GNU Linux-libre 4.19.211-gnu1
[releases.git] / fs / ext4 / extents.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
4  * Written by Alex Tomas <alex@clusterfs.com>
5  *
6  * Architecture independence:
7  *   Copyright (c) 2005, Bull S.A.
8  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
9  */
10
11 /*
12  * Extents support for EXT4
13  *
14  * TODO:
15  *   - ext4*_error() should be used in some situations
16  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
17  *   - smart tree reduction
18  */
19
20 #include <linux/fs.h>
21 #include <linux/time.h>
22 #include <linux/jbd2.h>
23 #include <linux/highuid.h>
24 #include <linux/pagemap.h>
25 #include <linux/quotaops.h>
26 #include <linux/string.h>
27 #include <linux/slab.h>
28 #include <linux/uaccess.h>
29 #include <linux/fiemap.h>
30 #include <linux/backing-dev.h>
31 #include "ext4_jbd2.h"
32 #include "ext4_extents.h"
33 #include "xattr.h"
34
35 #include <trace/events/ext4.h>
36
37 /*
38  * used by extent splitting.
39  */
40 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
41                                         due to ENOSPC */
42 #define EXT4_EXT_MARK_UNWRIT1   0x2  /* mark first half unwritten */
43 #define EXT4_EXT_MARK_UNWRIT2   0x4  /* mark second half unwritten */
44
45 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
46 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
47
48 static __le32 ext4_extent_block_csum(struct inode *inode,
49                                      struct ext4_extent_header *eh)
50 {
51         struct ext4_inode_info *ei = EXT4_I(inode);
52         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
53         __u32 csum;
54
55         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
56                            EXT4_EXTENT_TAIL_OFFSET(eh));
57         return cpu_to_le32(csum);
58 }
59
60 static int ext4_extent_block_csum_verify(struct inode *inode,
61                                          struct ext4_extent_header *eh)
62 {
63         struct ext4_extent_tail *et;
64
65         if (!ext4_has_metadata_csum(inode->i_sb))
66                 return 1;
67
68         et = find_ext4_extent_tail(eh);
69         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
70                 return 0;
71         return 1;
72 }
73
74 static void ext4_extent_block_csum_set(struct inode *inode,
75                                        struct ext4_extent_header *eh)
76 {
77         struct ext4_extent_tail *et;
78
79         if (!ext4_has_metadata_csum(inode->i_sb))
80                 return;
81
82         et = find_ext4_extent_tail(eh);
83         et->et_checksum = ext4_extent_block_csum(inode, eh);
84 }
85
86 static int ext4_split_extent(handle_t *handle,
87                                 struct inode *inode,
88                                 struct ext4_ext_path **ppath,
89                                 struct ext4_map_blocks *map,
90                                 int split_flag,
91                                 int flags);
92
93 static int ext4_split_extent_at(handle_t *handle,
94                              struct inode *inode,
95                              struct ext4_ext_path **ppath,
96                              ext4_lblk_t split,
97                              int split_flag,
98                              int flags);
99
100 static int ext4_find_delayed_extent(struct inode *inode,
101                                     struct extent_status *newes);
102
103 static int ext4_ext_truncate_extend_restart(handle_t *handle,
104                                             struct inode *inode,
105                                             int needed)
106 {
107         int err;
108
109         if (!ext4_handle_valid(handle))
110                 return 0;
111         if (handle->h_buffer_credits >= needed)
112                 return 0;
113         /*
114          * If we need to extend the journal get a few extra blocks
115          * while we're at it for efficiency's sake.
116          */
117         needed += 3;
118         err = ext4_journal_extend(handle, needed - handle->h_buffer_credits);
119         if (err <= 0)
120                 return err;
121         err = ext4_truncate_restart_trans(handle, inode, needed);
122         if (err == 0)
123                 err = -EAGAIN;
124
125         return err;
126 }
127
128 /*
129  * could return:
130  *  - EROFS
131  *  - ENOMEM
132  */
133 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
134                                 struct ext4_ext_path *path)
135 {
136         if (path->p_bh) {
137                 /* path points to block */
138                 BUFFER_TRACE(path->p_bh, "get_write_access");
139                 return ext4_journal_get_write_access(handle, path->p_bh);
140         }
141         /* path points to leaf/index in inode body */
142         /* we use in-core data, no need to protect them */
143         return 0;
144 }
145
146 /*
147  * could return:
148  *  - EROFS
149  *  - ENOMEM
150  *  - EIO
151  */
152 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
153                      struct inode *inode, struct ext4_ext_path *path)
154 {
155         int err;
156
157         WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
158         if (path->p_bh) {
159                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
160                 /* path points to block */
161                 err = __ext4_handle_dirty_metadata(where, line, handle,
162                                                    inode, path->p_bh);
163         } else {
164                 /* path points to leaf/index in inode body */
165                 err = ext4_mark_inode_dirty(handle, inode);
166         }
167         return err;
168 }
169
170 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
171                               struct ext4_ext_path *path,
172                               ext4_lblk_t block)
173 {
174         if (path) {
175                 int depth = path->p_depth;
176                 struct ext4_extent *ex;
177
178                 /*
179                  * Try to predict block placement assuming that we are
180                  * filling in a file which will eventually be
181                  * non-sparse --- i.e., in the case of libbfd writing
182                  * an ELF object sections out-of-order but in a way
183                  * the eventually results in a contiguous object or
184                  * executable file, or some database extending a table
185                  * space file.  However, this is actually somewhat
186                  * non-ideal if we are writing a sparse file such as
187                  * qemu or KVM writing a raw image file that is going
188                  * to stay fairly sparse, since it will end up
189                  * fragmenting the file system's free space.  Maybe we
190                  * should have some hueristics or some way to allow
191                  * userspace to pass a hint to file system,
192                  * especially if the latter case turns out to be
193                  * common.
194                  */
195                 ex = path[depth].p_ext;
196                 if (ex) {
197                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
198                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
199
200                         if (block > ext_block)
201                                 return ext_pblk + (block - ext_block);
202                         else
203                                 return ext_pblk - (ext_block - block);
204                 }
205
206                 /* it looks like index is empty;
207                  * try to find starting block from index itself */
208                 if (path[depth].p_bh)
209                         return path[depth].p_bh->b_blocknr;
210         }
211
212         /* OK. use inode's group */
213         return ext4_inode_to_goal_block(inode);
214 }
215
216 /*
217  * Allocation for a meta data block
218  */
219 static ext4_fsblk_t
220 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
221                         struct ext4_ext_path *path,
222                         struct ext4_extent *ex, int *err, unsigned int flags)
223 {
224         ext4_fsblk_t goal, newblock;
225
226         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
227         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
228                                         NULL, err);
229         return newblock;
230 }
231
232 static inline int ext4_ext_space_block(struct inode *inode, int check)
233 {
234         int size;
235
236         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
237                         / sizeof(struct ext4_extent);
238 #ifdef AGGRESSIVE_TEST
239         if (!check && size > 6)
240                 size = 6;
241 #endif
242         return size;
243 }
244
245 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
246 {
247         int size;
248
249         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
250                         / sizeof(struct ext4_extent_idx);
251 #ifdef AGGRESSIVE_TEST
252         if (!check && size > 5)
253                 size = 5;
254 #endif
255         return size;
256 }
257
258 static inline int ext4_ext_space_root(struct inode *inode, int check)
259 {
260         int size;
261
262         size = sizeof(EXT4_I(inode)->i_data);
263         size -= sizeof(struct ext4_extent_header);
264         size /= sizeof(struct ext4_extent);
265 #ifdef AGGRESSIVE_TEST
266         if (!check && size > 3)
267                 size = 3;
268 #endif
269         return size;
270 }
271
272 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
273 {
274         int size;
275
276         size = sizeof(EXT4_I(inode)->i_data);
277         size -= sizeof(struct ext4_extent_header);
278         size /= sizeof(struct ext4_extent_idx);
279 #ifdef AGGRESSIVE_TEST
280         if (!check && size > 4)
281                 size = 4;
282 #endif
283         return size;
284 }
285
286 static inline int
287 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
288                            struct ext4_ext_path **ppath, ext4_lblk_t lblk,
289                            int nofail)
290 {
291         struct ext4_ext_path *path = *ppath;
292         int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
293
294         return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
295                         EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
296                         EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
297                         (nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
298 }
299
300 /*
301  * Calculate the number of metadata blocks needed
302  * to allocate @blocks
303  * Worse case is one block per extent
304  */
305 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
306 {
307         struct ext4_inode_info *ei = EXT4_I(inode);
308         int idxs;
309
310         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
311                 / sizeof(struct ext4_extent_idx));
312
313         /*
314          * If the new delayed allocation block is contiguous with the
315          * previous da block, it can share index blocks with the
316          * previous block, so we only need to allocate a new index
317          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
318          * an additional index block, and at ldxs**3 blocks, yet
319          * another index blocks.
320          */
321         if (ei->i_da_metadata_calc_len &&
322             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
323                 int num = 0;
324
325                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
326                         num++;
327                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
328                         num++;
329                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
330                         num++;
331                         ei->i_da_metadata_calc_len = 0;
332                 } else
333                         ei->i_da_metadata_calc_len++;
334                 ei->i_da_metadata_calc_last_lblock++;
335                 return num;
336         }
337
338         /*
339          * In the worst case we need a new set of index blocks at
340          * every level of the inode's extent tree.
341          */
342         ei->i_da_metadata_calc_len = 1;
343         ei->i_da_metadata_calc_last_lblock = lblock;
344         return ext_depth(inode) + 1;
345 }
346
347 static int
348 ext4_ext_max_entries(struct inode *inode, int depth)
349 {
350         int max;
351
352         if (depth == ext_depth(inode)) {
353                 if (depth == 0)
354                         max = ext4_ext_space_root(inode, 1);
355                 else
356                         max = ext4_ext_space_root_idx(inode, 1);
357         } else {
358                 if (depth == 0)
359                         max = ext4_ext_space_block(inode, 1);
360                 else
361                         max = ext4_ext_space_block_idx(inode, 1);
362         }
363
364         return max;
365 }
366
367 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
368 {
369         ext4_fsblk_t block = ext4_ext_pblock(ext);
370         int len = ext4_ext_get_actual_len(ext);
371         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
372
373         /*
374          * We allow neither:
375          *  - zero length
376          *  - overflow/wrap-around
377          */
378         if (lblock + len <= lblock)
379                 return 0;
380         return ext4_inode_block_valid(inode, block, len);
381 }
382
383 static int ext4_valid_extent_idx(struct inode *inode,
384                                 struct ext4_extent_idx *ext_idx)
385 {
386         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
387
388         return ext4_inode_block_valid(inode, block, 1);
389 }
390
391 static int ext4_valid_extent_entries(struct inode *inode,
392                                 struct ext4_extent_header *eh,
393                                 int depth)
394 {
395         unsigned short entries;
396         if (eh->eh_entries == 0)
397                 return 1;
398
399         entries = le16_to_cpu(eh->eh_entries);
400
401         if (depth == 0) {
402                 /* leaf entries */
403                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
404                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
405                 ext4_fsblk_t pblock = 0;
406                 ext4_lblk_t lblock = 0;
407                 ext4_lblk_t prev = 0;
408                 int len = 0;
409                 while (entries) {
410                         if (!ext4_valid_extent(inode, ext))
411                                 return 0;
412
413                         /* Check for overlapping extents */
414                         lblock = le32_to_cpu(ext->ee_block);
415                         len = ext4_ext_get_actual_len(ext);
416                         if ((lblock <= prev) && prev) {
417                                 pblock = ext4_ext_pblock(ext);
418                                 es->s_last_error_block = cpu_to_le64(pblock);
419                                 return 0;
420                         }
421                         ext++;
422                         entries--;
423                         prev = lblock + len - 1;
424                 }
425         } else {
426                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
427                 while (entries) {
428                         if (!ext4_valid_extent_idx(inode, ext_idx))
429                                 return 0;
430                         ext_idx++;
431                         entries--;
432                 }
433         }
434         return 1;
435 }
436
437 static int __ext4_ext_check(const char *function, unsigned int line,
438                             struct inode *inode, struct ext4_extent_header *eh,
439                             int depth, ext4_fsblk_t pblk)
440 {
441         const char *error_msg;
442         int max = 0, err = -EFSCORRUPTED;
443
444         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
445                 error_msg = "invalid magic";
446                 goto corrupted;
447         }
448         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
449                 error_msg = "unexpected eh_depth";
450                 goto corrupted;
451         }
452         if (unlikely(eh->eh_max == 0)) {
453                 error_msg = "invalid eh_max";
454                 goto corrupted;
455         }
456         max = ext4_ext_max_entries(inode, depth);
457         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
458                 error_msg = "too large eh_max";
459                 goto corrupted;
460         }
461         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
462                 error_msg = "invalid eh_entries";
463                 goto corrupted;
464         }
465         if (!ext4_valid_extent_entries(inode, eh, depth)) {
466                 error_msg = "invalid extent entries";
467                 goto corrupted;
468         }
469         if (unlikely(depth > 32)) {
470                 error_msg = "too large eh_depth";
471                 goto corrupted;
472         }
473         /* Verify checksum on non-root extent tree nodes */
474         if (ext_depth(inode) != depth &&
475             !ext4_extent_block_csum_verify(inode, eh)) {
476                 error_msg = "extent tree corrupted";
477                 err = -EFSBADCRC;
478                 goto corrupted;
479         }
480         return 0;
481
482 corrupted:
483         ext4_error_inode(inode, function, line, 0,
484                          "pblk %llu bad header/extent: %s - magic %x, "
485                          "entries %u, max %u(%u), depth %u(%u)",
486                          (unsigned long long) pblk, error_msg,
487                          le16_to_cpu(eh->eh_magic),
488                          le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
489                          max, le16_to_cpu(eh->eh_depth), depth);
490         return err;
491 }
492
493 #define ext4_ext_check(inode, eh, depth, pblk)                  \
494         __ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
495
496 int ext4_ext_check_inode(struct inode *inode)
497 {
498         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
499 }
500
501 static void ext4_cache_extents(struct inode *inode,
502                                struct ext4_extent_header *eh)
503 {
504         struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
505         ext4_lblk_t prev = 0;
506         int i;
507
508         for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
509                 unsigned int status = EXTENT_STATUS_WRITTEN;
510                 ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
511                 int len = ext4_ext_get_actual_len(ex);
512
513                 if (prev && (prev != lblk))
514                         ext4_es_cache_extent(inode, prev, lblk - prev, ~0,
515                                              EXTENT_STATUS_HOLE);
516
517                 if (ext4_ext_is_unwritten(ex))
518                         status = EXTENT_STATUS_UNWRITTEN;
519                 ext4_es_cache_extent(inode, lblk, len,
520                                      ext4_ext_pblock(ex), status);
521                 prev = lblk + len;
522         }
523 }
524
525 static struct buffer_head *
526 __read_extent_tree_block(const char *function, unsigned int line,
527                          struct inode *inode, ext4_fsblk_t pblk, int depth,
528                          int flags)
529 {
530         struct buffer_head              *bh;
531         int                             err;
532
533         bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);
534         if (unlikely(!bh))
535                 return ERR_PTR(-ENOMEM);
536
537         if (!bh_uptodate_or_lock(bh)) {
538                 trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
539                 err = bh_submit_read(bh);
540                 if (err < 0)
541                         goto errout;
542         }
543         if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
544                 return bh;
545         err = __ext4_ext_check(function, line, inode,
546                                ext_block_hdr(bh), depth, pblk);
547         if (err)
548                 goto errout;
549         set_buffer_verified(bh);
550         /*
551          * If this is a leaf block, cache all of its entries
552          */
553         if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
554                 struct ext4_extent_header *eh = ext_block_hdr(bh);
555                 ext4_cache_extents(inode, eh);
556         }
557         return bh;
558 errout:
559         put_bh(bh);
560         return ERR_PTR(err);
561
562 }
563
564 #define read_extent_tree_block(inode, pblk, depth, flags)               \
565         __read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
566                                  (depth), (flags))
567
568 /*
569  * This function is called to cache a file's extent information in the
570  * extent status tree
571  */
572 int ext4_ext_precache(struct inode *inode)
573 {
574         struct ext4_inode_info *ei = EXT4_I(inode);
575         struct ext4_ext_path *path = NULL;
576         struct buffer_head *bh;
577         int i = 0, depth, ret = 0;
578
579         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
580                 return 0;       /* not an extent-mapped inode */
581
582         down_read(&ei->i_data_sem);
583         depth = ext_depth(inode);
584
585         path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
586                        GFP_NOFS);
587         if (path == NULL) {
588                 up_read(&ei->i_data_sem);
589                 return -ENOMEM;
590         }
591
592         /* Don't cache anything if there are no external extent blocks */
593         if (depth == 0)
594                 goto out;
595         path[0].p_hdr = ext_inode_hdr(inode);
596         ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
597         if (ret)
598                 goto out;
599         path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
600         while (i >= 0) {
601                 /*
602                  * If this is a leaf block or we've reached the end of
603                  * the index block, go up
604                  */
605                 if ((i == depth) ||
606                     path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
607                         brelse(path[i].p_bh);
608                         path[i].p_bh = NULL;
609                         i--;
610                         continue;
611                 }
612                 bh = read_extent_tree_block(inode,
613                                             ext4_idx_pblock(path[i].p_idx++),
614                                             depth - i - 1,
615                                             EXT4_EX_FORCE_CACHE);
616                 if (IS_ERR(bh)) {
617                         ret = PTR_ERR(bh);
618                         break;
619                 }
620                 i++;
621                 path[i].p_bh = bh;
622                 path[i].p_hdr = ext_block_hdr(bh);
623                 path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
624         }
625         ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
626 out:
627         up_read(&ei->i_data_sem);
628         ext4_ext_drop_refs(path);
629         kfree(path);
630         return ret;
631 }
632
633 #ifdef EXT_DEBUG
634 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
635 {
636         int k, l = path->p_depth;
637
638         ext_debug("path:");
639         for (k = 0; k <= l; k++, path++) {
640                 if (path->p_idx) {
641                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
642                             ext4_idx_pblock(path->p_idx));
643                 } else if (path->p_ext) {
644                         ext_debug("  %d:[%d]%d:%llu ",
645                                   le32_to_cpu(path->p_ext->ee_block),
646                                   ext4_ext_is_unwritten(path->p_ext),
647                                   ext4_ext_get_actual_len(path->p_ext),
648                                   ext4_ext_pblock(path->p_ext));
649                 } else
650                         ext_debug("  []");
651         }
652         ext_debug("\n");
653 }
654
655 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
656 {
657         int depth = ext_depth(inode);
658         struct ext4_extent_header *eh;
659         struct ext4_extent *ex;
660         int i;
661
662         if (!path)
663                 return;
664
665         eh = path[depth].p_hdr;
666         ex = EXT_FIRST_EXTENT(eh);
667
668         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
669
670         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
671                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
672                           ext4_ext_is_unwritten(ex),
673                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
674         }
675         ext_debug("\n");
676 }
677
678 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
679                         ext4_fsblk_t newblock, int level)
680 {
681         int depth = ext_depth(inode);
682         struct ext4_extent *ex;
683
684         if (depth != level) {
685                 struct ext4_extent_idx *idx;
686                 idx = path[level].p_idx;
687                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
688                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
689                                         le32_to_cpu(idx->ei_block),
690                                         ext4_idx_pblock(idx),
691                                         newblock);
692                         idx++;
693                 }
694
695                 return;
696         }
697
698         ex = path[depth].p_ext;
699         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
700                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
701                                 le32_to_cpu(ex->ee_block),
702                                 ext4_ext_pblock(ex),
703                                 ext4_ext_is_unwritten(ex),
704                                 ext4_ext_get_actual_len(ex),
705                                 newblock);
706                 ex++;
707         }
708 }
709
710 #else
711 #define ext4_ext_show_path(inode, path)
712 #define ext4_ext_show_leaf(inode, path)
713 #define ext4_ext_show_move(inode, path, newblock, level)
714 #endif
715
716 void ext4_ext_drop_refs(struct ext4_ext_path *path)
717 {
718         int depth, i;
719
720         if (!path)
721                 return;
722         depth = path->p_depth;
723         for (i = 0; i <= depth; i++, path++)
724                 if (path->p_bh) {
725                         brelse(path->p_bh);
726                         path->p_bh = NULL;
727                 }
728 }
729
730 /*
731  * ext4_ext_binsearch_idx:
732  * binary search for the closest index of the given block
733  * the header must be checked before calling this
734  */
735 static void
736 ext4_ext_binsearch_idx(struct inode *inode,
737                         struct ext4_ext_path *path, ext4_lblk_t block)
738 {
739         struct ext4_extent_header *eh = path->p_hdr;
740         struct ext4_extent_idx *r, *l, *m;
741
742
743         ext_debug("binsearch for %u(idx):  ", block);
744
745         l = EXT_FIRST_INDEX(eh) + 1;
746         r = EXT_LAST_INDEX(eh);
747         while (l <= r) {
748                 m = l + (r - l) / 2;
749                 if (block < le32_to_cpu(m->ei_block))
750                         r = m - 1;
751                 else
752                         l = m + 1;
753                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
754                                 m, le32_to_cpu(m->ei_block),
755                                 r, le32_to_cpu(r->ei_block));
756         }
757
758         path->p_idx = l - 1;
759         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
760                   ext4_idx_pblock(path->p_idx));
761
762 #ifdef CHECK_BINSEARCH
763         {
764                 struct ext4_extent_idx *chix, *ix;
765                 int k;
766
767                 chix = ix = EXT_FIRST_INDEX(eh);
768                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
769                   if (k != 0 &&
770                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
771                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
772                                        "first=0x%p\n", k,
773                                        ix, EXT_FIRST_INDEX(eh));
774                                 printk(KERN_DEBUG "%u <= %u\n",
775                                        le32_to_cpu(ix->ei_block),
776                                        le32_to_cpu(ix[-1].ei_block));
777                         }
778                         BUG_ON(k && le32_to_cpu(ix->ei_block)
779                                            <= le32_to_cpu(ix[-1].ei_block));
780                         if (block < le32_to_cpu(ix->ei_block))
781                                 break;
782                         chix = ix;
783                 }
784                 BUG_ON(chix != path->p_idx);
785         }
786 #endif
787
788 }
789
790 /*
791  * ext4_ext_binsearch:
792  * binary search for closest extent of the given block
793  * the header must be checked before calling this
794  */
795 static void
796 ext4_ext_binsearch(struct inode *inode,
797                 struct ext4_ext_path *path, ext4_lblk_t block)
798 {
799         struct ext4_extent_header *eh = path->p_hdr;
800         struct ext4_extent *r, *l, *m;
801
802         if (eh->eh_entries == 0) {
803                 /*
804                  * this leaf is empty:
805                  * we get such a leaf in split/add case
806                  */
807                 return;
808         }
809
810         ext_debug("binsearch for %u:  ", block);
811
812         l = EXT_FIRST_EXTENT(eh) + 1;
813         r = EXT_LAST_EXTENT(eh);
814
815         while (l <= r) {
816                 m = l + (r - l) / 2;
817                 if (block < le32_to_cpu(m->ee_block))
818                         r = m - 1;
819                 else
820                         l = m + 1;
821                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
822                                 m, le32_to_cpu(m->ee_block),
823                                 r, le32_to_cpu(r->ee_block));
824         }
825
826         path->p_ext = l - 1;
827         ext_debug("  -> %d:%llu:[%d]%d ",
828                         le32_to_cpu(path->p_ext->ee_block),
829                         ext4_ext_pblock(path->p_ext),
830                         ext4_ext_is_unwritten(path->p_ext),
831                         ext4_ext_get_actual_len(path->p_ext));
832
833 #ifdef CHECK_BINSEARCH
834         {
835                 struct ext4_extent *chex, *ex;
836                 int k;
837
838                 chex = ex = EXT_FIRST_EXTENT(eh);
839                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
840                         BUG_ON(k && le32_to_cpu(ex->ee_block)
841                                           <= le32_to_cpu(ex[-1].ee_block));
842                         if (block < le32_to_cpu(ex->ee_block))
843                                 break;
844                         chex = ex;
845                 }
846                 BUG_ON(chex != path->p_ext);
847         }
848 #endif
849
850 }
851
852 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
853 {
854         struct ext4_extent_header *eh;
855
856         eh = ext_inode_hdr(inode);
857         eh->eh_depth = 0;
858         eh->eh_entries = 0;
859         eh->eh_magic = EXT4_EXT_MAGIC;
860         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
861         eh->eh_generation = 0;
862         ext4_mark_inode_dirty(handle, inode);
863         return 0;
864 }
865
866 struct ext4_ext_path *
867 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
868                  struct ext4_ext_path **orig_path, int flags)
869 {
870         struct ext4_extent_header *eh;
871         struct buffer_head *bh;
872         struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
873         short int depth, i, ppos = 0;
874         int ret;
875
876         eh = ext_inode_hdr(inode);
877         depth = ext_depth(inode);
878         if (depth < 0 || depth > EXT4_MAX_EXTENT_DEPTH) {
879                 EXT4_ERROR_INODE(inode, "inode has invalid extent depth: %d",
880                                  depth);
881                 ret = -EFSCORRUPTED;
882                 goto err;
883         }
884
885         if (path) {
886                 ext4_ext_drop_refs(path);
887                 if (depth > path[0].p_maxdepth) {
888                         kfree(path);
889                         *orig_path = path = NULL;
890                 }
891         }
892         if (!path) {
893                 /* account possible depth increase */
894                 path = kcalloc(depth + 2, sizeof(struct ext4_ext_path),
895                                 GFP_NOFS);
896                 if (unlikely(!path))
897                         return ERR_PTR(-ENOMEM);
898                 path[0].p_maxdepth = depth + 1;
899         }
900         path[0].p_hdr = eh;
901         path[0].p_bh = NULL;
902
903         i = depth;
904         if (!(flags & EXT4_EX_NOCACHE) && depth == 0)
905                 ext4_cache_extents(inode, eh);
906         /* walk through the tree */
907         while (i) {
908                 ext_debug("depth %d: num %d, max %d\n",
909                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
910
911                 ext4_ext_binsearch_idx(inode, path + ppos, block);
912                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
913                 path[ppos].p_depth = i;
914                 path[ppos].p_ext = NULL;
915
916                 bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
917                                             flags);
918                 if (IS_ERR(bh)) {
919                         ret = PTR_ERR(bh);
920                         goto err;
921                 }
922
923                 eh = ext_block_hdr(bh);
924                 ppos++;
925                 path[ppos].p_bh = bh;
926                 path[ppos].p_hdr = eh;
927         }
928
929         path[ppos].p_depth = i;
930         path[ppos].p_ext = NULL;
931         path[ppos].p_idx = NULL;
932
933         /* find extent */
934         ext4_ext_binsearch(inode, path + ppos, block);
935         /* if not an empty leaf */
936         if (path[ppos].p_ext)
937                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
938
939         ext4_ext_show_path(inode, path);
940
941         return path;
942
943 err:
944         ext4_ext_drop_refs(path);
945         kfree(path);
946         if (orig_path)
947                 *orig_path = NULL;
948         return ERR_PTR(ret);
949 }
950
951 /*
952  * ext4_ext_insert_index:
953  * insert new index [@logical;@ptr] into the block at @curp;
954  * check where to insert: before @curp or after @curp
955  */
956 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
957                                  struct ext4_ext_path *curp,
958                                  int logical, ext4_fsblk_t ptr)
959 {
960         struct ext4_extent_idx *ix;
961         int len, err;
962
963         err = ext4_ext_get_access(handle, inode, curp);
964         if (err)
965                 return err;
966
967         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
968                 EXT4_ERROR_INODE(inode,
969                                  "logical %d == ei_block %d!",
970                                  logical, le32_to_cpu(curp->p_idx->ei_block));
971                 return -EFSCORRUPTED;
972         }
973
974         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
975                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
976                 EXT4_ERROR_INODE(inode,
977                                  "eh_entries %d >= eh_max %d!",
978                                  le16_to_cpu(curp->p_hdr->eh_entries),
979                                  le16_to_cpu(curp->p_hdr->eh_max));
980                 return -EFSCORRUPTED;
981         }
982
983         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
984                 /* insert after */
985                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
986                 ix = curp->p_idx + 1;
987         } else {
988                 /* insert before */
989                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
990                 ix = curp->p_idx;
991         }
992
993         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
994         BUG_ON(len < 0);
995         if (len > 0) {
996                 ext_debug("insert new index %d: "
997                                 "move %d indices from 0x%p to 0x%p\n",
998                                 logical, len, ix, ix + 1);
999                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
1000         }
1001
1002         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
1003                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
1004                 return -EFSCORRUPTED;
1005         }
1006
1007         ix->ei_block = cpu_to_le32(logical);
1008         ext4_idx_store_pblock(ix, ptr);
1009         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1010
1011         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1012                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1013                 return -EFSCORRUPTED;
1014         }
1015
1016         err = ext4_ext_dirty(handle, inode, curp);
1017         ext4_std_error(inode->i_sb, err);
1018
1019         return err;
1020 }
1021
1022 /*
1023  * ext4_ext_split:
1024  * inserts new subtree into the path, using free index entry
1025  * at depth @at:
1026  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1027  * - makes decision where to split
1028  * - moves remaining extents and index entries (right to the split point)
1029  *   into the newly allocated blocks
1030  * - initializes subtree
1031  */
1032 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1033                           unsigned int flags,
1034                           struct ext4_ext_path *path,
1035                           struct ext4_extent *newext, int at)
1036 {
1037         struct buffer_head *bh = NULL;
1038         int depth = ext_depth(inode);
1039         struct ext4_extent_header *neh;
1040         struct ext4_extent_idx *fidx;
1041         int i = at, k, m, a;
1042         ext4_fsblk_t newblock, oldblock;
1043         __le32 border;
1044         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1045         int err = 0;
1046         size_t ext_size = 0;
1047
1048         /* make decision: where to split? */
1049         /* FIXME: now decision is simplest: at current extent */
1050
1051         /* if current leaf will be split, then we should use
1052          * border from split point */
1053         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1054                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1055                 return -EFSCORRUPTED;
1056         }
1057         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1058                 border = path[depth].p_ext[1].ee_block;
1059                 ext_debug("leaf will be split."
