Attempt to transfer newly initialized blocks from the currently
3250          * uninitialized extent to its neighbor. This is much cheaper
3251          * than an insertion followed by a merge as those involve costly
3252          * memmove() calls. Transferring to the left is the common case in
3253          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3254          * followed by append writes.
3255          *
3256          * Limitations of the current logic:
3257          *  - L1: we do not deal with writes covering the whole extent.
3258          *    This would require removing the extent if the transfer
3259          *    is possible.
3260          *  - L2: we only attempt to merge with an extent stored in the
3261          *    same extent tree node.
3262          */
3263         if ((map->m_lblk == ee_block) &&
3264                 /* See if we can merge left */
3265                 (map_len < ee_len) &&           /*L1*/
3266                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3267                 ext4_lblk_t prev_lblk;
3268                 ext4_fsblk_t prev_pblk, ee_pblk;
3269                 unsigned int prev_len;
3270
3271                 abut_ex = ex - 1;
3272                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3273                 prev_len = ext4_ext_get_actual_len(abut_ex);
3274                 prev_pblk = ext4_ext_pblock(abut_ex);
3275                 ee_pblk = ext4_ext_pblock(ex);
3276
3277                 /*
3278                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3279                  * upon those conditions:
3280                  * - C1: abut_ex is initialized,
3281                  * - C2: abut_ex is logically abutting ex,
3282                  * - C3: abut_ex is physically abutting ex,
3283                  * - C4: abut_ex can receive the additional blocks without
3284                  *   overflowing the (initialized) length limit.
3285                  */
3286                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3287                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3288                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3289                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3290                         err = ext4_ext_get_access(handle, inode, path + depth);
3291                         if (err)
3292                                 goto out;
3293
3294                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3295                                 map, ex, abut_ex);
3296
3297                         /* Shift the start of ex by 'map_len' blocks */
3298                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3299                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3300                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3301                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3302
3303                         /* Extend abut_ex by 'map_len' blocks */
3304                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3305
3306                         /* Result: number of initialized blocks past m_lblk */
3307                         allocated = map_len;
3308                 }
3309         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3310                    (map_len < ee_len) &&        /*L1*/
3311                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3312                 /* See if we can merge right */
3313                 ext4_lblk_t next_lblk;
3314                 ext4_fsblk_t next_pblk, ee_pblk;
3315                 unsigned int next_len;
3316
3317                 abut_ex = ex + 1;
3318                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3319                 next_len = ext4_ext_get_actual_len(abut_ex);
3320                 next_pblk = ext4_ext_pblock(abut_ex);
3321                 ee_pblk = ext4_ext_pblock(ex);
3322
3323                 /*
3324                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3325                  * upon those conditions:
3326                  * - C1: abut_ex is initialized,
3327                  * - C2: abut_ex is logically abutting ex,
3328                  * - C3: abut_ex is physically abutting ex,
3329                  * - C4: abut_ex can receive the additional blocks without
3330                  *   overflowing the (initialized) length limit.
3331                  */
3332                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3333                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3334                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3335                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3336                         err = ext4_ext_get_access(handle, inode, path + depth);
3337                         if (err)
3338                                 goto out;
3339
3340                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3341                                 map, ex, abut_ex);
3342
3343                         /* Shift the start of abut_ex by 'map_len' blocks */
3344                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3345                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3346                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3347                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3348
3349                         /* Extend abut_ex by 'map_len' blocks */
3350                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3351
3352                         /* Result: number of initialized blocks past m_lblk */
3353                         allocated = map_len;
3354                 }
3355         }
3356         if (allocated) {
3357                 /* Mark the block containing both extents as dirty */
3358                 ext4_ext_dirty(handle, inode, path + depth);
3359
3360                 /* Update path to point to the right extent */
3361                 path[depth].p_ext = abut_ex;
3362                 goto out;
3363         } else
3364                 allocated = ee_len - (map->m_lblk - ee_block);
3365
3366         WARN_ON(map->m_lblk < ee_block);
3367         /*
3368          * It is safe to convert extent to initialized via explicit
3369          * zeroout only if extent is fully insde i_size or new_size.
3370          */
3371         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3372
3373         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3374                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3375                         (inode->i_sb->s_blocksize_bits - 10);
3376
3377         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3378         if (max_zeroout && (ee_len <= max_zeroout)) {
3379                 err = ext4_ext_zeroout(inode, ex);
3380                 if (err)
3381                         goto out;
3382                 zero_ex.ee_block = ex->ee_block;
3383                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3384                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3385
3386                 err = ext4_ext_get_access(handle, inode, path + depth);
3387                 if (err)
3388                         goto out;
3389                 ext4_ext_mark_initialized(ex);
3390                 ext4_ext_try_to_merge(handle, inode, path, ex);
3391                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3392                 goto out;
3393         }
3394
3395         /*
3396          * four cases:
3397          * 1. split the extent into three extents.
3398          * 2. split the extent into two extents, zeroout the first half.
3399          * 3. split the extent into two extents, zeroout the second half.
3400          * 4. split the extent into two extents with out zeroout.
3401          */
3402         split_map.m_lblk = map->m_lblk;
3403         split_map.m_len = map->m_len;
3404
3405         if (max_zeroout && (allocated > map->m_len)) {
3406                 if (allocated <= max_zeroout) {
3407                         /* case 3 */
3408                         zero_ex.ee_block =
3409                                          cpu_to_le32(map->m_lblk);
3410                         zero_ex.ee_len = cpu_to_le16(allocated);
3411                         ext4_ext_store_pblock(&zero_ex,
3412                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3413                         err = ext4_ext_zeroout(inode, &zero_ex);
3414                         if (err)
3415                                 goto out;
3416                         split_map.m_lblk = map->m_lblk;
3417                         split_map.m_len = allocated;
3418                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3419                         /* case 2 */
3420                         if (map->m_lblk != ee_block) {
3421                                 zero_ex.ee_block = ex->ee_block;
3422                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3423                                                         ee_block);
3424                                 ext4_ext_store_pblock(&zero_ex,
3425                                                       ext4_ext_pblock(ex));
3426                                 err = ext4_ext_zeroout(inode, &zero_ex);
3427                                 if (err)
3428                                         goto out;
3429                         }
3430
3431                         split_map.m_lblk = ee_block;
3432                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3433                         allocated = map->m_len;
3434                 }
3435         }
3436
3437         allocated = ext4_split_extent(handle, inode, path,
3438                                       &split_map, split_flag, flags);
3439         if (allocated < 0)
3440                 err = allocated;
3441
3442 out:
3443         /* If we have gotten a failure, don't zero out status tree */
3444         if (!err)
3445                 err = ext4_es_zeroout(inode, &zero_ex);
3446         return err ? err : allocated;
3447 }
3448
3449 /*
3450  * This function is called by ext4_ext_map_blocks() from
3451  * ext4_get_blocks_dio_write() when DIO to write
3452  * to an uninitialized extent.