1060                                 " next leaf starts at %d\n",
1061                                   le32_to_cpu(border));
1062         } else {
1063                 border = newext->ee_block;
1064                 ext_debug("leaf will be added."
1065                                 " next leaf starts at %d\n",
1066                                 le32_to_cpu(border));
1067         }
1068
1069         /*
1070          * If error occurs, then we break processing
1071          * and mark filesystem read-only. index won't
1072          * be inserted and tree will be in consistent
1073          * state. Next mount will repair buffers too.
1074          */
1075
1076         /*
1077          * Get array to track all allocated blocks.
1078          * We need this to handle errors and free blocks
1079          * upon them.
1080          */
1081         ablocks = kcalloc(depth, sizeof(ext4_fsblk_t), GFP_NOFS);
1082         if (!ablocks)
1083                 return -ENOMEM;
1084
1085         /* allocate all needed blocks */
1086         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1087         for (a = 0; a < depth - at; a++) {
1088                 newblock = ext4_ext_new_meta_block(handle, inode, path,
1089                                                    newext, &err, flags);
1090                 if (newblock == 0)
1091                         goto cleanup;
1092                 ablocks[a] = newblock;
1093         }
1094
1095         /* initialize new leaf */
1096         newblock = ablocks[--a];
1097         if (unlikely(newblock == 0)) {
1098                 EXT4_ERROR_INODE(inode, "newblock == 0!");
1099                 err = -EFSCORRUPTED;
1100                 goto cleanup;
1101         }
1102         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1103         if (unlikely(!bh)) {
1104                 err = -ENOMEM;
1105                 goto cleanup;
1106         }
1107         lock_buffer(bh);
1108
1109         err = ext4_journal_get_create_access(handle, bh);
1110         if (err)
1111                 goto cleanup;
1112
1113         neh = ext_block_hdr(bh);
1114         neh->eh_entries = 0;
1115         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1116         neh->eh_magic = EXT4_EXT_MAGIC;
1117         neh->eh_depth = 0;
1118         neh->eh_generation = 0;
1119
1120         /* move remainder of path[depth] to the new leaf */
1121         if (unlikely(path[depth].p_hdr->eh_entries !=
1122                      path[depth].p_hdr->eh_max)) {
1123                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1124                                  path[depth].p_hdr->eh_entries,
1125                                  path[depth].p_hdr->eh_max);
1126                 err = -EFSCORRUPTED;
1127                 goto cleanup;
1128         }
1129         /* start copy from next extent */
1130         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1131         ext4_ext_show_move(inode, path, newblock, depth);
1132         if (m) {
1133                 struct ext4_extent *ex;
1134                 ex = EXT_FIRST_EXTENT(neh);
1135                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1136                 le16_add_cpu(&neh->eh_entries, m);
1137         }
1138
1139         /* zero out unused area in the extent block */
1140         ext_size = sizeof(struct ext4_extent_header) +
1141                 sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries);
1142         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1143         ext4_extent_block_csum_set(inode, neh);
1144         set_buffer_uptodate(bh);
1145         unlock_buffer(bh);
1146
1147         err = ext4_handle_dirty_metadata(handle, inode, bh);
1148         if (err)
1149                 goto cleanup;
1150         brelse(bh);
1151         bh = NULL;
1152
1153         /* correct old leaf */
1154         if (m) {
1155                 err = ext4_ext_get_access(handle, inode, path + depth);
1156                 if (err)
1157                         goto cleanup;
1158                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1159                 err = ext4_ext_dirty(handle, inode, path + depth);
1160                 if (err)
1161                         goto cleanup;
1162
1163         }
1164
1165         /* create intermediate indexes */
1166         k = depth - at - 1;
1167         if (unlikely(k < 0)) {
1168                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1169                 err = -EFSCORRUPTED;
1170                 goto cleanup;
1171         }
1172         if (k)
1173                 ext_debug("create %d intermediate indices\n", k);
1174         /* insert new index into current index block */
1175         /* current depth stored in i var */
1176         i = depth - 1;
1177         while (k--) {
1178                 oldblock = newblock;
1179                 newblock = ablocks[--a];
1180                 bh = sb_getblk(inode->i_sb, newblock);
1181                 if (unlikely(!bh)) {
1182                         err = -ENOMEM;
1183                         goto cleanup;
1184                 }
1185                 lock_buffer(bh);
1186
1187                 err = ext4_journal_get_create_access(handle, bh);
1188                 if (err)
1189                         goto cleanup;
1190
1191                 neh = ext_block_hdr(bh);
1192                 neh->eh_entries = cpu_to_le16(1);
1193                 neh->eh_magic = EXT4_EXT_MAGIC;
1194                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1195                 neh->eh_depth = cpu_to_le16(depth - i);
1196                 neh->eh_generation = 0;
1197                 fidx = EXT_FIRST_INDEX(neh);
1198                 fidx->ei_block = border;
1199                 ext4_idx_store_pblock(fidx, oldblock);
1200
1201                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1202                                 i, newblock, le32_to_cpu(border), oldblock);
1203
1204                 /* move remainder of path[i] to the new index block */
1205                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1206                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1207                         EXT4_ERROR_INODE(inode,
1208                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1209                                          le32_to_cpu(path[i].p_ext->ee_block));
1210                         err = -EFSCORRUPTED;
1211                         goto cleanup;
1212                 }
1213                 /* start copy indexes */
1214                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1215                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1216                                 EXT_MAX_INDEX(path[i].p_hdr));
1217                 ext4_ext_show_move(inode, path, newblock, i);
1218                 if (m) {
1219                         memmove(++fidx, path[i].p_idx,
1220                                 sizeof(struct ext4_extent_idx) * m);
1221                         le16_add_cpu(&neh->eh_entries, m);
1222                 }
1223                 /* zero out unused area in the extent block */
1224                 ext_size = sizeof(struct ext4_extent_header) +
1225                    (sizeof(struct ext4_extent) * le16_to_cpu(neh->eh_entries));
1226                 memset(bh->b_data + ext_size, 0,
1227                         inode->i_sb->s_blocksize - ext_size);
1228                 ext4_extent_block_csum_set(inode, neh);
1229                 set_buffer_uptodate(bh);
1230                 unlock_buffer(bh);
1231
1232                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1233                 if (err)
1234                         goto cleanup;
1235                 brelse(bh);
1236                 bh = NULL;
1237
1238                 /* correct old index */
1239                 if (m) {
1240                         err = ext4_ext_get_access(handle, inode, path + i);
1241                         if (err)
1242                                 goto cleanup;
1243                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1244                         err = ext4_ext_dirty(handle, inode, path + i);
1245                         if (err)
1246                                 goto cleanup;
1247                 }
1248
1249                 i--;
1250         }
1251
1252         /* insert new index */
1253         err = ext4_ext_insert_index(handle, inode, path + at,
1254                                     le32_to_cpu(border), newblock);
1255
1256 cleanup:
1257         if (bh) {
1258                 if (buffer_locked(bh))
1259                         unlock_buffer(bh);
1260                 brelse(bh);
1261         }
1262
1263         if (err) {
1264                 /* free all allocated blocks in error case */
1265                 for (i = 0; i < depth; i++) {
1266                         if (!ablocks[i])
1267                                 continue;
1268                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1269                                          EXT4_FREE_BLOCKS_METADATA);
1270                 }
1271         }
1272         kfree(ablocks);
1273
1274         return err;
1275 }
1276
1277 /*
1278  * ext4_ext_grow_indepth:
1279  * implements tree growing procedure:
1280  * - allocates new block
1281  * - moves top-level data (index block or leaf) into the new block
1282  * - initializes new top-level, creating index that points to the
1283  *   just created block
1284  */
1285 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1286                                  unsigned int flags)
1287 {
1288         struct ext4_extent_header *neh;
1289         struct buffer_head *bh;
1290         ext4_fsblk_t newblock, goal = 0;
1291         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1292         int err = 0;
1293         size_t ext_size = 0;
1294
1295         /* Try to prepend new index to old one */
1296         if (ext_depth(inode))
1297                 goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1298         if (goal > le32_to_cpu(es->s_first_data_block)) {
1299                 flags |= EXT4_MB_HINT_TRY_GOAL;
1300                 goal--;
1301         } else
1302                 goal = ext4_inode_to_goal_block(inode);
1303         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1304                                         NULL, &err);
1305         if (newblock == 0)
1306                 return err;
1307
1308         bh = sb_getblk_gfp(inode->i_sb, newblock, __GFP_MOVABLE | GFP_NOFS);
1309         if (unlikely(!bh))
1310                 return -ENOMEM;
1311         lock_buffer(bh);
1312
1313         err = ext4_journal_get_create_access(handle, bh);
1314         if (err) {
1315                 unlock_buffer(bh);
1316                 goto out;
1317         }
1318
1319         ext_size = sizeof(EXT4_I(inode)->i_data);
1320         /* move top-level index/leaf into new block */
1321         memmove(bh->b_data, EXT4_I(inode)->i_data, ext_size);
1322         /* zero out unused area in the extent block */
1323         memset(bh->b_data + ext_size, 0, inode->i_sb->s_blocksize - ext_size);
1324
1325         /* set size of new block */
1326         neh = ext_block_hdr(bh);
1327         /* old root could have indexes or leaves
1328          * so calculate e_max right way */
1329         if (ext_depth(inode))
1330                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1331         else
1332                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1333         neh->eh_magic = EXT4_EXT_MAGIC;
1334         ext4_extent_block_csum_set(inode, neh);
1335         set_buffer_uptodate(bh);
1336         unlock_buffer(bh);
1337
1338         err = ext4_handle_dirty_metadata(handle, inode, bh);
1339         if (err)
1340                 goto out;
1341
1342         /* Update top-level index: num,max,pointer */
1343         neh = ext_inode_hdr(inode);
1344         neh->eh_entries = cpu_to_le16(1);
1345         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1346         if (neh->eh_depth == 0) {
1347                 /* Root extent block becomes index block */
1348                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1349                 EXT_FIRST_INDEX(neh)->ei_block =
1350                         EXT_FIRST_EXTENT(neh)->ee_block;
1351         }
1352         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1353                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1354                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1355                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1356
1357         le16_add_cpu(&neh->eh_depth, 1);
1358         ext4_mark_inode_dirty(handle, inode);
1359 out:
1360         brelse(bh);
1361
1362         return err;
1363 }
1364
1365 /*
1366  * ext4_ext_create_new_leaf:
1367  * finds empty index and adds new leaf.
1368  * if no free index is found, then it requests in-depth growing.
1369  */
1370 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1371                                     unsigned int mb_flags,
1372                                     unsigned int gb_flags,
1373                                     struct ext4_ext_path **ppath,
1374                                     struct ext4_extent *newext)
1375 {
1376         struct ext4_ext_path *path = *ppath;
1377         struct ext4_ext_path *curp;
1378         int depth, i, err = 0;
1379
1380 repeat:
1381         i = depth = ext_depth(inode);
1382
1383         /* walk up to the tree and look for free index entry */
1384         curp = path + depth;
1385         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1386                 i--;
1387                 curp--;
1388         }
1389
1390         /* we use already allocated block for index block,
1391          * so subsequent data blocks should be contiguous */
1392         if (EXT_HAS_FREE_INDEX(curp)) {
1393                 /* if we found index with free entry, then use that
1394                  * entry: create all needed subtree and add new leaf */
1395                 err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1396                 if (err)
1397                         goto out;
1398
1399                 /* refill path */
1400                 path = ext4_find_extent(inode,
1401                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1402                                     ppath, gb_flags);
1403                 if (IS_ERR(path))
1404                         err = PTR_ERR(path);
1405         } else {
1406                 /* tree is full, time to grow in depth */
1407                 err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1408                 if (err)
1409                         goto out;
1410
1411                 /* refill path */
1412                 path = ext4_find_extent(inode,
1413                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1414                                     ppath, gb_flags);
1415                 if (IS_ERR(path)) {
1416                         err = PTR_ERR(path);
1417                         goto out;
1418                 }
1419
1420                 /*
1421                  * only first (depth 0 -> 1) produces free space;
1422                  * in all other cases we have to split the grown tree
1423                  */
1424                 depth = ext_depth(inode);
1425                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1426                         /* now we need to split */
1427                         goto repeat;
1428                 }
1429         }
1430
1431 out:
1432         return err;
1433 }
1434
1435 /*
1436  * search the closest allocated block to the left for *logical
1437  * and returns it at @logical + it's physical address at @phys
1438  * if *logical is the smallest allocated block, the function
1439  * returns 0 at @phys
1440  * return value contains 0 (success) or error code
1441  */
1442 static int ext4_ext_search_left(struct inode *inode,
1443                                 struct ext4_ext_path *path,
1444                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1445 {
1446         struct ext4_extent_idx *ix;
1447         struct ext4_extent *ex;
1448         int depth, ee_len;
1449
1450         if (unlikely(path == NULL)) {
1451                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1452                 return -EFSCORRUPTED;
1453         }
1454         depth = path->p_depth;
1455         *phys = 0;
1456
1457         if (depth == 0 && path->p_ext == NULL)
1458                 return 0;
1459
1460         /* usually extent in the path covers blocks smaller
1461          * then *logical, but it can be that extent is the
1462          * first one in the file */
1463
1464         ex = path[depth].p_ext;
1465         ee_len = ext4_ext_get_actual_len(ex);
1466         if (*logical < le32_to_cpu(ex->ee_block)) {
1467                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1468                         EXT4_ERROR_INODE(inode,
1469                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1470                                          *logical, le32_to_cpu(ex->ee_block));
1471                         return -EFSCORRUPTED;
1472                 }
1473                 while (--depth >= 0) {
1474                         ix = path[depth].p_idx;
1475                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1476                                 EXT4_ERROR_INODE(inode,
1477                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1478                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1479                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1480                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1481                                   depth);
1482                                 return -EFSCORRUPTED;
1483                         }
1484                 }
1485                 return 0;
1486         }
1487
1488         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1489                 EXT4_ERROR_INODE(inode,
1490                                  "logical %d < ee_block %d + ee_len %d!",
1491                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1492                 return -EFSCORRUPTED;
1493         }
1494
1495         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1496         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1497         return 0;
1498 }
1499
1500 /*
1501  * search the closest allocated block to the right for *logical
1502  * and returns it at @logical + it's physical address at @phys
1503  * if *logical is the largest allocated block, the function
1504  * returns 0 at @phys
1505  * return value contains 0 (success) or error code
1506  */
1507 static int ext4_ext_search_right(struct inode *inode,
1508                                  struct ext4_ext_path *path,
1509                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1510                                  struct ext4_extent **ret_ex)
1511 {
1512         struct buffer_head *bh = NULL;
1513         struct ext4_extent_header *eh;
1514         struct ext4_extent_idx *ix;
1515         struct ext4_extent *ex;
1516         ext4_fsblk_t block;
1517         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1518         int ee_len;
1519
1520         if (unlikely(path == NULL)) {
1521                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1522                 return -EFSCORRUPTED;
1523         }
1524         depth = path->p_depth;
1525         *phys = 0;
1526
1527         if (depth == 0 && path->p_ext == NULL)
1528                 return 0;
1529
1530         /* usually extent in the path covers blocks smaller
1531          * then *logical, but it can be that extent is the
1532          * first one in the file */
1533
1534         ex = path[depth].p_ext;
1535         ee_len = ext4_ext_get_actual_len(ex);
1536         if (*logical < le32_to_cpu(ex->ee_block)) {
1537                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1538                         EXT4_ERROR_INODE(inode,
1539                                          "first_extent(path[%d].p_hdr) != ex",
1540                                          depth);
1541                         return -EFSCORRUPTED;
1542                 }
1543                 while (--depth >= 0) {
1544                         ix = path[depth].p_idx;
1545                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1546                                 EXT4_ERROR_INODE(inode,
1547                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1548                                                  *logical);
1549                                 return -EFSCORRUPTED;
1550                         }
1551                 }
1552                 goto found_extent;
1553         }
1554
1555         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1556                 EXT4_ERROR_INODE(inode,
1557                                  "logical %d < ee_block %d + ee_len %d!",
1558                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1559                 return -EFSCORRUPTED;
1560         }
1561
1562         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1563                 /* next allocated block in this leaf */
1564                 ex++;
1565                 goto found_extent;
1566         }
1567
1568         /* go up and search for index to the right */
1569         while (--depth >= 0) {
1570                 ix = path[depth].p_idx;
1571                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1572                         goto got_index;
1573         }
1574
1575         /* we've gone up to the root and found no index to the right */
1576         return 0;
1577
1578 got_index:
1579         /* we've found index to the right, let's
1580          * follow it and find the closest allocated
1581          * block to the right */
1582         ix++;
1583         block = ext4_idx_pblock(ix);
1584         while (++depth < path->p_depth) {
1585                 /* subtract from p_depth to get proper eh_depth */
1586                 bh = read_extent_tree_block(inode, block,
1587                                             path->p_depth - depth, 0);
1588                 if (IS_ERR(bh))
1589                         return PTR_ERR(bh);
1590                 eh = ext_block_hdr(bh);
1591                 ix = EXT_FIRST_INDEX(eh);
1592                 block = ext4_idx_pblock(ix);
1593                 put_bh(bh);
1594         }
1595
1596         bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1597         if (IS_ERR(bh))
1598                 return PTR_ERR(bh);
1599         eh = ext_block_hdr(bh);
1600         ex = EXT_FIRST_EXTENT(eh);
1601 found_extent:
1602         *logical = le32_to_cpu(ex->ee_block);
1603         *phys = ext4_ext_pblock(ex);
1604         *ret_ex = ex;
1605         if (bh)
1606                 put_bh(bh);
1607         return 0;
1608 }
1609
1610 /*
1611  * ext4_ext_next_allocated_block:
1612  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1613  * NOTE: it considers block number from index entry as
1614  * allocated block. Thus, index entries have to be consistent
1615  * with leaves.
1616  */
1617 ext4_lblk_t
1618 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1619 {
1620         int depth;
1621
1622         BUG_ON(path == NULL);
1623         depth = path->p_depth;
1624
1625         if (depth == 0 && path->p_ext == NULL)
1626                 return EXT_MAX_BLOCKS;
1627
1628         while (depth >= 0) {
1629                 if (depth == path->p_depth) {
1630                         /* leaf */
1631                         if (path[depth].p_ext &&
1632                                 path[depth].p_ext !=
1633                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1634                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1635                 } else {
1636                         /* index */
1637                         if (path[depth].p_idx !=
1638                                         EXT_LAST_INDEX(path[depth].p_hdr))
1639                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1640                 }
1641                 depth--;
1642         }
1643
1644         return EXT_MAX_BLOCKS;
1645 }
1646
1647 /*
1648  * ext4_ext_next_leaf_block:
1649  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1650  */
1651 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1652 {
1653         int depth;
1654
1655         BUG_ON(path == NULL);
1656         depth = path->p_depth;
1657
1658         /* zero-tree has no leaf blocks at all */
1659         if (depth == 0)
1660                 return EXT_MAX_BLOCKS;
1661
1662         /* go to index block */
1663         depth--;
1664
1665         while (depth >= 0) {
1666                 if (path[depth].p_idx !=
1667                                 EXT_LAST_INDEX(path[depth].p_hdr))
1668                         return (ext4_lblk_t)
1669                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1670                 depth--;
1671         }
1672
1673         return EXT_MAX_BLOCKS;
1674 }
1675
1676 /*
1677  * ext4_ext_correct_indexes:
1678  * if leaf gets modified and modified extent is first in the leaf,
1679  * then we have to correct all indexes above.
1680  * TODO: do we need to correct tree in all cases?
1681  */
1682 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1683                                 struct ext4_ext_path *path)
1684 {
1685         struct ext4_extent_header *eh;
1686         int depth = ext_depth(inode);
1687         struct ext4_extent *ex;
1688         __le32 border;
1689         int k, err = 0;
1690
1691         eh = path[depth].p_hdr;
1692         ex = path[depth].p_ext;
1693
1694         if (unlikely(ex == NULL || eh == NULL)) {
1695                 EXT4_ERROR_INODE(inode,
1696                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1697                 return -EFSCORRUPTED;
1698         }
1699
1700         if (depth == 0) {
1701                 /* there is no tree at all */
1702                 return 0;
1703         }
1704
1705         if (ex != EXT_FIRST_EXTENT(eh)) {
1706                 /* we correct tree if first leaf got modified only */
1707                 return 0;
1708         }
1709
1710         /*
1711          * TODO: we need correction if border is smaller than current one
1712          */
1713         k = depth - 1;
1714         border = path[depth].p_ext->ee_block;
1715         err = ext4_ext_get_access(handle, inode, path + k);
1716         if (err)
1717                 return err;
1718         path[k].p_idx->ei_block = border;
1719         err = ext4_ext_dirty(handle, inode, path + k);
1720         if (err)
1721                 return err;
1722
1723         while (k--) {
1724                 /* change all left-side indexes */
1725                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1726                         break;
1727                 err = ext4_ext_get_access(handle, inode, path + k);
1728                 if (err)
1729                         break;
1730                 path[k].p_idx->ei_block = border;
1731                 err = ext4_ext_dirty(handle, inode, path + k);
1732                 if (err)
1733                         break;
1734         }
1735
1736         return err;
1737 }
1738
1739 int
1740 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1741                                 struct ext4_extent *ex2)
1742 {
1743         unsigned short ext1_ee_len, ext2_ee_len;
1744
1745         if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1746                 return 0;
1747
1748         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1749         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1750
1751         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1752                         le32_to_cpu(ex2->ee_block))
1753                 return 0;
1754
1755         /*
1756          * To allow future support for preallocated extents to be added
1757          * as an RO_COMPAT feature, refuse to merge to extents if
1758          * this can result in the top bit of ee_len being set.
1759          */
1760         if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1761                 return 0;
1762         /*
1763          * The check for IO to unwritten extent is somewhat racy as we
1764          * increment i_unwritten / set EXT4_STATE_DIO_UNWRITTEN only after
1765          * dropping i_data_sem. But reserved blocks should save us in that
1766          * case.
1767          */
1768         if (ext4_ext_is_unwritten(ex1) &&
1769             (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1770              atomic_read(&EXT4_I(inode)->i_unwritten) ||
1771              (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1772                 return 0;
1773 #ifdef AGGRESSIVE_TEST
1774         if (ext1_ee_len >= 4)
1775                 return 0;
1776 #endif
1777
1778         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1779                 return 1;
1780         return 0;
1781 }
1782
1783 /*
1784  * This function tries to merge the "ex" extent to the next extent in the tree.
1785  * It always tries to merge towards right. If you want to merge towards
1786  * left, pass "ex - 1" as argument instead of "ex".
1787  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1788  * 1 if they got merged.
1789  */
1790 static int ext4_ext_try_to_merge_right(struct inode *inode,
1791                                  struct ext4_ext_path *path,
1792                                  struct ext4_extent *ex)
1793 {
1794         struct ext4_extent_header *eh;
1795         unsigned int depth, len;
1796         int merge_done = 0, unwritten;
1797
1798         depth = ext_depth(inode);
1799         BUG_ON(path[depth].p_hdr == NULL);
1800         eh = path[depth].p_hdr;
1801
1802         while (ex < EXT_LAST_EXTENT(eh)) {
1803                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1804                         break;
1805                 /* merge with next extent! */
1806                 unwritten = ext4_ext_is_unwritten(ex);
1807                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1808                                 + ext4_ext_get_actual_len(ex + 1));
1809                 if (unwritten)
1810                         ext4_ext_mark_unwritten(ex);
1811
1812                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1813                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1814                                 * sizeof(struct ext4_extent);
1815                         memmove(ex + 1, ex + 2, len);
1816                 }
1817                 le16_add_cpu(&eh->eh_entries, -1);
1818                 merge_done = 1;
1819                 WARN_ON(eh->eh_entries == 0);
1820                 if (!eh->eh_entries)
1821                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1822         }
1823
1824         return merge_done;
1825 }
1826
1827 /*
1828  * This function does a very simple check to see if we can collapse
1829  * an extent tree with a single extent tree leaf block into the inode.
1830  */
1831 static void ext4_ext_try_to_merge_up(handle_t *handle,
1832                                      struct inode *inode,
1833                                      struct ext4_ext_path *path)
1834 {
1835         size_t s;
1836         unsigned max_root = ext4_ext_space_root(inode, 0);
1837         ext4_fsblk_t blk;
1838
1839         if ((path[0].p_depth != 1) ||
1840             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1841             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1842                 return;
1843
1844         /*
1845          * We need to modify the block allocation bitmap and the block
1846          * group descriptor to release the extent tree block.  If we
1847          * can't get the journal credits, give up.
1848          */
1849         if (ext4_journal_extend(handle, 2))
1850                 return;
1851
1852         /*
1853          * Copy the extent data up to the inode
1854          */
1855         blk = ext4_idx_pblock(path[0].p_idx);
1856         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1857                 sizeof(struct ext4_extent_idx);
1858         s += sizeof(struct ext4_extent_header);
1859
1860         path[1].p_maxdepth = path[0].p_maxdepth;
1861         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1862         path[0].p_depth = 0;
1863         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1864                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1865         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1866
1867         brelse(path[1].p_bh);
1868         ext4_free_blocks(handle, inode, NULL, blk, 1,
1869                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1870 }
1871
1872 /*
1873  * This function tries to merge the @ex extent to neighbours in the tree.
1874  * return 1 if merge left else 0.
1875  */
1876 static void ext4_ext_try_to_merge(handle_t *handle,
1877                                   struct inode *inode,
1878                                   struct ext4_ext_path *path,
1879                                   struct ext4_extent *ex) {
1880         struct ext4_extent_header *eh;
1881         unsigned int depth;
1882         int merge_done = 0;
1883
1884         depth = ext_depth(inode);
1885         BUG_ON(path[depth].p_hdr == NULL);
1886         eh = path[depth].p_hdr;
1887
1888         if (ex > EXT_FIRST_EXTENT(eh))
1889                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1890
1891         if (!merge_done)
1892                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1893
1894         ext4_ext_try_to_merge_up(handle, inode, path);
1895 }
1896
1897 /*
1898  * check if a portion of the "newext" extent overlaps with an
1899  * existing extent.
1900  *
1901  * If there is an overlap discovered, it updates the length of the newext
1902  * such that there will be no overlap, and then returns 1.
1903  * If there is no overlap found, it returns 0.
1904  */
1905 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1906                                            struct inode *inode,
1907                                            struct ext4_extent *newext,
1908                                            struct ext4_ext_path *path)
1909 {
1910         ext4_lblk_t b1, b2;
1911         unsigned int depth, len1;
1912         unsigned int ret = 0;
1913
1914         b1 = le32_to_cpu(newext->ee_block);
1915         len1 = ext4_ext_get_actual_len(newext);
1916         depth = ext_depth(inode);
1917         if (!path[depth].p_ext)
1918                 goto out;
1919         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1920
1921         /*
1922          * get the next allocated block if the extent in the path
1923          * is before the requested block(s)
1924          */
1925         if (b2 < b1) {
1926                 b2 = ext4_ext_next_allocated_block(path);
1927                 if (b2 == EXT_MAX_BLOCKS)
1928                         goto out;
1929                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1930         }
1931
1932         /* check for wrap through zero on extent logical start block*/
1933         if (b1 + len1 < b1) {
1934                 len1 = EXT_MAX_BLOCKS - b1;
1935                 newext->ee_len = cpu_to_le16(len1);
1936                 ret = 1;
1937         }
1938
1939         /* check for overlap */
1940         if (b1 + len1 > b2) {
1941                 newext->ee_len = cpu_to_le16(b2 - b1);
1942                 ret = 1;
1943         }
1944 out:
1945         return ret;
1946 }
1947
1948 /*
1949  * ext4_ext_insert_extent:
1950  * tries to merge requsted extent into the existing extent or
1951  * inserts requested extent as new one into the tree,
1952  * creating new leaf in the no-space case.
1953  */
1954 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1955                                 struct ext4_ext_path **ppath,
1956                                 struct ext4_extent *newext, int gb_flags)
1957 {
1958         struct ext4_ext_path *path = *ppath;
1959         struct ext4_extent_header *eh;
1960         struct ext4_extent *ex, *fex;
1961         struct ext4_extent *nearex; /* nearest extent */
1962         struct ext4_ext_path *npath = NULL;
1963         int depth, len, err;
1964         ext4_lblk_t next;
1965         int mb_flags = 0, unwritten;
1966
1967         if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1968                 mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1969         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1970                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1971                 return -EFSCORRUPTED;
1972         }
1973         depth = ext_depth(inode);
1974         ex = path[depth].p_ext;
1975         eh = path[depth].p_hdr;
1976         if (unlikely(path[depth].p_hdr == NULL)) {
1977                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1978                 return -EFSCORRUPTED;
1979         }
1980
1981         /* try to insert block into found extent and return */
1982         if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1983
1984                 /*
1985                  * Try to see whether we should rather test the extent on
1986                  * right from ex, or from the left of ex. This is because
1987                  * ext4_find_extent() can return either extent on the
1988                  * left, or on the right from the searched position. This
1989                  * will make merging more effective.
1990                  */
1991                 if (ex < EXT_LAST_EXTENT(eh) &&
1992                     (le32_to_cpu(ex->ee_block) +
1993                     ext4_ext_get_actual_len(ex) <
1994                     le32_to_cpu(newext->ee_block))) {
1995                         ex += 1;
1996                         goto prepend;
1997                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1998                            (le32_to_cpu(newext->ee_block) +
1999                            ext4_ext_get_actual_len(newext) <
2000                            le32_to_cpu(ex->ee_block)))
2001                         ex -= 1;
2002
2003                 /* Try to append newex to the ex */
2004                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
2005                         ext_debug("append [%d]%d block to %u:[%d]%d"
2006                                   "(from %llu)\n",
2007                                   ext4_ext_is_unwritten(newext),
2008                                   ext4_ext_get_actual_len(newext),
2009                                   le32_to_cpu(ex->ee_block),
2010                                   ext4_ext_is_unwritten(ex),
2011                                   ext4_ext_get_actual_len(ex),
2012                                   ext4_ext_pblock(ex));
2013                         err = ext4_ext_get_access(handle, inode,
2014                                                   path + depth);
2015                         if (err)
2016                                 return err;
2017                         unwritten = ext4_ext_is_unwritten(ex);
2018                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2019                                         + ext4_ext_get_actual_len(newext));
2020                         if (unwritten)
2021                                 ext4_ext_mark_unwritten(ex);
2022                         eh = path[depth].p_hdr;
2023                         nearex = ex;
2024                         goto merge;
2025                 }
2026
2027 prepend:
2028                 /* Try to prepend newex to the ex */
2029                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
2030                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2031                                   "(from %llu)\n",
2032                                   le32_to_cpu(newext->ee_block),
2033                                   ext4_ext_is_unwritten(newext),
2034                                   ext4_ext_get_actual_len(newext),
2035                                   le32_to_cpu(ex->ee_block),
2036                                   ext4_ext_is_unwritten(ex),
2037                                   ext4_ext_get_actual_len(ex),
2038                                   ext4_ext_pblock(ex));
2039                         err = ext4_ext_get_access(handle, inode,
2040                                                   path + depth);
2041                         if (err)
2042                                 return err;
2043
2044                         unwritten = ext4_ext_is_unwritten(ex);
2045                         ex->ee_block = newext->ee_block;
2046                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2047                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2048                                         + ext4_ext_get_actual_len(newext));
2049                         if (unwritten)
2050                                 ext4_ext_mark_unwritten(ex);
2051                         eh = path[depth].p_hdr;
2052                         nearex = ex;
2053                         goto merge;
2054                 }
2055         }
2056
2057         depth = ext_depth(inode);
2058         eh = path[depth].p_hdr;
2059         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2060                 goto has_space;
2061
2062         /* probably next leaf has space for us? */
2063         fex = EXT_LAST_EXTENT(eh);
2064         next = EXT_MAX_BLOCKS;
2065         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2066                 next = ext4_ext_next_leaf_block(path);
2067         if (next != EXT_MAX_BLOCKS) {
2068                 ext_debug("next leaf block - %u\n", next);
2069                 BUG_ON(npath != NULL);
2070                 npath = ext4_find_extent(inode, next, NULL, 0);
2071                 if (IS_ERR(npath))
2072                         return PTR_ERR(npath);
2073                 BUG_ON(npath->p_depth != path->p_depth);
2074                 eh = npath[depth].p_hdr;
2075                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2076                         ext_debug("next leaf isn't full(%d)\n",
2077                                   le16_to_cpu(eh->eh_entries));
2078                         path = npath;
2079                         goto has_space;
2080                 }
2081                 ext_debug("next leaf has no free space(%d,%d)\n",
2082                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2083         }
2084
2085         /*
2086          * There is no free space in the found leaf.
2087          * We're gonna add a new leaf in the tree.
2088          */
2089         if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2090                 mb_flags |= EXT4_MB_USE_RESERVED;
2091         err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2092                                        ppath, newext);
2093         if (err)
2094                 goto cleanup;
2095         depth = ext_depth(inode);
2096         eh = path[depth].p_hdr;
2097
2098 has_space:
2099         nearex = path[depth].p_ext;
2100
2101         err = ext4_ext_get_access(handle, inode, path + depth);
2102         if (err)
2103                 goto cleanup;
2104
2105         if (!nearex) {
2106                 /* there is no extent in this leaf, create first one */
2107                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2108                                 le32_to_cpu(newext->ee_block),
2109                                 ext4_ext_pblock(newext),
2110                                 ext4_ext_is_unwritten(newext),
2111                                 ext4_ext_get_actual_len(newext));
2112                 nearex = EXT_FIRST_EXTENT(eh);
2113         } else {
2114                 if (le32_to_cpu(newext->ee_block)
2115                            > le32_to_cpu(nearex->ee_block)) {
2116                         /* Insert after */
2117                         ext_debug("insert %u:%llu:[%d]%d before: "
2118                                         "nearest %p\n",
2119                                         le32_to_cpu(newext->ee_block),
2120                                         ext4_ext_pblock(newext),
2121                                         ext4_ext_is_unwritten(newext),
2122                                         ext4_ext_get_actual_len(newext),
2123                                         nearex);
2124                         nearex++;
2125                 } else {
2126                         /* Insert before */
2127                         BUG_ON(newext->ee_block == nearex->ee_block);
2128                         ext_debug("insert %u:%llu:[%d]%d after: "
2129                                         "nearest %p\n",
2130                                         le32_to_cpu(newext->ee_block),
2131                                         ext4_ext_pblock(newext),
2132                                         ext4_ext_is_unwritten(newext),
2133                                         ext4_ext_get_actual_len(newext),
2134                                         nearex);
2135                 }
2136                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2137                 if (len > 0) {
2138                         ext_debug("insert %u:%llu:[%d]%d: "
2139                                         "move %d extents from 0x%p to 0x%p\n",
2140                                         le32_to_cpu(newext->ee_block),
2141                                         ext4_ext_pblock(newext),
2142                                         ext4_ext_is_unwritten(newext),
2143                                         ext4_ext_get_actual_len(newext),
2144                                         len, nearex, nearex + 1);
2145                         memmove(nearex + 1, nearex,
2146                                 len * sizeof(struct ext4_extent));
2147                 }
2148         }
2149
2150         le16_add_cpu(&eh->eh_entries, 1);
2151         path[depth].p_ext = nearex;
2152         nearex->ee_block = newext->ee_block;
2153         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2154         nearex->ee_len = newext->ee_len;
2155
2156 merge:
2157         /* try to merge extents */
2158         if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2159                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2160
2161
2162         /* time to correct all indexes above */
2163         err = ext4_ext_correct_indexes(handle, inode, path);
2164         if (err)
2165                 goto cleanup;
2166
2167         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2168
2169 cleanup:
2170         ext4_ext_drop_refs(npath);
2171         kfree(npath);
2172         return err;
2173 }
2174
2175 static int ext4_fill_fiemap_extents(struct inode *inode,
2176                                     ext4_lblk_t block, ext4_lblk_t num,
2177                                     struct fiemap_extent_info *fieinfo)
2178 {
2179         struct ext4_ext_path *path = NULL;
2180         struct ext4_extent *ex;
2181         struct extent_status es;
2182         ext4_lblk_t next, next_del, start = 0, end = 0;
2183         ext4_lblk_t last = block + num;
2184         int exists, depth = 0, err = 0;
2185         unsigned int flags = 0;
2186         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2187
2188         while (block < last && block != EXT_MAX_BLOCKS) {
2189                 num = last - block;
2190                 /* find extent for this block */
2191                 down_read(&EXT4_I(inode)->i_data_sem);
2192
2193                 path = ext4_find_extent(inode, block, &path, 0);
2194                 if (IS_ERR(path)) {
2195                         up_read(&EXT4_I(inode)->i_data_sem);
2196                         err = PTR_ERR(path);
2197                         path = NULL;
2198                         break;
2199                 }
2200
2201                 depth = ext_depth(inode);
2202                 if (unlikely(path[depth].p_hdr == NULL)) {
2203                         up_read(&EXT4_I(inode)->i_data_sem);
2204                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2205                         err = -EFSCORRUPTED;
2206                         break;
2207                 }
2208                 ex = path[depth].p_ext;
2209                 next = ext4_ext_next_allocated_block(path);
2210
2211                 flags = 0;
2212                 exists = 0;
2213                 if (!ex) {
2214                         /* there is no extent yet, so try to allocate
2215                          * all requested space */
2216                         start = block;
2217                         end = block + num;
2218                 } else if (le32_to_cpu(ex->ee_block) > block) {
2219                         /* need to allocate space before found extent */
2220                         start = block;
2221                         end = le32_to_cpu(ex->ee_block);
2222                         if (block + num < end)
2223                                 end = block + num;
2224                 } else if (block >= le32_to_cpu(ex->ee_block)
2225                                         + ext4_ext_get_actual_len(ex)) {
2226                         /* need to allocate space after found extent */
2227                         start = block;
2228                         end = block + num;
2229                         if (end >= next)
2230                                 end = next;
2231                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2232                         /*
2233                          * some part of requested space is covered
2234                          * by found extent
2235                          */
2236                         start = block;
2237                         end = le32_to_cpu(ex->ee_block)
2238                                 + ext4_ext_get_actual_len(ex);
2239                         if (block + num < end)
2240                                 end = block + num;
2241                         exists = 1;
2242                 } else {
2243                         BUG();
2244                 }
2245                 BUG_ON(end <= start);
2246
2247                 if (!exists) {
2248                         es.es_lblk = start;
2249                         es.es_len = end - start;
2250                         es.es_pblk = 0;
2251                 } else {
2252                         es.es_lblk = le32_to_cpu(ex->ee_block);
2253                         es.es_len = ext4_ext_get_actual_len(ex);
2254                         es.es_pblk = ext4_ext_pblock(ex);
2255                         if (ext4_ext_is_unwritten(ex))
2256                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2257                 }
2258
2259                 /*
2260                  * Find delayed extent and update es accordingly. We call
2261                  * it even in !exists case to find out whether es is the
2262                  * last existing extent or not.
2263                  */
2264                 next_del = ext4_find_delayed_extent(inode, &es);
2265                 if (!exists && next_del) {
2266                         exists = 1;
2267                         flags |= (FIEMAP_EXTENT_DELALLOC |
2268                                   FIEMAP_EXTENT_UNKNOWN);
2269                 }
2270                 up_read(&EXT4_I(inode)->i_data_sem);
2271
2272                 if (unlikely(es.es_len == 0)) {
2273                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2274                         err = -EFSCORRUPTED;
2275                         break;
2276                 }
2277
2278                 /*
2279                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2280                  * we need to check next == EXT_MAX_BLOCKS because it is
2281                  * possible that an extent is with unwritten and delayed
2282                  * status due to when an extent is delayed allocated and
2283                  * is allocated by fallocate status tree will track both of
2284                  * them in a extent.
2285                  *
2286                  * So we could return a unwritten and delayed extent, and
2287                  * its block is equal to 'next'.
2288                  */
2289                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2290                         flags |= FIEMAP_EXTENT_LAST;
2291                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2292                                      next != EXT_MAX_BLOCKS)) {
2293                                 EXT4_ERROR_INODE(inode,
2294                                                  "next extent == %u, next "
2295                                                  "delalloc extent = %u",
2296                                                  next, next_del);
2297                                 err = -EFSCORRUPTED;
2298                                 break;
2299                         }
2300                 }
2301
2302                 if (exists) {
2303                         err = fiemap_fill_next_extent(fieinfo,
2304                                 (__u64)es.es_lblk << blksize_bits,
2305                                 (__u64)es.es_pblk << blksize_bits,
2306                                 (__u64)es.es_len << blksize_bits,
2307                                 flags);
2308                         if (err < 0)
2309                                 break;
2310                         if (err == 1) {
2311                                 err = 0;
2312                                 break;
2313                         }
2314                 }
2315
2316                 block = es.es_lblk + es.es_len;
2317         }
2318
2319         ext4_ext_drop_refs(path);
2320         kfree(path);
2321         return err;
2322 }
2323
2324 /*
2325  * ext4_ext_determine_hole - determine hole around given block
2326  * @inode:      inode we lookup in
2327  * @path:       path in extent tree to @lblk
2328  * @lblk:       pointer to logical block around which we want to determine hole
2329  *
2330  * Determine hole length (and start if easily possible) around given logical
2331  * block. We don't try too hard to find the beginning of the hole but @path
2332  * actually points to extent before @lblk, we provide it.
2333  *
2334  * The function returns the length of a hole starting at @lblk. We update @lblk
2335  * to the beginning of the hole if we managed to find it.
2336  */
2337 static ext4_lblk_t ext4_ext_determine_hole(struct inode *inode,
2338                                            struct ext4_ext_path *path,
2339                                            ext4_lblk_t *lblk)
2340 {
2341         int depth = ext_depth(inode);
2342         struct ext4_extent *ex;
2343         ext4_lblk_t len;
2344
2345         ex = path[depth].p_ext;
2346         if (ex == NULL) {
2347                 /* there is no extent yet, so gap is [0;-] */
2348                 *lblk = 0;
2349                 len = EXT_MAX_BLOCKS;
2350         } else if (*lblk < le32_to_cpu(ex->ee_block)) {
2351                 len = le32_to_cpu(ex->ee_block) - *lblk;
2352         } else if (*lblk >= le32_to_cpu(ex->ee_block)
2353                         + ext4_ext_get_actual_len(ex)) {
2354                 ext4_lblk_t next;
2355
2356                 *lblk = le32_to_cpu(ex->ee_block) + ext4_ext_get_actual_len(ex);
2357                 next = ext4_ext_next_allocated_block(path);
2358                 BUG_ON(next == *lblk);
2359                 len = next - *lblk;
2360         } else {
2361                 BUG();
2362         }
2363         return len;
2364 }
2365
2366 /*
2367  * ext4_ext_put_gap_in_cache:
2368  * calculate boundaries of the gap that the requested block fits into
2369  * and cache this gap
2370  */
2371 static void
2372 ext4_ext_put_gap_in_cache(struct inode *inode, ext4_lblk_t hole_start,
2373                           ext4_lblk_t hole_len)
2374 {
2375         struct extent_status es;
2376
2377         ext4_es_find_delayed_extent_range(inode, hole_start,
2378                                           hole_start + hole_len - 1, &es);
2379         if (es.es_len) {
2380                 /* There's delayed extent containing lblock? */
2381                 if (es.es_lblk <= hole_start)
2382                         return;
2383                 hole_len = min(es.es_lblk - hole_start, hole_len);
2384         }
2385         ext_debug(" -> %u:%u\n", hole_start, hole_len);
2386         ext4_es_insert_extent(inode, hole_start, hole_len, ~0,
2387                               EXTENT_STATUS_HOLE);
2388 }
2389
2390 /*
2391  * ext4_ext_rm_idx:
2392  * removes index from the index block.