3453  *
3454  * Writing to an uninitialized extent may result in splitting the uninitialized
3455  * extent into multiple initialized/uninitialized extents (up to three)
3456  * There are three possibilities:
3457  *   a> There is no split required: Entire extent should be uninitialized
3458  *   b> Splits in two extents: Write is happening at either end of the extent
3459  *   c> Splits in three extents: Somone is writing in middle of the extent
3460  *
3461  * One of more index blocks maybe needed if the extent tree grow after
3462  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3463  * complete, we need to split the uninitialized extent before DIO submit
3464  * the IO. The uninitialized extent called at this time will be split
3465  * into three uninitialized extent(at most). After IO complete, the part
3466  * being filled will be convert to initialized by the end_io callback function
3467  * via ext4_convert_unwritten_extents().
3468  *
3469  * Returns the size of uninitialized extent to be written on success.
3470  */
3471 static int ext4_split_unwritten_extents(handle_t *handle,
3472                                         struct inode *inode,
3473                                         struct ext4_map_blocks *map,
3474                                         struct ext4_ext_path *path,
3475                                         int flags)
3476 {
3477         ext4_lblk_t eof_block;
3478         ext4_lblk_t ee_block;
3479         struct ext4_extent *ex;
3480         unsigned int ee_len;
3481         int split_flag = 0, depth;
3482
3483         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3484                 "block %llu, max_blocks %u\n", inode->i_ino,
3485                 (unsigned long long)map->m_lblk, map->m_len);
3486
3487         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3488                 inode->i_sb->s_blocksize_bits;
3489         if (eof_block < map->m_lblk + map->m_len)
3490                 eof_block = map->m_lblk + map->m_len;
3491         /*
3492          * It is safe to convert extent to initialized via explicit
3493          * zeroout only if extent is fully insde i_size or new_size.
3494          */
3495         depth = ext_depth(inode);
3496         ex = path[depth].p_ext;
3497         ee_block = le32_to_cpu(ex->ee_block);
3498         ee_len = ext4_ext_get_actual_len(ex);
3499
3500         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3501         split_flag |= EXT4_EXT_MARK_UNINIT2;
3502         if (flags & EXT4_GET_BLOCKS_CONVERT)
3503                 split_flag |= EXT4_EXT_DATA_VALID2;
3504         flags |= EXT4_GET_BLOCKS_PRE_IO;
3505         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3506 }
3507
3508 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3509                                                 struct inode *inode,
3510                                                 struct ext4_map_blocks *map,
3511                                                 struct ext4_ext_path *path)
3512 {
3513         struct ext4_extent *ex;
3514         ext4_lblk_t ee_block;
3515         unsigned int ee_len;
3516         int depth;
3517         int err = 0;
3518
3519         depth = ext_depth(inode);
3520         ex = path[depth].p_ext;
3521         ee_block = le32_to_cpu(ex->ee_block);
3522         ee_len = ext4_ext_get_actual_len(ex);
3523
3524         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3525                 "block %llu, max_blocks %u\n", inode->i_ino,
3526                   (unsigned long long)ee_block, ee_len);
3527
3528         /* If extent is larger than requested it is a clear sign that we still
3529          * have some extent state machine issues left. So extent_split is still
3530          * required.
3531          * TODO: Once all related issues will be fixed this situation should be
3532          * illegal.
3533          */
3534         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3535 #ifdef EXT4_DEBUG
3536                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3537                              " len %u; IO logical block %llu, len %u\n",
3538                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3539                              (unsigned long long)map->m_lblk, map->m_len);
3540 #endif
3541                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3542                                                    EXT4_GET_BLOCKS_CONVERT);
3543                 if (err < 0)
3544                         goto out;
3545                 ext4_ext_drop_refs(path);
3546                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3547                 if (IS_ERR(path)) {
3548                         err = PTR_ERR(path);
3549                         goto out;
3550                 }
3551                 depth = ext_depth(inode);
3552                 ex = path[depth].p_ext;
3553         }
3554
3555         err = ext4_ext_get_access(handle, inode, path + depth);
3556         if (err)
3557                 goto out;
3558         /* first mark the extent as initialized */
3559         ext4_ext_mark_initialized(ex);
3560
3561         /* note: ext4_ext_correct_indexes() isn't needed here because
3562          * borders are not changed
3563          */
3564         ext4_ext_try_to_merge(handle, inode, path, ex);
3565
3566         /* Mark modified extent as dirty */
3567         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3568 out:
3569         ext4_ext_show_leaf(inode, path);
3570         return err;
3571 }
3572
3573 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3574                         sector_t block, int count)
3575 {
3576         int i;
3577         for (i = 0; i < count; i++)
3578                 unmap_underlying_metadata(bdev, block + i);
3579 }
3580
3581 /*
3582  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3583  */
3584 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3585                               ext4_lblk_t lblk,
3586                               struct ext4_ext_path *path,
3587                               unsigned int len)
3588 {
3589         int i, depth;
3590         struct ext4_extent_header *eh;
3591         struct ext4_extent *last_ex;
3592
3593         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3594                 return 0;
3595
3596         depth = ext_depth(inode);
3597         eh = path[depth].p_hdr;
3598
3599         /*
3600          * We're going to remove EOFBLOCKS_FL entirely in future so we
3601          * do not care for this case anymore. Simply remove the flag
3602          * if there are no extents.
3603          */
3604         if (unlikely(!eh->eh_entries))
3605                 goto out;
3606         last_ex = EXT_LAST_EXTENT(eh);
3607         /*
3608          * We should clear the EOFBLOCKS_FL flag if we are writing the
3609          * last block in the last extent in the file.  We test this by
3610          * first checking to see if the caller to
3611          * ext4_ext_get_blocks() was interested in the last block (or
3612          * a block beyond the last block) in the current extent.  If
3613          * this turns out to be false, we can bail out from this
3614          * function immediately.
3615          */
3616         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3617             ext4_ext_get_actual_len(last_ex))
3618                 return 0;
3619         /*
3620          * If the caller does appear to be planning to write at or
3621          * beyond the end of the current extent, we then test to see
3622          * if the current extent is the last extent in the file, by
3623          * checking to make sure it was reached via the rightmost node
3624          * at each level of the tree.
3625          */
3626         for (i = depth-1; i >= 0; i--)
3627                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3628                         return 0;
3629 out:
3630         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3631         return ext4_mark_inode_dirty(handle, inode);
3632 }
3633
3634 /**
3635  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3636  *
3637  * Return 1 if there is a delalloc block in the range, otherwise 0.
3638  */
3639 int ext4_find_delalloc_range(struct inode *inode,
3640                              ext4_lblk_t lblk_start,
3641                              ext4_lblk_t lblk_end)
3642 {
3643         struct extent_status es;
3644
3645         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3646         if (es.es_len == 0)
3647                 return 0; /* there is no delay extent in this tree */
3648         else if (es.es_lblk <= lblk_start &&
3649                  lblk_start < es.es_lblk + es.es_len)
3650                 return 1;
3651         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3652                 return 1;
3653         else
3654                 return 0;
3655 }
3656
3657 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3658 {
3659         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3660         ext4_lblk_t lblk_start, lblk_end;
3661         lblk_start = lblk & (~(sbi->s_cluster_ratio - 1));
3662         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3663
3664         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3665 }
3666
3667 /**
3668  * Determines how many complete clusters (out of those specified by the 'map')
3669  * are under delalloc and were reserved quota for.
3670  * This function is called when we are writing out the blocks that were
3671  * originally written with their allocation delayed, but then the space was
3672  * allocated using fallocate() before the delayed allocation could be resolved.
3673  * The cases to look for are:
3674  * ('=' indicated delayed allocated blocks
3675  *  '-' indicates non-delayed allocated blocks)
3676  * (a) partial clusters towards beginning and/or end outside of allocated range
3677  *     are not delalloc'ed.
3678  *      Ex:
3679  *      |----c---=|====c====|====c====|===-c----|
3680  *               |++++++ allocated ++++++|
3681  *      ==> 4 complete clusters in above example
3682  *
3683  * (b) partial cluster (outside of allocated range) towards either end is
3684  *     marked for delayed allocation. In this case, we will exclude that
3685  *     cluster.
3686  *      Ex:
3687  *      |----====c========|========c========|
3688  *           |++++++ allocated ++++++|
3689  *      ==> 1 complete clusters in above example
3690  *
3691  *      Ex:
3692  *      |================c================|
3693  *            |++++++ allocated ++++++|
3694  *      ==> 0 complete clusters in above example
3695  *
3696  * The ext4_da_update_reserve_space will be called only if we
3697  * determine here that there were some "entire" clusters that span
3698  * this 'allocated' range.
3699  * In the non-bigalloc case, this function will just end up returning num_blks
3700  * without ever calling ext4_find_delalloc_range.
3701  */
3702 static unsigned int
3703 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3704                            unsigned int num_blks)
3705 {
3706         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3707         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3708         ext4_lblk_t lblk_from, lblk_to, c_offset;
3709         unsigned int allocated_clusters = 0;
3710
3711         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3712         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3713
3714         /* max possible clusters for this allocation */
3715         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3716
3717         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3718
3719         /* Check towards left side */
3720         c_offset = lblk_start & (sbi->s_cluster_ratio - 1);
3721         if (c_offset) {
3722                 lblk_from = lblk_start & (~(sbi->s_cluster_ratio - 1));
3723                 lblk_to = lblk_from + c_offset - 1;
3724
3725                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3726                         allocated_clusters--;
3727         }
3728
3729         /* Now check towards right. */
3730         c_offset = (lblk_start + num_blks) & (sbi->s_cluster_ratio - 1);
3731         if (allocated_clusters && c_offset) {
3732                 lblk_from = lblk_start + num_blks;
3733                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3734
3735                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3736                         allocated_clusters--;
3737         }
3738
3739         return allocated_clusters;
3740 }
3741
3742 static int
3743 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3744                         struct ext4_map_blocks *map,
3745                         struct ext4_ext_path *path, int flags,
3746                         unsigned int allocated, ext4_fsblk_t newblock)
3747 {
3748         int ret = 0;
3749         int err = 0;
3750         ext4_io_end_t *io = ext4_inode_aio(inode);
3751
3752         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3753                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3754                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3755                   flags, allocated);
3756         ext4_ext_show_leaf(inode, path);
3757
3758         /*
3759          * When writing into uninitialized space, we should not fail to
3760          * allocate metadata blocks for the new extent block if needed.
3761          */
3762         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
3763
3764         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3765                                                     allocated, newblock);
3766
3767         /* get_block() before submit the IO, split the extent */
3768         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3769                 ret = ext4_split_unwritten_extents(handle, inode, map,
3770                                                    path, flags);
3771                 if (ret <= 0)
3772                         goto out;
3773                 /*
3774                  * Flag the inode(non aio case) or end_io struct (aio case)
3775                  * that this IO needs to conversion to written when IO is
3776                  * completed
3777                  */
3778                 if (io)
3779                         ext4_set_io_unwritten_flag(inode, io);
3780                 else
3781                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3782                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3783                 if (ext4_should_dioread_nolock(inode))
3784                         map->m_flags |= EXT4_MAP_UNINIT;
3785                 goto out;
3786         }
3787         /* IO end_io complete, convert the filled extent to written */
3788         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3789                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3790                                                         path);
3791                 if (ret >= 0) {
3792                         ext4_update_inode_fsync_trans(handle, inode, 1);
3793                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3794                                                  path, map->m_len);
3795                 } else
3796                         err = ret;
3797                 map->m_flags |= EXT4_MAP_MAPPED;
3798                 if (allocated > map->m_len)
3799                         allocated = map->m_len;
3800                 map->m_len = allocated;
3801                 goto out2;
3802         }
3803         /* buffered IO case */
3804         /*
3805          * repeat fallocate creation request
3806          * we already have an unwritten extent
3807          */
3808         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT) {
3809                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3810                 goto map_out;
3811         }
3812
3813         /* buffered READ or buffered write_begin() lookup */
3814         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3815                 /*
3816                  * We have blocks reserved already.  We
3817                  * return allocated blocks so that delalloc
3818                  * won't do block reservation for us.  But
3819                  * the buffer head will be unmapped so that
3820                  * a read from the block returns 0s.