2393  */
2394 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2395                         struct ext4_ext_path *path, int depth)
2396 {
2397         int err;
2398         ext4_fsblk_t leaf;
2399
2400         /* free index block */
2401         depth--;
2402         path = path + depth;
2403         leaf = ext4_idx_pblock(path->p_idx);
2404         if (unlikely(path->p_hdr->eh_entries == 0)) {
2405                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2406                 return -EFSCORRUPTED;
2407         }
2408         err = ext4_ext_get_access(handle, inode, path);
2409         if (err)
2410                 return err;
2411
2412         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2413                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2414                 len *= sizeof(struct ext4_extent_idx);
2415                 memmove(path->p_idx, path->p_idx + 1, len);
2416         }
2417
2418         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2419         err = ext4_ext_dirty(handle, inode, path);
2420         if (err)
2421                 return err;
2422         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2423         trace_ext4_ext_rm_idx(inode, leaf);
2424
2425         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2426                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2427
2428         while (--depth >= 0) {
2429                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2430                         break;
2431                 path--;
2432                 err = ext4_ext_get_access(handle, inode, path);
2433                 if (err)
2434                         break;
2435                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2436                 err = ext4_ext_dirty(handle, inode, path);
2437                 if (err)
2438                         break;
2439         }
2440         return err;
2441 }
2442
2443 /*
2444  * ext4_ext_calc_credits_for_single_extent:
2445  * This routine returns max. credits that needed to insert an extent
2446  * to the extent tree.
2447  * When pass the actual path, the caller should calculate credits
2448  * under i_data_sem.
2449  */
2450 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2451                                                 struct ext4_ext_path *path)
2452 {
2453         if (path) {
2454                 int depth = ext_depth(inode);
2455                 int ret = 0;
2456
2457                 /* probably there is space in leaf? */
2458                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2459                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2460
2461                         /*
2462                          *  There are some space in the leaf tree, no
2463                          *  need to account for leaf block credit
2464                          *
2465                          *  bitmaps and block group descriptor blocks
2466                          *  and other metadata blocks still need to be
2467                          *  accounted.
2468                          */
2469                         /* 1 bitmap, 1 block group descriptor */
2470                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2471                         return ret;
2472                 }
2473         }
2474
2475         return ext4_chunk_trans_blocks(inode, nrblocks);
2476 }
2477
2478 /*
2479  * How many index/leaf blocks need to change/allocate to add @extents extents?
2480  *
2481  * If we add a single extent, then in the worse case, each tree level
2482  * index/leaf need to be changed in case of the tree split.
2483  *
2484  * If more extents are inserted, they could cause the whole tree split more
2485  * than once, but this is really rare.
2486  */
2487 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2488 {
2489         int index;
2490         int depth;
2491
2492         /* If we are converting the inline data, only one is needed here. */
2493         if (ext4_has_inline_data(inode))
2494                 return 1;
2495
2496         depth = ext_depth(inode);
2497
2498         if (extents <= 1)
2499                 index = depth * 2;
2500         else
2501                 index = depth * 3;
2502
2503         return index;
2504 }
2505
2506 static inline int get_default_free_blocks_flags(struct inode *inode)
2507 {
2508         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
2509             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
2510                 return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2511         else if (ext4_should_journal_data(inode))
2512                 return EXT4_FREE_BLOCKS_FORGET;
2513         return 0;
2514 }
2515
2516 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2517                               struct ext4_extent *ex,
2518                               long long *partial_cluster,
2519                               ext4_lblk_t from, ext4_lblk_t to)
2520 {
2521         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2522         unsigned short ee_len = ext4_ext_get_actual_len(ex);
2523         ext4_fsblk_t pblk;
2524         int flags = get_default_free_blocks_flags(inode);
2525
2526         /*
2527          * For bigalloc file systems, we never free a partial cluster
2528          * at the beginning of the extent.  Instead, we make a note
2529          * that we tried freeing the cluster, and check to see if we
2530          * need to free it on a subsequent call to ext4_remove_blocks,
2531          * or at the end of ext4_ext_rm_leaf or ext4_ext_remove_space.
2532          */
2533         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2534
2535         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2536         /*
2537          * If we have a partial cluster, and it's different from the
2538          * cluster of the last block, we need to explicitly free the
2539          * partial cluster here.
2540          */
2541         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2542         if (*partial_cluster > 0 &&
2543             *partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2544                 ext4_free_blocks(handle, inode, NULL,
2545                                  EXT4_C2B(sbi, *partial_cluster),
2546                                  sbi->s_cluster_ratio, flags);
2547                 *partial_cluster = 0;
2548         }
2549
2550 #ifdef EXTENTS_STATS
2551         {
2552                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2553                 spin_lock(&sbi->s_ext_stats_lock);
2554                 sbi->s_ext_blocks += ee_len;
2555                 sbi->s_ext_extents++;
2556                 if (ee_len < sbi->s_ext_min)
2557                         sbi->s_ext_min = ee_len;
2558                 if (ee_len > sbi->s_ext_max)
2559                         sbi->s_ext_max = ee_len;
2560                 if (ext_depth(inode) > sbi->s_depth_max)
2561                         sbi->s_depth_max = ext_depth(inode);
2562                 spin_unlock(&sbi->s_ext_stats_lock);
2563         }
2564 #endif
2565         if (from >= le32_to_cpu(ex->ee_block)
2566             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2567                 /* tail removal */
2568                 ext4_lblk_t num;
2569                 long long first_cluster;
2570
2571                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2572                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2573                 /*
2574                  * Usually we want to free partial cluster at the end of the
2575                  * extent, except for the situation when the cluster is still
2576                  * used by any other extent (partial_cluster is negative).
2577                  */
2578                 if (*partial_cluster < 0 &&
2579                     *partial_cluster == -(long long) EXT4_B2C(sbi, pblk+num-1))
2580                         flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2581
2582                 ext_debug("free last %u blocks starting %llu partial %lld\n",
2583                           num, pblk, *partial_cluster);
2584                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2585                 /*
2586                  * If the block range to be freed didn't start at the
2587                  * beginning of a cluster, and we removed the entire
2588                  * extent and the cluster is not used by any other extent,
2589                  * save the partial cluster here, since we might need to
2590                  * delete if we determine that the truncate or punch hole
2591                  * operation has removed all of the blocks in the cluster.
2592                  * If that cluster is used by another extent, preserve its
2593                  * negative value so it isn't freed later on.
2594                  *
2595                  * If the whole extent wasn't freed, we've reached the
2596                  * start of the truncated/punched region and have finished
2597                  * removing blocks.  If there's a partial cluster here it's
2598                  * shared with the remainder of the extent and is no longer
2599                  * a candidate for removal.
2600                  */
2601                 if (EXT4_PBLK_COFF(sbi, pblk) && ee_len == num) {
2602                         first_cluster = (long long) EXT4_B2C(sbi, pblk);
2603                         if (first_cluster != -*partial_cluster)
2604                                 *partial_cluster = first_cluster;
2605                 } else {
2606                         *partial_cluster = 0;
2607                 }
2608         } else
2609                 ext4_error(sbi->s_sb, "strange request: removal(2) "
2610                            "%u-%u from %u:%u",
2611                            from, to, le32_to_cpu(ex->ee_block), ee_len);
2612         return 0;
2613 }
2614
2615
2616 /*
2617  * ext4_ext_rm_leaf() Removes the extents associated with the
2618  * blocks appearing between "start" and "end".  Both "start"
2619  * and "end" must appear in the same extent or EIO is returned.
2620  *
2621  * @handle: The journal handle
2622  * @inode:  The files inode
2623  * @path:   The path to the leaf
2624  * @partial_cluster: The cluster which we'll have to free if all extents
2625  *                   has been released from it.  However, if this value is
2626  *                   negative, it's a cluster just to the right of the
2627  *                   punched region and it must not be freed.
2628  * @start:  The first block to remove
2629  * @end:   The last block to remove
2630  */
2631 static int
2632 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2633                  struct ext4_ext_path *path,
2634                  long long *partial_cluster,
2635                  ext4_lblk_t start, ext4_lblk_t end)
2636 {
2637         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2638         int err = 0, correct_index = 0;
2639         int depth = ext_depth(inode), credits;
2640         struct ext4_extent_header *eh;
2641         ext4_lblk_t a, b;
2642         unsigned num;
2643         ext4_lblk_t ex_ee_block;
2644         unsigned short ex_ee_len;
2645         unsigned unwritten = 0;
2646         struct ext4_extent *ex;
2647         ext4_fsblk_t pblk;
2648
2649         /* the header must be checked already in ext4_ext_remove_space() */
2650         ext_debug("truncate since %u in leaf to %u\n", start, end);
2651         if (!path[depth].p_hdr)
2652                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2653         eh = path[depth].p_hdr;
2654         if (unlikely(path[depth].p_hdr == NULL)) {
2655                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2656                 return -EFSCORRUPTED;
2657         }
2658         /* find where to start removing */
2659         ex = path[depth].p_ext;
2660         if (!ex)
2661                 ex = EXT_LAST_EXTENT(eh);
2662
2663         ex_ee_block = le32_to_cpu(ex->ee_block);
2664         ex_ee_len = ext4_ext_get_actual_len(ex);
2665
2666         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2667
2668         while (ex >= EXT_FIRST_EXTENT(eh) &&
2669                         ex_ee_block + ex_ee_len > start) {
2670
2671                 if (ext4_ext_is_unwritten(ex))
2672                         unwritten = 1;
2673                 else
2674                         unwritten = 0;
2675
2676                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2677                           unwritten, ex_ee_len);
2678                 path[depth].p_ext = ex;
2679
2680                 a = ex_ee_block > start ? ex_ee_block : start;
2681                 b = ex_ee_block+ex_ee_len - 1 < end ?
2682                         ex_ee_block+ex_ee_len - 1 : end;
2683
2684                 ext_debug("  border %u:%u\n", a, b);
2685
2686                 /* If this extent is beyond the end of the hole, skip it */
2687                 if (end < ex_ee_block) {
2688                         /*
2689                          * We're going to skip this extent and move to another,
2690                          * so note that its first cluster is in use to avoid
2691                          * freeing it when removing blocks.  Eventually, the
2692                          * right edge of the truncated/punched region will
2693                          * be just to the left.
2694                          */
2695                         if (sbi->s_cluster_ratio > 1) {
2696                                 pblk = ext4_ext_pblock(ex);
2697                                 *partial_cluster =
2698                                         -(long long) EXT4_B2C(sbi, pblk);
2699                         }
2700                         ex--;
2701                         ex_ee_block = le32_to_cpu(ex->ee_block);
2702                         ex_ee_len = ext4_ext_get_actual_len(ex);
2703                         continue;
2704                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2705                         EXT4_ERROR_INODE(inode,
2706                                          "can not handle truncate %u:%u "
2707                                          "on extent %u:%u",
2708                                          start, end, ex_ee_block,
2709                                          ex_ee_block + ex_ee_len - 1);
2710                         err = -EFSCORRUPTED;
2711                         goto out;
2712                 } else if (a != ex_ee_block) {
2713                         /* remove tail of the extent */
2714                         num = a - ex_ee_block;
2715                 } else {
2716                         /* remove whole extent: excellent! */
2717                         num = 0;
2718                 }
2719                 /*
2720                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2721                  * descriptor) for each block group; assume two block
2722                  * groups plus ex_ee_len/blocks_per_block_group for
2723                  * the worst case
2724                  */
2725                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2726                 if (ex == EXT_FIRST_EXTENT(eh)) {
2727                         correct_index = 1;
2728                         credits += (ext_depth(inode)) + 1;
2729                 }
2730                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2731
2732                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2733                 if (err)
2734                         goto out;
2735
2736                 err = ext4_ext_get_access(handle, inode, path + depth);
2737                 if (err)
2738                         goto out;
2739
2740                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2741                                          a, b);
2742                 if (err)
2743                         goto out;
2744
2745                 if (num == 0)
2746                         /* this extent is removed; mark slot entirely unused */
2747                         ext4_ext_store_pblock(ex, 0);
2748
2749                 ex->ee_len = cpu_to_le16(num);
2750                 /*
2751                  * Do not mark unwritten if all the blocks in the
2752                  * extent have been removed.
2753                  */
2754                 if (unwritten && num)
2755                         ext4_ext_mark_unwritten(ex);
2756                 /*
2757                  * If the extent was completely released,
2758                  * we need to remove it from the leaf
2759                  */
2760                 if (num == 0) {
2761                         if (end != EXT_MAX_BLOCKS - 1) {
2762                                 /*
2763                                  * For hole punching, we need to scoot all the
2764                                  * extents up when an extent is removed so that
2765                                  * we dont have blank extents in the middle
2766                                  */
2767                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2768                                         sizeof(struct ext4_extent));
2769
2770                                 /* Now get rid of the one at the end */
2771                                 memset(EXT_LAST_EXTENT(eh), 0,
2772                                         sizeof(struct ext4_extent));
2773                         }
2774                         le16_add_cpu(&eh->eh_entries, -1);
2775                 }
2776
2777                 err = ext4_ext_dirty(handle, inode, path + depth);
2778                 if (err)
2779                         goto out;
2780
2781                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2782                                 ext4_ext_pblock(ex));
2783                 ex--;
2784                 ex_ee_block = le32_to_cpu(ex->ee_block);
2785                 ex_ee_len = ext4_ext_get_actual_len(ex);
2786         }
2787
2788         if (correct_index && eh->eh_entries)
2789                 err = ext4_ext_correct_indexes(handle, inode, path);
2790
2791         /*
2792          * If there's a partial cluster and at least one extent remains in
2793          * the leaf, free the partial cluster if it isn't shared with the
2794          * current extent.  If it is shared with the current extent
2795          * we zero partial_cluster because we've reached the start of the
2796          * truncated/punched region and we're done removing blocks.
2797          */
2798         if (*partial_cluster > 0 && ex >= EXT_FIRST_EXTENT(eh)) {
2799                 pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2800                 if (*partial_cluster != (long long) EXT4_B2C(sbi, pblk)) {
2801                         ext4_free_blocks(handle, inode, NULL,
2802                                          EXT4_C2B(sbi, *partial_cluster),
2803                                          sbi->s_cluster_ratio,
2804                                          get_default_free_blocks_flags(inode));
2805                 }
2806                 *partial_cluster = 0;
2807         }
2808
2809         /* if this leaf is free, then we should
2810          * remove it from index block above */
2811         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2812                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2813
2814 out:
2815         return err;
2816 }
2817
2818 /*
2819  * ext4_ext_more_to_rm:
2820  * returns 1 if current index has to be freed (even partial)
2821  */
2822 static int
2823 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2824 {
2825         BUG_ON(path->p_idx == NULL);
2826
2827         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2828                 return 0;
2829
2830         /*
2831          * if truncate on deeper level happened, it wasn't partial,
2832          * so we have to consider current index for truncation
2833          */
2834         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2835                 return 0;
2836         return 1;
2837 }
2838
2839 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2840                           ext4_lblk_t end)
2841 {
2842         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2843         int depth = ext_depth(inode);
2844         struct ext4_ext_path *path = NULL;
2845         long long partial_cluster = 0;
2846         handle_t *handle;
2847         int i = 0, err = 0;
2848
2849         ext_debug("truncate since %u to %u\n", start, end);
2850
2851         /* probably first extent we're gonna free will be last in block */
2852         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2853         if (IS_ERR(handle))
2854                 return PTR_ERR(handle);
2855
2856 again:
2857         trace_ext4_ext_remove_space(inode, start, end, depth);
2858
2859         /*
2860          * Check if we are removing extents inside the extent tree. If that
2861          * is the case, we are going to punch a hole inside the extent tree
2862          * so we have to check whether we need to split the extent covering
2863          * the last block to remove so we can easily remove the part of it
2864          * in ext4_ext_rm_leaf().
2865          */
2866         if (end < EXT_MAX_BLOCKS - 1) {
2867                 struct ext4_extent *ex;
2868                 ext4_lblk_t ee_block, ex_end, lblk;
2869                 ext4_fsblk_t pblk;
2870
2871                 /* find extent for or closest extent to this block */
2872                 path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2873                 if (IS_ERR(path)) {
2874                         ext4_journal_stop(handle);
2875                         return PTR_ERR(path);
2876                 }
2877                 depth = ext_depth(inode);
2878                 /* Leaf not may not exist only if inode has no blocks at all */
2879                 ex = path[depth].p_ext;
2880                 if (!ex) {
2881                         if (depth) {
2882                                 EXT4_ERROR_INODE(inode,
2883                                                  "path[%d].p_hdr == NULL",
2884                                                  depth);
2885                                 err = -EFSCORRUPTED;
2886                         }
2887                         goto out;
2888                 }
2889
2890                 ee_block = le32_to_cpu(ex->ee_block);
2891                 ex_end = ee_block + ext4_ext_get_actual_len(ex) - 1;
2892
2893                 /*
2894                  * See if the last block is inside the extent, if so split
2895                  * the extent at 'end' block so we can easily remove the
2896                  * tail of the first part of the split extent in
2897                  * ext4_ext_rm_leaf().
2898                  */
2899                 if (end >= ee_block && end < ex_end) {
2900
2901                         /*
2902                          * If we're going to split the extent, note that
2903                          * the cluster containing the block after 'end' is
2904                          * in use to avoid freeing it when removing blocks.
2905                          */
2906                         if (sbi->s_cluster_ratio > 1) {
2907                                 pblk = ext4_ext_pblock(ex) + end - ee_block + 1;
2908                                 partial_cluster =
2909                                         -(long long) EXT4_B2C(sbi, pblk);
2910                         }
2911
2912                         /*
2913                          * Split the extent in two so that 'end' is the last
2914                          * block in the first new extent. Also we should not
2915                          * fail removing space due to ENOSPC so try to use
2916                          * reserved block if that happens.
2917                          */
2918                         err = ext4_force_split_extent_at(handle, inode, &path,
2919                                                          end + 1, 1);
2920                         if (err < 0)
2921                                 goto out;
2922
2923                 } else if (sbi->s_cluster_ratio > 1 && end >= ex_end) {
2924                         /*
2925                          * If there's an extent to the right its first cluster
2926                          * contains the immediate right boundary of the
2927                          * truncated/punched region.  Set partial_cluster to
2928                          * its negative value so it won't be freed if shared
2929                          * with the current extent.  The end < ee_block case
2930                          * is handled in ext4_ext_rm_leaf().
2931                          */
2932                         lblk = ex_end + 1;
2933                         err = ext4_ext_search_right(inode, path, &lblk, &pblk,
2934                                                     &ex);
2935                         if (err)
2936                                 goto out;
2937                         if (pblk)
2938                                 partial_cluster =
2939                                         -(long long) EXT4_B2C(sbi, pblk);
2940                 }
2941         }
2942         /*
2943          * We start scanning from right side, freeing all the blocks
2944          * after i_size and walking into the tree depth-wise.
2945          */
2946         depth = ext_depth(inode);
2947         if (path) {
2948                 int k = i = depth;
2949                 while (--k > 0)
2950                         path[k].p_block =
2951                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2952         } else {
2953                 path = kcalloc(depth + 1, sizeof(struct ext4_ext_path),
2954                                GFP_NOFS);
2955                 if (path == NULL) {
2956                         ext4_journal_stop(handle);
2957                         return -ENOMEM;
2958                 }
2959                 path[0].p_maxdepth = path[0].p_depth = depth;
2960                 path[0].p_hdr = ext_inode_hdr(inode);
2961                 i = 0;
2962
2963                 if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2964                         err = -EFSCORRUPTED;
2965                         goto out;
2966                 }
2967         }
2968         err = 0;
2969
2970         while (i >= 0 && err == 0) {
2971                 if (i == depth) {
2972                         /* this is leaf block */
2973                         err = ext4_ext_rm_leaf(handle, inode, path,
2974                                                &partial_cluster, start,
2975                                                end);
2976                         /* root level has p_bh == NULL, brelse() eats this */
2977                         brelse(path[i].p_bh);
2978                         path[i].p_bh = NULL;
2979                         i--;
2980                         continue;
2981                 }
2982
2983                 /* this is index block */
2984                 if (!path[i].p_hdr) {
2985                         ext_debug("initialize header\n");
2986                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2987                 }
2988
2989                 if (!path[i].p_idx) {
2990                         /* this level hasn't been touched yet */
2991                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2992                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2993                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2994                                   path[i].p_hdr,
2995                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2996                 } else {
2997                         /* we were already here, see at next index */
2998                         path[i].p_idx--;
2999                 }
3000
3001                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
3002                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
3003                                 path[i].p_idx);
3004                 if (ext4_ext_more_to_rm(path + i)) {
3005                         struct buffer_head *bh;
3006                         /* go to the next level */
3007                         ext_debug("move to level %d (block %llu)\n",
3008                                   i + 1, ext4_idx_pblock(path[i].p_idx));
3009                         memset(path + i + 1, 0, sizeof(*path));
3010                         bh = read_extent_tree_block(inode,
3011                                 ext4_idx_pblock(path[i].p_idx), depth - i - 1,
3012                                 EXT4_EX_NOCACHE);
3013                         if (IS_ERR(bh)) {
3014                                 /* should we reset i_size? */
3015                                 err = PTR_ERR(bh);
3016                                 break;
3017                         }
3018                         /* Yield here to deal with large extent trees.
3019                          * Should be a no-op if we did IO above. */
3020                         cond_resched();
3021                         if (WARN_ON(i + 1 > depth)) {
3022                                 err = -EFSCORRUPTED;
3023                                 break;
3024                         }
3025                         path[i + 1].p_bh = bh;
3026
3027                         /* save actual number of indexes since this
3028                          * number is changed at the next iteration */
3029                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
3030                         i++;
3031                 } else {
3032                         /* we finished processing this index, go up */
3033                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
3034                                 /* index is empty, remove it;
3035                                  * handle must be already prepared by the
3036                                  * truncatei_leaf() */
3037                                 err = ext4_ext_rm_idx(handle, inode, path, i);
3038                         }
3039                         /* root level has p_bh == NULL, brelse() eats this */
3040                         brelse(path[i].p_bh);
3041                         path[i].p_bh = NULL;
3042                         i--;
3043                         ext_debug("return to level %d\n", i);
3044                 }
3045         }
3046
3047         trace_ext4_ext_remove_space_done(inode, start, end, depth,
3048                         partial_cluster, path->p_hdr->eh_entries);
3049
3050         /*
3051          * If we still have something in the partial cluster and we have removed
3052          * even the first extent, then we should free the blocks in the partial
3053          * cluster as well.  (This code will only run when there are no leaves
3054          * to the immediate left of the truncated/punched region.)