3821                  */
3822                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3823                 goto out1;
3824         }
3825
3826         /* buffered write, writepage time, convert*/
3827         ret = ext4_ext_convert_to_initialized(handle, inode, map, path, flags);
3828         if (ret >= 0)
3829                 ext4_update_inode_fsync_trans(handle, inode, 1);
3830 out:
3831         if (ret <= 0) {
3832                 err = ret;
3833                 goto out2;
3834         } else
3835                 allocated = ret;
3836         map->m_flags |= EXT4_MAP_NEW;
3837         /*
3838          * if we allocated more blocks than requested
3839          * we need to make sure we unmap the extra block
3840          * allocated. The actual needed block will get
3841          * unmapped later when we find the buffer_head marked
3842          * new.
3843          */
3844         if (allocated > map->m_len) {
3845                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3846                                         newblock + map->m_len,
3847                                         allocated - map->m_len);
3848                 allocated = map->m_len;
3849         }
3850         map->m_len = allocated;
3851
3852         /*
3853          * If we have done fallocate with the offset that is already
3854          * delayed allocated, we would have block reservation
3855          * and quota reservation done in the delayed write path.
3856          * But fallocate would have already updated quota and block
3857          * count for this offset. So cancel these reservation
3858          */
3859         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3860                 unsigned int reserved_clusters;
3861                 reserved_clusters = get_reserved_cluster_alloc(inode,
3862                                 map->m_lblk, map->m_len);
3863                 if (reserved_clusters)
3864                         ext4_da_update_reserve_space(inode,
3865                                                      reserved_clusters,
3866                                                      0);
3867         }
3868
3869 map_out:
3870         map->m_flags |= EXT4_MAP_MAPPED;
3871         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3872                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3873                                          map->m_len);
3874                 if (err < 0)
3875                         goto out2;
3876         }
3877 out1:
3878         if (allocated > map->m_len)
3879                 allocated = map->m_len;
3880         ext4_ext_show_leaf(inode, path);
3881         map->m_pblk = newblock;
3882         map->m_len = allocated;
3883 out2:
3884         if (path) {
3885                 ext4_ext_drop_refs(path);
3886                 kfree(path);
3887         }
3888         return err ? err : allocated;
3889 }
3890
3891 /*
3892  * get_implied_cluster_alloc - check to see if the requested
3893  * allocation (in the map structure) overlaps with a cluster already
3894  * allocated in an extent.
3895  *      @sb     The filesystem superblock structure
3896  *      @map    The requested lblk->pblk mapping
3897  *      @ex     The extent structure which might contain an implied
3898  *                      cluster allocation
3899  *
3900  * This function is called by ext4_ext_map_blocks() after we failed to
3901  * find blocks that were already in the inode's extent tree.  Hence,
3902  * we know that the beginning of the requested region cannot overlap
3903  * the extent from the inode's extent tree.  There are three cases we
3904  * want to catch.  The first is this case:
3905  *
3906  *               |--- cluster # N--|
3907  *    |--- extent ---|  |---- requested region ---|
3908  *                      |==========|
3909  *
3910  * The second case that we need to test for is this one:
3911  *
3912  *   |--------- cluster # N ----------------|
3913  *         |--- requested region --|   |------- extent ----|
3914  *         |=======================|
3915  *
3916  * The third case is when the requested region lies between two extents
3917  * within the same cluster:
3918  *          |------------- cluster # N-------------|
3919  * |----- ex -----|                  |---- ex_right ----|
3920  *                  |------ requested region ------|
3921  *                  |================|
3922  *
3923  * In each of the above cases, we need to set the map->m_pblk and
3924  * map->m_len so it corresponds to the return the extent labelled as
3925  * "|====|" from cluster #N, since it is already in use for data in
3926  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3927  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3928  * as a new "allocated" block region.  Otherwise, we will return 0 and
3929  * ext4_ext_map_blocks() will then allocate one or more new clusters
3930  * by calling ext4_mb_new_blocks().
3931  */
3932 static int get_implied_cluster_alloc(struct super_block *sb,
3933                                      struct ext4_map_blocks *map,
3934                                      struct ext4_extent *ex,
3935                                      struct ext4_ext_path *path)
3936 {
3937         struct ext4_sb_info *sbi = EXT4_SB(sb);
3938         ext4_lblk_t c_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
3939         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3940         ext4_lblk_t rr_cluster_start;
3941         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3942         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3943         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3944
3945         /* The extent passed in that we are trying to match */
3946         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3947         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3948
3949         /* The requested region passed into ext4_map_blocks() */
3950         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3951
3952         if ((rr_cluster_start == ex_cluster_end) ||
3953             (rr_cluster_start == ex_cluster_start)) {
3954                 if (rr_cluster_start == ex_cluster_end)
3955                         ee_start += ee_len - 1;
3956                 map->m_pblk = (ee_start & ~(sbi->s_cluster_ratio - 1)) +
3957                         c_offset;
3958                 map->m_len = min(map->m_len,
3959                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3960                 /*
3961                  * Check for and handle this case:
3962                  *
3963                  *   |--------- cluster # N-------------|
3964                  *                     |------- extent ----|
3965                  *         |--- requested region ---|
3966                  *         |===========|
3967                  */
3968
3969                 if (map->m_lblk < ee_block)
3970                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
3971
3972                 /*
3973                  * Check for the case where there is already another allocated
3974                  * block to the right of 'ex' but before the end of the cluster.