3055          */
3056         if (partial_cluster > 0 && err == 0) {
3057                 /* don't zero partial_cluster since it's not used afterwards */
3058                 ext4_free_blocks(handle, inode, NULL,
3059                                  EXT4_C2B(sbi, partial_cluster),
3060                                  sbi->s_cluster_ratio,
3061                                  get_default_free_blocks_flags(inode));
3062         }
3063
3064         /* TODO: flexible tree reduction should be here */
3065         if (path->p_hdr->eh_entries == 0) {
3066                 /*
3067                  * truncate to zero freed all the tree,
3068                  * so we need to correct eh_depth
3069                  */
3070                 err = ext4_ext_get_access(handle, inode, path);
3071                 if (err == 0) {
3072                         ext_inode_hdr(inode)->eh_depth = 0;
3073                         ext_inode_hdr(inode)->eh_max =
3074                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
3075                         err = ext4_ext_dirty(handle, inode, path);
3076                 }
3077         }
3078 out:
3079         ext4_ext_drop_refs(path);
3080         kfree(path);
3081         path = NULL;
3082         if (err == -EAGAIN)
3083                 goto again;
3084         ext4_journal_stop(handle);
3085
3086         return err;
3087 }
3088
3089 /*
3090  * called at mount time
3091  */
3092 void ext4_ext_init(struct super_block *sb)
3093 {
3094         /*
3095          * possible initialization would be here
3096          */
3097
3098         if (ext4_has_feature_extents(sb)) {
3099 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3100                 printk(KERN_INFO "EXT4-fs: file extents enabled"
3101 #ifdef AGGRESSIVE_TEST
3102                        ", aggressive tests"
3103 #endif
3104 #ifdef CHECK_BINSEARCH
3105                        ", check binsearch"
3106 #endif
3107 #ifdef EXTENTS_STATS
3108                        ", stats"
3109 #endif
3110                        "\n");
3111 #endif
3112 #ifdef EXTENTS_STATS
3113                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3114                 EXT4_SB(sb)->s_ext_min = 1 << 30;
3115                 EXT4_SB(sb)->s_ext_max = 0;
3116 #endif
3117         }
3118 }
3119
3120 /*
3121  * called at umount time
3122  */
3123 void ext4_ext_release(struct super_block *sb)
3124 {
3125         if (!ext4_has_feature_extents(sb))
3126                 return;
3127
3128 #ifdef EXTENTS_STATS
3129         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3130                 struct ext4_sb_info *sbi = EXT4_SB(sb);
3131                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3132                         sbi->s_ext_blocks, sbi->s_ext_extents,
3133                         sbi->s_ext_blocks / sbi->s_ext_extents);
3134                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3135                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3136         }
3137 #endif
3138 }
3139
3140 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3141 {
3142         ext4_lblk_t  ee_block;
3143         ext4_fsblk_t ee_pblock;
3144         unsigned int ee_len;
3145
3146         ee_block  = le32_to_cpu(ex->ee_block);
3147         ee_len    = ext4_ext_get_actual_len(ex);
3148         ee_pblock = ext4_ext_pblock(ex);
3149
3150         if (ee_len == 0)
3151                 return 0;
3152
3153         return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3154                                      EXTENT_STATUS_WRITTEN);
3155 }
3156
3157 /* FIXME!! we need to try to merge to left or right after zero-out  */
3158 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3159 {
3160         ext4_fsblk_t ee_pblock;
3161         unsigned int ee_len;
3162
3163         ee_len    = ext4_ext_get_actual_len(ex);
3164         ee_pblock = ext4_ext_pblock(ex);
3165         return ext4_issue_zeroout(inode, le32_to_cpu(ex->ee_block), ee_pblock,
3166                                   ee_len);
3167 }
3168
3169 /*
3170  * ext4_split_extent_at() splits an extent at given block.
3171  *
3172  * @handle: the journal handle
3173  * @inode: the file inode
3174  * @path: the path to the extent
3175  * @split: the logical block where the extent is splitted.
3176  * @split_flags: indicates if the extent could be zeroout if split fails, and
3177  *               the states(init or unwritten) of new extents.
3178  * @flags: flags used to insert new extent to extent tree.
3179  *
3180  *
3181  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3182  * of which are deterimined by split_flag.
3183  *
3184  * There are two cases:
3185  *  a> the extent are splitted into two extent.
3186  *  b> split is not needed, and just mark the extent.
3187  *
3188  * return 0 on success.
3189  */
3190 static int ext4_split_extent_at(handle_t *handle,
3191                              struct inode *inode,
3192                              struct ext4_ext_path **ppath,
3193                              ext4_lblk_t split,
3194                              int split_flag,
3195                              int flags)
3196 {
3197         struct ext4_ext_path *path = *ppath;
3198         ext4_fsblk_t newblock;
3199         ext4_lblk_t ee_block;
3200         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3201         struct ext4_extent *ex2 = NULL;
3202         unsigned int ee_len, depth;
3203         int err = 0;
3204
3205         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3206                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3207
3208         ext_debug("ext4_split_extents_at: inode %lu, logical"
3209                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3210
3211         ext4_ext_show_leaf(inode, path);
3212
3213         depth = ext_depth(inode);
3214         ex = path[depth].p_ext;
3215         ee_block = le32_to_cpu(ex->ee_block);
3216         ee_len = ext4_ext_get_actual_len(ex);
3217         newblock = split - ee_block + ext4_ext_pblock(ex);
3218
3219         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3220         BUG_ON(!ext4_ext_is_unwritten(ex) &&
3221                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3222                              EXT4_EXT_MARK_UNWRIT1 |
3223                              EXT4_EXT_MARK_UNWRIT2));
3224
3225         err = ext4_ext_get_access(handle, inode, path + depth);
3226         if (err)
3227                 goto out;
3228
3229         if (split == ee_block) {
3230                 /*
3231                  * case b: block @split is the block that the extent begins with
3232                  * then we just change the state of the extent, and splitting
3233                  * is not needed.
3234                  */
3235                 if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3236                         ext4_ext_mark_unwritten(ex);
3237                 else
3238                         ext4_ext_mark_initialized(ex);
3239
3240                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3241                         ext4_ext_try_to_merge(handle, inode, path, ex);
3242
3243                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3244                 goto out;
3245         }
3246
3247         /* case a */
3248         memcpy(&orig_ex, ex, sizeof(orig_ex));
3249         ex->ee_len = cpu_to_le16(split - ee_block);
3250         if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3251                 ext4_ext_mark_unwritten(ex);
3252
3253         /*
3254          * path may lead to new leaf, not to original leaf any more
3255          * after ext4_ext_insert_extent() returns,
3256          */
3257         err = ext4_ext_dirty(handle, inode, path + depth);
3258         if (err)
3259                 goto fix_extent_len;
3260
3261         ex2 = &newex;
3262         ex2->ee_block = cpu_to_le32(split);
3263         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3264         ext4_ext_store_pblock(ex2, newblock);
3265         if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3266                 ext4_ext_mark_unwritten(ex2);
3267
3268         err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3269         if (err != -ENOSPC && err != -EDQUOT)
3270                 goto out;
3271
3272         if (EXT4_EXT_MAY_ZEROOUT & split_flag) {
3273                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3274                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3275                                 err = ext4_ext_zeroout(inode, ex2);
3276                                 zero_ex.ee_block = ex2->ee_block;
3277                                 zero_ex.ee_len = cpu_to_le16(
3278                                                 ext4_ext_get_actual_len(ex2));
3279                                 ext4_ext_store_pblock(&zero_ex,
3280                                                       ext4_ext_pblock(ex2));
3281                         } else {
3282                                 err = ext4_ext_zeroout(inode, ex);
3283                                 zero_ex.ee_block = ex->ee_block;
3284                                 zero_ex.ee_len = cpu_to_le16(
3285                                                 ext4_ext_get_actual_len(ex));
3286                                 ext4_ext_store_pblock(&zero_ex,
3287                                                       ext4_ext_pblock(ex));
3288                         }
3289                 } else {
3290                         err = ext4_ext_zeroout(inode, &orig_ex);
3291                         zero_ex.ee_block = orig_ex.ee_block;
3292                         zero_ex.ee_len = cpu_to_le16(
3293                                                 ext4_ext_get_actual_len(&orig_ex));
3294                         ext4_ext_store_pblock(&zero_ex,
3295                                               ext4_ext_pblock(&orig_ex));
3296                 }
3297
3298                 if (!err) {
3299                         /* update the extent length and mark as initialized */
3300                         ex->ee_len = cpu_to_le16(ee_len);
3301                         ext4_ext_try_to_merge(handle, inode, path, ex);
3302                         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3303                         if (!err)
3304                                 /* update extent status tree */
3305                                 err = ext4_zeroout_es(inode, &zero_ex);
3306                         /* If we failed at this point, we don't know in which
3307                          * state the extent tree exactly is so don't try to fix
3308                          * length of the original extent as it may do even more
3309                          * damage.
3310                          */
3311                         goto out;
3312                 }
3313         }
3314
3315 fix_extent_len:
3316         ex->ee_len = orig_ex.ee_len;
3317         ext4_ext_dirty(handle, inode, path + path->p_depth);
3318         return err;
3319 out:
3320         ext4_ext_show_leaf(inode, path);
3321         return err;
3322 }
3323
3324 /*
3325  * ext4_split_extents() splits an extent and mark extent which is covered
3326  * by @map as split_flags indicates
3327  *
3328  * It may result in splitting the extent into multiple extents (up to three)
3329  * There are three possibilities:
3330  *   a> There is no split required
3331  *   b> Splits in two extents: Split is happening at either end of the extent
3332  *   c> Splits in three extents: Somone is splitting in middle of the extent
3333  *
3334  */
3335 static int ext4_split_extent(handle_t *handle,
3336                               struct inode *inode,
3337                               struct ext4_ext_path **ppath,
3338                               struct ext4_map_blocks *map,
3339                               int split_flag,
3340                               int flags)
3341 {
3342         struct ext4_ext_path *path = *ppath;
3343         ext4_lblk_t ee_block;
3344         struct ext4_extent *ex;
3345         unsigned int ee_len, depth;
3346         int err = 0;
3347         int unwritten;
3348         int split_flag1, flags1;
3349         int allocated = map->m_len;
3350
3351         depth = ext_depth(inode);
3352         ex = path[depth].p_ext;
3353         ee_block = le32_to_cpu(ex->ee_block);
3354         ee_len = ext4_ext_get_actual_len(ex);
3355         unwritten = ext4_ext_is_unwritten(ex);
3356
3357         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3358                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3359                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3360                 if (unwritten)
3361                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3362                                        EXT4_EXT_MARK_UNWRIT2;
3363                 if (split_flag & EXT4_EXT_DATA_VALID2)
3364                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3365                 err = ext4_split_extent_at(handle, inode, ppath,
3366                                 map->m_lblk + map->m_len, split_flag1, flags1);
3367                 if (err)
3368                         goto out;
3369         } else {
3370                 allocated = ee_len - (map->m_lblk - ee_block);
3371         }
3372         /*
3373          * Update path is required because previous ext4_split_extent_at() may
3374          * result in split of original leaf or extent zeroout.
3375          */
3376         path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3377         if (IS_ERR(path))
3378                 return PTR_ERR(path);
3379         depth = ext_depth(inode);
3380         ex = path[depth].p_ext;
3381         if (!ex) {
3382                 EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3383                                  (unsigned long) map->m_lblk);
3384                 return -EFSCORRUPTED;
3385         }
3386         unwritten = ext4_ext_is_unwritten(ex);
3387         split_flag1 = 0;
3388
3389         if (map->m_lblk >= ee_block) {
3390                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3391                 if (unwritten) {
3392                         split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3393                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3394                                                      EXT4_EXT_MARK_UNWRIT2);
3395                 }
3396                 err = ext4_split_extent_at(handle, inode, ppath,
3397                                 map->m_lblk, split_flag1, flags);
3398                 if (err)
3399                         goto out;
3400         }
3401
3402         ext4_ext_show_leaf(inode, path);
3403 out:
3404         return err ? err : allocated;
3405 }
3406
3407 /*
3408  * This function is called by ext4_ext_map_blocks() if someone tries to write
3409  * to an unwritten extent. It may result in splitting the unwritten
3410  * extent into multiple extents (up to three - one initialized and two
3411  * unwritten).
3412  * There are three possibilities:
3413  *   a> There is no split required: Entire extent should be initialized
3414  *   b> Splits in two extents: Write is happening at either end of the extent
3415  *   c> Splits in three extents: Somone is writing in middle of the extent
3416  *
3417  * Pre-conditions:
3418  *  - The extent pointed to by 'path' is unwritten.
3419  *  - The extent pointed to by 'path' contains a superset
3420  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3421  *
3422  * Post-conditions on success:
3423  *  - the returned value is the number of blocks beyond map->l_lblk
3424  *    that are allocated and initialized.
3425  *    It is guaranteed to be >= map->m_len.
3426  */
3427 static int ext4_ext_convert_to_initialized(handle_t *handle,
3428                                            struct inode *inode,
3429                                            struct ext4_map_blocks *map,
3430                                            struct ext4_ext_path **ppath,
3431                                            int flags)
3432 {
3433         struct ext4_ext_path *path = *ppath;
3434         struct ext4_sb_info *sbi;
3435         struct ext4_extent_header *eh;
3436         struct ext4_map_blocks split_map;
3437         struct ext4_extent zero_ex1, zero_ex2;
3438         struct ext4_extent *ex, *abut_ex;
3439         ext4_lblk_t ee_block, eof_block;
3440         unsigned int ee_len, depth, map_len = map->m_len;
3441         int allocated = 0, max_zeroout = 0;
3442         int err = 0;
3443         int split_flag = EXT4_EXT_DATA_VALID2;
3444
3445         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3446                 "block %llu, max_blocks %u\n", inode->i_ino,
3447                 (unsigned long long)map->m_lblk, map_len);
3448
3449         sbi = EXT4_SB(inode->i_sb);
3450         eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1)
3451                         >> inode->i_sb->s_blocksize_bits;
3452         if (eof_block < map->m_lblk + map_len)
3453                 eof_block = map->m_lblk + map_len;
3454
3455         depth = ext_depth(inode);
3456         eh = path[depth].p_hdr;
3457         ex = path[depth].p_ext;
3458         ee_block = le32_to_cpu(ex->ee_block);
3459         ee_len = ext4_ext_get_actual_len(ex);
3460         zero_ex1.ee_len = 0;
3461         zero_ex2.ee_len = 0;
3462
3463         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3464
3465         /* Pre-conditions */
3466         BUG_ON(!ext4_ext_is_unwritten(ex));
3467         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3468
3469         /*
3470          * Attempt to transfer newly initialized blocks from the currently
3471          * unwritten extent to its neighbor. This is much cheaper
3472          * than an insertion followed by a merge as those involve costly
3473          * memmove() calls. Transferring to the left is the common case in
3474          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3475          * followed by append writes.
3476          *
3477          * Limitations of the current logic:
3478          *  - L1: we do not deal with writes covering the whole extent.
3479          *    This would require removing the extent if the transfer
3480          *    is possible.
3481          *  - L2: we only attempt to merge with an extent stored in the
3482          *    same extent tree node.
3483          */
3484         if ((map->m_lblk == ee_block) &&
3485                 /* See if we can merge left */
3486                 (map_len < ee_len) &&           /*L1*/
3487                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3488                 ext4_lblk_t prev_lblk;
3489                 ext4_fsblk_t prev_pblk, ee_pblk;
3490                 unsigned int prev_len;
3491
3492                 abut_ex = ex - 1;
3493                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3494                 prev_len = ext4_ext_get_actual_len(abut_ex);
3495                 prev_pblk = ext4_ext_pblock(abut_ex);
3496                 ee_pblk = ext4_ext_pblock(ex);
3497
3498                 /*
3499                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3500                  * upon those conditions:
3501                  * - C1: abut_ex is initialized,
3502                  * - C2: abut_ex is logically abutting ex,
3503                  * - C3: abut_ex is physically abutting ex,
3504                  * - C4: abut_ex can receive the additional blocks without
3505                  *   overflowing the (initialized) length limit.
3506                  */
3507                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3508                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3509                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3510                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3511                         err = ext4_ext_get_access(handle, inode, path + depth);
3512                         if (err)
3513                                 goto out;
3514
3515                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3516                                 map, ex, abut_ex);
3517
3518                         /* Shift the start of ex by 'map_len' blocks */
3519                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3520                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3521                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3522                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3523
3524                         /* Extend abut_ex by 'map_len' blocks */
3525                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3526
3527                         /* Result: number of initialized blocks past m_lblk */
3528                         allocated = map_len;
3529                 }
3530         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3531                    (map_len < ee_len) &&        /*L1*/
3532                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3533                 /* See if we can merge right */
3534                 ext4_lblk_t next_lblk;
3535                 ext4_fsblk_t next_pblk, ee_pblk;
3536                 unsigned int next_len;
3537
3538                 abut_ex = ex + 1;
3539                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3540                 next_len = ext4_ext_get_actual_len(abut_ex);
3541                 next_pblk = ext4_ext_pblock(abut_ex);
3542                 ee_pblk = ext4_ext_pblock(ex);
3543
3544                 /*
3545                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3546                  * upon those conditions:
3547                  * - C1: abut_ex is initialized,
3548                  * - C2: abut_ex is logically abutting ex,
3549                  * - C3: abut_ex is physically abutting ex,
3550                  * - C4: abut_ex can receive the additional blocks without
3551                  *   overflowing the (initialized) length limit.
3552                  */
3553                 if ((!ext4_ext_is_unwritten(abut_ex)) &&                /*C1*/
3554                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3555                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3556                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3557                         err = ext4_ext_get_access(handle, inode, path + depth);
3558                         if (err)
3559                                 goto out;
3560
3561                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3562                                 map, ex, abut_ex);
3563
3564                         /* Shift the start of abut_ex by 'map_len' blocks */
3565                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3566                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3567                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3568                         ext4_ext_mark_unwritten(ex); /* Restore the flag */
3569
3570                         /* Extend abut_ex by 'map_len' blocks */
3571                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3572
3573                         /* Result: number of initialized blocks past m_lblk */
3574                         allocated = map_len;
3575                 }
3576         }
3577         if (allocated) {
3578                 /* Mark the block containing both extents as dirty */
3579                 ext4_ext_dirty(handle, inode, path + depth);
3580
3581                 /* Update path to point to the right extent */
3582                 path[depth].p_ext = abut_ex;
3583                 goto out;
3584         } else
3585                 allocated = ee_len - (map->m_lblk - ee_block);
3586
3587         WARN_ON(map->m_lblk < ee_block);
3588         /*
3589          * It is safe to convert extent to initialized via explicit
3590          * zeroout only if extent is fully inside i_size or new_size.
3591          */
3592         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3593
3594         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3595                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3596                         (inode->i_sb->s_blocksize_bits - 10);
3597
3598         if (ext4_encrypted_inode(inode))
3599                 max_zeroout = 0;
3600
3601         /*
3602          * five cases:
3603          * 1. split the extent into three extents.
3604          * 2. split the extent into two extents, zeroout the head of the first
3605          *    extent.
3606          * 3. split the extent into two extents, zeroout the tail of the second
3607          *    extent.
3608          * 4. split the extent into two extents with out zeroout.
3609          * 5. no splitting needed, just possibly zeroout the head and / or the
3610          *    tail of the extent.
3611          */
3612         split_map.m_lblk = map->m_lblk;
3613         split_map.m_len = map->m_len;
3614
3615         if (max_zeroout && (allocated > split_map.m_len)) {
3616                 if (allocated <= max_zeroout) {
3617                         /* case 3 or 5 */
3618                         zero_ex1.ee_block =
3619                                  cpu_to_le32(split_map.m_lblk +
3620                                              split_map.m_len);
3621                         zero_ex1.ee_len =
3622                                 cpu_to_le16(allocated - split_map.m_len);
3623                         ext4_ext_store_pblock(&zero_ex1,
3624                                 ext4_ext_pblock(ex) + split_map.m_lblk +
3625                                 split_map.m_len - ee_block);
3626                         err = ext4_ext_zeroout(inode, &zero_ex1);
3627                         if (err)
3628                                 goto out;
3629                         split_map.m_len = allocated;
3630                 }
3631                 if (split_map.m_lblk - ee_block + split_map.m_len <
3632                                                                 max_zeroout) {
3633                         /* case 2 or 5 */
3634                         if (split_map.m_lblk != ee_block) {
3635                                 zero_ex2.ee_block = ex->ee_block;
3636                                 zero_ex2.ee_len = cpu_to_le16(split_map.m_lblk -
3637                                                         ee_block);
3638                                 ext4_ext_store_pblock(&zero_ex2,
3639                                                       ext4_ext_pblock(ex));
3640                                 err = ext4_ext_zeroout(inode, &zero_ex2);
3641                                 if (err)
3642                                         goto out;
3643                         }
3644
3645                         split_map.m_len += split_map.m_lblk - ee_block;
3646                         split_map.m_lblk = ee_block;
3647                         allocated = map->m_len;
3648                 }
3649         }
3650
3651         err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3652                                 flags);
3653         if (err > 0)
3654                 err = 0;
3655 out:
3656         /* If we have gotten a failure, don't zero out status tree */
3657         if (!err) {
3658                 err = ext4_zeroout_es(inode, &zero_ex1);
3659                 if (!err)
3660                         err = ext4_zeroout_es(inode, &zero_ex2);
3661         }
3662         return err ? err : allocated;
3663 }
3664
3665 /*
3666  * This function is called by ext4_ext_map_blocks() from
3667  * ext4_get_blocks_dio_write() when DIO to write
3668  * to an unwritten extent.
3669  *
3670  * Writing to an unwritten extent may result in splitting the unwritten
3671  * extent into multiple initialized/unwritten extents (up to three)
3672  * There are three possibilities:
3673  *   a> There is no split required: Entire extent should be unwritten
3674  *   b> Splits in two extents: Write is happening at either end of the extent
3675  *   c> Splits in three extents: Somone is writing in middle of the extent
3676  *
3677  * This works the same way in the case of initialized -> unwritten conversion.
3678  *
3679  * One of more index blocks maybe needed if the extent tree grow after
3680  * the unwritten extent split. To prevent ENOSPC occur at the IO
3681  * complete, we need to split the unwritten extent before DIO submit
3682  * the IO. The unwritten extent called at this time will be split
3683  * into three unwritten extent(at most). After IO complete, the part
3684  * being filled will be convert to initialized by the end_io callback function
3685  * via ext4_convert_unwritten_extents().
3686  *
3687  * Returns the size of unwritten extent to be written on success.
3688  */
3689 static int ext4_split_convert_extents(handle_t *handle,
3690                                         struct inode *inode,
3691                                         struct ext4_map_blocks *map,
3692                                         struct ext4_ext_path **ppath,
3693                                         int flags)
3694 {
3695         struct ext4_ext_path *path = *ppath;
3696         ext4_lblk_t eof_block;
3697         ext4_lblk_t ee_block;
3698         struct ext4_extent *ex;
3699         unsigned int ee_len;
3700         int split_flag = 0, depth;
3701
3702         ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3703                   __func__, inode->i_ino,
3704                   (unsigned long long)map->m_lblk, map->m_len);
3705
3706         eof_block = (EXT4_I(inode)->i_disksize + inode->i_sb->s_blocksize - 1)
3707                         >> inode->i_sb->s_blocksize_bits;
3708         if (eof_block < map->m_lblk + map->m_len)
3709                 eof_block = map->m_lblk + map->m_len;
3710         /*
3711          * It is safe to convert extent to initialized via explicit
3712          * zeroout only if extent is fully insde i_size or new_size.
3713          */
3714         depth = ext_depth(inode);
3715         ex = path[depth].p_ext;
3716         ee_block = le32_to_cpu(ex->ee_block);
3717         ee_len = ext4_ext_get_actual_len(ex);
3718
3719         /* Convert to unwritten */
3720         if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3721                 split_flag |= EXT4_EXT_DATA_VALID1;
3722         /* Convert to initialized */
3723         } else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3724                 split_flag |= ee_block + ee_len <= eof_block ?
3725                               EXT4_EXT_MAY_ZEROOUT : 0;
3726                 split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3727         }
3728         flags |= EXT4_GET_BLOCKS_PRE_IO;
3729         return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3730 }
3731
3732 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3733                                                 struct inode *inode,
3734                                                 struct ext4_map_blocks *map,
3735                                                 struct ext4_ext_path **ppath)
3736 {
3737         struct ext4_ext_path *path = *ppath;
3738         struct ext4_extent *ex;
3739         ext4_lblk_t ee_block;
3740         unsigned int ee_len;
3741         int depth;
3742         int err = 0;
3743
3744         depth = ext_depth(inode);
3745         ex = path[depth].p_ext;
3746         ee_block = le32_to_cpu(ex->ee_block);
3747         ee_len = ext4_ext_get_actual_len(ex);
3748
3749         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3750                 "block %llu, max_blocks %u\n", inode->i_ino,
3751                   (unsigned long long)ee_block, ee_len);
3752
3753         /* If extent is larger than requested it is a clear sign that we still
3754          * have some extent state machine issues left. So extent_split is still
3755          * required.
3756          * TODO: Once all related issues will be fixed this situation should be
3757          * illegal.
3758          */
3759         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3760 #ifdef CONFIG_EXT4_DEBUG
3761                 ext4_warning(inode->i_sb, "Inode (%ld) finished: extent logical block %llu,"
3762                              " len %u; IO logical block %llu, len %u",
3763                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3764                              (unsigned long long)map->m_lblk, map->m_len);
3765 #endif
3766                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3767                                                  EXT4_GET_BLOCKS_CONVERT);
3768                 if (err < 0)
3769                         return err;
3770                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3771                 if (IS_ERR(path))
3772                         return PTR_ERR(path);
3773                 depth = ext_depth(inode);
3774                 ex = path[depth].p_ext;
3775         }
3776
3777         err = ext4_ext_get_access(handle, inode, path + depth);
3778         if (err)
3779                 goto out;
3780         /* first mark the extent as initialized */
3781         ext4_ext_mark_initialized(ex);
3782
3783         /* note: ext4_ext_correct_indexes() isn't needed here because
3784          * borders are not changed
3785          */
3786         ext4_ext_try_to_merge(handle, inode, path, ex);
3787
3788         /* Mark modified extent as dirty */
3789         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3790 out:
3791         ext4_ext_show_leaf(inode, path);
3792         return err;
3793 }
3794
3795 /*
3796  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3797  */
3798 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3799                               ext4_lblk_t lblk,
3800                               struct ext4_ext_path *path,
3801                               unsigned int len)
3802 {
3803         int i, depth;
3804         struct ext4_extent_header *eh;
3805         struct ext4_extent *last_ex;
3806
3807         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3808                 return 0;
3809
3810         depth = ext_depth(inode);
3811         eh = path[depth].p_hdr;
3812
3813         /*
3814          * We're going to remove EOFBLOCKS_FL entirely in future so we
3815          * do not care for this case anymore. Simply remove the flag
3816          * if there are no extents.