3975                  *
3976                  *          |------------- cluster # N-------------|
3977                  * |----- ex -----|                  |---- ex_right ----|
3978                  *                  |------ requested region ------|
3979                  *                  |================|
3980                  */
3981                 if (map->m_lblk > ee_block) {
3982                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
3983                         map->m_len = min(map->m_len, next - map->m_lblk);
3984                 }
3985
3986                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
3987                 return 1;
3988         }
3989
3990         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
3991         return 0;
3992 }
3993
3994
3995 /*
3996  * Block allocation/map/preallocation routine for extents based files
3997  *
3998  *
3999  * Need to be called with
4000  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4001  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4002  *
4003  * return > 0, number of of blocks already mapped/allocated
4004  *          if create == 0 and these are pre-allocated blocks
4005  *              buffer head is unmapped
4006  *          otherwise blocks are mapped
4007  *
4008  * return = 0, if plain look up failed (blocks have not been allocated)
4009  *          buffer head is unmapped
4010  *
4011  * return < 0, error case.
4012  */
4013 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4014                         struct ext4_map_blocks *map, int flags)
4015 {
4016         struct ext4_ext_path *path = NULL;
4017         struct ext4_extent newex, *ex, *ex2;
4018         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4019         ext4_fsblk_t newblock = 0;
4020         int free_on_err = 0, err = 0, depth;
4021         unsigned int allocated = 0, offset = 0;
4022         unsigned int allocated_clusters = 0;
4023         struct ext4_allocation_request ar;
4024         ext4_io_end_t *io = ext4_inode_aio(inode);
4025         ext4_lblk_t cluster_offset;
4026         int set_unwritten = 0;
4027
4028         ext_debug("blocks %u/%u requested for inode %lu\n",
4029                   map->m_lblk, map->m_len, inode->i_ino);
4030         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4031
4032         /* find extent for this block */
4033         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
4034         if (IS_ERR(path)) {
4035                 err = PTR_ERR(path);
4036                 path = NULL;
4037                 goto out2;
4038         }
4039
4040         depth = ext_depth(inode);
4041
4042         /*
4043          * consistent leaf must not be empty;
4044          * this situation is possible, though, _during_ tree modification;
4045          * this is why assert can't be put in ext4_ext_find_extent()
4046          */
4047         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4048                 EXT4_ERROR_INODE(inode, "bad extent address "
4049                                  "lblock: %lu, depth: %d pblock %lld",
4050                                  (unsigned long) map->m_lblk, depth,
4051                                  path[depth].p_block);
4052                 err = -EIO;
4053                 goto out2;
4054         }
4055
4056         ex = path[depth].p_ext;
4057         if (ex) {
4058                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4059                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4060                 unsigned short ee_len;
4061
4062                 /*
4063                  * Uninitialized extents are treated as holes, except that
4064                  * we split out initialized portions during a write.
4065                  */
4066                 ee_len = ext4_ext_get_actual_len(ex);
4067
4068                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4069
4070                 /* if found extent covers block, simply return it */
4071                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4072                         newblock = map->m_lblk - ee_block + ee_start;
4073                         /* number of remaining blocks in the extent */
4074                         allocated = ee_len - (map->m_lblk - ee_block);
4075                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4076                                   ee_block, ee_len, newblock);
4077
4078                         if (!ext4_ext_is_uninitialized(ex))
4079                                 goto out;
4080
4081                         allocated = ext4_ext_handle_uninitialized_extents(
4082                                 handle, inode, map, path, flags,
4083                                 allocated, newblock);
4084                         goto out3;
4085                 }
4086         }
4087
4088         if ((sbi->s_cluster_ratio > 1) &&
4089             ext4_find_delalloc_cluster(inode, map->m_lblk))
4090                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4091
4092         /*
4093          * requested block isn't allocated yet;
4094          * we couldn't try to create block if create flag is zero
4095          */
4096         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4097                 /*
4098                  * put just found gap into cache to speed up
4099                  * subsequent requests
4100                  */
4101                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4102                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4103                 goto out2;
4104         }
4105
4106         /*
4107          * Okay, we need to do block allocation.
4108          */
4109         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4110         newex.ee_block = cpu_to_le32(map->m_lblk);
4111         cluster_offset = map->m_lblk & (sbi->s_cluster_ratio-1);
4112
4113         /*
4114          * If we are doing bigalloc, check to see if the extent returned
4115          * by ext4_ext_find_extent() implies a cluster we can use.
4116          */
4117         if (cluster_offset && ex &&
4118             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4119                 ar.len = allocated = map->m_len;
4120                 newblock = map->m_pblk;
4121                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4122                 goto got_allocated_blocks;
4123         }
4124
4125         /* find neighbour allocated blocks */
4126         ar.lleft = map->m_lblk;
4127         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4128         if (err)
4129                 goto out2;
4130         ar.lright = map->m_lblk;
4131         ex2 = NULL;
4132         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4133         if (err)
4134                 goto out2;
4135
4136         /* Check if the extent after searching to the right implies a
4137          * cluster we can use. */
4138         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4139             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4140                 ar.len = allocated = map->m_len;
4141                 newblock = map->m_pblk;
4142                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4143                 goto got_allocated_blocks;
4144         }
4145
4146         /*
4147          * See if request is beyond maximum number of blocks we can have in
4148          * a single extent. For an initialized extent this limit is
4149          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4150          * EXT_UNINIT_MAX_LEN.
4151          */
4152         if (map->m_len > EXT_INIT_MAX_LEN &&
4153             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4154                 map->m_len = EXT_INIT_MAX_LEN;
4155         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4156                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4157                 map->m_len = EXT_UNINIT_MAX_LEN;
4158
4159         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4160         newex.ee_len = cpu_to_le16(map->m_len);
4161         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4162         if (err)
4163                 allocated = ext4_ext_get_actual_len(&newex);
4164         else
4165                 allocated = map->m_len;
4166
4167         /* allocate new block */
4168         ar.inode = inode;
4169         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4170         ar.logical = map->m_lblk;
4171         /*
4172          * We calculate the offset from the beginning of the cluster
4173          * for the logical block number, since when we allocate a
4174          * physical cluster, the physical block should start at the
4175          * same offset from the beginning of the cluster.  This is
4176          * needed so that future calls to get_implied_cluster_alloc()
4177          * work correctly.