3817          */
3818         if (unlikely(!eh->eh_entries))
3819                 goto out;
3820         last_ex = EXT_LAST_EXTENT(eh);
3821         /*
3822          * We should clear the EOFBLOCKS_FL flag if we are writing the
3823          * last block in the last extent in the file.  We test this by
3824          * first checking to see if the caller to
3825          * ext4_ext_get_blocks() was interested in the last block (or
3826          * a block beyond the last block) in the current extent.  If
3827          * this turns out to be false, we can bail out from this
3828          * function immediately.
3829          */
3830         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3831             ext4_ext_get_actual_len(last_ex))
3832                 return 0;
3833         /*
3834          * If the caller does appear to be planning to write at or
3835          * beyond the end of the current extent, we then test to see
3836          * if the current extent is the last extent in the file, by
3837          * checking to make sure it was reached via the rightmost node
3838          * at each level of the tree.
3839          */
3840         for (i = depth-1; i >= 0; i--)
3841                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3842                         return 0;
3843 out:
3844         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3845         return ext4_mark_inode_dirty(handle, inode);
3846 }
3847
3848 /**
3849  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3850  *
3851  * Return 1 if there is a delalloc block in the range, otherwise 0.
3852  */
3853 int ext4_find_delalloc_range(struct inode *inode,
3854                              ext4_lblk_t lblk_start,
3855                              ext4_lblk_t lblk_end)
3856 {
3857         struct extent_status es;
3858
3859         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3860         if (es.es_len == 0)
3861                 return 0; /* there is no delay extent in this tree */
3862         else if (es.es_lblk <= lblk_start &&
3863                  lblk_start < es.es_lblk + es.es_len)
3864                 return 1;
3865         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3866                 return 1;
3867         else
3868                 return 0;
3869 }
3870
3871 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3872 {
3873         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3874         ext4_lblk_t lblk_start, lblk_end;
3875         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3876         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3877
3878         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3879 }
3880
3881 /**
3882  * Determines how many complete clusters (out of those specified by the 'map')
3883  * are under delalloc and were reserved quota for.
3884  * This function is called when we are writing out the blocks that were
3885  * originally written with their allocation delayed, but then the space was
3886  * allocated using fallocate() before the delayed allocation could be resolved.
3887  * The cases to look for are:
3888  * ('=' indicated delayed allocated blocks
3889  *  '-' indicates non-delayed allocated blocks)
3890  * (a) partial clusters towards beginning and/or end outside of allocated range
3891  *     are not delalloc'ed.
3892  *      Ex:
3893  *      |----c---=|====c====|====c====|===-c----|
3894  *               |++++++ allocated ++++++|
3895  *      ==> 4 complete clusters in above example
3896  *
3897  * (b) partial cluster (outside of allocated range) towards either end is
3898  *     marked for delayed allocation. In this case, we will exclude that
3899  *     cluster.
3900  *      Ex:
3901  *      |----====c========|========c========|
3902  *           |++++++ allocated ++++++|
3903  *      ==> 1 complete clusters in above example
3904  *
3905  *      Ex:
3906  *      |================c================|
3907  *            |++++++ allocated ++++++|
3908  *      ==> 0 complete clusters in above example
3909  *
3910  * The ext4_da_update_reserve_space will be called only if we
3911  * determine here that there were some "entire" clusters that span
3912  * this 'allocated' range.
3913  * In the non-bigalloc case, this function will just end up returning num_blks
3914  * without ever calling ext4_find_delalloc_range.
3915  */
3916 static unsigned int
3917 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3918                            unsigned int num_blks)
3919 {
3920         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3921         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3922         ext4_lblk_t lblk_from, lblk_to, c_offset;
3923         unsigned int allocated_clusters = 0;
3924
3925         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3926         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3927
3928         /* max possible clusters for this allocation */
3929         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3930
3931         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3932
3933         /* Check towards left side */
3934         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3935         if (c_offset) {
3936                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3937                 lblk_to = lblk_from + c_offset - 1;
3938
3939                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3940                         allocated_clusters--;
3941         }
3942
3943         /* Now check towards right. */
3944         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3945         if (allocated_clusters && c_offset) {
3946                 lblk_from = lblk_start + num_blks;
3947                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3948
3949                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3950                         allocated_clusters--;
3951         }
3952
3953         return allocated_clusters;
3954 }
3955
3956 static int
3957 convert_initialized_extent(handle_t *handle, struct inode *inode,
3958                            struct ext4_map_blocks *map,
3959                            struct ext4_ext_path **ppath,
3960                            unsigned int allocated)
3961 {
3962         struct ext4_ext_path *path = *ppath;
3963         struct ext4_extent *ex;
3964         ext4_lblk_t ee_block;
3965         unsigned int ee_len;
3966         int depth;
3967         int err = 0;
3968
3969         /*
3970          * Make sure that the extent is no bigger than we support with
3971          * unwritten extent
3972          */
3973         if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3974                 map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3975
3976         depth = ext_depth(inode);
3977         ex = path[depth].p_ext;
3978         ee_block = le32_to_cpu(ex->ee_block);
3979         ee_len = ext4_ext_get_actual_len(ex);
3980
3981         ext_debug("%s: inode %lu, logical"
3982                 "block %llu, max_blocks %u\n", __func__, inode->i_ino,
3983                   (unsigned long long)ee_block, ee_len);
3984
3985         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3986                 err = ext4_split_convert_extents(handle, inode, map, ppath,
3987                                 EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3988                 if (err < 0)
3989                         return err;
3990                 path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3991                 if (IS_ERR(path))
3992                         return PTR_ERR(path);
3993                 depth = ext_depth(inode);
3994                 ex = path[depth].p_ext;
3995                 if (!ex) {
3996                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3997                                          (unsigned long) map->m_lblk);
3998                         return -EFSCORRUPTED;
3999                 }
4000         }
4001
4002         err = ext4_ext_get_access(handle, inode, path + depth);
4003         if (err)
4004                 return err;
4005         /* first mark the extent as unwritten */
4006         ext4_ext_mark_unwritten(ex);
4007
4008         /* note: ext4_ext_correct_indexes() isn't needed here because
4009          * borders are not changed
4010          */
4011         ext4_ext_try_to_merge(handle, inode, path, ex);
4012
4013         /* Mark modified extent as dirty */
4014         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
4015         if (err)
4016                 return err;
4017         ext4_ext_show_leaf(inode, path);
4018
4019         ext4_update_inode_fsync_trans(handle, inode, 1);
4020         err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
4021         if (err)
4022                 return err;
4023         map->m_flags |= EXT4_MAP_UNWRITTEN;
4024         if (allocated > map->m_len)
4025                 allocated = map->m_len;
4026         map->m_len = allocated;
4027         return allocated;
4028 }
4029
4030 static int
4031 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
4032                         struct ext4_map_blocks *map,
4033                         struct ext4_ext_path **ppath, int flags,
4034                         unsigned int allocated, ext4_fsblk_t newblock)
4035 {
4036         struct ext4_ext_path *path = *ppath;
4037         int ret = 0;
4038         int err = 0;
4039
4040         ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4041                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
4042                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4043                   flags, allocated);
4044         ext4_ext_show_leaf(inode, path);
4045
4046         /*
4047          * When writing into unwritten space, we should not fail to
4048          * allocate metadata blocks for the new extent block if needed.
4049          */
4050         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4051
4052         trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4053                                                     allocated, newblock);
4054
4055         /* get_block() before submit the IO, split the extent */
4056         if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4057                 ret = ext4_split_convert_extents(handle, inode, map, ppath,
4058                                          flags | EXT4_GET_BLOCKS_CONVERT);
4059                 if (ret <= 0)
4060                         goto out;
4061                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4062                 goto out;
4063         }
4064         /* IO end_io complete, convert the filled extent to written */
4065         if (flags & EXT4_GET_BLOCKS_CONVERT) {
4066                 if (flags & EXT4_GET_BLOCKS_ZERO) {
4067                         if (allocated > map->m_len)
4068                                 allocated = map->m_len;
4069                         err = ext4_issue_zeroout(inode, map->m_lblk, newblock,
4070                                                  allocated);
4071                         if (err < 0)
4072                                 goto out2;
4073                 }
4074                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4075                                                            ppath);
4076                 if (ret >= 0) {
4077                         ext4_update_inode_fsync_trans(handle, inode, 1);
4078                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
4079                                                  path, map->m_len);
4080                 } else
4081                         err = ret;
4082                 map->m_flags |= EXT4_MAP_MAPPED;
4083                 map->m_pblk = newblock;
4084                 if (allocated > map->m_len)
4085                         allocated = map->m_len;
4086                 map->m_len = allocated;
4087                 goto out2;
4088         }
4089         /* buffered IO case */
4090         /*
4091          * repeat fallocate creation request
4092          * we already have an unwritten extent
4093          */
4094         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4095                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4096                 goto map_out;
4097         }
4098
4099         /* buffered READ or buffered write_begin() lookup */
4100         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4101                 /*
4102                  * We have blocks reserved already.  We
4103                  * return allocated blocks so that delalloc
4104                  * won't do block reservation for us.  But
4105                  * the buffer head will be unmapped so that
4106                  * a read from the block returns 0s.
4107                  */
4108                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4109                 goto out1;
4110         }
4111
4112         /* buffered write, writepage time, convert*/
4113         ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4114         if (ret >= 0)
4115                 ext4_update_inode_fsync_trans(handle, inode, 1);
4116 out:
4117         if (ret <= 0) {
4118                 err = ret;
4119                 goto out2;
4120         } else
4121                 allocated = ret;
4122         map->m_flags |= EXT4_MAP_NEW;
4123         /*
4124          * if we allocated more blocks than requested
4125          * we need to make sure we unmap the extra block
4126          * allocated. The actual needed block will get
4127          * unmapped later when we find the buffer_head marked
4128          * new.
4129          */
4130         if (allocated > map->m_len) {
4131                 clean_bdev_aliases(inode->i_sb->s_bdev, newblock + map->m_len,
4132                                    allocated - map->m_len);
4133                 allocated = map->m_len;
4134         }
4135         map->m_len = allocated;
4136
4137         /*
4138          * If we have done fallocate with the offset that is already
4139          * delayed allocated, we would have block reservation
4140          * and quota reservation done in the delayed write path.
4141          * But fallocate would have already updated quota and block
4142          * count for this offset. So cancel these reservation
4143          */
4144         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4145                 unsigned int reserved_clusters;
4146                 reserved_clusters = get_reserved_cluster_alloc(inode,
4147                                 map->m_lblk, map->m_len);
4148                 if (reserved_clusters)
4149                         ext4_da_update_reserve_space(inode,
4150                                                      reserved_clusters,
4151                                                      0);
4152         }
4153
4154 map_out:
4155         map->m_flags |= EXT4_MAP_MAPPED;
4156         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4157                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4158                                          map->m_len);
4159                 if (err < 0)
4160                         goto out2;
4161         }
4162 out1:
4163         if (allocated > map->m_len)
4164                 allocated = map->m_len;
4165         ext4_ext_show_leaf(inode, path);
4166         map->m_pblk = newblock;
4167         map->m_len = allocated;
4168 out2:
4169         return err ? err : allocated;
4170 }
4171
4172 /*
4173  * get_implied_cluster_alloc - check to see if the requested
4174  * allocation (in the map structure) overlaps with a cluster already
4175  * allocated in an extent.
4176  *      @sb     The filesystem superblock structure
4177  *      @map    The requested lblk->pblk mapping
4178  *      @ex     The extent structure which might contain an implied
4179  *                      cluster allocation
4180  *
4181  * This function is called by ext4_ext_map_blocks() after we failed to
4182  * find blocks that were already in the inode's extent tree.  Hence,
4183  * we know that the beginning of the requested region cannot overlap
4184  * the extent from the inode's extent tree.  There are three cases we
4185  * want to catch.  The first is this case:
4186  *
4187  *               |--- cluster # N--|
4188  *    |--- extent ---|  |---- requested region ---|
4189  *                      |==========|
4190  *
4191  * The second case that we need to test for is this one:
4192  *
4193  *   |--------- cluster # N ----------------|
4194  *         |--- requested region --|   |------- extent ----|
4195  *         |=======================|
4196  *
4197  * The third case is when the requested region lies between two extents
4198  * within the same cluster:
4199  *          |------------- cluster # N-------------|
4200  * |----- ex -----|                  |---- ex_right ----|
4201  *                  |------ requested region ------|
4202  *                  |================|
4203  *
4204  * In each of the above cases, we need to set the map->m_pblk and
4205  * map->m_len so it corresponds to the return the extent labelled as
4206  * "|====|" from cluster #N, since it is already in use for data in
4207  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
4208  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4209  * as a new "allocated" block region.  Otherwise, we will return 0 and
4210  * ext4_ext_map_blocks() will then allocate one or more new clusters
4211  * by calling ext4_mb_new_blocks().
4212  */
4213 static int get_implied_cluster_alloc(struct super_block *sb,
4214                                      struct ext4_map_blocks *map,
4215                                      struct ext4_extent *ex,
4216                                      struct ext4_ext_path *path)
4217 {
4218         struct ext4_sb_info *sbi = EXT4_SB(sb);
4219         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4220         ext4_lblk_t ex_cluster_start, ex_cluster_end;
4221         ext4_lblk_t rr_cluster_start;
4222         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4223         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4224         unsigned short ee_len = ext4_ext_get_actual_len(ex);
4225
4226         /* The extent passed in that we are trying to match */
4227         ex_cluster_start = EXT4_B2C(sbi, ee_block);
4228         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4229
4230         /* The requested region passed into ext4_map_blocks() */
4231         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4232
4233         if ((rr_cluster_start == ex_cluster_end) ||
4234             (rr_cluster_start == ex_cluster_start)) {
4235                 if (rr_cluster_start == ex_cluster_end)
4236                         ee_start += ee_len - 1;
4237                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4238                 map->m_len = min(map->m_len,
4239                                  (unsigned) sbi->s_cluster_ratio - c_offset);
4240                 /*
4241                  * Check for and handle this case:
4242                  *
4243                  *   |--------- cluster # N-------------|
4244                  *                     |------- extent ----|
4245                  *         |--- requested region ---|
4246                  *         |===========|
4247                  */
4248
4249                 if (map->m_lblk < ee_block)
4250                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4251
4252                 /*
4253                  * Check for the case where there is already another allocated
4254                  * block to the right of 'ex' but before the end of the cluster.
4255                  *
4256                  *          |------------- cluster # N-------------|
4257                  * |----- ex -----|                  |---- ex_right ----|
4258                  *                  |------ requested region ------|
4259                  *                  |================|
4260                  */
4261                 if (map->m_lblk > ee_block) {
4262                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4263                         map->m_len = min(map->m_len, next - map->m_lblk);
4264                 }
4265
4266                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4267                 return 1;
4268         }
4269
4270         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4271         return 0;
4272 }
4273
4274
4275 /*
4276  * Block allocation/map/preallocation routine for extents based files
4277  *
4278  *
4279  * Need to be called with
4280  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4281  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4282  *
4283  * return > 0, number of of blocks already mapped/allocated
4284  *          if create == 0 and these are pre-allocated blocks
4285  *              buffer head is unmapped
4286  *          otherwise blocks are mapped
4287  *
4288  * return = 0, if plain look up failed (blocks have not been allocated)
4289  *          buffer head is unmapped
4290  *
4291  * return < 0, error case.
4292  */
4293 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4294                         struct ext4_map_blocks *map, int flags)
4295 {
4296         struct ext4_ext_path *path = NULL;
4297         struct ext4_extent newex, *ex, *ex2;
4298         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4299         ext4_fsblk_t newblock = 0;
4300         int free_on_err = 0, err = 0, depth, ret;
4301         unsigned int allocated = 0, offset = 0;
4302         unsigned int allocated_clusters = 0;
4303         struct ext4_allocation_request ar;
4304         ext4_lblk_t cluster_offset;
4305         bool map_from_cluster = false;
4306
4307         ext_debug("blocks %u/%u requested for inode %lu\n",
4308                   map->m_lblk, map->m_len, inode->i_ino);
4309         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4310
4311         /* find extent for this block */
4312         path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4313         if (IS_ERR(path)) {
4314                 err = PTR_ERR(path);
4315                 path = NULL;
4316                 goto out2;
4317         }
4318
4319         depth = ext_depth(inode);
4320
4321         /*
4322          * consistent leaf must not be empty;
4323          * this situation is possible, though, _during_ tree modification;
4324          * this is why assert can't be put in ext4_find_extent()
4325          */
4326         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4327                 EXT4_ERROR_INODE(inode, "bad extent address "
4328                                  "lblock: %lu, depth: %d pblock %lld",
4329                                  (unsigned long) map->m_lblk, depth,
4330                                  path[depth].p_block);
4331                 err = -EFSCORRUPTED;
4332                 goto out2;
4333         }
4334
4335         ex = path[depth].p_ext;
4336         if (ex) {
4337                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4338                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4339                 unsigned short ee_len;
4340
4341
4342                 /*
4343                  * unwritten extents are treated as holes, except that
4344                  * we split out initialized portions during a write.
4345                  */
4346                 ee_len = ext4_ext_get_actual_len(ex);
4347
4348                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4349
4350                 /* if found extent covers block, simply return it */
4351                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4352                         newblock = map->m_lblk - ee_block + ee_start;
4353                         /* number of remaining blocks in the extent */
4354                         allocated = ee_len - (map->m_lblk - ee_block);
4355                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4356                                   ee_block, ee_len, newblock);
4357
4358                         /*
4359                          * If the extent is initialized check whether the
4360                          * caller wants to convert it to unwritten.
4361                          */
4362                         if ((!ext4_ext_is_unwritten(ex)) &&
4363                             (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4364                                 allocated = convert_initialized_extent(
4365                                                 handle, inode, map, &path,
4366                                                 allocated);
4367                                 goto out2;
4368                         } else if (!ext4_ext_is_unwritten(ex))
4369                                 goto out;
4370
4371                         ret = ext4_ext_handle_unwritten_extents(
4372                                 handle, inode, map, &path, flags,
4373                                 allocated, newblock);
4374                         if (ret < 0)
4375                                 err = ret;
4376                         else
4377                                 allocated = ret;
4378                         goto out2;
4379                 }
4380         }
4381
4382         /*
4383          * requested block isn't allocated yet;
4384          * we couldn't try to create block if create flag is zero
4385          */
4386         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4387                 ext4_lblk_t hole_start, hole_len;
4388
4389                 hole_start = map->m_lblk;
4390                 hole_len = ext4_ext_determine_hole(inode, path, &hole_start);
4391                 /*
4392                  * put just found gap into cache to speed up
4393                  * subsequent requests
4394                  */
4395                 ext4_ext_put_gap_in_cache(inode, hole_start, hole_len);
4396
4397                 /* Update hole_len to reflect hole size after map->m_lblk */
4398                 if (hole_start != map->m_lblk)
4399                         hole_len -= map->m_lblk - hole_start;
4400                 map->m_pblk = 0;
4401                 map->m_len = min_t(unsigned int, map->m_len, hole_len);
4402
4403                 goto out2;
4404         }
4405
4406         /*
4407          * Okay, we need to do block allocation.
4408          */
4409         newex.ee_block = cpu_to_le32(map->m_lblk);
4410         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4411
4412         /*
4413          * If we are doing bigalloc, check to see if the extent returned
4414          * by ext4_find_extent() implies a cluster we can use.
4415          */
4416         if (cluster_offset && ex &&
4417             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4418                 ar.len = allocated = map->m_len;
4419                 newblock = map->m_pblk;
4420                 map_from_cluster = true;
4421                 goto got_allocated_blocks;
4422         }
4423
4424         /* find neighbour allocated blocks */
4425         ar.lleft = map->m_lblk;
4426         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4427         if (err)
4428                 goto out2;
4429         ar.lright = map->m_lblk;
4430         ex2 = NULL;
4431         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4432         if (err)
4433                 goto out2;
4434
4435         /* Check if the extent after searching to the right implies a
4436          * cluster we can use. */
4437         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4438             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4439                 ar.len = allocated = map->m_len;
4440                 newblock = map->m_pblk;
4441                 map_from_cluster = true;
4442                 goto got_allocated_blocks;
4443         }
4444
4445         /*
4446          * See if request is beyond maximum number of blocks we can have in
4447          * a single extent. For an initialized extent this limit is
4448          * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4449          * EXT_UNWRITTEN_MAX_LEN.
4450          */
4451         if (map->m_len > EXT_INIT_MAX_LEN &&
4452             !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4453                 map->m_len = EXT_INIT_MAX_LEN;
4454         else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4455                  (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4456                 map->m_len = EXT_UNWRITTEN_MAX_LEN;
4457
4458         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4459         newex.ee_len = cpu_to_le16(map->m_len);
4460         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4461         if (err)
4462                 allocated = ext4_ext_get_actual_len(&newex);
4463         else
4464                 allocated = map->m_len;
4465
4466         /* allocate new block */
4467         ar.inode = inode;
4468         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4469         ar.logical = map->m_lblk;
4470         /*
4471          * We calculate the offset from the beginning of the cluster
4472          * for the logical block number, since when we allocate a
4473          * physical cluster, the physical block should start at the
4474          * same offset from the beginning of the cluster.  This is
4475          * needed so that future calls to get_implied_cluster_alloc()
4476          * work correctly.
4477          */
4478         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4479         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4480         ar.goal -= offset;
4481         ar.logical -= offset;
4482         if (S_ISREG(inode->i_mode))
4483                 ar.flags = EXT4_MB_HINT_DATA;
4484         else
4485                 /* disable in-core preallocation for non-regular files */
4486                 ar.flags = 0;
4487         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4488                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4489         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4490                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4491         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
4492                 ar.flags |= EXT4_MB_USE_RESERVED;
4493         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4494         if (!newblock)
4495                 goto out2;
4496         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4497                   ar.goal, newblock, allocated);
4498         free_on_err = 1;
4499         allocated_clusters = ar.len;
4500         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4501         if (ar.len > allocated)
4502                 ar.len = allocated;
4503
4504 got_allocated_blocks:
4505         /* try to insert new extent into found leaf and return */
4506         ext4_ext_store_pblock(&newex, newblock + offset);
4507         newex.ee_len = cpu_to_le16(ar.len);
4508         /* Mark unwritten */
4509         if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4510                 ext4_ext_mark_unwritten(&newex);
4511                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4512         }
4513
4514         err = 0;
4515         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4516                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4517                                          path, ar.len);
4518         if (!err)
4519                 err = ext4_ext_insert_extent(handle, inode, &path,
4520                                              &newex, flags);
4521
4522         if (err && free_on_err) {
4523                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4524                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4525                 /* free data blocks we just allocated */
4526                 /* not a good idea to call discard here directly,
4527                  * but otherwise we'd need to call it every free() */
4528                 ext4_discard_preallocations(inode);
4529                 ext4_free_blocks(handle, inode, NULL, newblock,
4530                                  EXT4_C2B(sbi, allocated_clusters), fb_flags);
4531                 goto out2;
4532         }
4533
4534         /* previous routine could use block we allocated */
4535         newblock = ext4_ext_pblock(&newex);
4536         allocated = ext4_ext_get_actual_len(&newex);
4537         if (allocated > map->m_len)
4538                 allocated = map->m_len;
4539         map->m_flags |= EXT4_MAP_NEW;
4540
4541         /*
4542          * Update reserved blocks/metadata blocks after successful
4543          * block allocation which had been deferred till now.
4544          */
4545         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4546                 unsigned int reserved_clusters;
4547                 /*
4548                  * Check how many clusters we had reserved this allocated range
4549                  */
4550                 reserved_clusters = get_reserved_cluster_alloc(inode,
4551                                                 map->m_lblk, allocated);
4552                 if (!map_from_cluster) {
4553                         BUG_ON(allocated_clusters < reserved_clusters);
4554                         if (reserved_clusters < allocated_clusters) {
4555                                 struct ext4_inode_info *ei = EXT4_I(inode);
4556                                 int reservation = allocated_clusters -
4557                                                   reserved_clusters;
4558                                 /*
4559                                  * It seems we claimed few clusters outside of
4560                                  * the range of this allocation. We should give
4561                                  * it back to the reservation pool. This can
4562                                  * happen in the following case:
4563                                  *
4564                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4565                                  *   cluster has 4 blocks. Thus, the clusters
4566                                  *   are [0-3],[4-7],[8-11]...
4567                                  * * First comes delayed allocation write for
4568                                  *   logical blocks 10 & 11. Since there were no
4569                                  *   previous delayed allocated blocks in the
4570                                  *   range [8-11], we would reserve 1 cluster
4571                                  *   for this write.
4572                                  * * Next comes write for logical blocks 3 to 8.
4573                                  *   In this case, we will reserve 2 clusters
4574                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4575                                  *   that range has a delayed allocated blocks.
4576                                  *   Thus total reserved clusters now becomes 3.
4577                                  * * Now, during the delayed allocation writeout
4578                                  *   time, we will first write blocks [3-8] and
4579                                  *   allocate 3 clusters for writing these
4580                                  *   blocks. Also, we would claim all these
4581                                  *   three clusters above.
4582                                  * * Now when we come here to writeout the
4583                                  *   blocks [10-11], we would expect to claim
4584                                  *   the reservation of 1 cluster we had made
4585                                  *   (and we would claim it since there are no
4586                                  *   more delayed allocated blocks in the range
4587                                  *   [8-11]. But our reserved cluster count had
4588                                  *   already gone to 0.
4589                                  *
4590                                  *   Thus, at the step 4 above when we determine
4591                                  *   that there are still some unwritten delayed
4592                                  *   allocated blocks outside of our current
4593                                  *   block range, we should increment the
4594                                  *   reserved clusters count so that when the
4595                                  *   remaining blocks finally gets written, we
4596                                  *   could claim them.
4597                                  */
4598                                 dquot_reserve_block(inode,
4599                                                 EXT4_C2B(sbi, reservation));
4600                                 spin_lock(&ei->i_block_reservation_lock);
4601                                 ei->i_reserved_data_blocks += reservation;
4602                                 spin_unlock(&ei->i_block_reservation_lock);
4603                         }
4604                         /*
4605                          * We will claim quota for all newly allocated blocks.
4606                          * We're updating the reserved space *after* the
4607                          * correction above so we do not accidentally free
4608                          * all the metadata reservation because we might
4609                          * actually need it later on.
4610                          */
4611                         ext4_da_update_reserve_space(inode, allocated_clusters,
4612                                                         1);
4613                 }
4614         }
4615
4616         /*
4617          * Cache the extent and update transaction to commit on fdatasync only
4618          * when it is _not_ an unwritten extent.
4619          */
4620         if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4621                 ext4_update_inode_fsync_trans(handle, inode, 1);
4622         else
4623                 ext4_update_inode_fsync_trans(handle, inode, 0);
4624 out:
4625         if (allocated > map->m_len)
4626                 allocated = map->m_len;
4627         ext4_ext_show_leaf(inode, path);
4628         map->m_flags |= EXT4_MAP_MAPPED;
4629         map->m_pblk = newblock;
4630         map->m_len = allocated;
4631 out2:
4632         ext4_ext_drop_refs(path);
4633         kfree(path);
4634
4635         trace_ext4_ext_map_blocks_exit(inode, flags, map,
4636                                        err ? err : allocated);
4637         return err ? err : allocated;
4638 }
4639
4640 int ext4_ext_truncate(handle_t *handle, struct inode *inode)
4641 {
4642         struct super_block *sb = inode->i_sb;
4643         ext4_lblk_t last_block;
4644         int err = 0;
4645
4646         /*
4647          * TODO: optimization is possible here.
4648          * Probably we need not scan at all,
4649          * because page truncation is enough.
4650          */
4651
4652         /* we have to know where to truncate from in crash case */
4653         EXT4_I(inode)->i_disksize = inode->i_size;
4654         err = ext4_mark_inode_dirty(handle, inode);
4655         if (err)
4656                 return err;
4657
4658         last_block = (inode->i_size + sb->s_blocksize - 1)
4659                         >> EXT4_BLOCK_SIZE_BITS(sb);
4660 retry:
4661         err = ext4_es_remove_extent(inode, last_block,
4662                                     EXT_MAX_BLOCKS - last_block);
4663         if (err == -ENOMEM) {
4664                 cond_resched();
4665                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4666                 goto retry;
4667         }
4668         if (err)
4669                 return err;
4670         return ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4671 }
4672
4673 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4674                                   ext4_lblk_t len, loff_t new_size,
4675                                   int flags)
4676 {
4677         struct inode *inode = file_inode(file);
4678         handle_t *handle;
4679         int ret = 0;
4680         int ret2 = 0;
4681         int retries = 0;
4682         int depth = 0;
4683         struct ext4_map_blocks map;
4684         unsigned int credits;
4685         loff_t epos;
4686
4687         BUG_ON(!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS));
4688         map.m_lblk = offset;
4689         map.m_len = len;
4690         /*
4691          * Don't normalize the request if it can fit in one extent so
4692          * that it doesn't get unnecessarily split into multiple
4693          * extents.
4694          */
4695         if (len <= EXT_UNWRITTEN_MAX_LEN)
4696                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4697
4698         /*
4699          * credits to insert 1 extent into extent tree
4700          */
4701         credits = ext4_chunk_trans_blocks(inode, len);
4702         depth = ext_depth(inode);
4703
4704 retry:
4705         while (ret >= 0 && len) {
4706                 /*
4707                  * Recalculate credits when extent tree depth changes.
4708                  */
4709                 if (depth != ext_depth(inode)) {
4710                         credits = ext4_chunk_trans_blocks(inode, len);
4711                         depth = ext_depth(inode);
4712                 }
4713
4714                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4715                                             credits);
4716                 if (IS_ERR(handle)) {
4717                         ret = PTR_ERR(handle);
4718                         break;
4719                 }
4720                 ret = ext4_map_blocks(handle, inode, &map, flags);
4721                 if (ret <= 0) {
4722                         ext4_debug("inode #%lu: block %u: len %u: "
4723                                    "ext4_ext_map_blocks returned %d",
4724                                    inode->i_ino, map.m_lblk,
4725                                    map.m_len, ret);
4726                         ext4_mark_inode_dirty(handle, inode);
4727                         ret2 = ext4_journal_stop(handle);
4728                         break;
4729                 }
4730                 map.m_lblk += ret;
4731                 map.m_len = len = len - ret;
4732                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
4733                 inode->i_ctime = current_time(inode);
4734                 if (new_size) {
4735                         if (epos > new_size)
4736                                 epos = new_size;
4737                         if (ext4_update_inode_size(inode, epos) & 0x1)
4738                                 inode->i_mtime = inode->i_ctime;
4739                 } else {
4740                         if (epos > inode->i_size)
4741                                 ext4_set_inode_flag(inode,
4742                                                     EXT4_INODE_EOFBLOCKS);
4743                 }
4744                 ext4_mark_inode_dirty(handle, inode);
4745                 ext4_update_inode_fsync_trans(handle, inode, 1);
4746                 ret2 = ext4_journal_stop(handle);
4747                 if (ret2)
4748                         break;
4749         }
4750         if (ret == -ENOSPC &&
4751                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4752                 ret = 0;
4753                 goto retry;
4754         }
4755
4756         return ret > 0 ? ret2 : ret;
4757 }
4758
4759 static long ext4_zero_range(struct file *file, loff_t offset,
4760                             loff_t len, int mode)
4761 {
4762         struct inode *inode = file_inode(file);
4763         handle_t *handle = NULL;
4764         unsigned int max_blocks;
4765         loff_t new_size = 0;
4766         int ret = 0;
4767         int flags;
4768         int credits;
4769         int partial_begin, partial_end;
4770         loff_t start, end;
4771         ext4_lblk_t lblk;
4772         unsigned int blkbits = inode->i_blkbits;
4773
4774         trace_ext4_zero_range(inode, offset, len, mode);
4775
4776         if (!S_ISREG(inode->i_mode))
4777                 return -EINVAL;
4778
4779         /* Call ext4_force_commit to flush all data in case of data=journal. */
4780         if (ext4_should_journal_data(inode)) {
4781                 ret = ext4_force_commit(inode->i_sb);
4782                 if (ret)
4783                         return ret;
4784         }
4785
4786         /*
4787          * Round up offset. This is not fallocate, we neet to zero out
4788          * blocks, so convert interior block aligned part of the range to
4789          * unwritten and possibly manually zero out unaligned parts of the
4790          * range.
4791          */
4792         start = round_up(offset, 1 << blkbits);
4793         end = round_down((offset + len), 1 << blkbits);
4794
4795         if (start < offset || end > offset + len)
4796                 return -EINVAL;
4797         partial_begin = offset & ((1 << blkbits) - 1);
4798         partial_end = (offset + len) & ((1 << blkbits) - 1);
4799
4800         lblk = start >> blkbits;
4801         max_blocks = (end >> blkbits);
4802         if (max_blocks < lblk)
4803                 max_blocks = 0;
4804         else
4805                 max_blocks -= lblk;
4806
4807         inode_lock(inode);
4808
4809         /*
4810          * Indirect files do not support unwritten extnets
4811          */
4812         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4813                 ret = -EOPNOTSUPP;
4814                 goto out_mutex;
4815         }
4816
4817         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4818             (offset + len > i_size_read(inode) ||
4819              offset + len > EXT4_I(inode)->i_disksize)) {
4820                 new_size = offset + len;
4821                 ret = inode_newsize_ok(inode, new_size);
4822                 if (ret)
4823                         goto out_mutex;
4824         }
4825
4826         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4827         if (mode & FALLOC_FL_KEEP_SIZE)
4828                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4829
4830         /* Wait all existing dio workers, newcomers will block on i_mutex */
4831         inode_dio_wait(inode);
4832
4833         /* Preallocate the range including the unaligned edges */
4834         if (partial_begin || partial_end) {
4835                 ret = ext4_alloc_file_blocks(file,
4836                                 round_down(offset, 1 << blkbits) >> blkbits,
4837                                 (round_up((offset + len), 1 << blkbits) -
4838                                  round_down(offset, 1 << blkbits)) >> blkbits,
4839                                 new_size, flags);
4840                 if (ret)
4841                         goto out_mutex;
4842
4843         }
4844
4845         /* Zero range excluding the unaligned edges */
4846         if (max_blocks > 0) {
4847                 flags |= (EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4848                           EXT4_EX_NOCACHE);
4849
4850                 /*
4851                  * Prevent page faults from reinstantiating pages we have
4852                  * released from page cache.
4853                  */
4854                 down_write(&EXT4_I(inode)->i_mmap_sem);
4855
4856                 ret = ext4_break_layouts(inode);
4857                 if (ret) {
4858                         up_write(&EXT4_I(inode)->i_mmap_sem);
4859                         goto out_mutex;
4860                 }
4861
4862                 ret = ext4_update_disksize_before_punch(inode, offset, len);
4863                 if (ret) {
4864                         up_write(&EXT4_I(inode)->i_mmap_sem);
4865                         goto out_mutex;
4866                 }
4867                 /* Now release the pages and zero block aligned part of pages */
4868                 truncate_pagecache_range(inode, start, end - 1);
4869                 inode->i_mtime = inode->i_ctime = current_time(inode);
4870
4871                 ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4872                                              flags);
4873                 up_write(&EXT4_I(inode)->i_mmap_sem);
4874                 if (ret)
4875                         goto out_mutex;
4876         }
4877         if (!partial_begin && !partial_end)
4878                 goto out_mutex;
4879
4880         /*
4881          * In worst case we have to writeout two nonadjacent unwritten
4882          * blocks and update the inode
4883          */
4884         credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4885         if (ext4_should_journal_data(inode))
4886                 credits += 2;
4887         handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4888         if (IS_ERR(handle)) {
4889                 ret = PTR_ERR(handle);
4890                 ext4_std_error(inode->i_sb, ret);
4891                 goto out_mutex;
4892         }
4893
4894         inode->i_mtime = inode->i_ctime = current_time(inode);
4895         if (new_size) {
4896                 ext4_update_inode_size(inode, new_size);
4897         } else {
4898                 /*
4899                 * Mark that we allocate beyond EOF so the subsequent truncate
4900                 * can proceed even if the new size is the same as i_size.
4901                 */
4902                 if ((offset + len) > i_size_read(inode))
4903                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4904         }
4905         ext4_mark_inode_dirty(handle, inode);
4906
4907         /* Zero out partial block at the edges of the range */
4908         ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4909         if (ret >= 0)
4910                 ext4_update_inode_fsync_trans(handle, inode, 1);
4911
4912         if (file->f_flags & O_SYNC)
4913                 ext4_handle_sync(handle);
4914
4915         ext4_journal_stop(handle);
4916 out_mutex:
4917         inode_unlock(inode);
4918         return ret;
4919 }
4920
4921 /*
4922  * preallocate space for a file. This implements ext4's fallocate file
4923  * operation, which gets called from sys_fallocate system call.
4924  * For block-mapped files, posix_fallocate should fall back to the method
4925  * of writing zeroes to the required new blocks (the same behavior which is
4926  * expected for file systems which do not support fallocate() system call).
4927  */
4928 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4929 {
4930         struct inode *inode = file_inode(file);
4931         loff_t new_size = 0;
4932         unsigned int max_blocks;
4933         int ret = 0;
4934         int flags;
4935         ext4_lblk_t lblk;
4936         unsigned int blkbits = inode->i_blkbits;
4937
4938         /*
4939          * Encrypted inodes can't handle collapse range or insert
4940          * range since we would need to re-encrypt blocks with a
4941          * different IV or XTS tweak (which are based on the logical
4942          * block number).
4943          *
4944          * XXX It's not clear why zero range isn't working, but we'll
4945          * leave it disabled for encrypted inodes for now.  This is a
4946          * bug we should fix....
4947          */
4948         if (ext4_encrypted_inode(inode) &&
4949             (mode & (FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_INSERT_RANGE |
4950                      FALLOC_FL_ZERO_RANGE)))
4951                 return -EOPNOTSUPP;
4952
4953         /* Return error if mode is not supported */
4954         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4955                      FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |
4956                      FALLOC_FL_INSERT_RANGE))
4957                 return -EOPNOTSUPP;
4958
4959         if (mode & FALLOC_FL_PUNCH_HOLE)
4960                 return ext4_punch_hole(inode, offset, len);
4961
4962         ret = ext4_convert_inline_data(inode);
4963         if (ret)
4964                 return ret;
4965
4966         if (mode & FALLOC_FL_COLLAPSE_RANGE)
4967                 return ext4_collapse_range(inode, offset, len);
4968
4969         if (mode & FALLOC_FL_INSERT_RANGE)
4970                 return ext4_insert_range(inode, offset, len);
4971
4972         if (mode & FALLOC_FL_ZERO_RANGE)
4973                 return ext4_zero_range(file, offset, len, mode);
4974
4975         trace_ext4_fallocate_enter(inode, offset, len, mode);
4976         lblk = offset >> blkbits;
4977
4978         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
4979         flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4980         if (mode & FALLOC_FL_KEEP_SIZE)
4981                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4982
4983         inode_lock(inode);
4984
4985         /*
4986          * We only support preallocation for extent-based files only
4987          */
4988         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4989                 ret = -EOPNOTSUPP;
4990                 goto out;
4991         }
4992
4993         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4994             (offset + len > i_size_read(inode) ||
4995              offset + len > EXT4_I(inode)->i_disksize)) {
4996                 new_size = offset + len;
4997                 ret = inode_newsize_ok(inode, new_size);
4998                 if (ret)
4999                         goto out;
5000         }
5001
5002         /* Wait all existing dio workers, newcomers will block on i_mutex */
5003         inode_dio_wait(inode);
5004
5005         ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size, flags);
5006         if (ret)
5007                 goto out;
5008
5009         if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
5010                 ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
5011                                                 EXT4_I(inode)->i_sync_tid);
5012         }
5013 out:
5014         inode_unlock(inode);
5015         trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
5016         return ret;
5017 }
5018
5019 /*
5020  * This function convert a range of blocks to written extents
5021  * The caller of this function will pass the start offset and the size.
5022  * all unwritten extents within this range will be converted to
5023  * written extents.
5024  *
5025  * This function is called from the direct IO end io call back
5026  * function, to convert the fallocated extents after IO is completed.
5027  * Returns 0 on success.
5028  */
5029 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
5030                                    loff_t offset, ssize_t len)
5031 {
5032         unsigned int max_blocks;
5033         int ret = 0;
5034         int ret2 = 0;
5035         struct ext4_map_blocks map;
5036         unsigned int credits, blkbits = inode->i_blkbits;
5037
5038         map.m_lblk = offset >> blkbits;
5039         max_blocks = EXT4_MAX_BLOCKS(len, offset, blkbits);
5040
5041         /*
5042          * This is somewhat ugly but the idea is clear: When transaction is
5043          * reserved, everything goes into it. Otherwise we rather start several
5044          * smaller transactions for conversion of each extent separately.
5045          */
5046         if (handle) {
5047                 handle = ext4_journal_start_reserved(handle,
5048                                                      EXT4_HT_EXT_CONVERT);
5049                 if (IS_ERR(handle))
5050                         return PTR_ERR(handle);
5051                 credits = 0;
5052         } else {
5053                 /*
5054                  * credits to insert 1 extent into extent tree
5055                  */
5056                 credits = ext4_chunk_trans_blocks(inode, max_blocks);
5057         }
5058         while (ret >= 0 && ret < max_blocks) {
5059                 map.m_lblk += ret;
5060                 map.m_len = (max_blocks -= ret);
5061                 if (credits) {
5062                         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5063                                                     credits);
5064                         if (IS_ERR(handle)) {
5065                                 ret = PTR_ERR(handle);
5066                                 break;
5067                         }
5068                 }
5069                 ret = ext4_map_blocks(handle, inode, &map,
5070                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5071                 if (ret <= 0)
5072                         ext4_warning(inode->i_sb,
5073                                      "inode #%lu: block %u: len %u: "
5074                                      "ext4_ext_map_blocks returned %d",
5075                                      inode->i_ino, map.m_lblk,
5076                                      map.m_len, ret);
5077                 ext4_mark_inode_dirty(handle, inode);
5078                 if (credits)
5079                         ret2 = ext4_journal_stop(handle);
5080                 if (ret <= 0 || ret2)
5081                         break;
5082         }
5083         if (!credits)
5084                 ret2 = ext4_journal_stop(handle);
5085         return ret > 0 ? ret2 : ret;
5086 }
5087
5088 /*
5089  * If newes is not existing extent (newes->ec_pblk equals zero) find
5090  * delayed extent at start of newes and update newes accordingly and
5091  * return start of the next delayed extent.
5092  *
5093  * If newes is existing extent (newes->ec_pblk is not equal zero)
5094  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5095  * extent found. Leave newes unmodified.
5096  */
5097 static int ext4_find_delayed_extent(struct inode *inode,
5098                                     struct extent_status *newes)
5099 {
5100         struct extent_status es;
5101         ext4_lblk_t block, next_del;
5102
5103         if (newes->es_pblk == 0) {
5104                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5105                                 newes->es_lblk + newes->es_len - 1, &es);
5106
5107                 /*
5108                  * No extent in extent-tree contains block @newes->es_pblk,
5109                  * then the block may stay in 1)a hole or 2)delayed-extent.
5110                  */
5111                 if (es.es_len == 0)
5112                         /* A hole found. */
5113                         return 0;
5114
5115                 if (es.es_lblk > newes->es_lblk) {
5116                         /* A hole found. */
5117                         newes->es_len = min(es.es_lblk - newes->es_lblk,
5118                                             newes->es_len);
5119                         return 0;
5120                 }
5121
5122                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5123         }
5124
5125         block = newes->es_lblk + newes->es_len;
5126         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5127         if (es.es_len == 0)
5128                 next_del = EXT_MAX_BLOCKS;
5129         else
5130                 next_del = es.es_lblk;
5131
5132         return next_del;
5133 }
5134 /* fiemap flags we can handle specified here */
5135 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5136
5137 static int ext4_xattr_fiemap(struct inode *inode,
5138                                 struct fiemap_extent_info *fieinfo)
5139 {
5140         __u64 physical = 0;
5141         __u64 length;
5142         __u32 flags = FIEMAP_EXTENT_LAST;
5143         int blockbits = inode->i_sb->s_blocksize_bits;
5144         int error = 0;
5145
5146         /* in-inode? */
5147         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5148                 struct ext4_iloc iloc;
5149                 int offset;     /* offset of xattr in inode */
5150
5151                 error = ext4_get_inode_loc(inode, &iloc);
5152                 if (error)
5153                         return error;
5154                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
5155                 offset = EXT4_GOOD_OLD_INODE_SIZE +
5156                                 EXT4_I(inode)->i_extra_isize;
5157                 physical += offset;
5158                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5159                 flags |= FIEMAP_EXTENT_DATA_INLINE;
5160                 brelse(iloc.bh);
5161         } else { /* external block */
5162                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5163                 length = inode->i_sb->s_blocksize;
5164         }
5165
5166         if (physical)
5167                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
5168                                                 length, flags);
5169         return (error < 0 ? error : 0);
5170 }
5171
5172 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5173                 __u64 start, __u64 len)
5174 {
5175         ext4_lblk_t start_blk;
5176         int error = 0;
5177
5178         if (ext4_has_inline_data(inode)) {
5179                 int has_inline = 1;
5180
5181                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline,
5182                                                 start, len);
5183
5184                 if (has_inline)
5185                         return error;
5186         }
5187
5188         if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5189                 error = ext4_ext_precache(inode);
5190                 if (error)
5191                         return error;
5192         }
5193
5194         /* fallback to generic here if not in extents fmt */
5195         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5196                 return generic_block_fiemap(inode, fieinfo, start, len,
5197                         ext4_get_block);
5198
5199         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5200                 return -EBADR;
5201
5202         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5203                 error = ext4_xattr_fiemap(inode, fieinfo);
5204         } else {
5205                 ext4_lblk_t len_blks;
5206                 __u64 last_blk;
5207
5208                 start_blk = start >> inode->i_sb->s_blocksize_bits;
5209                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5210                 if (last_blk >= EXT_MAX_BLOCKS)
5211                         last_blk = EXT_MAX_BLOCKS-1;
5212                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5213
5214                 /*
5215                  * Walk the extent tree gathering extent information
5216                  * and pushing extents back to the user.
5217                  */
5218                 error = ext4_fill_fiemap_extents(inode, start_blk,
5219                                                  len_blks, fieinfo);
5220         }
5221         return error;
5222 }
5223
5224 /*
5225  * ext4_access_path:
5226  * Function to access the path buffer for marking it dirty.
5227  * It also checks if there are sufficient credits left in the journal handle
5228  * to update path.
5229  */
5230 static int
5231 ext4_access_path(handle_t *handle, struct inode *inode,
5232                 struct ext4_ext_path *path)
5233 {
5234         int credits, err;
5235
5236         if (!ext4_handle_valid(handle))
5237                 return 0;
5238
5239         /*
5240          * Check if need to extend journal credits
5241          * 3 for leaf, sb, and inode plus 2 (bmap and group
5242          * descriptor) for each block group; assume two block
5243          * groups
5244          */
5245         if (handle->h_buffer_credits < 7) {
5246                 credits = ext4_writepage_trans_blocks(inode);
5247                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5248                 /* EAGAIN is success */
5249                 if (err && err != -EAGAIN)
5250                         return err;
5251         }
5252
5253         err = ext4_ext_get_access(handle, inode, path);
5254         return err;
5255 }
5256
5257 /*
5258  * ext4_ext_shift_path_extents:
5259  * Shift the extents of a path structure lying between path[depth].p_ext
5260  * and EXT_LAST_EXTENT(path[depth].p_hdr), by @shift blocks. @SHIFT tells
5261  * if it is right shift or left shift operation.