4178          */
4179         offset = map->m_lblk & (sbi->s_cluster_ratio - 1);
4180         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4181         ar.goal -= offset;
4182         ar.logical -= offset;
4183         if (S_ISREG(inode->i_mode))
4184                 ar.flags = EXT4_MB_HINT_DATA;
4185         else
4186                 /* disable in-core preallocation for non-regular files */
4187                 ar.flags = 0;
4188         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4189                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4190         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4191         if (!newblock)
4192                 goto out2;
4193         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4194                   ar.goal, newblock, allocated);
4195         free_on_err = 1;
4196         allocated_clusters = ar.len;
4197         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4198         if (ar.len > allocated)
4199                 ar.len = allocated;
4200
4201 got_allocated_blocks:
4202         /* try to insert new extent into found leaf and return */
4203         ext4_ext_store_pblock(&newex, newblock + offset);
4204         newex.ee_len = cpu_to_le16(ar.len);
4205         /* Mark uninitialized */
4206         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4207                 ext4_ext_mark_uninitialized(&newex);
4208                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4209                 /*
4210                  * io_end structure was created for every IO write to an
4211                  * uninitialized extent. To avoid unnecessary conversion,
4212                  * here we flag the IO that really needs the conversion.
4213                  * For non asycn direct IO case, flag the inode state
4214                  * that we need to perform conversion when IO is done.
4215                  */
4216                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4217                         set_unwritten = 1;
4218                 if (ext4_should_dioread_nolock(inode))
4219                         map->m_flags |= EXT4_MAP_UNINIT;
4220         }
4221
4222         err = 0;
4223         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4224                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4225                                          path, ar.len);
4226         if (!err)
4227                 err = ext4_ext_insert_extent(handle, inode, path,
4228                                              &newex, flags);
4229
4230         if (!err && set_unwritten) {
4231                 if (io)
4232                         ext4_set_io_unwritten_flag(inode, io);
4233                 else
4234                         ext4_set_inode_state(inode,
4235                                              EXT4_STATE_DIO_UNWRITTEN);
4236         }
4237
4238         if (err && free_on_err) {
4239                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4240                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4241                 /* free data blocks we just allocated */
4242                 /* not a good idea to call discard here directly,
4243                  * but otherwise we'd need to call it every free() */
4244                 ext4_discard_preallocations(inode);
4245                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4246                                  ext4_ext_get_actual_len(&newex), fb_flags);
4247                 goto out2;
4248         }
4249
4250         /* previous routine could use block we allocated */
4251         newblock = ext4_ext_pblock(&newex);
4252         allocated = ext4_ext_get_actual_len(&newex);
4253         if (allocated > map->m_len)
4254                 allocated = map->m_len;
4255         map->m_flags |= EXT4_MAP_NEW;
4256
4257         /*
4258          * Update reserved blocks/metadata blocks after successful
4259          * block allocation which had been deferred till now.
4260          */
4261         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4262                 unsigned int reserved_clusters;
4263                 /*
4264                  * Check how many clusters we had reserved this allocated range
4265                  */
4266                 reserved_clusters = get_reserved_cluster_alloc(inode,
4267                                                 map->m_lblk, allocated);
4268                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4269                         if (reserved_clusters) {
4270                                 /*
4271                                  * We have clusters reserved for this range.
4272                                  * But since we are not doing actual allocation
4273                                  * and are simply using blocks from previously
4274                                  * allocated cluster, we should release the
4275                                  * reservation and not claim quota.
4276                                  */
4277                                 ext4_da_update_reserve_space(inode,
4278                                                 reserved_clusters, 0);
4279                         }
4280                 } else {
4281                         BUG_ON(allocated_clusters < reserved_clusters);
4282                         if (reserved_clusters < allocated_clusters) {
4283                                 struct ext4_inode_info *ei = EXT4_I(inode);
4284                                 int reservation = allocated_clusters -
4285                                                   reserved_clusters;
4286                                 /*
4287                                  * It seems we claimed few clusters outside of
4288                                  * the range of this allocation. We should give
4289                                  * it back to the reservation pool. This can
4290                                  * happen in the following case:
4291                                  *
4292                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4293                                  *   cluster has 4 blocks. Thus, the clusters
4294                                  *   are [0-3],[4-7],[8-11]...
4295                                  * * First comes delayed allocation write for
4296                                  *   logical blocks 10 & 11. Since there were no
4297                                  *   previous delayed allocated blocks in the
4298                                  *   range [8-11], we would reserve 1 cluster
4299                                  *   for this write.
4300                                  * * Next comes write for logical blocks 3 to 8.
4301                                  *   In this case, we will reserve 2 clusters
4302                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4303                                  *   that range has a delayed allocated blocks.
4304                                  *   Thus total reserved clusters now becomes 3.
4305                                  * * Now, during the delayed allocation writeout
4306                                  *   time, we will first write blocks [3-8] and
4307                                  *   allocate 3 clusters for writing these
4308                                  *   blocks. Also, we would claim all these
4309                                  *   three clusters above.
4310                                  * * Now when we come here to writeout the
4311                                  *   blocks [10-11], we would expect to claim
4312                                  *   the reservation of 1 cluster we had made
4313                                  *   (and we would claim it since there are no
4314                                  *   more delayed allocated blocks in the range
4315                                  *   [8-11]. But our reserved cluster count had
4316                                  *   already gone to 0.
4317                                  *
4318                                  *   Thus, at the step 4 above when we determine
4319                                  *   that there are still some unwritten delayed
4320                                  *   allocated blocks outside of our current
4321                                  *   block range, we should increment the
4322                                  *   reserved clusters count so that when the
4323                                  *   remaining blocks finally gets written, we
4324                                  *   could claim them.
4325                                  */
4326                                 dquot_reserve_block(inode,
4327                                                 EXT4_C2B(sbi, reservation));
4328                                 spin_lock(&ei->i_block_reservation_lock);
4329                                 ei->i_reserved_data_blocks += reservation;
4330                                 spin_unlock(&ei->i_block_reservation_lock);
4331                         }
4332                         /*
4333                          * We will claim quota for all newly allocated blocks.
4334                          * We're updating the reserved space *after* the
4335                          * correction above so we do not accidentally free
4336                          * all the metadata reservation because we might
4337                          * actually need it later on.
4338                          */
4339                         ext4_da_update_reserve_space(inode, allocated_clusters,
4340                                                         1);
4341                 }
4342         }
4343
4344         /*
4345          * Cache the extent and update transaction to commit on fdatasync only
4346          * when it is _not_ an uninitialized extent.
4347          */
4348         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0)
4349                 ext4_update_inode_fsync_trans(handle, inode, 1);
4350         else
4351                 ext4_update_inode_fsync_trans(handle, inode, 0);
4352 out:
4353         if (allocated > map->m_len)
4354                 allocated = map->m_len;
4355         ext4_ext_show_leaf(inode, path);
4356         map->m_flags |= EXT4_MAP_MAPPED;
4357         map->m_pblk = newblock;
4358         map->m_len = allocated;
4359 out2:
4360         if (path) {
4361                 ext4_ext_drop_refs(path);
4362                 kfree(path);
4363         }
4364
4365 out3:
4366         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4367
4368         return err ? err : allocated;
4369 }
4370
4371 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4372 {
4373         struct super_block *sb = inode->i_sb;
4374         ext4_lblk_t last_block;
4375         int err = 0;
4376
4377         /*
4378          * TODO: optimization is possible here.
4379          * Probably we need not scan at all,
4380          * because page truncation is enough.
4381          */
4382
4383         /* we have to know where to truncate from in crash case */
4384         EXT4_I(inode)->i_disksize = inode->i_size;
4385         ext4_mark_inode_dirty(handle, inode);
4386
4387         last_block = (inode->i_size + sb->s_blocksize - 1)
4388                         >> EXT4_BLOCK_SIZE_BITS(sb);
4389         err = ext4_es_remove_extent(inode, last_block,
4390                                     EXT_MAX_BLOCKS - last_block);
4391         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4392 }
4393
4394 static void ext4_falloc_update_inode(struct inode *inode,
4395                                 int mode, loff_t new_size, int update_ctime)
4396 {
4397         struct timespec now;
4398
4399         if (update_ctime) {
4400                 now = current_fs_time(inode->i_sb);
4401                 if (!timespec_equal(&inode->i_ctime, &now))
4402                         inode->i_ctime = now;
4403         }
4404         /*
4405          * Update only when preallocation was requested beyond
4406          * the file size.
4407          */
4408         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4409                 if (new_size > i_size_read(inode))
4410                         i_size_write(inode, new_size);
4411                 if (new_size > EXT4_I(inode)->i_disksize)
4412                         ext4_update_i_disksize(inode, new_size);
4413         } else {
4414                 /*
4415                  * Mark that we allocate beyond EOF so the subsequent truncate
4416                  * can proceed even if the new size is the same as i_size.
4417                  */
4418                 if (new_size > i_size_read(inode))
4419                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4420         }
4421
4422 }
4423
4424 /*
4425  * preallocate space for a file. This implements ext4's fallocate file
4426  * operation, which gets called from sys_fallocate system call.
4427  * For block-mapped files, posix_fallocate should fall back to the method
4428  * of writing zeroes to the required new blocks (the same behavior which is
4429  * expected for file systems which do not support fallocate() system call).
4430  */
4431 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4432 {
4433         struct inode *inode = file_inode(file);
4434         handle_t *handle;
4435         loff_t new_size;
4436         unsigned int max_blocks;
4437         int ret = 0;
4438         int ret2 = 0;
4439         int retries = 0;
4440         int flags;
4441         struct ext4_map_blocks map;
4442         unsigned int credits, blkbits = inode->i_blkbits;
4443
4444         /* Return error if mode is not supported */
4445         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4446                 return -EOPNOTSUPP;
4447
4448         if (mode & FALLOC_FL_PUNCH_HOLE)
4449                 return ext4_punch_hole(file, offset, len);
4450
4451         ret = ext4_convert_inline_data(inode);
4452         if (ret)
4453                 return ret;
4454
4455         /*
4456          * currently supporting (pre)allocate mode for extent-based
4457          * files _only_
4458          */
4459         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4460                 return -EOPNOTSUPP;
4461
4462         trace_ext4_fallocate_enter(inode, offset, len, mode);
4463         map.m_lblk = offset >> blkbits;
4464         /*
4465          * We can't just convert len to max_blocks because
4466          * If blocksize = 4096 offset = 3072 and len = 2048
4467          */
4468         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4469                 - map.m_lblk;
4470         /*
4471          * credits to insert 1 extent into extent tree
4472          */
4473         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4474         mutex_lock(&inode->i_mutex);
4475         ret = inode_newsize_ok(inode, (len + offset));
4476         if (ret) {
4477                 mutex_unlock(&inode->i_mutex);
4478                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4479                 return ret;
4480         }
4481         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4482         if (mode & FALLOC_FL_KEEP_SIZE)
4483                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4484         /*
4485          * Don't normalize the request if it can fit in one extent so
4486          * that it doesn't get unnecessarily split into multiple
4487          * extents.
4488          */
4489         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4490                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4491
4492 retry:
4493         while (ret >= 0 && ret < max_blocks) {
4494                 map.m_lblk = map.m_lblk + ret;
4495                 map.m_len = max_blocks = max_blocks - ret;
4496                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4497                                             credits);
4498                 if (IS_ERR(handle)) {
4499                         ret = PTR_ERR(handle);
4500                         break;
4501                 }
4502                 ret = ext4_map_blocks(handle, inode, &map, flags);
4503                 if (ret <= 0) {
4504 #ifdef EXT4FS_DEBUG
4505                         ext4_warning(inode->i_sb,
4506                                      "inode #%lu: block %u: len %u: "
4507                                      "ext4_ext_map_blocks returned %d",
4508                                      inode->i_ino, map.m_lblk,
4509                                      map.m_len, ret);
4510 #endif
4511                         ext4_mark_inode_dirty(handle, inode);
4512                         ret2 = ext4_journal_stop(handle);
4513                         break;
4514                 }
4515                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4516                                                 blkbits) >> blkbits))
4517                         new_size = offset + len;
4518                 else
4519                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4520
4521                 ext4_falloc_update_inode(inode, mode, new_size,
4522                                          (map.m_flags & EXT4_MAP_NEW));
4523                 ext4_mark_inode_dirty(handle, inode);
4524                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4525                         ext4_handle_sync(handle);
4526                 ret2 = ext4_journal_stop(handle);
4527                 if (ret2)
4528                         break;
4529         }
4530         if (ret == -ENOSPC &&
4531                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4532                 ret = 0;
4533                 goto retry;
4534         }
4535         mutex_unlock(&inode->i_mutex);
4536         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4537                                 ret > 0 ? ret2 : ret);
4538         return ret > 0 ? ret2 : ret;
4539 }
4540
4541 /*
4542  * This function convert a range of blocks to written extents
4543  * The caller of this function will pass the start offset and the size.