5262  */
5263 static int
5264 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5265                             struct inode *inode, handle_t *handle,
5266                             enum SHIFT_DIRECTION SHIFT)
5267 {
5268         int depth, err = 0;
5269         struct ext4_extent *ex_start, *ex_last;
5270         bool update = 0;
5271         depth = path->p_depth;
5272
5273         while (depth >= 0) {
5274                 if (depth == path->p_depth) {
5275                         ex_start = path[depth].p_ext;
5276                         if (!ex_start)
5277                                 return -EFSCORRUPTED;
5278
5279                         ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5280
5281                         err = ext4_access_path(handle, inode, path + depth);
5282                         if (err)
5283                                 goto out;
5284
5285                         if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5286                                 update = 1;
5287
5288                         while (ex_start <= ex_last) {
5289                                 if (SHIFT == SHIFT_LEFT) {
5290                                         le32_add_cpu(&ex_start->ee_block,
5291                                                 -shift);
5292                                         /* Try to merge to the left. */
5293                                         if ((ex_start >
5294                                             EXT_FIRST_EXTENT(path[depth].p_hdr))
5295                                             &&
5296                                             ext4_ext_try_to_merge_right(inode,
5297                                             path, ex_start - 1))
5298                                                 ex_last--;
5299                                         else
5300                                                 ex_start++;
5301                                 } else {
5302                                         le32_add_cpu(&ex_last->ee_block, shift);
5303                                         ext4_ext_try_to_merge_right(inode, path,
5304                                                 ex_last);
5305                                         ex_last--;
5306                                 }
5307                         }
5308                         err = ext4_ext_dirty(handle, inode, path + depth);
5309                         if (err)
5310                                 goto out;
5311
5312                         if (--depth < 0 || !update)
5313                                 break;
5314                 }
5315
5316                 /* Update index too */
5317                 err = ext4_access_path(handle, inode, path + depth);
5318                 if (err)
5319                         goto out;
5320
5321                 if (SHIFT == SHIFT_LEFT)
5322                         le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5323                 else
5324                         le32_add_cpu(&path[depth].p_idx->ei_block, shift);
5325                 err = ext4_ext_dirty(handle, inode, path + depth);
5326                 if (err)
5327                         goto out;
5328
5329                 /* we are done if current index is not a starting index */
5330                 if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5331                         break;
5332
5333                 depth--;
5334         }
5335
5336 out:
5337         return err;
5338 }
5339
5340 /*
5341  * ext4_ext_shift_extents:
5342  * All the extents which lies in the range from @start to the last allocated
5343  * block for the @inode are shifted either towards left or right (depending
5344  * upon @SHIFT) by @shift blocks.
5345  * On success, 0 is returned, error otherwise.
5346  */
5347 static int
5348 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5349                        ext4_lblk_t start, ext4_lblk_t shift,
5350                        enum SHIFT_DIRECTION SHIFT)
5351 {
5352         struct ext4_ext_path *path;
5353         int ret = 0, depth;
5354         struct ext4_extent *extent;
5355         ext4_lblk_t stop, *iterator, ex_start, ex_end;
5356
5357         /* Let path point to the last extent */
5358         path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
5359                                 EXT4_EX_NOCACHE);
5360         if (IS_ERR(path))
5361                 return PTR_ERR(path);
5362
5363         depth = path->p_depth;
5364         extent = path[depth].p_ext;
5365         if (!extent)
5366                 goto out;
5367
5368         stop = le32_to_cpu(extent->ee_block);
5369
5370        /*
5371         * For left shifts, make sure the hole on the left is big enough to
5372         * accommodate the shift.  For right shifts, make sure the last extent
5373         * won't be shifted beyond EXT_MAX_BLOCKS.
5374         */
5375         if (SHIFT == SHIFT_LEFT) {
5376                 path = ext4_find_extent(inode, start - 1, &path,
5377                                         EXT4_EX_NOCACHE);
5378                 if (IS_ERR(path))
5379                         return PTR_ERR(path);
5380                 depth = path->p_depth;
5381                 extent =  path[depth].p_ext;
5382                 if (extent) {
5383                         ex_start = le32_to_cpu(extent->ee_block);
5384                         ex_end = le32_to_cpu(extent->ee_block) +
5385                                 ext4_ext_get_actual_len(extent);
5386                 } else {
5387                         ex_start = 0;
5388                         ex_end = 0;
5389                 }
5390
5391                 if ((start == ex_start && shift > ex_start) ||
5392                     (shift > start - ex_end)) {
5393                         ret = -EINVAL;
5394                         goto out;
5395                 }
5396         } else {
5397                 if (shift > EXT_MAX_BLOCKS -
5398                     (stop + ext4_ext_get_actual_len(extent))) {
5399                         ret = -EINVAL;
5400                         goto out;
5401                 }
5402         }
5403
5404         /*
5405          * In case of left shift, iterator points to start and it is increased
5406          * till we reach stop. In case of right shift, iterator points to stop
5407          * and it is decreased till we reach start.
5408          */
5409         if (SHIFT == SHIFT_LEFT)
5410                 iterator = &start;
5411         else
5412                 iterator = &stop;
5413
5414         /*
5415          * Its safe to start updating extents.  Start and stop are unsigned, so
5416          * in case of right shift if extent with 0 block is reached, iterator
5417          * becomes NULL to indicate the end of the loop.
5418          */
5419         while (iterator && start <= stop) {
5420                 path = ext4_find_extent(inode, *iterator, &path,
5421                                         EXT4_EX_NOCACHE);
5422                 if (IS_ERR(path))
5423                         return PTR_ERR(path);
5424                 depth = path->p_depth;
5425                 extent = path[depth].p_ext;
5426                 if (!extent) {
5427                         EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5428                                          (unsigned long) *iterator);
5429                         return -EFSCORRUPTED;
5430                 }
5431                 if (SHIFT == SHIFT_LEFT && *iterator >
5432                     le32_to_cpu(extent->ee_block)) {
5433                         /* Hole, move to the next extent */
5434                         if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5435                                 path[depth].p_ext++;
5436                         } else {
5437                                 *iterator = ext4_ext_next_allocated_block(path);
5438                                 continue;
5439                         }
5440                 }
5441
5442                 if (SHIFT == SHIFT_LEFT) {
5443                         extent = EXT_LAST_EXTENT(path[depth].p_hdr);
5444                         *iterator = le32_to_cpu(extent->ee_block) +
5445                                         ext4_ext_get_actual_len(extent);
5446                 } else {
5447                         extent = EXT_FIRST_EXTENT(path[depth].p_hdr);
5448                         if (le32_to_cpu(extent->ee_block) > 0)
5449                                 *iterator = le32_to_cpu(extent->ee_block) - 1;
5450                         else
5451                                 /* Beginning is reached, end of the loop */
5452                                 iterator = NULL;
5453                         /* Update path extent in case we need to stop */
5454                         while (le32_to_cpu(extent->ee_block) < start)
5455                                 extent++;
5456                         path[depth].p_ext = extent;
5457                 }
5458                 ret = ext4_ext_shift_path_extents(path, shift, inode,
5459                                 handle, SHIFT);
5460                 if (ret)
5461                         break;
5462         }
5463 out:
5464         ext4_ext_drop_refs(path);
5465         kfree(path);
5466         return ret;
5467 }
5468
5469 /*
5470  * ext4_collapse_range:
5471  * This implements the fallocate's collapse range functionality for ext4
5472  * Returns: 0 and non-zero on error.
5473  */
5474 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5475 {
5476         struct super_block *sb = inode->i_sb;
5477         ext4_lblk_t punch_start, punch_stop;
5478         handle_t *handle;
5479         unsigned int credits;
5480         loff_t new_size, ioffset;
5481         int ret;
5482
5483         /*
5484          * We need to test this early because xfstests assumes that a
5485          * collapse range of (0, 1) will return EOPNOTSUPP if the file
5486          * system does not support collapse range.
5487          */
5488         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5489                 return -EOPNOTSUPP;
5490
5491         /* Collapse range works only on fs block size aligned offsets. */
5492         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5493             len & (EXT4_CLUSTER_SIZE(sb) - 1))
5494                 return -EINVAL;
5495
5496         if (!S_ISREG(inode->i_mode))
5497                 return -EINVAL;
5498
5499         trace_ext4_collapse_range(inode, offset, len);
5500
5501         punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5502         punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5503
5504         /* Call ext4_force_commit to flush all data in case of data=journal. */
5505         if (ext4_should_journal_data(inode)) {
5506                 ret = ext4_force_commit(inode->i_sb);
5507                 if (ret)
5508                         return ret;
5509         }
5510
5511         inode_lock(inode);
5512         /*
5513          * There is no need to overlap collapse range with EOF, in which case
5514          * it is effectively a truncate operation
5515          */
5516         if (offset + len >= i_size_read(inode)) {
5517                 ret = -EINVAL;
5518                 goto out_mutex;
5519         }
5520
5521         /* Currently just for extent based files */
5522         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5523                 ret = -EOPNOTSUPP;
5524                 goto out_mutex;
5525         }
5526
5527         /* Wait for existing dio to complete */
5528         inode_dio_wait(inode);
5529
5530         /*
5531          * Prevent page faults from reinstantiating pages we have released from
5532          * page cache.
5533          */
5534         down_write(&EXT4_I(inode)->i_mmap_sem);
5535
5536         ret = ext4_break_layouts(inode);
5537         if (ret)
5538                 goto out_mmap;
5539
5540         /*
5541          * Need to round down offset to be aligned with page size boundary
5542          * for page size > block size.
5543          */
5544         ioffset = round_down(offset, PAGE_SIZE);
5545         /*
5546          * Write tail of the last page before removed range since it will get
5547          * removed from the page cache below.
5548          */
5549         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset, offset);
5550         if (ret)
5551                 goto out_mmap;
5552         /*
5553          * Write data that will be shifted to preserve them when discarding
5554          * page cache below. We are also protected from pages becoming dirty
5555          * by i_mmap_sem.
5556          */
5557         ret = filemap_write_and_wait_range(inode->i_mapping, offset + len,
5558                                            LLONG_MAX);
5559         if (ret)
5560                 goto out_mmap;
5561         truncate_pagecache(inode, ioffset);
5562
5563         credits = ext4_writepage_trans_blocks(inode);
5564         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5565         if (IS_ERR(handle)) {
5566                 ret = PTR_ERR(handle);
5567                 goto out_mmap;
5568         }
5569
5570         down_write(&EXT4_I(inode)->i_data_sem);
5571         ext4_discard_preallocations(inode);
5572
5573         ret = ext4_es_remove_extent(inode, punch_start,
5574                                     EXT_MAX_BLOCKS - punch_start);
5575         if (ret) {
5576                 up_write(&EXT4_I(inode)->i_data_sem);
5577                 goto out_stop;
5578         }
5579
5580         ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5581         if (ret) {
5582                 up_write(&EXT4_I(inode)->i_data_sem);
5583                 goto out_stop;
5584         }
5585         ext4_discard_preallocations(inode);
5586
5587         ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5588                                      punch_stop - punch_start, SHIFT_LEFT);
5589         if (ret) {
5590                 up_write(&EXT4_I(inode)->i_data_sem);
5591                 goto out_stop;
5592         }
5593
5594         new_size = i_size_read(inode) - len;
5595         i_size_write(inode, new_size);
5596         EXT4_I(inode)->i_disksize = new_size;
5597
5598         up_write(&EXT4_I(inode)->i_data_sem);
5599         if (IS_SYNC(inode))
5600                 ext4_handle_sync(handle);
5601         inode->i_mtime = inode->i_ctime = current_time(inode);
5602         ext4_mark_inode_dirty(handle, inode);
5603         ext4_update_inode_fsync_trans(handle, inode, 1);
5604
5605 out_stop:
5606         ext4_journal_stop(handle);
5607 out_mmap:
5608         up_write(&EXT4_I(inode)->i_mmap_sem);
5609 out_mutex:
5610         inode_unlock(inode);
5611         return ret;
5612 }
5613
5614 /*
5615  * ext4_insert_range:
5616  * This function implements the FALLOC_FL_INSERT_RANGE flag of fallocate.
5617  * The data blocks starting from @offset to the EOF are shifted by @len
5618  * towards right to create a hole in the @inode. Inode size is increased
5619  * by len bytes.
5620  * Returns 0 on success, error otherwise.
5621  */
5622 int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
5623 {
5624         struct super_block *sb = inode->i_sb;
5625         handle_t *handle;
5626         struct ext4_ext_path *path;
5627         struct ext4_extent *extent;
5628         ext4_lblk_t offset_lblk, len_lblk, ee_start_lblk = 0;
5629         unsigned int credits, ee_len;
5630         int ret = 0, depth, split_flag = 0;
5631         loff_t ioffset;
5632
5633         /*
5634          * We need to test this early because xfstests assumes that an
5635          * insert range of (0, 1) will return EOPNOTSUPP if the file
5636          * system does not support insert range.
5637          */
5638         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
5639                 return -EOPNOTSUPP;
5640
5641         /* Insert range works only on fs block size aligned offsets. */
5642         if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5643                         len & (EXT4_CLUSTER_SIZE(sb) - 1))
5644                 return -EINVAL;
5645
5646         if (!S_ISREG(inode->i_mode))
5647                 return -EOPNOTSUPP;
5648
5649         trace_ext4_insert_range(inode, offset, len);
5650
5651         offset_lblk = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5652         len_lblk = len >> EXT4_BLOCK_SIZE_BITS(sb);
5653
5654         /* Call ext4_force_commit to flush all data in case of data=journal */
5655         if (ext4_should_journal_data(inode)) {
5656                 ret = ext4_force_commit(inode->i_sb);
5657                 if (ret)
5658                         return ret;
5659         }
5660
5661         inode_lock(inode);
5662         /* Currently just for extent based files */
5663         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5664                 ret = -EOPNOTSUPP;
5665                 goto out_mutex;
5666         }
5667
5668         /* Check for wrap through zero */
5669         if (inode->i_size + len > inode->i_sb->s_maxbytes) {
5670                 ret = -EFBIG;
5671                 goto out_mutex;
5672         }
5673
5674         /* Offset should be less than i_size */
5675         if (offset >= i_size_read(inode)) {
5676                 ret = -EINVAL;
5677                 goto out_mutex;
5678         }
5679
5680         /* Wait for existing dio to complete */
5681         inode_dio_wait(inode);
5682
5683         /*
5684          * Prevent page faults from reinstantiating pages we have released from
5685          * page cache.
5686          */
5687         down_write(&EXT4_I(inode)->i_mmap_sem);
5688
5689         ret = ext4_break_layouts(inode);
5690         if (ret)
5691                 goto out_mmap;
5692
5693         /*
5694          * Need to round down to align start offset to page size boundary
5695          * for page size > block size.
5696          */
5697         ioffset = round_down(offset, PAGE_SIZE);
5698         /* Write out all dirty pages */
5699         ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5700                         LLONG_MAX);
5701         if (ret)
5702                 goto out_mmap;
5703         truncate_pagecache(inode, ioffset);
5704
5705         credits = ext4_writepage_trans_blocks(inode);
5706         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5707         if (IS_ERR(handle)) {
5708                 ret = PTR_ERR(handle);
5709                 goto out_mmap;
5710         }
5711
5712         /* Expand file to avoid data loss if there is error while shifting */
5713         inode->i_size += len;
5714         EXT4_I(inode)->i_disksize += len;
5715         inode->i_mtime = inode->i_ctime = current_time(inode);
5716         ret = ext4_mark_inode_dirty(handle, inode);
5717         if (ret)
5718                 goto out_stop;
5719
5720         down_write(&EXT4_I(inode)->i_data_sem);
5721         ext4_discard_preallocations(inode);
5722
5723         path = ext4_find_extent(inode, offset_lblk, NULL, 0);
5724         if (IS_ERR(path)) {
5725                 up_write(&EXT4_I(inode)->i_data_sem);
5726                 goto out_stop;
5727         }
5728
5729         depth = ext_depth(inode);
5730         extent = path[depth].p_ext;
5731         if (extent) {
5732                 ee_start_lblk = le32_to_cpu(extent->ee_block);
5733                 ee_len = ext4_ext_get_actual_len(extent);
5734
5735                 /*
5736                  * If offset_lblk is not the starting block of extent, split
5737                  * the extent @offset_lblk
5738                  */
5739                 if ((offset_lblk > ee_start_lblk) &&
5740                                 (offset_lblk < (ee_start_lblk + ee_len))) {
5741                         if (ext4_ext_is_unwritten(extent))
5742                                 split_flag = EXT4_EXT_MARK_UNWRIT1 |
5743                                         EXT4_EXT_MARK_UNWRIT2;
5744                         ret = ext4_split_extent_at(handle, inode, &path,
5745                                         offset_lblk, split_flag,
5746                                         EXT4_EX_NOCACHE |
5747                                         EXT4_GET_BLOCKS_PRE_IO |
5748                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
5749                 }
5750
5751                 ext4_ext_drop_refs(path);
5752                 kfree(path);
5753                 if (ret < 0) {
5754                         up_write(&EXT4_I(inode)->i_data_sem);
5755                         goto out_stop;
5756                 }
5757         } else {
5758                 ext4_ext_drop_refs(path);
5759                 kfree(path);
5760         }
5761
5762         ret = ext4_es_remove_extent(inode, offset_lblk,
5763                         EXT_MAX_BLOCKS - offset_lblk);
5764         if (ret) {
5765                 up_write(&EXT4_I(inode)->i_data_sem);
5766                 goto out_stop;
5767         }
5768
5769         /*
5770          * if offset_lblk lies in a hole which is at start of file, use
5771          * ee_start_lblk to shift extents
5772          */
5773         ret = ext4_ext_shift_extents(inode, handle,
5774                 ee_start_lblk > offset_lblk ? ee_start_lblk : offset_lblk,
5775                 len_lblk, SHIFT_RIGHT);
5776
5777         up_write(&EXT4_I(inode)->i_data_sem);
5778         if (IS_SYNC(inode))
5779                 ext4_handle_sync(handle);
5780         if (ret >= 0)
5781                 ext4_update_inode_fsync_trans(handle, inode, 1);
5782
5783 out_stop:
5784         ext4_journal_stop(handle);
5785 out_mmap:
5786         up_write(&EXT4_I(inode)->i_mmap_sem);
5787 out_mutex:
5788         inode_unlock(inode);
5789         return ret;
5790 }
5791
5792 /**
5793  * ext4_swap_extents - Swap extents between two inodes
5794  *
5795  * @inode1:     First inode
5796  * @inode2:     Second inode
5797  * @lblk1:      Start block for first inode
5798  * @lblk2:      Start block for second inode
5799  * @count:      Number of blocks to swap
5800  * @unwritten: Mark second inode's extents as unwritten after swap
5801  * @erp:        Pointer to save error value
5802  *
5803  * This helper routine does exactly what is promise "swap extents". All other
5804  * stuff such as page-cache locking consistency, bh mapping consistency or
5805  * extent's data copying must be performed by caller.
5806  * Locking:
5807  *              i_mutex is held for both inodes
5808  *              i_data_sem is locked for write for both inodes
5809  * Assumptions:
5810  *              All pages from requested range are locked for both inodes
5811  */
5812 int
5813 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5814                   struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5815                   ext4_lblk_t count, int unwritten, int *erp)
5816 {
5817         struct ext4_ext_path *path1 = NULL;
5818         struct ext4_ext_path *path2 = NULL;
5819         int replaced_count = 0;
5820
5821         BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5822         BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5823         BUG_ON(!inode_is_locked(inode1));
5824         BUG_ON(!inode_is_locked(inode2));
5825
5826         *erp = ext4_es_remove_extent(inode1, lblk1, count);
5827         if (unlikely(*erp))
5828                 return 0;
5829         *erp = ext4_es_remove_extent(inode2, lblk2, count);
5830         if (unlikely(*erp))
5831                 return 0;
5832
5833         while (count) {
5834                 struct ext4_extent *ex1, *ex2, tmp_ex;
5835                 ext4_lblk_t e1_blk, e2_blk;
5836                 int e1_len, e2_len, len;
5837                 int split = 0;
5838
5839                 path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5840                 if (IS_ERR(path1)) {
5841                         *erp = PTR_ERR(path1);
5842                         path1 = NULL;
5843                 finish:
5844                         count = 0;
5845                         goto repeat;
5846                 }
5847                 path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5848                 if (IS_ERR(path2)) {
5849                         *erp = PTR_ERR(path2);
5850                         path2 = NULL;
5851                         goto finish;
5852                 }
5853                 ex1 = path1[path1->p_depth].p_ext;
5854                 ex2 = path2[path2->p_depth].p_ext;
5855                 /* Do we have somthing to swap ? */
5856                 if (unlikely(!ex2 || !ex1))
5857                         goto finish;
5858
5859                 e1_blk = le32_to_cpu(ex1->ee_block);
5860                 e2_blk = le32_to_cpu(ex2->ee_block);
5861                 e1_len = ext4_ext_get_actual_len(ex1);
5862                 e2_len = ext4_ext_get_actual_len(ex2);
5863
5864                 /* Hole handling */
5865                 if (!in_range(lblk1, e1_blk, e1_len) ||
5866                     !in_range(lblk2, e2_blk, e2_len)) {
5867                         ext4_lblk_t next1, next2;
5868
5869                         /* if hole after extent, then go to next extent */
5870                         next1 = ext4_ext_next_allocated_block(path1);
5871                         next2 = ext4_ext_next_allocated_block(path2);
5872                         /* If hole before extent, then shift to that extent */
5873                         if (e1_blk > lblk1)
5874                                 next1 = e1_blk;
5875                         if (e2_blk > lblk2)
5876                                 next2 = e2_blk;
5877                         /* Do we have something to swap */
5878                         if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5879                                 goto finish;
5880                         /* Move to the rightest boundary */
5881                         len = next1 - lblk1;
5882                         if (len < next2 - lblk2)
5883                                 len = next2 - lblk2;
5884                         if (len > count)
5885                                 len = count;
5886                         lblk1 += len;
5887                         lblk2 += len;
5888                         count -= len;
5889                         goto repeat;
5890                 }
5891
5892                 /* Prepare left boundary */
5893                 if (e1_blk < lblk1) {
5894                         split = 1;
5895                         *erp = ext4_force_split_extent_at(handle, inode1,
5896                                                 &path1, lblk1, 0);
5897                         if (unlikely(*erp))
5898                                 goto finish;
5899                 }
5900                 if (e2_blk < lblk2) {
5901                         split = 1;
5902                         *erp = ext4_force_split_extent_at(handle, inode2,
5903                                                 &path2,  lblk2, 0);
5904                         if (unlikely(*erp))
5905                                 goto finish;
5906                 }
5907                 /* ext4_split_extent_at() may result in leaf extent split,
5908                  * path must to be revalidated. */
5909                 if (split)
5910                         goto repeat;
5911
5912                 /* Prepare right boundary */
5913                 len = count;
5914                 if (len > e1_blk + e1_len - lblk1)
5915                         len = e1_blk + e1_len - lblk1;
5916                 if (len > e2_blk + e2_len - lblk2)
5917                         len = e2_blk + e2_len - lblk2;
5918
5919                 if (len != e1_len) {
5920                         split = 1;
5921                         *erp = ext4_force_split_extent_at(handle, inode1,
5922                                                 &path1, lblk1 + len, 0);
5923                         if (unlikely(*erp))
5924                                 goto finish;
5925                 }
5926                 if (len != e2_len) {
5927                         split = 1;
5928                         *erp = ext4_force_split_extent_at(handle, inode2,
5929                                                 &path2, lblk2 + len, 0);
5930                         if (*erp)
5931                                 goto finish;
5932                 }
5933                 /* ext4_split_extent_at() may result in leaf extent split,
5934                  * path must to be revalidated. */
5935                 if (split)
5936                         goto repeat;
5937
5938                 BUG_ON(e2_len != e1_len);
5939                 *erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5940                 if (unlikely(*erp))
5941                         goto finish;
5942                 *erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5943                 if (unlikely(*erp))
5944                         goto finish;
5945
5946                 /* Both extents are fully inside boundaries. Swap it now */
5947                 tmp_ex = *ex1;
5948                 ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5949                 ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5950                 ex1->ee_len = cpu_to_le16(e2_len);
5951                 ex2->ee_len = cpu_to_le16(e1_len);
5952                 if (unwritten)
5953                         ext4_ext_mark_unwritten(ex2);
5954                 if (ext4_ext_is_unwritten(&tmp_ex))
5955                         ext4_ext_mark_unwritten(ex1);
5956
5957                 ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5958                 ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5959                 *erp = ext4_ext_dirty(handle, inode2, path2 +
5960                                       path2->p_depth);
5961                 if (unlikely(*erp))
5962                         goto finish;
5963                 *erp = ext4_ext_dirty(handle, inode1, path1 +
5964                                       path1->p_depth);
5965                 /*
5966                  * Looks scarry ah..? second inode already points to new blocks,
5967                  * and it was successfully dirtied. But luckily error may happen
5968                  * only due to journal error, so full transaction will be
5969                  * aborted anyway.
5970                  */
5971                 if (unlikely(*erp))
5972                         goto finish;
5973                 lblk1 += len;
5974                 lblk2 += len;
5975                 replaced_count += len;
5976                 count -= len;
5977
5978         repeat:
5979                 ext4_ext_drop_refs(path1);
5980                 kfree(path1);
5981                 ext4_ext_drop_refs(path2);
5982                 kfree(path2);
5983                 path1 = path2 = NULL;
5984         }
5985         return replaced_count;
5986 }