4544  * all unwritten extents within this range will be converted to
4545  * written extents.
4546  *
4547  * This function is called from the direct IO end io call back
4548  * function, to convert the fallocated extents after IO is completed.
4549  * Returns 0 on success.
4550  */
4551 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4552                                     ssize_t len)
4553 {
4554         handle_t *handle;
4555         unsigned int max_blocks;
4556         int ret = 0;
4557         int ret2 = 0;
4558         struct ext4_map_blocks map;
4559         unsigned int credits, blkbits = inode->i_blkbits;
4560
4561         map.m_lblk = offset >> blkbits;
4562         /*
4563          * We can't just convert len to max_blocks because
4564          * If blocksize = 4096 offset = 3072 and len = 2048
4565          */
4566         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4567                       map.m_lblk);
4568         /*
4569          * credits to insert 1 extent into extent tree
4570          */
4571         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4572         while (ret >= 0 && ret < max_blocks) {
4573                 map.m_lblk += ret;
4574                 map.m_len = (max_blocks -= ret);
4575                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
4576                 if (IS_ERR(handle)) {
4577                         ret = PTR_ERR(handle);
4578                         break;
4579                 }
4580                 ret = ext4_map_blocks(handle, inode, &map,
4581                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4582                 if (ret <= 0)
4583                         ext4_warning(inode->i_sb,
4584                                      "inode #%lu: block %u: len %u: "
4585                                      "ext4_ext_map_blocks returned %d",
4586                                      inode->i_ino, map.m_lblk,
4587                                      map.m_len, ret);
4588                 ext4_mark_inode_dirty(handle, inode);
4589                 ret2 = ext4_journal_stop(handle);
4590                 if (ret <= 0 || ret2 )
4591                         break;
4592         }
4593         return ret > 0 ? ret2 : ret;
4594 }
4595
4596 /*
4597  * If newes is not existing extent (newes->ec_pblk equals zero) find
4598  * delayed extent at start of newes and update newes accordingly and
4599  * return start of the next delayed extent.
4600  *
4601  * If newes is existing extent (newes->ec_pblk is not equal zero)
4602  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4603  * extent found. Leave newes unmodified.
4604  */
4605 static int ext4_find_delayed_extent(struct inode *inode,
4606                                     struct extent_status *newes)
4607 {
4608         struct extent_status es;
4609         ext4_lblk_t block, next_del;
4610
4611         if (newes->es_pblk == 0) {
4612                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
4613                                 newes->es_lblk + newes->es_len - 1, &es);
4614
4615                 /*
4616                  * No extent in extent-tree contains block @newes->es_pblk,
4617                  * then the block may stay in 1)a hole or 2)delayed-extent.
4618                  */
4619                 if (es.es_len == 0)
4620                         /* A hole found. */
4621                         return 0;
4622
4623                 if (es.es_lblk > newes->es_lblk) {
4624                         /* A hole found. */
4625                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4626                                             newes->es_len);
4627                         return 0;
4628                 }
4629
4630                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
4631         }
4632
4633         block = newes->es_lblk + newes->es_len;
4634         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
4635         if (es.es_len == 0)
4636                 next_del = EXT_MAX_BLOCKS;
4637         else
4638                 next_del = es.es_lblk;
4639
4640         return next_del;
4641 }
4642 /* fiemap flags we can handle specified here */
4643 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4644
4645 static int ext4_xattr_fiemap(struct inode *inode,
4646                                 struct fiemap_extent_info *fieinfo)
4647 {
4648         __u64 physical = 0;
4649         __u64 length;
4650         __u32 flags = FIEMAP_EXTENT_LAST;
4651         int blockbits = inode->i_sb->s_blocksize_bits;
4652         int error = 0;
4653
4654         /* in-inode? */
4655         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4656                 struct ext4_iloc iloc;
4657                 int offset;     /* offset of xattr in inode */
4658
4659                 error = ext4_get_inode_loc(inode, &iloc);
4660                 if (error)
4661                         return error;
4662                 physical = iloc.bh->b_blocknr << blockbits;
4663                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4664                                 EXT4_I(inode)->i_extra_isize;
4665                 physical += offset;
4666                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4667                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4668                 brelse(iloc.bh);
4669         } else { /* external block */
4670                 physical = EXT4_I(inode)->i_file_acl << blockbits;
4671                 length = inode->i_sb->s_blocksize;
4672         }
4673
4674         if (physical)
4675                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4676                                                 length, flags);
4677         return (error < 0 ? error : 0);
4678 }
4679
4680 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4681                 __u64 start, __u64 len)
4682 {
4683         ext4_lblk_t start_blk;
4684         int error = 0;
4685
4686         if (ext4_has_inline_data(inode)) {
4687                 int has_inline = 1;
4688
4689                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4690
4691                 if (has_inline)
4692                         return error;
4693         }
4694
4695         /* fallback to generic here if not in extents fmt */
4696         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4697                 return generic_block_fiemap(inode, fieinfo, start, len,
4698                         ext4_get_block);
4699
4700         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4701                 return -EBADR;
4702
4703         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4704                 error = ext4_xattr_fiemap(inode, fieinfo);
4705         } else {
4706                 ext4_lblk_t len_blks;
4707                 __u64 last_blk;
4708
4709                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4710                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4711                 if (last_blk >= EXT_MAX_BLOCKS)
4712                         last_blk = EXT_MAX_BLOCKS-1;
4713                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4714
4715                 /*
4716                  * Walk the extent tree gathering extent information
4717                  * and pushing extents back to the user.
4718                  */
4719                 error = ext4_fill_fiemap_extents(inode, start_blk,
4720                                                  len_blks, fieinfo);
4721         }
4722
4723         return error;
4724 }